best research papers on artificial intelligence

The present and future of AI

Finale doshi-velez on how ai is shaping our lives and how we can shape ai.

Finale Doshi-Velez, the John L. Loeb Professor of Engineering and Applied Sciences. (Photo courtesy of Eliza Grinnell/Harvard SEAS)

How has artificial intelligence changed and shaped our world over the last five years? How will AI continue to impact our lives in the coming years? Those were the questions addressed in the most recent report from the One Hundred Year Study on Artificial Intelligence (AI100), an ongoing project hosted at Stanford University, that will study the status of AI technology and its impacts on the world over the next 100 years.

The 2021 report is the second in a series that will be released every five years until 2116. Titled “Gathering Strength, Gathering Storms,” the report explores the various ways AI is  increasingly touching people’s lives in settings that range from  movie recommendations  and  voice assistants  to  autonomous driving  and  automated medical diagnoses .

Barbara Grosz , the Higgins Research Professor of Natural Sciences at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) is a member of the standing committee overseeing the AI100 project and Finale Doshi-Velez , Gordon McKay Professor of Computer Science, is part of the panel of interdisciplinary researchers who wrote this year’s report. 

We spoke with Doshi-Velez about the report, what it says about the role AI is currently playing in our lives, and how it will change in the future.  

Q: Let's start with a snapshot: What is the current state of AI and its potential?

Doshi-Velez: Some of the biggest changes in the last five years have been how well AIs now perform in large data regimes on specific types of tasks.  We've seen [DeepMind’s] AlphaZero become the best Go player entirely through self-play, and everyday uses of AI such as grammar checks and autocomplete, automatic personal photo organization and search, and speech recognition become commonplace for large numbers of people.  

In terms of potential, I'm most excited about AIs that might augment and assist people.  They can be used to drive insights in drug discovery, help with decision making such as identifying a menu of likely treatment options for patients, and provide basic assistance, such as lane keeping while driving or text-to-speech based on images from a phone for the visually impaired.  In many situations, people and AIs have complementary strengths. I think we're getting closer to unlocking the potential of people and AI teams.

There's a much greater recognition that we should not be waiting for AI tools to become mainstream before making sure they are ethical.

Q: Over the course of 100 years, these reports will tell the story of AI and its evolving role in society. Even though there have only been two reports, what's the story so far?

There's actually a lot of change even in five years.  The first report is fairly rosy.  For example, it mentions how algorithmic risk assessments may mitigate the human biases of judges.  The second has a much more mixed view.  I think this comes from the fact that as AI tools have come into the mainstream — both in higher stakes and everyday settings — we are appropriately much less willing to tolerate flaws, especially discriminatory ones. There's also been questions of information and disinformation control as people get their news, social media, and entertainment via searches and rankings personalized to them. So, there's a much greater recognition that we should not be waiting for AI tools to become mainstream before making sure they are ethical.

Q: What is the responsibility of institutes of higher education in preparing students and the next generation of computer scientists for the future of AI and its impact on society?

First, I'll say that the need to understand the basics of AI and data science starts much earlier than higher education!  Children are being exposed to AIs as soon as they click on videos on YouTube or browse photo albums. They need to understand aspects of AI such as how their actions affect future recommendations.

But for computer science students in college, I think a key thing that future engineers need to realize is when to demand input and how to talk across disciplinary boundaries to get at often difficult-to-quantify notions of safety, equity, fairness, etc.  I'm really excited that Harvard has the Embedded EthiCS program to provide some of this education.  Of course, this is an addition to standard good engineering practices like building robust models, validating them, and so forth, which is all a bit harder with AI.

I think a key thing that future engineers need to realize is when to demand input and how to talk across disciplinary boundaries to get at often difficult-to-quantify notions of safety, equity, fairness, etc. 

Q: Your work focuses on machine learning with applications to healthcare, which is also an area of focus of this report. What is the state of AI in healthcare? 

A lot of AI in healthcare has been on the business end, used for optimizing billing, scheduling surgeries, that sort of thing.  When it comes to AI for better patient care, which is what we usually think about, there are few legal, regulatory, and financial incentives to do so, and many disincentives. Still, there's been slow but steady integration of AI-based tools, often in the form of risk scoring and alert systems.

In the near future, two applications that I'm really excited about are triage in low-resource settings — having AIs do initial reads of pathology slides, for example, if there are not enough pathologists, or get an initial check of whether a mole looks suspicious — and ways in which AIs can help identify promising treatment options for discussion with a clinician team and patient.

Q: Any predictions for the next report?

I'll be keen to see where currently nascent AI regulation initiatives have gotten to. Accountability is such a difficult question in AI,  it's tricky to nurture both innovation and basic protections.  Perhaps the most important innovation will be in approaches for AI accountability.

Topics: AI / Machine Learning , Computer Science

Cutting-edge science delivered direct to your inbox.

Join the Harvard SEAS mailing list.

Scientist Profiles

Finale Doshi-Velez

Herchel Smith Professor of Computer Science

Press Contact

Leah Burrows | 617-496-1351 | [email protected]

Related News

Seven SEAS teams named President’s Innovation Challenge finalists

Start-ups will vie for up to $75,000 in prize money

Computer Science , Design , Electrical Engineering , Entrepreneurship , Events , Master of Design Engineering , Materials Science & Mechanical Engineering , MS/MBA

Exploring the depths of AI

 New SEAS club spends Spring Break meeting AI technology professionals in San Francisco

AI / Machine Learning , Computer Science , Student Organizations

Alumni profile: Jacomo Corbo, Ph.D. '08

Racing into the future of machine learning 

AI / Machine Learning , Computer Science

The Journal of Artificial Intelligence Research (JAIR) is dedicated to the rapid dissemination of important research results to the global artificial intelligence (AI) community. The journal’s scope encompasses all areas of AI, including agents and multi-agent systems, automated reasoning, constraint processing and search, knowledge representation, machine learning, natural language, planning and scheduling, robotics and vision, and uncertainty in AI.

Current Issue

Vol. 79 (2024)

Published: 2024-01-10

Bt-GAN: Generating Fair Synthetic Healthdata via Bias-transforming Generative Adversarial Networks

Collision avoiding max-sum for mobile sensor teams, usn: a robust imitation learning method against diverse action noise, structure in deep reinforcement learning: a survey and open problems, a map of diverse synthetic stable matching instances, digcn: a dynamic interaction graph convolutional network based on learnable proposals for object detection, iterative train scheduling under disruption with maximum satisfiability, removing bias and incentivizing precision in peer-grading, cultural bias in explainable ai research: a systematic analysis, learning to resolve social dilemmas: a survey, a principled distributional approach to trajectory similarity measurement and its application to anomaly detection, multi-modal attentive prompt learning for few-shot emotion recognition in conversations, condense: conditional density estimation for time series anomaly detection, performative ethics from within the ivory tower: how cs practitioners uphold systems of oppression, learning logic specifications for policy guidance in pomdps: an inductive logic programming approach, multi-objective reinforcement learning based on decomposition: a taxonomy and framework, can fairness be automated guidelines and opportunities for fairness-aware automl, practical and parallelizable algorithms for non-monotone submodular maximization with size constraint, exploring the tradeoff between system profit and income equality among ride-hailing drivers, on mitigating the utility-loss in differentially private learning: a new perspective by a geometrically inspired kernel approach, an algorithm with improved complexity for pebble motion/multi-agent path finding on trees, weighted, circular and semi-algebraic proofs, reinforcement learning for generative ai: state of the art, opportunities and open research challenges, human-in-the-loop reinforcement learning: a survey and position on requirements, challenges, and opportunities, boolean observation games, detecting change intervals with isolation distributional kernel, query-driven qualitative constraint acquisition, visually grounded language learning: a review of language games, datasets, tasks, and models, right place, right time: proactive multi-robot task allocation under spatiotemporal uncertainty, principles and their computational consequences for argumentation frameworks with collective attacks, the ai race: why current neural network-based architectures are a poor basis for artificial general intelligence, undesirable biases in nlp: addressing challenges of measurement.

> cs > cs.AI

Help | Advanced Search

Artificial Intelligence

Authors and titles for recent submissions.

• Fri, 26 Apr 2024
• Thu, 25 Apr 2024
• Wed, 24 Apr 2024
• Tue, 23 Apr 2024
• Mon, 22 Apr 2024

Fri, 26 Apr 2024 (showing first 25 of 106 entries)

Links to: arXiv , form interface , find , cs , new , 2404 , contact , h elp   ( Access key information)

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

• Event calendar

The Top 17 ‘Must-Read’ AI Papers in 2022

We caught up with experts in the RE•WORK community to find out what the top 17 AI papers are for 2022 so far that you can add to your Summer must reads. The papers cover a wide range of topics including AI in social media and how AI can benefit humanity and are free to access.

Interested in learning more? Check out all the upcoming RE•WORK events to find out about the latest trends and industry updates in AI here .

Max Li, Staff Data Scientist – Tech Lead at Wish

Max is a Staff Data Scientist at Wish where he focuses on experimentation (A/B testing) and machine learning.  His passion is to empower data-driven decision-making through the rigorous use of data. View Max’s presentation, ‘Assign Experiment Variants at Scale in A/B Tests’, from our Deep Learning Summit in February 2022 here .

1. Boostrapped Meta-Learning (2022) – Sebastian Flennerhag et al.

The first paper selected by Max proposes an algorithm in which allows the meta-learner teach itself, allowing to overcome the meta-optimisation challenge. The algorithm focuses meta-learning with gradients, which guarantees improvements in performance. The paper also looks at how bootstrapping opens up possibilities. Read the full paper here .

2. Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces (2022) – Samuel Daulton et al.

Another paper selected by Max proposes MORBO, a scalable method for multiple-objective BO as it performs better than that of high-dimensional search spaces. MORBO significantly improves the sample efficiency, and where BO algorithms fail, MORBO provides improved sample efficiencies to the current BO approach used. Read the full paper here .

3. Tabular Data: Deep Learning is Not All You Need (2021) – Ravid Shwartz-Ziv, Amitai Armon

To solve real-life data science problems, selecting the right model to use is crucial. This final paper selected by Max explores whether deep models should be recommended as an option for tabular data. Read the full paper here .

Jigyasa Grover, Senior Machine Learning Engineer at Twitter

Jigyasa Grover is a Senior Machine Learning Engineer at Twitter working in the performance ads ranking domain. Recently, she was honoured with the 'Outstanding in AI: Young Role Model Award' by Women in AI across North America. She is one of the few ML Google Developer Experts globally. Jigyasa has previously presented at our Deep Learning Summit and MLOps event in San Fransisco earlier this year.

4. Privacy for Free: How does Dataset Condensation Help Privacy? (2022) – Tian Dong et al.

Jigyasa’s first recommendation concentrates on Privacy Preserving Machine Learning, specifically mitigating the leakage of sensitive data in Machine Learning. The paper provides one of the first propositions of using dataset condensation techniques to preserve the data efficiency during model training and furnish membership privacy. This paper was published by Sony AI and won the Outstanding Paper Award at ICML 2022. Read the full paper here .

5. Affective Signals in a Social Media Recommender System (2022) – Jane Dwivedi-Yu et al.

The second paper recommended by Jigyasa talks about operationalising Affective Computing, also known as Emotional AI, for an improved personalised feed on social media. The paper discusses the design of an affective taxonomy customised to user needs on social media. It further lays out the curation of suitable training data by combining engagement data and data from a human-labelling task to enable the identification of the affective response a user might exhibit for a particular post. Read the full paper here .

6. ItemSage: Learning Product Embeddings for Shopping Recommendations at Pinterest (2022) – Paul Baltescu et al.

Jigyasa’s last recommendation is a paper by Pinterest that illustrates the aggregation of both textual and visual information to build a unified set of product embeddings to enhance recommendation results on e-commerce websites. By applying multi-task learning, the proposed embeddings can optimise for multiple engagement types and ensures that the shopping recommendation stack is efficient with respect to all objectives. Read the full article here .

Asmita Poddar, Software Development Engineer at Amazon Alexa

Asmita is a Software Development Engineer at Amazon Alexa, where she works on developing and productionising natural language processing and speech models. Asmita also has prior experience in applying machine learning in diverse domains. Asmita will be presenting at our London AI Summit , in September, where she will discuss AI for Spoken Communication.

7. Competition-Level Code Generation with AlphaCode (2022) – Yujia Li et al.

Systems can help programmers become more productive. Asmita has selected this paper which addresses the problems with incorporating innovations in AI into these systems. AlphaCode is a system that creates solutions for problems that requires deeper reasoning. Read the full paper here .

8. A Commonsense Knowledge Enhanced Network with Retrospective Loss for Emotion Recognition in Spoken Dialog (2022) – Yunhe Xie et al.

There are limits to model’s reasoning in regards to the existing ERSD datasets. The final paper selected by Asmita proposes a Commonsense Knowledge Enhanced Network with a backward-looking loss to perform dialog modelling, external knowledge integration and historical state retrospect. The model used has been shown to outperform other models. Read the full paper here .

Discover the speakers we have lined up and the topics we will cover at the London AI Summit.

Sergei Bobrovskyi, Expert in Anomaly Detection for Root Cause Analysis at Airbus

Dr. Sergei Bobrovskyi is a Data Scientist within the Analytics Accelerator team of the Airbus Digital Transformation Office. His work focuses on applications of AI for anomaly detection in time series, spanning various use-cases across Airbus. Sergei will be presenting at our Berlin AI Summit in October about Anomaly Detection, Root Cause Analysis and Explainability.

9. LaMDA: Language Models for Dialog Applications (2022) – Romal Thoppilan et al.

The paper chosen by Sergei describes the LaMDA system, which caused the furor this summer, when a former Google engineer claimed it has shown signs of being sentient. LaMDA is a family of large language models for dialog applications based on Transformer architecture. The interesting feature of the model is their fine-tuning with human annotated data and possibility to consult external sources. In any case, this is a very interesting model family, which we might encounter in many of the applications we use daily. Read the full paper here .

10. A Path Towards Autonomous Machine Intelligence Version 0.9.2, 2022-06-27 (2022) – Yann LeCun

The second paper chosen by Sergei provides a vision on how to progress towards general AI. The study combines a number of concepts including configurable predictive world model, behaviour driven through intrinsic motivation, and hierarchical joint embedding architectures. Read the full paper here .

11. Coordination Among Neural Modules Through a Shared Global Workpace (2022) – Anirudh Goyal et al.

This paper chosen by Sergei combines the Transformer architecture underlying most of the recent successes of deep learning with ideas from the Global Workspace Theory from cognitive sciences. This is an interesting read to broaden the understanding of why certain model architectures perform well and in which direction we might go in the future to further improve performance on challenging tasks. Read the full paper here .

12. Magnetic control of tokamak plasmas through deep reinforcement learning (2022) – Jonas Degrave et al.

Sergei chose the next paper, which asks the question of ‘how can AI research benefit humanity?’. The use of AI to enable safe, reliable and scalable deployment of fusion energy could contribute to the solution of pression problems of climate change. Sergei has said that this is an extremely interesting application of AI technology for engineering. Read the full paper here .

13. TranAd: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data (2022) – Shreshth Tuli, Giuliano Casale and Nicholas R. Jennings

The final paper chosen by Sergei is a specialised paper applying transformer architecture to the problem of unsupervised anomaly detection in multivariate time-series. Many architectures which were successful in other fields are at some points being also applied to time-series. The paper shows an improved performance on some known data sets. Read the full paper here .

Abdullahi Adamu, Senior Software Engineer at Sony

Abdullahi has worked in various industries including working at a market research start-up where he developed models that could extract insights from human conversations about products or services. He moved to Publicis, where he became Data Engineer and Data Scientist in 2018. Abdullahi will be part of our panel discussion at the London AI Summit in September, where he will discuss Harnessing the Power of Deep Learning.

14. Self-Supervision for Learning from the Bottom Up (2022) – Alexei Efros

This paper chosen by Abdullahi makes compelling arguments for why self-supervision is the next step in the evolution of AI/ML for building more robust models. Overall, these compelling arguments justify even further why self-supervised learning is important on our journey towards more robust models that generalise better in the wild. Read the full paper here .

15. Neural Architecture Search Survey: A Hardware Perspective (2022) – Krishna Teja Chitty-Venkata and Arun K. Somani

Another paper chosen by Abdullahi understands that as we move towards edge computing and federated learning, neural architecture search that takes into account hardware constraints which will be more critical in ensuring that we have leaner neural network models that balance latency and generalisation performance. This survey gives a birds eye view of the various neural architecture search algorithms that take into account hardware constraints to design artificial neural networks that give the best tradeoff of performance and accuracy. Read the full paper here .

16. What Should Not Be Contrastive In Contrastive Learning (2021) – Tete Xiao et al.

In the paper chosen by Abdullahi highlights the underlying assumptions behind data augmentation methods and how these can be counter productive in the context of contrastive learning; for example colour augmentation whilst a downstream task is meant to differentiate colours of objects. The result reported show promising results in the wild. Overall, it presents an elegant solution to using data augmentation for contrastive learning. Read the full paper here .

17. Why do tree-based models still outperform deep learning on tabular data? (2022) – Leo Grinsztajn, Edouard Oyallon and Gael Varoquaux

The final paper selected by Abdulliah works on answering the question of why deep learning models still find it hard to compete on tabular data compared to tree-based models. It is shown that MLP-like architectures are more sensitive to uninformative features in data, compared to their tree-based counterparts. Read the full paper here .

Sign up to the RE•WORK monthly newsletter for the latest AI news, trends and events.

Join us at our upcoming events this year:

·       London AI Summit – 14-15 September 2022

·       Berlin AI Summit – 4-5 October 2022

·       AI in Healthcare Summit Boston – 13-14 October 2022

·       Sydney Deep Learning and Enterprise AI Summits – 17-18 October 2022

·       MLOps Summit – 9-10 November 2022

·       Toronto AI Summit – 9-10 November 2022

·       Nordics AI Summit - 7-8 December 2022

AI Index Report

Welcome to the seventh edition of the AI Index report. The 2024 Index is our most comprehensive to date and arrives at an important moment when AI’s influence on society has never been more pronounced. This year, we have broadened our scope to more extensively cover essential trends such as technical advancements in AI, public perceptions of the technology, and the geopolitical dynamics surrounding its development. Featuring more original data than ever before, this edition introduces new estimates on AI training costs, detailed analyses of the responsible AI landscape, and an entirely new chapter dedicated to AI’s impact on science and medicine.

Read the 2024 AI Index Report

The AI Index report tracks, collates, distills, and visualizes data related to artificial intelligence (AI). Our mission is to provide unbiased, rigorously vetted, broadly sourced data in order for policymakers, researchers, executives, journalists, and the general public to develop a more thorough and nuanced understanding of the complex field of AI.

The AI Index is recognized globally as one of the most credible and authoritative sources for data and insights on artificial intelligence. Previous editions have been cited in major newspapers, including the The New York Times, Bloomberg, and The Guardian, have amassed hundreds of academic citations, and been referenced by high-level policymakers in the United States, the United Kingdom, and the European Union, among other places. This year’s edition surpasses all previous ones in size, scale, and scope, reflecting the growing significance that AI is coming to hold in all of our lives.

Steering Committee Co-Directors

Ray Perrault

Steering committee members.

Erik Brynjolfsson

John Etchemendy

Katrina Ligett

Terah Lyons

James Manyika

Juan Carlos Niebles

Vanessa Parli

Yoav Shoham

Russell Wald

Staff members.

Loredana Fattorini

Nestor Maslej

Letter from the co-directors.

A decade ago, the best AI systems in the world were unable to classify objects in images at a human level. AI struggled with language comprehension and could not solve math problems. Today, AI systems routinely exceed human performance on standard benchmarks.

Progress accelerated in 2023. New state-of-the-art systems like GPT-4, Gemini, and Claude 3 are impressively multimodal: They can generate fluent text in dozens of languages, process audio, and even explain memes. As AI has improved, it has increasingly forced its way into our lives. Companies are racing to build AI-based products, and AI is increasingly being used by the general public. But current AI technology still has significant problems. It cannot reliably deal with facts, perform complex reasoning, or explain its conclusions.

AI faces two interrelated futures. First, technology continues to improve and is increasingly used, having major consequences for productivity and employment. It can be put to both good and bad uses. In the second future, the adoption of AI is constrained by the limitations of the technology. Regardless of which future unfolds, governments are increasingly concerned. They are stepping in to encourage the upside, such as funding university R&D and incentivizing private investment. Governments are also aiming to manage the potential downsides, such as impacts on employment, privacy concerns, misinformation, and intellectual property rights.

As AI rapidly evolves, the AI Index aims to help the AI community, policymakers, business leaders, journalists, and the general public navigate this complex landscape. It provides ongoing, objective snapshots tracking several key areas: technical progress in AI capabilities, the community and investments driving AI development and deployment, public opinion on current and potential future impacts, and policy measures taken to stimulate AI innovation while managing its risks and challenges. By comprehensively monitoring the AI ecosystem, the Index serves as an important resource for understanding this transformative technological force.

On the technical front, this year’s AI Index reports that the number of new large language models released worldwide in 2023 doubled over the previous year. Two-thirds were open-source, but the highest-performing models came from industry players with closed systems. Gemini Ultra became the first LLM to reach human-level performance on the Massive Multitask Language Understanding (MMLU) benchmark; performance on the benchmark has improved by 15 percentage points since last year. Additionally, GPT-4 achieved an impressive 0.97 mean win rate score on the comprehensive Holistic Evaluation of Language Models (HELM) benchmark, which includes MMLU among other evaluations.

Although global private investment in AI decreased for the second consecutive year, investment in generative AI skyrocketed. More Fortune 500 earnings calls mentioned AI than ever before, and new studies show that AI tangibly boosts worker productivity. On the policymaking front, global mentions of AI in legislative proceedings have never been higher. U.S. regulators passed more AI-related regulations in 2023 than ever before. Still, many expressed concerns about AI’s ability to generate deepfakes and impact elections. The public became more aware of AI, and studies suggest that they responded with nervousness.

Ray Perrault Co-director, AI Index

Our Supporting Partners

Analytics & Research Partners

Stay up to date on the AI Index by subscribing to the  Stanford HAI newsletter.

Artificial Intelligence Review

An International Science and Engineering Journal

• Publishes critical evaluations of applications, techniques, and algorithms in the field.
• Provides a platform for researchers and application developers.
• Presents refereed survey and tutorial articles.
• Offers reviews and commentary on significant developments.

Latest articles

Revolutionizing personalized medicine with generative ai: a systematic review.

• Isaias Ghebrehiwet
• Mohd Saberi Mohamad

Modified crayfish optimization algorithm for solving multiple engineering application problems

• Xuelian Zhou
• Abdelazim G. Hussien

A review of nature-inspired algorithms on single-objective optimization problems from 2019 to 2023

• Sarika Jain
• Harish Garg

UAV control in autonomous object-goal navigation: a systematic literature review

• Angel Ayala
• Leticia Portela
• Francisco Cruz

Deep reinforcement learning-based methods for resource scheduling in cloud computing: a review and future directions

• Guangyao Zhou
• Wenhong Tian

Journal updates

Artificial intelligence review is now fully open access.

We are excited to announce that Artificial Intelligence Review  has now become a fully open access (OA) journal as of January 2024. This means that we will only be publishing articles as Open Access meaning content will be and freely available to readers worldwide, enabling the widest possible dissemination and reuse. 

Journal information

• ACM Digital Library
• Current Contents/Engineering, Computing and Technology
• EI Compendex
• Google Scholar
• Japanese Science and Technology Agency (JST)
• Norwegian Register for Scientific Journals and Series
• OCLC WorldCat Discovery Service
• Science Citation Index Expanded (SCIE)
• TD Net Discovery Service
• UGC-CARE List (India)

Rights and permissions

Springer policies

© Springer Nature B.V.

• Find a journal
• Publish with us
• Track your research

• Conferences
• Last updated February 4, 2020
• In AI Origins & Evolution

10 most impressive Research Papers around Artificial Intelligence

Illustration by Progress in AI research is propelling the technology ahead

• Published on May 31, 2017
• by Amit Paul Chowdhury

Artificial Intelligence research advances are transforming technology as we know it. The AI research community is solving some of the most technology problems related to software and hardware infrastructure, theory and algorithms. Interestingly, the field of AI AI research has drawn acolytes from the non-tech field as well. Case in point — prolific Hollywood actor Kristen Stewart’s highly publicized paper on Artificial Intelligence, originally published at Cornell University library’s open access site .  Stewart co-authored the paper , titled “ Bringing Impressionism to Life with Neural Style Transfer in Come Swim ” with the American poet and literary critic David Shapiro and Adobe Research Engineer Bhautik Joshi .

Essentially, the AI-based paper talks about the style transfer techniques used in her short film Come Swim . However, Stewart’s detractors dismissed it as another “high-level case study.”

Meanwhile, the community is awash with ground-breaking research papers around AI.   Analytics India Magazine lists down the most cited scientific papers around AI, machine intelligence, and computer vision , that will give a perspective on the technology and its applications.

Most of these papers have been chosen on the basis of citation value for each. Some of these papers take into account a Highly Influential Citation count (HIC) and Citation Velocity (CV). Citation Velocity is the weighted average number of citations per year over the last 3 years.

A Computational Approach to Edge Detection : Originally published in 1986 and authored by John Canny this paper, on the computational approach to edge detection, has approximately 9724 citations . The success of this approach is defined by a comprehensive set of goals for the computation of edge points. These goals must be precise enough to delimit the desired behavior of the detector while making minimal assumptions about the form of the solution.

Besides, the paper also presents a general method, called feature synthesis, for the fine-to-coarse integration of information from operators at different scales. This helps in establishing the fact that edge detector performance improves considerably as the operator point spread function is extended along the edge.

A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence : This research paper was co-written by John McCarthy, Marvin L. Minsky, Nathaniel Rochester, Claude E. Shannon, and published in the year 1955. This summer research proposal defined the field, and has another first to its name — it is the first paper to use the term Artificial Intelligence. The proposal invited researchers to the Dartmouth conference , which is widely considered the “birth of AI”.

A Threshold Selection Method from Gray-Level Histograms : The paper was authored by Nobuyuki Otsu and published in 1979 . It has received 7849 paper citations so far. Through this paper, Otsu discusses a nonparametric and unsupervised method of automatic threshold selection for picture segmentation.

The paper delves into how an optimal threshold is selected by the discriminant criterion to maximize the separability of the resultant classes in gray levels. The procedure utilizes only the zeroth- and first-order cumulative moments of the gray-level histogram. The method can be easily applied across multi threshold problems. The paper validates the method by presenting several experimental results.

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift : This 2015 article was co-written by Sergey Ioffe and Christian Szegedy . The paper received 946 citations and reflects on a HIC score of 56.

The paper talks about how training deep neural networks is complicated by the fact that the distribution of each layer’s inputs changes during training. This is a result of change in parameters of the previous layers. The phenomenon is termed as internal covariate shift. This issue is addressed by normalizing layer inputs.

Batch normalization achieves the same accuracy with 14 times fewer training steps when applied to a state-of-the-art image classification model. In other words, Batch Normalization beats the original model by a significant margin.

Deep Residual Learning for Image Recognition : The 2016 paper was co-authored by Kaiming He, Xiangyu Zhang, and Shaoqing Ren. The paper has been cited 1436 times, reflecting on a HIC value of 137 and a CV of 582 . The authors have delved into residual learning framework to ease the training of deep neural networks that are substantially deeper than those used previously.

Besides, the research paper explicitly reformulates the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. The research also delves into how comprehensive empirical evidence show that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth.

Distinctive Image Features from Scale-Invariant Keypoints : This article was authored by David G. Lowe in 2004 . The paper received 21528 citations  and explores the method for extracting distinctive invariant features from images. These can be utilized to perform reliable matching between different views of an object or scene. The features are invariant to image scale and rotation, and are shown to provide robust matching across a substantial range of affine distortion, change in 3D viewpoint, addition of noise, and change in illumination.

The paper additionally delves into an approach which leverages these features for image recognition. This approach can help identify objects among clutter and occlusion while achieving near real-time performance.

Dropout: a simple way to prevent neural networks from overfitting : The 2014 paper was co-authored by Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov . The paper has been cited around 2084 times , with a HIC and CV value of 142 and 536 respectively . Deep neural nets with a large number of parameters are very powerful machine learning systems. However, overfitting is a serious problem in such networks.

The central premise of the paper is to drop units (along with their connections) from the neural network during training, thus preventing units from co-adapting too much. This helps in significantly reducing overfitting, while furnishing major improvements over other regularization methods.

Induction of decision trees : Authored by J. R. Quinlan , this scientific paper was originally published in 1986 and summarizes an approach to synthesizing decision trees that has been used in a variety of systems. The paper specifically describes one such system, ID3, in detail. Additionally, the paper discusses a reported shortcoming of the basic algorithm , besides comparing the two methods of overcoming it. To conclude the paper, the author presents illustrations of current research directions.

Large-Scale Video Classification with Convolutional Neural Networks : This 2014 paper was co-written by 6 authors, Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. The paper has been cited over 865 times , and reflects on a HIC score of 24 , and a CV of 239 .

Convolutional Neural Networks (CNNs) are proven to stand as a powerful class of models for image recognition problems. These results encouraged the authors to provide an extensive empirical evaluation of CNNs on large-scale video classification. This was accomplished using a new dataset of 1 million YouTube videos belonging to 487 classes.

Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference : The paper was published in 1988 . Judea Pearl is the author to this article. The paper presents a complete and accessible account of the theoretical foundations and computational methods that underlie plausible reasoning under uncertainty.

Pearl furnishes a provides a coherent explication of probability as a language for reasoning with partial belief and offers a unifying perspective on other AI approaches to uncertainty, such as the Dempster-Shafer formalism , truth maintenance systems, and nonmonotonic logic.

Access all our open Survey & Awards Nomination forms in one place

Amit Paul Chowdhury

This Indian Company is Harnessing AI for Traffic Regulation

Can India’s Billion-Dollar Funding Agency Propel AI Research?

6 Innovative AI Models for Weather Forecasting

Apollo Brings AI-Powered Healthcare For Indians with Microsoft

Yann LeCun’s Version of Autonomous Machine Intelligence

What can AI Learn from Animals

What is pruning in tree based ML models and why is it done?

Different edge detection techniques with implementation in OpenCV

CORPORATE TRAINING PROGRAMS ON GENERATIVE AI

Generative ai skilling for enterprises, our customized corporate training program on generative ai provides a unique opportunity to empower, retain, and advance your talent., upcoming large format conference, data engineering summit 2024, may 30 and 31, 2024 | 📍 bangalore, india, download the easiest way to stay informed.

Zerodha CTO Warns Companies to Not Look at AI as a Solution Chasing a Problem

10 Deepfake AI Tools to Help You Create Content within Minutes

Commvault’s Arlie Teams Up with Microsoft to Elevate Cyber Resilience Globally

Top editorial picks, sbi to leverage hcl unica to digitally transform customer engagement, african tech companies prefer zoho enterprise over google workspace, reid hoffman creates a deepfake of himself, reid ai, github copilot rival, augment secures $252 mn at $1 bn valuation to boost ai for developers, subscribe to the belamy: our weekly newsletter, biggest ai stories, delivered to your inbox every week., also in news, ready or not, ai agents are coming.

GPU Poor, But Luxury Rich

Why IBM is Acquiring HashiCorp

Healthify Uses OpenAI’s GPTs to Help Indians Make Better Health Choices

Why NVIDIA is Acquiring Run:ai

Zerodha CTO Says He Stopped Googling Technical Stuff Over the Past Year

Smartphones Will Soon be Dead

Adobe Unveils World’s First Large-Scale GAN-based Model for Video Super-Resolution 

Ai courses & careers.

India is a Goldmine for AI Talent 

Top 10 LMS Platforms for Enterprise AI Training and Development

AI Clock is Ticking: Wake Up Call for Education Institutions 

Become a certified generative ai engineer, industry insights.

Healthtech AI startup Endimension Technology raises INR 6 Crore in Pre-Series A Round

New Relic Enhances AI Monitoring, Industry’s First APM for AI

BCG Predicts AI to Drive 20% of 2024 Revenues, Doubling to 40% by 2026

Check our industry research reports.

AI Forum for India

Our discord community for ai ecosystem, in collaboration with nvidia. .

Zerodha CTO Dr. Kailash Nadh Decodes AI Culture in Tech

"> "> Flagship Events

Rising 2024 | de&i in tech summit, april 4 and 5, 2024 | 📍 hilton convention center, manyata tech park, bangalore, machinecon gcc summit 2024, june 28 2024 | 📍bangalore, india, machinecon usa 2024, 26 july 2024 | 583 park avenue, new york, cypher india 2024, september 25-27, 2024 | 📍bangalore, india, cypher usa 2024, nov 21-22 2024 | 📍santa clara convention center, california, usa, genai corner.

7 AI Startups that Featured on Shark Tank India Season 3

Top 9 Semiconductor GCCs in India

Top 6 Devin Alternatives to Automate Your Coding Tasks 

10 Free AI Courses by NVIDIA

Top 6 AI/ML Hackathons to Participate in 2024 

What’s Devin Up to?

10 Underrated Women in AI to Watchout For

10 AI Startups Run by Incredible Women Entrepreneurs 

Data dialogues.

Fibe Leverages Amazon Bedrock to Increase Customer Support Efficiency by 30% 

This 18-Year-Old Programmer is Creating an Open Source Alternative to Redis

Automation Anywhere Wants to Augment Humans with AI, Not Replace Them

Father of Computational Theory Wins 2023 Turing Award

Building Open Source LLMs is Not for Everyone 

This 20-year-old AI Researcher Created the much-needed Indic LLM Leaderboard

NPCI is Exploring AI-Powered Futuristic Payment Frontiers: CTO

Prisma AI Has an ‘Eye on You’ at Adani Airports

Future talks.

T-Hub Supported MATH is Launching AI Career Finder to Create AI Jobs 

Quora’s Poe Eats Google’s Lunch

Zoho Collaborates with Intel to Optimise & Accelerate Video AI Workloads

Rakuten Certified as Best Firm for Data Scientists for the 2nd Time

This Indian Logistics Company Developed an LLM to Enhance Last-Mile Delivery 

Perplexity AI Reviews with Pro Access

What to Expect at the ‘Absolutely Incredible’ Apple WWDC 2024

Will StarCoder 2 Win Over Enterprises?

Developer’s corner.

Japan is the Next Big Hub for Indian Tech Talent

Will TypeScript Wipe Out JavaScript? 

Meta Forces Developers Cite ‘Llama 3’ in their AI Development

Why Developers Hate Jira

In case you missed it.

Which is the Most Frustrating Programming Language?

AI4Bharat Rolls Out IndicLLMSuite for Building LLMs in Indian Languages

Google Introduces Synth^2 to Enhance the Training of Visual Language Models  

Infosys Founder Funds Meta’s Llama 2 Project with 22 Indian Languages

9 Best AI Tools for Excel and Google Spread Sheet Automation

8 Best Generative AI Courses for Executives and Managers

Top 8 AI Browser Extensions for Chrome Users in 2024

Top 5 Devin AI Alternatives for Coders and Developers

10 Best AI Code Generator Tools to Use for Free in 2024

What are the Responsibility of Developers Using Generative AI

Also in trends.

Synology Launches Advanced Data Management & Security Solutions Against Ransomware in India

PyTorch Releases Version 2.3 with Focus on Large Language Models and Sparse Inference

GitHub Secures Millions of Developers Through Two-Factor Authentication

90% of Indian Internet Users are already using AI, says Report

Jensen Huang Personally Delivers First NVIDIA DGX H200 to OpenAI 

Six Months Old Cognition Labs Raises $175 Mn from Founders Fund at $2 Bn Valuation

Apple Releases Four Open Source LLMs with OpenELM Series of Models

Snowflake Releases Open Enterprise LLM, Arctic with 480 Billion Parameters

World's biggest media & analyst firm specializing in ai, advertise with us, aim publishes every day, and we believe in quality over quantity, honesty over spin. we offer a wide variety of branding and targeting options to make it easy for you to propagate your brand., branded content, aim brand solutions, a marketing division within aim, specializes in creating diverse content such as documentaries, public artworks, podcasts, videos, articles, and more to effectively tell compelling stories., corporate upskilling, adasci corporate training program on generative ai provides a unique opportunity to empower, retain and advance your talent, with machinehack you can not only find qualified developers with hiring challenges but can also engage the developer community and your internal workforce by hosting hackathons., talent assessment, conduct customized online assessments on our powerful cloud-based platform, secured with best-in-class proctoring, research & advisory, aim research produces a series of annual reports on ai & data science covering every aspect of the industry. request customised reports & aim surveys for a study on topics of your interest., conferences & events, immerse yourself in ai and business conferences tailored to your role, designed to elevate your performance and empower you to accomplish your organization’s vital objectives., aim launches the 3rd edition of data engineering summit. may 30-31, bengaluru.

© Analytics India Magazine Pvt Ltd & AIM Media House LLC 2024

• Terms of use
• Privacy Policy

Subscribe to Our Newsletter

The Belamy, our weekly Newsletter is a rage. Just enter your email below.

April 18, 2024

AI Report Shows ‘Startlingly Rapid’ Progress—And Ballooning Costs

A new report finds that AI matches or outperforms people at tasks such as competitive math and reading comprehension

By Nicola Jones & Nature magazine

xijian/Getty Images

Artificial intelligence (AI) systems, such as the chatbot ChatGPT , have become so advanced that they now very nearly match or exceed human performance in tasks including reading comprehension, image classification and competition-level mathematics, according to a new report. Rapid progress in the development of these systems also means that many common benchmarks and tests for assessing them are quickly becoming obsolete.

These are just a few of the top-line findings from the Artificial Intelligence Index Report 2024 , which was published on 15 April by the Institute for Human-Centered Artificial Intelligence at Stanford University in California. The report charts the meteoric progress in machine-learning systems over the past decade.

In particular, the report says, new ways of assessing AI — for example, evaluating their performance on complex tasks, such as abstraction and reasoning — are more and more necessary. “A decade ago, benchmarks would serve the community for 5–10 years” whereas now they often become irrelevant in just a few years, says Nestor Maslej, a social scientist at Stanford and editor-in-chief of the AI Index. “The pace of gain has been startlingly rapid.”

On supporting science journalism

If you're enjoying this article, consider supporting our award-winning journalism by subscribing . By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

Stanford’s annual AI Index, first published in 2017, is compiled by a group of academic and industry specialists to assess the field’s technical capabilities, costs, ethics and more — with an eye towards informing researchers, policymakers and the public. This year’s report, which is more than 400 pages long and was copy-edited and tightened with the aid of AI tools, notes that AI-related regulation in the United States is sharply rising. But the lack of standardized assessments for responsible use of AI makes it difficult to compare systems in terms of the risks that they pose.

The rising use of AI in science is also highlighted in this year’s edition: for the first time, it dedicates an entire chapter to science applications, highlighting projects including Graph Networks for Materials Exploration (GNoME), a project from Google DeepMind that aims to help chemists discover materials , and GraphCast, another DeepMind tool, which does rapid weather forecasting .

The current AI boom — built on neural networks and machine-learning algorithms — dates back to the early 2010s . The field has since rapidly expanded. For example, the number of AI coding projects on GitHub, a common platform for sharing code, increased from about 800 in 2011 to 1.8 million last year. And journal publications about AI roughly tripled over this period, the report says.

Much of the cutting-edge work on AI is being done in industry: that sector produced 51 notable machine-learning systems last year, whereas academic researchers contributed 15. “Academic work is shifting to analysing the models coming out of companies — doing a deeper dive into their weaknesses,” says Raymond Mooney, director of the AI Lab at the University of Texas at Austin, who wasn’t involved in the report.

That includes developing tougher tests to assess the visual, mathematical and even moral-reasoning capabilities of large language models (LLMs), which power chatbots. One of the latest tests is the Graduate-Level Google-Proof Q&A Benchmark (GPQA), developed last year by a team including machine-learning researcher David Rein at New York University.

The GPQA, consisting of more than 400 multiple-choice questions, is tough: PhD-level scholars could correctly answer questions in their field 65% of the time. The same scholars, when attempting to answer questions outside their field, scored only 34%, despite having access to the Internet during the test (randomly selecting answers would yield a score of 25%). As of last year, AI systems scored about 30–40%. This year, Rein says, Claude 3 — the latest chatbot released by AI company Anthropic, based in San Francisco, California — scored about 60%. “The rate of progress is pretty shocking to a lot of people, me included,” Rein adds. “It’s quite difficult to make a benchmark that survives for more than a few years.”

Cost of business

As performance is skyrocketing, so are costs. GPT-4 — the LLM that powers ChatGPT and that was released in March 2023 by San Francisco-based firm OpenAI — reportedly cost US$78 million to train. Google’s chatbot Gemini Ultra, launched in December, cost $191 million. Many people are concerned about the energy use of these systems, as well as the amount of water needed to cool the data centres that help to run them. “These systems are impressive, but they’re also very inefficient,” Maslej says.

Costs and energy use for AI models are high in large part because one of the main ways to make current systems better is to make them bigger. This means training them on ever-larger stocks of text and images. The AI Index notes that some researchers now worry about running out of training data. Last year, according to the report, the non-profit research institute Epoch projected that we might exhaust supplies of high-quality language data as soon as this year. (However, the institute’s most recent analysis suggests that 2028 is a better estimate.)

Ethical concerns about how AI is built and used are also mounting. “People are way more nervous about AI than ever before, both in the United States and across the globe,” says Maslej, who sees signs of a growing international divide. “There are now some countries very excited about AI, and others that are very pessimistic.”

In the United States, the report notes a steep rise in regulatory interest. In 2016, there was just one US regulation that mentioned AI; last year, there were 25. “After 2022, there’s a massive spike in the number of AI-related bills that have been proposed” by policymakers, Maslej says.

Regulatory action is increasingly focused on promoting responsible AI use. Although benchmarks are emerging that can score metrics such as an AI tool’s truthfulness, bias and even likability, not everyone is using the same models, Maslej says, which makes cross-comparisons hard. “This is a really important topic,” he says. “We need to bring the community together on this.”

This article is reproduced with permission and was first published on April 15, 2024 .

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Innovation (Camb)
• v.2(4); 2021 Nov 28

Artificial intelligence: A powerful paradigm for scientific research

1 Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China

35 University of Chinese Academy of Sciences, Beijing 100049, China

5 Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China

10 Zhongshan Hospital Institute of Clinical Science, Fudan University, Shanghai 200032, China

Changping Huang

18 Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China

11 Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China

37 Songshan Lake Materials Laboratory, Dongguan, Guangdong 523808, China

26 Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China

Xingchen Liu

28 Institute of Coal Chemistry, Chinese Academy of Sciences, Taiyuan 030001, China

2 Institute of Software, Chinese Academy of Sciences, Beijing 100190, China

Fengliang Dong

3 National Center for Nanoscience and Technology, Beijing 100190, China

Cheng-Wei Qiu

4 Department of Electrical and Computer Engineering, National University of Singapore, Singapore 117583, Singapore

6 Department of Gynaecology, Obstetrics and Gynaecology Hospital, Fudan University, Shanghai 200011, China

36 Shanghai Key Laboratory of Female Reproductive Endocrine-Related Diseases, Shanghai 200011, China

7 School of Food Science and Technology, Dalian Polytechnic University, Dalian 116034, China

41 Second Affiliated Hospital School of Medicine, and School of Public Health, Zhejiang University, Hangzhou 310058, China

8 Department of Obstetrics and Gynecology, Peking University Third Hospital, Beijing 100191, China

9 Zhejiang Provincial People’s Hospital, Hangzhou 310014, China

Chenguang Fu

12 School of Materials Science and Engineering, Zhejiang University, Hangzhou 310027, China

Zhigang Yin

13 Fujian Institute of Research on the Structure of Matter, Chinese Academy of Sciences, Fuzhou 350002, China

Ronald Roepman

14 Medical Center, Radboud University, 6500 Nijmegen, the Netherlands

Sabine Dietmann

15 Institute for Informatics, Washington University School of Medicine, St. Louis, MO 63110, USA

Marko Virta

16 Department of Microbiology, University of Helsinki, 00014 Helsinki, Finland

Fredrick Kengara

17 School of Pure and Applied Sciences, Bomet University College, Bomet 20400, Kenya

19 Agriculture College of Shihezi University, Xinjiang 832000, China

Taolan Zhao

20 Institute of Genetics and Developmental Biology, Chinese Academy of Sciences, Beijing 100101, China

21 The Brain Cognition and Brain Disease Institute, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China

38 Shenzhen-Hong Kong Institute of Brain Science-Shenzhen Fundamental Research Institutions, Shenzhen 518055, China

Jialiang Yang

22 Geneis (Beijing) Co., Ltd, Beijing 100102, China

23 Department of Communication Studies, Hong Kong Baptist University, Hong Kong, China

24 South China Botanical Garden, Chinese Academy of Sciences, Guangzhou 510650, China

39 Center of Economic Botany, Core Botanical Gardens, Chinese Academy of Sciences, Guangzhou 510650, China

Zhaofeng Liu

27 Shanghai Astronomical Observatory, Chinese Academy of Sciences, Shanghai 200030, China

29 Suzhou Institute of Nano-Tech and Nano-Bionics, Chinese Academy of Sciences, Suzhou 215123, China

Xiaohong Liu

30 Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing 400714, China

James P. Lewis

James m. tiedje.

34 Center for Microbial Ecology, Department of Plant, Soil and Microbial Sciences, Michigan State University, East Lansing, MI 48824, USA

40 Zhejiang Lab, Hangzhou 311121, China

25 Shanghai Institute of Nutrition and Health, Chinese Academy of Sciences, Shanghai 200031, China

31 Department of Computer Science, Aberystwyth University, Aberystwyth, Ceredigion SY23 3FL, UK

Zhipeng Cai

32 Department of Computer Science, Georgia State University, Atlanta, GA 30303, USA

33 Institute of Soil Science, Chinese Academy of Sciences, Nanjing 210008, China

Jiabao Zhang

Artificial intelligence (AI) coupled with promising machine learning (ML) techniques well known from computer science is broadly affecting many aspects of various fields including science and technology, industry, and even our day-to-day life. The ML techniques have been developed to analyze high-throughput data with a view to obtaining useful insights, categorizing, predicting, and making evidence-based decisions in novel ways, which will promote the growth of novel applications and fuel the sustainable booming of AI. This paper undertakes a comprehensive survey on the development and application of AI in different aspects of fundamental sciences, including information science, mathematics, medical science, materials science, geoscience, life science, physics, and chemistry. The challenges that each discipline of science meets, and the potentials of AI techniques to handle these challenges, are discussed in detail. Moreover, we shed light on new research trends entailing the integration of AI into each scientific discipline. The aim of this paper is to provide a broad research guideline on fundamental sciences with potential infusion of AI, to help motivate researchers to deeply understand the state-of-the-art applications of AI-based fundamental sciences, and thereby to help promote the continuous development of these fundamental sciences.

Graphical abstract

Public summary

• • “Can machines think?” The goal of artificial intelligence (AI) is to enable machines to mimic human thoughts and behaviors, including learning, reasoning, predicting, and so on.
• • “Can AI do fundamental research?” AI coupled with machine learning techniques is impacting a wide range of fundamental sciences, including mathematics, medical science, physics, etc.
• • “How does AI accelerate fundamental research?” New research and applications are emerging rapidly with the support by AI infrastructure, including data storage, computing power, AI algorithms, and frameworks.

Introduction

“Can machines think?” Alan Turing posed this question in his famous paper “Computing Machinery and Intelligence.” 1 He believes that to answer this question, we need to define what thinking is. However, it is difficult to define thinking clearly, because thinking is a subjective behavior. Turing then introduced an indirect method to verify whether a machine can think, the Turing test, which examines a machine's ability to show intelligence indistinguishable from that of human beings. A machine that succeeds in the test is qualified to be labeled as artificial intelligence (AI).

AI refers to the simulation of human intelligence by a system or a machine. The goal of AI is to develop a machine that can think like humans and mimic human behaviors, including perceiving, reasoning, learning, planning, predicting, and so on. Intelligence is one of the main characteristics that distinguishes human beings from animals. With the interminable occurrence of industrial revolutions, an increasing number of types of machine types continuously replace human labor from all walks of life, and the imminent replacement of human resources by machine intelligence is the next big challenge to be overcome. Numerous scientists are focusing on the field of AI, and this makes the research in the field of AI rich and diverse. AI research fields include search algorithms, knowledge graphs, natural languages processing, expert systems, evolution algorithms, machine learning (ML), deep learning (DL), and so on.

The general framework of AI is illustrated in Figure 1 . The development process of AI includes perceptual intelligence, cognitive intelligence, and decision-making intelligence. Perceptual intelligence means that a machine has the basic abilities of vision, hearing, touch, etc., which are familiar to humans. Cognitive intelligence is a higher-level ability of induction, reasoning and acquisition of knowledge. It is inspired by cognitive science, brain science, and brain-like intelligence to endow machines with thinking logic and cognitive ability similar to human beings. Once a machine has the abilities of perception and cognition, it is often expected to make optimal decisions as human beings, to improve the lives of people, industrial manufacturing, etc. Decision intelligence requires the use of applied data science, social science, decision theory, and managerial science to expand data science, so as to make optimal decisions. To achieve the goal of perceptual intelligence, cognitive intelligence, and decision-making intelligence, the infrastructure layer of AI, supported by data, storage and computing power, ML algorithms, and AI frameworks is required. Then by training models, it is able to learn the internal laws of data for supporting and realizing AI applications. The application layer of AI is becoming more and more extensive, and deeply integrated with fundamental sciences, industrial manufacturing, human life, social governance, and cyberspace, which has a profound impact on our work and lifestyle.

The general framework of AI

History of AI

The beginning of modern AI research can be traced back to John McCarthy, who coined the term “artificial intelligence (AI),” during at a conference at Dartmouth College in 1956. This symbolized the birth of the AI scientific field. Progress in the following years was astonishing. Many scientists and researchers focused on automated reasoning and applied AI for proving of mathematical theorems and solving of algebraic problems. One of the famous examples is Logic Theorist, a computer program written by Allen Newell, Herbert A. Simon, and Cliff Shaw, which proves 38 of the first 52 theorems in “Principia Mathematica” and provides more elegant proofs for some. 2 These successes made many AI pioneers wildly optimistic, and underpinned the belief that fully intelligent machines would be built in the near future. However, they soon realized that there was still a long way to go before the end goals of human-equivalent intelligence in machines could come true. Many nontrivial problems could not be handled by the logic-based programs. Another challenge was the lack of computational resources to compute more and more complicated problems. As a result, organizations and funders stopped supporting these under-delivering AI projects.

AI came back to popularity in the 1980s, as several research institutions and universities invented a type of AI systems that summarizes a series of basic rules from expert knowledge to help non-experts make specific decisions. These systems are “expert systems.” Examples are the XCON designed by Carnegie Mellon University and the MYCIN designed by Stanford University. The expert system derived logic rules from expert knowledge to solve problems in the real world for the first time. The core of AI research during this period is the knowledge that made machines “smarter.” However, the expert system gradually revealed several disadvantages, such as privacy technologies, lack of flexibility, poor versatility, expensive maintenance cost, and so on. At the same time, the Fifth Generation Computer Project, heavily funded by the Japanese government, failed to meet most of its original goals. Once again, the funding for AI research ceased, and AI was at the second lowest point of its life.

In 2006, Geoffrey Hinton and coworkers 3 , 4 made a breakthrough in AI by proposing an approach of building deeper neural networks, as well as a way to avoid gradient vanishing during training. This reignited AI research, and DL algorithms have become one of the most active fields of AI research. DL is a subset of ML based on multiple layers of neural networks with representation learning, 5 while ML is a part of AI that a computer or a program can use to learn and acquire intelligence without human intervention. Thus, “learn” is the keyword of this era of AI research. Big data technologies, and the improvement of computing power have made deriving features and information from massive data samples more efficient. An increasing number of new neural network structures and training methods have been proposed to improve the representative learning ability of DL, and to further expand it into general applications. Current DL algorithms match and exceed human capabilities on specific datasets in the areas of computer vision (CV) and natural language processing (NLP). AI technologies have achieved remarkable successes in all walks of life, and continued to show their value as backbones in scientific research and real-world applications.

Within AI, ML is having a substantial broad effect across many aspects of technology and science: from computer science to geoscience to materials science, from life science to medical science to chemistry to mathematics and to physics, from management science to economics to psychology, and other data-intensive empirical sciences, as ML methods have been developed to analyze high-throughput data to obtain useful insights, categorize, predict, and make evidence-based decisions in novel ways. To train a system by presenting it with examples of desired input-output behavior, could be far easier than to program it manually by predicting the desired response for all potential inputs. The following sections survey eight fundamental sciences, including information science (informatics), mathematics, medical science, materials science, geoscience, life science, physics, and chemistry, which develop or exploit AI techniques to promote the development of sciences and accelerate their applications to benefit human beings, society, and the world.

AI in information science

AI aims to provide the abilities of perception, cognition, and decision-making for machines. At present, new research and applications in information science are emerging at an unprecedented rate, which is inseparable from the support by the AI infrastructure. As shown in Figure 2 , the AI infrastructure layer includes data, storage and computing power, ML algorithms, and the AI framework. The perception layer enables machines have the basic ability of vision, hearing, etc. For instance, CV enables machines to “see” and identify objects, while speech recognition and synthesis helps machines to “hear” and recognize speech elements. The cognitive layer provides higher ability levels of induction, reasoning, and acquiring knowledge with the help of NLP, 6 knowledge graphs, 7 and continual learning. 8 In the decision-making layer, AI is capable of making optimal decisions, such as automatic planning, expert systems, and decision-supporting systems. Numerous applications of AI have had a profound impact on fundamental sciences, industrial manufacturing, human life, social governance, and cyberspace. The following subsections provide an overview of the AI framework, automatic machine learning (AutoML) technology, and several state-of-the-art AI/ML applications in the information field.

The knowledge graph of the AI framework

The AI framework provides basic tools for AI algorithm implementation

In the past 10 years, applications based on AI algorithms have played a significant role in various fields and subjects, on the basis of which the prosperity of the DL framework and platform has been founded. AI frameworks and platforms reduce the requirement of accessing AI technology by integrating the overall process of algorithm development, which enables researchers from different areas to use it across other fields, allowing them to focus on designing the structure of neural networks, thus providing better solutions to problems in their fields. At the beginning of the 21st century, only a few tools, such as MATLAB, OpenNN, and Torch, were capable of describing and developing neural networks. However, these tools were not originally designed for AI models, and thus faced problems, such as complicated user API and lacking GPU support. During this period, using these frameworks demanded professional computer science knowledge and tedious work on model construction. As a solution, early frameworks of DL, such as Caffe, Chainer, and Theano, emerged, allowing users to conveniently construct complex deep neural networks (DNNs), such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and LSTM conveniently, and this significantly reduced the cost of applying AI models. Tech giants then joined the march in researching AI frameworks. 9 Google developed the famous open-source framework, TensorFlow, while Facebook's AI research team released another popular platform, PyTorch, which is based on Torch; Microsoft Research published CNTK, and Amazon announced MXNet. Among them, TensorFlow, also the most representative framework, referred to Theano's declarative programming style, offering a larger space for graph-based optimization, while PyTorch inherited the imperative programming style of Torch, which is intuitive, user friendly, more flexible, and easier to be traced. As modern AI frameworks and platforms are being widely applied, practitioners can now assemble models swiftly and conveniently by adopting various building block sets and languages specifically suitable for given fields. Polished over time, these platforms gradually developed a clearly defined user API, the ability for multi-GPU training and distributed training, as well as a variety of model zoos and tool kits for specific tasks. 10 Looking forward, there are a few trends that may become the mainstream of next-generation framework development. (1) Capability of super-scale model training. With the emergence of models derived from Transformer, such as BERT and GPT-3, the ability of training large models has become an ideal feature of the DL framework. It requires AI frameworks to train effectively under the scale of hundreds or even thousands of devices. (2) Unified API standard. The APIs of many frameworks are generally similar but slightly different at certain points. This leads to some difficulties and unnecessary learning efforts, when the user attempts to shift from one framework to another. The API of some frameworks, such as JAX, has already become compatible with Numpy standard, which is familiar to most practitioners. Therefore, a unified API standard for AI frameworks may gradually come into being in the future. (3) Universal operator optimization. At present, kernels of DL operator are implemented either manually or based on third-party libraries. Most third-party libraries are developed to suit certain hardware platforms, causing large unnecessary spending when models are trained or deployed on different hardware platforms. The development speed of new DL algorithms is usually much faster than the update rate of libraries, which often makes new algorithms to be beyond the range of libraries' support. 11

To improve the implementation speed of AI algorithms, much research focuses on how to use hardware for acceleration. The DianNao family is one of the earliest research innovations on AI hardware accelerators. 12 It includes DianNao, DaDianNao, ShiDianNao, and PuDianNao, which can be used to accelerate the inference speed of neural networks and other ML algorithms. Of these, the best performance of a 64-chip DaDianNao system can achieve a speed up of 450.65× over a GPU, and reduce the energy by 150.31×. Prof. Chen and his team in the Institute of Computing Technology also designed an Instruction Set Architecture for a broad range of neural network accelerators, called Cambricon, which developed into a serial DL accelerator. After Cambricon, many AI-related companies, such as Apple, Google, HUAWEI, etc., developed their own DL accelerators, and AI accelerators became an important research field of AI.

AI for AI—AutoML

AutoML aims to study how to use evolutionary computing, reinforcement learning (RL), and other AI algorithms, to automatically generate specified AI algorithms. Research on the automatic generation of neural networks has existed before the emergence of DL, e.g., neural evolution. 13 The main purpose of neural evolution is to allow neural networks to evolve according to the principle of survival of the fittest in the biological world. Through selection, crossover, mutation, and other evolutionary operators, the individual quality in a population is continuously improved and, finally, the individual with the greatest fitness represents the best neural network. The biological inspiration in this field lies in the evolutionary process of human brain neurons. The human brain has such developed learning and memory functions that it cannot do without the complex neural network system in the brain. The whole neural network system of the human brain benefits from a long evolutionary process rather than gradient descent and back propagation. In the era of DL, the application of AI algorithms to automatically generate DNN has attracted more attention and, gradually, developed into an important direction of AutoML research: neural architecture search. The implementation methods of neural architecture search are usually divided into the RL-based method and the evolutionary algorithm-based method. In the RL-based method, an RNN is used as a controller to generate a neural network structure layer by layer, and then the network is trained, and the accuracy of the verification set is used as the reward signal of the RNN to calculate the strategy gradient. During the iteration, the controller will give the neural network, with higher accuracy, a higher probability value, so as to ensure that the strategy function can output the optimal network structure. 14 The method of neural architecture search through evolution is similar to the neural evolution method, which is based on a population and iterates continuously according to the principle of survival of the fittest, so as to obtain a high-quality neural network. 15 Through the application of neural architecture search technology, the design of neural networks is more efficient and automated, and the accuracy of the network gradually outperforms that of the networks designed by AI experts. For example, Google's SOTA network EfficientNet was realized through the baseline network based on neural architecture search. 16

AI enabling networking design adaptive to complex network conditions

The application of DL in the networking field has received strong interest. Network design often relies on initial network conditions and/or theoretical assumptions to characterize real network environments. However, traditional network modeling and design, regulated by mathematical models, are unlikely to deal with complex scenarios with many imperfect and high dynamic network environments. Integrating DL into network research allows for a better representation of complex network environments. Furthermore, DL could be combined with the Markov decision process and evolve into the deep reinforcement learning (DRL) model, which finds an optimal policy based on the reward function and the states of the system. Taken together, these techniques could be used to make better decisions to guide proper network design, thereby improving the network quality of service and quality of experience. With regard to the aspect of different layers of the network protocol stack, DL/DRL can be adopted for network feature extraction, decision-making, etc. In the physical layer, DL can be used for interference alignment. It can also be used to classify the modulation modes, design efficient network coding 17 and error correction codes, etc. In the data link layer, DL can be used for resource (such as channels) allocation, medium access control, traffic prediction, 18 link quality evaluation, and so on. In the network (routing) layer, routing establishment and routing optimization 19 can help to obtain an optimal routing path. In higher layers (such as the application layer), enhanced data compression and task allocation is used. Besides the above protocol stack, one critical area of using DL is network security. DL can be used to classify the packets into benign/malicious types, and how it can be integrated with other ML schemes, such as unsupervised clustering, to achieve a better anomaly detection effect.

AI enabling more powerful and intelligent nanophotonics

Nanophotonic components have recently revolutionized the field of optics via metamaterials/metasurfaces by enabling the arbitrary manipulation of light-matter interactions with subwavelength meta-atoms or meta-molecules. 20 , 21 , 22 The conventional design of such components involves generally forward modeling, i.e., solving Maxwell's equations based on empirical and intuitive nanostructures to find corresponding optical properties, as well as the inverse design of nanophotonic devices given an on-demand optical response. The trans-dimensional feature of macro-optical components consisting of complex nano-antennas makes the design process very time consuming, computationally expensive, and even numerically prohibitive, such as device size and complexity increase. DL is an efficient and automatic platform, enabling novel efficient approaches to designing nanophotonic devices with high-performance and versatile functions. Here, we present briefly the recent progress of DL-based nanophotonics and its wide-ranging applications. DL was exploited for forward modeling at first using a DNN. 23 The transmission or reflection coefficients can be well predicted after training on huge datasets. To improve the prediction accuracy of DNN in case of small datasets, transfer learning was introduced to migrate knowledge between different physical scenarios, which greatly reduced the relative error. Furthermore, a CNN and an RNN were developed for the prediction of optical properties from arbitrary structures using images. 24 The CNN-RNN combination successfully predicted the absorption spectra from the given input structural images. In inverse design of nanophotonic devices, there are three different paradigms of DL methods, i.e., supervised, unsupervised, and RL. 25 Supervised learning has been utilized to design structural parameters for the pre-defined geometries, such as tandem DNN and bidirectional DNNs. Unsupervised learning methods learn by themselves without a specific target, and thus are more accessible to discovering new and arbitrary patterns 26 in completely new data than supervised learning. A generative adversarial network (GAN)-based approach, combining conditional GANs and Wasserstein GANs, was proposed to design freeform all-dielectric multifunctional metasurfaces. RL, especially double-deep Q-learning, powers up the inverse design of high-performance nanophotonic devices. 27 DL has endowed nanophotonic devices with better performance and more emerging applications. 28 , 29 For instance, an intelligent microwave cloak driven by DL exhibits millisecond and self-adaptive response to an ever-changing incident wave and background. 28 Another example is that a DL-augmented infrared nanoplasmonic metasurface is developed for monitoring dynamics between four major classes of bio-molecules, which could impact the fields of biology, bioanalytics, and pharmacology from fundamental research, to disease diagnostics, to drug development. 29 The potential of DL in the wide arena of nanophotonics has been unfolding. Even end-users without optics and photonics background could exploit the DL as a black box toolkit to design powerful optical devices. Nevertheless, how to interpret/mediate the intermediate DL process and determine the most dominant factors in the search for optimal solutions, are worthy of being investigated in depth. We optimistically envisage that the advancements in DL algorithms and computation/optimization infrastructures would enable us to realize more efficient and reliable training approaches, more complex nanostructures with unprecedented shapes and sizes, and more intelligent and reconfigurable optic/optoelectronic systems.

AI in other fields of information science

We believe that AI has great potential in the following directions:

• • AI-based risk control and management in utilities can prevent costly or hazardous equipment failures by using sensors that detect and send information regarding the machine's health to the manufacturer, predicting possible issues that could occur so as to ensure timely maintenance or automated shutdown.
• • AI could be used to produce simulations of real-world objects, called digital twins. When applied to the field of engineering, digital twins allow engineers and technicians to analyze the performance of an equipment virtually, thus avoiding safety and budget issues associated with traditional testing methods.
• • Combined with AI, intelligent robots are playing an important role in industry and human life. Different from traditional robots working according to the procedures specified by humans, intelligent robots have the ability of perception, recognition, and even automatic planning and decision-making, based on changes in environmental conditions.
• • AI of things (AIoT) or AI-empowered IoT applications. 30 have become a promising development trend. AI can empower the connected IoT devices, embedded in various physical infrastructures, to perceive, recognize, learn, and act. For instance, smart cities constantly collect data regarding quality-of-life factors, such as the status of power supply, public transportation, air pollution, and water use, to manage and optimize systems in cities. Due to these data, especially personal data being collected from informed or uninformed participants, data security, and privacy 31 require protection.

AI in mathematics

Mathematics always plays a crucial and indispensable role in AI. Decades ago, quite a few classical AI-related approaches, such as k-nearest neighbor, 32 support vector machine, 33 and AdaBoost, 34 were proposed and developed after their rigorous mathematical formulations had been established. In recent years, with the rapid development of DL, 35 AI has been gaining more and more attention in the mathematical community. Equipped with the Markov process, minimax optimization, and Bayesian statistics, RL, 36 GANs, 37 and Bayesian learning 38 became the most favorable tools in many AI applications. Nevertheless, there still exist plenty of open problems in mathematics for ML, including the interpretability of neural networks, the optimization problems of parameter estimation, and the generalization ability of learning models. In the rest of this section, we discuss these three questions in turn.

The interpretability of neural networks

From a mathematical perspective, ML usually constructs nonlinear models, with neural networks as a typical case, to approximate certain functions. The well-known Universal Approximation Theorem suggests that, under very mild conditions, any continuous function can be uniformly approximated on compact domains by neural networks, 39 which serves a vital function in the interpretability of neural networks. However, in real applications, ML models seem to admit accurate approximations of many extremely complicated functions, sometimes even black boxes, which are far beyond the scope of continuous functions. To understand the effectiveness of ML models, many researchers have investigated the function spaces that can be well approximated by them, and the corresponding quantitative measures. This issue is closely related to the classical approximation theory, but the approximation scheme is distinct. For example, Bach 40 finds that the random feature model is naturally associated with the corresponding reproducing kernel Hilbert space. In the same way, the Barron space is identified as the natural function space associated with two-layer neural networks, and the approximation error is measured using the Barron norm. 41 The corresponding quantities of residual networks (ResNets) are defined for the flow-induced spaces. For multi-layer networks, the natural function spaces for the purposes of approximation theory are the tree-like function spaces introduced in Wojtowytsch. 42 There are several works revealing the relationship between neural networks and numerical algorithms for solving partial differential equations. For example, He and Xu 43 discovered that CNNs for image classification have a strong connection with multi-grid (MG) methods. In fact, the pooling operation and feature extraction in CNNs correspond directly to restriction operation and iterative smoothers in MG, respectively. Hence, various convolution and pooling operations used in CNNs can be better understood.

The optimization problems of parameter estimation

In general, the optimization problem of estimating parameters of certain DNNs is in practice highly nonconvex and often nonsmooth. Can the global minimizers be expected? What is the landscape of local minimizers? How does one handle the nonsmoothness? All these questions are nontrivial from an optimization perspective. Indeed, numerous works and experiments demonstrate that the optimization for parameter estimation in DL is itself a much nicer problem than once thought; see, e.g., Goodfellow et al. 44 As a consequence, the study on the solution landscape ( Figure 3 ), also known as loss surface of neural networks, is no longer supposed to be inaccessible and can even in turn provide guidance for global optimization. Interested readers can refer to the survey paper (Sun et al. 45 ) for recent progress in this aspect.

Recent studies indicate that nonsmooth activation functions, e.g., rectified linear units, are better than smooth ones in finding sparse solutions. However, the chain rule does not work in the case that the activation functions are nonsmooth, which then makes the widely used stochastic gradient (SG)-based approaches not feasible in theory. Taking approximated gradients at nonsmooth iterates as a remedy ensures that SG-type methods are still in extensive use, but that the numerical evidence has also exposed their limitations. Also, the penalty-based approaches proposed by Cui et al. 46 and Liu et al. 47 provide a new direction to solve the nonsmooth optimization problems efficiently.

The generalization ability of learning models

A small training error does not always lead to a small test error. This gap is caused by the generalization ability of learning models. A key finding in statistical learning theory states that the generalization error is bounded by a quantity that grows with the increase of the model capacity, but shrinks as the number of training examples increases. 48 A common conjecture relating generalization to solution landscape is that flat and wide minima generalize better than sharp ones. Thus, regularization techniques, including the dropout approach, 49 have emerged to force the algorithms to bypass the sharp minima. However, the mechanism behind this has not been fully explored. Recently, some researchers have focused on the ResNet-type architecture, with dropout being inserted after the last convolutional layer of each modular building. They thus managed to explain the stochastic dropout training process and the ensuing dropout regularization effect from the perspective of optimal control. 50

AI in medical science

There is a great trend for AI technology to grow more and more significant in daily operations, including medical fields. With the growing needs of healthcare for patients, hospital needs are evolving from informationization networking to the Internet Hospital and eventually to the Smart Hospital. At the same time, AI tools and hardware performance are also growing rapidly with each passing day. Eventually, common AI algorithms, such as CV, NLP, and data mining, will begin to be embedded in the medical equipment market ( Figure 4 ).

AI doctor based on electronic medical records

For medical history data, it is inevitable to mention Doctor Watson, developed by the Watson platform of IBM, and Modernizing Medicine, which aims to solve oncology, and is now adopted by CVS & Walgreens in the US and various medical organizations in China as well. Doctor Watson takes advantage of the NLP performance of the IBM Watson platform, which already collected vast data of medical history, as well as prior knowledge in the literature for reference. After inputting the patients' case, Doctor Watson searches the medical history reserve and forms an elementary treatment proposal, which will be further ranked by prior knowledge reserves. With the multiple models stored, Doctor Watson gives the final proposal as well as the confidence of the proposal. However, there are still problems for such AI doctors because, 51 as they rely on prior experience from US hospitals, the proposal may not be suitable for other regions with different medical insurance policies. Besides, the knowledge updating of the Watson platform also relies highly on the updating of the knowledge reserve, which still needs manual work.

AI for public health: Outbreak detection and health QR code for COVID-19

AI can be used for public health purposes in many ways. One classical usage is to detect disease outbreaks using search engine query data or social media data, as Google did for prediction of influenza epidemics 52 and the Chinese Academy of Sciences did for modeling the COVID-19 outbreak through multi-source information fusion. 53 After the COVID-19 outbreak, a digital health Quick Response (QR) code system has been developed by China, first to detect potential contact with confirmed COVID-19 cases and, secondly, to indicate the person's health status using mobile big data. 54 Different colors indicate different health status: green means healthy and is OK for daily life, orange means risky and requires quarantine, and red means confirmed COVID-19 patient. It is easy to use for the general public, and has been adopted by many other countries. The health QR code has made great contributions to the worldwide prevention and control of the COVID-19 pandemic.

Biomarker discovery with AI

High-dimensional data, including multi-omics data, patient characteristics, medical laboratory test data, etc., are often used for generating various predictive or prognostic models through DL or statistical modeling methods. For instance, the COVID-19 severity evaluation model was built through ML using proteomic and metabolomic profiling data of sera 55 ; using integrated genetic, clinical, and demographic data, Taliaz et al. built an ML model to predict patient response to antidepressant medications 56 ; prognostic models for multiple cancer types (such as liver cancer, lung cancer, breast cancer, gastric cancer, colorectal cancer, pancreatic cancer, prostate cancer, ovarian cancer, lymphoma, leukemia, sarcoma, melanoma, bladder cancer, renal cancer, thyroid cancer, head and neck cancer, etc.) were constructed through DL or statistical methods, such as least absolute shrinkage and selection operator （LASSO）, combined with Cox proportional hazards regression model using genomic data. 57

Image-based medical AI

Medical image AI is one of the most developed mature areas as there are numerous models for classification, detection, and segmentation tasks in CV. For the clinical area, CV algorithms can also be used for computer-aided diagnosis and treatment with ECG, CT, eye fundus imaging, etc. As human doctors may be tired and prone to make mistakes after viewing hundreds and hundreds of images for diagnosis, AI doctors can outperform a human medical image viewer due to their specialty at repeated work without fatigue. The first medical AI product approved by FDA is IDx-DR, which uses an AI model to make predictions of diabetic retinopathy. The smartphone app SkinVision can accurately detect melanomas. 58 It uses “fractal analysis” to identify moles and their surrounding skin, based on size, diameter, and many other parameters, and to detect abnormal growth trends. AI-ECG of LEPU Medical can automatically detect heart disease with ECG images. Lianying Medical takes advantage of their hardware equipment to produce real-time high-definition image-guided all-round radiotherapy technology, which successfully achieves precise treatment.

Wearable devices for surveillance and early warning

For wearable devices, AliveCor has developed an algorithm to automatically predict the presence of atrial fibrillation, which is an early warning sign of stroke and heart failure. The 23andMe company can also test saliva samples at a small cost, and a customer can be provided with information based on their genes, including who their ancestors were or potential diseases they may be prone to later in life. It provides accurate health management solutions based on individual and family genetic data. In the 20–30 years of the near feature, we believe there are several directions for further research: (1) causal inference for real-time in-hospital risk prediction. Clinical doctors usually acquire reasonable explanations for certain medical decisions, but the current AI models nowadays are usually black box models. The casual inference will help doctors to explain certain AI decisions and even discover novel ground truths. (2) Devices, including wearable instruments for multi-dimensional health monitoring. The multi-modality model is now a trend for AI research. With various devices to collect multi-modality data and a central processor to fuse all these data, the model can monitor the user's overall real-time health condition and give precautions more precisely. (3) Automatic discovery of clinical markers for diseases that are difficult to diagnose. Diseases, such as ALS, are still difficult for clinical doctors to diagnose because they lack any effective general marker. It may be possible for AI to discover common phenomena for these patients and find an effective marker for early diagnosis.

AI-aided drug discovery

Today we have come into the precision medicine era, and the new targeted drugs are the cornerstones for precision therapy. However, over the past decades, it takes an average of over one billion dollars and 10 years to bring a new drug into the market. How to accelerate the drug discovery process, and avoid late-stage failure, are key concerns for all the big and fiercely competitive pharmaceutical companies. The highlighted emerging role of AI, including ML, DL, expert systems, and artificial neural networks (ANNs), has brought new insights and high efficiency into the new drug discovery processes. AI has been adopted in many aspects of drug discovery, including de novo molecule design, structure-based modeling for proteins and ligands, quantitative structure-activity relationship research, and druggable property judgments. DL-based AI appliances demonstrate superior merits in addressing some challenging problems in drug discovery. Of course, prediction of chemical synthesis routes and chemical process optimization are also valuable in accelerating new drug discovery, as well as lowering production costs.

There has been notable progress in the AI-aided new drug discovery in recent years, for both new chemical entity discovery and the relating business area. Based on DNNs, DeepMind built the AlphaFold platform to predict 3D protein structures that outperformed other algorithms. As an illustration of great achievement, AlphaFold successfully and accurately predicted 25 scratch protein structures from a 43 protein panel without using previously built proteins models. Accordingly, AlphaFold won the CASP13 protein-folding competition in December 2018. 59 Based on the GANs and other ML methods, Insilico constructed a modular drug design platform GENTRL system. In September 2019, they reported the discovery of the first de novo active DDR1 kinase inhibitor developed by the GENTRL system. It took the team only 46 days from target selection to get an active drug candidate using in vivo data. 60 Exscientia and Sumitomo Dainippon Pharma developed a new drug candidate, DSP-1181, for the treatment of obsessive-compulsive disorder on the Centaur Chemist AI platform. In January 2020, DSP-1181 started its phase I clinical trials, which means that, from program initiation to phase I study, the comprehensive exploration took less than 12 months. In contrast, comparable drug discovery using traditional methods usually needs 4–5 years with traditional methods.

How AI transforms medical practice: A case study of cervical cancer

As the most common malignant tumor in women, cervical cancer is a disease that has a clear cause and can be prevented, and even treated, if detected early. Conventionally, the screening strategy for cervical cancer mainly adopts the “three-step” model of “cervical cytology-colposcopy-histopathology.” 61 However, limited by the level of testing methods, the efficiency of cervical cancer screening is not high. In addition, owing to the lack of knowledge by doctors in some primary hospitals, patients cannot be provided with the best diagnosis and treatment decisions. In recent years, with the advent of the era of computer science and big data, AI has gradually begun to extend and blend into various fields. In particular, AI has been widely used in a variety of cancers as a new tool for data mining. For cervical cancer, a clinical database with millions of medical records and pathological data has been built, and an AI medical tool set has been developed. 62 Such an AI analysis algorithm supports doctors to access the ability of rapid iterative AI model training. In addition, a prognostic prediction model established by ML and a web-based prognostic result calculator have been developed, which can accurately predict the risk of postoperative recurrence and death in cervical cancer patients, and thereby better guide decision-making in postoperative adjuvant treatment. 63

AI in materials science

As the cornerstone of modern industry, materials have played a crucial role in the design of revolutionary forms of matter, with targeted properties for broad applications in energy, information, biomedicine, construction, transportation, national security, spaceflight, and so forth. Traditional strategies rely on the empirical trial and error experimental approaches as well as the theoretical simulation methods, e.g., density functional theory, thermodynamics, or molecular dynamics, to discover novel materials. 64 These methods often face the challenges of long research cycles, high costs, and low success rates, and thus cannot meet the increasingly growing demands of current materials science. Accelerating the speed of discovery and deployment of advanced materials will therefore be essential in the coming era.

With the rapid development of data processing and powerful algorithms, AI-based methods, such as ML and DL, are emerging with good potentials in the search for and design of new materials prior to actually manufacturing them. 65 , 66 By integrating material property data, such as the constituent element, lattice symmetry, atomic radius, valence, binding energy, electronegativity, magnetism, polarization, energy band, structure-property relation, and functionalities, the machine can be trained to “think” about how to improve material design and even predict the properties of new materials in a cost-effective manner ( Figure 5 ).

AI is expected to power the development of materials science

AI in discovery and design of new materials

Recently, AI techniques have made significant advances in rational design and accelerated discovery of various materials, such as piezoelectric materials with large electrostrains, 67 organic-inorganic perovskites for photovoltaics, 68 molecular emitters for efficient light-emitting diodes, 69 inorganic solid materials for thermoelectrics, 70 and organic electronic materials for renewable-energy applications. 66 , 71 The power of data-driven computing and algorithmic optimization can promote comprehensive applications of simulation and ML (i.e., high-throughput virtual screening, inverse molecular design, Bayesian optimization, and supervised learning, etc.), in material discovery and property prediction in various fields. 72 For instance, using a DL Bayesian framework, the attribute-driven inverse materials design has been demonstrated for efficient and accurate prediction of functional molecular materials, with desired semiconducting properties or redox stability for applications in organic thin-film transistors, organic solar cells, or lithium-ion batteries. 73 It is meaningful to adopt automation tools for quick experimental testing of potential materials and utilize high-performance computing to calculate their bulk, interface, and defect-related properties. 74 The effective convergence of automation, computing, and ML can greatly speed up the discovery of materials. In the future, with the aid of AI techniques, it will be possible to accomplish the design of superconductors, metallic glasses, solder alloys, high-entropy alloys, high-temperature superalloys, thermoelectric materials, two-dimensional materials, magnetocaloric materials, polymeric bio-inspired materials, sensitive composite materials, and topological (electronic and phonon) materials, and so on. In the past decade, topological materials have ignited the research enthusiasm of condensed matter physicists, materials scientists, and chemists, as they exhibit exotic physical properties with potential applications in electronics, thermoelectrics, optics, catalysis, and energy-related fields. From the most recent predictions, more than a quarter of all inorganic materials in nature are topologically nontrivial. The establishment of topological electronic materials databases 75 , 76 , 77 and topological phononic materials databases 78 using high-throughput methods will help to accelerate the screening and experimental discovery of new topological materials for functional applications. It is recognized that large-scale high-quality datasets are required to practice AI. Great efforts have also been expended in building high-quality materials science databases. As one of the top-ranking databases of its kind, the “atomly.net” materials data infrastructure, 79 has calculated the properties of more than 180,000 inorganic compounds, including their equilibrium structures, electron energy bands, dielectric properties, simulated diffraction patterns, elasticity tensors, etc. As such, the atomly.net database has set a solid foundation for extending AI into the area of materials science research. The X-ray diffraction (XRD)-matcher model of atomly.net uses ML to match and classify the experimental XRD to the simulated patterns. Very recently, by using the dataset from atomly.net, an accurate AI model was built to rapidly predict the formation energy of almost any given compound to yield a fairly good predictive ability. 80

AI-powered Materials Genome Initiative

The Materials Genome Initiative (MGI) is a great plan for rational realization of new materials and related functions, and it aims to discover, manufacture, and deploy advanced materials efficiently, cost-effectively, and intelligently. The initiative creates policy, resources, and infrastructure for accelerating materials development at a high level. This is a new paradigm for the discovery and design of next-generation materials, and runs from a view point of fundamental building blocks toward general materials developments, and accelerates materials development through efforts in theory, computation, and experiment, in a highly integrated high-throughput manner. MGI raises an ultimately high goal and high level for materials development and materials science for humans in the future. The spirit of MGI is to design novel materials by using data pools and powerful computation once the requirements or aspirations of functional usages appear. The theory, computation, and algorithm are the primary and substantial factors in the establishment and implementation of MGI. Advances in theories, computations, and experiments in materials science and engineering provide the footstone to not only accelerate the speed at which new materials are realized but to also shorten the time needed to push new products into the market. These AI techniques bring a great promise to the developing MGI. The applications of new technologies, such as ML and DL, directly accelerate materials research and the establishment of MGI. The model construction and application to science and engineering, as well as the data infrastructure, are of central importance. When the AI-powered MGI approaches are coupled with the ongoing autonomy of manufacturing methods, the potential impact to society and the economy in the future is profound. We are now beginning to see that the AI-aided MGI, among other things, integrates experiments, computation, and theory, and facilitates access to materials data, equips the next generation of the materials workforce, and enables a paradigm shift in materials development. Furthermore, the AI-powdered MGI could also design operational procedures and control the equipment to execute experiments, and to further realize autonomous experimentation in future material research.

Advanced functional materials for generation upgrade of AI

The realization and application of AI techniques depend on the computational capability and computer hardware, and this bases physical functionality on the performance of computers or supercomputers. For our current technology, the electric currents or electric carriers for driving electric chips and devices consist of electrons with ordinary characteristics, such as heavy mass and low mobility. All chips and devices emit relatively remarkable heat levels, consuming too much energy and lowering the efficiency of information transmission. Benefiting from the rapid development of modern physics, a series of advanced materials with exotic functional effects have been discovered or designed, including superconductors, quantum anomalous Hall insulators, and topological fermions. In particular, the superconducting state or topologically nontrivial electrons will promote the next-generation AI techniques once the (near) room temperature applications of these states are realized and implanted in integrated circuits. 81 In this case, the central processing units, signal circuits, and power channels will be driven based on the electronic carriers that show massless, energy-diffusionless, ultra-high mobility, or chiral-protection characteristics. The ordinary electrons will be removed from the physical circuits of future-generation chips and devices, leaving superconducting and topological chiral electrons running in future AI chips and supercomputers. The efficiency of transmission, for information and logic computing will be improved on a vast scale and at a very low cost.

AI for materials and materials for AI

The coming decade will continue to witness the development of advanced ML algorithms, newly emerging data-driven AI methodologies, and integrated technologies for facilitating structure design and property prediction, as well as to accelerate the discovery, design, development, and deployment of advanced materials into existing and emerging industrial sectors. At this moment, we are facing challenges in achieving accelerated materials research through the integration of experiment, computation, and theory. The great MGI, proposed for high-level materials research, helps to promote this process, especially when it is assisted by AI techniques. Still, there is a long way to go for the usage of these advanced functional materials in future-generation electric chips and devices to be realized. More materials and functional effects need to be discovered or improved by the developing AI techniques. Meanwhile, it is worth noting that materials are the core components of devices and chips that are used for construction of computers or machines for advanced AI systems. The rapid development of new materials, especially the emergence of flexible, sensitive, and smart materials, is of great importance for a broad range of attractive technologies, such as flexible circuits, stretchable tactile sensors, multifunctional actuators, transistor-based artificial synapses, integrated networks of semiconductor/quantum devices, intelligent robotics, human-machine interactions, simulated muscles, biomimetic prostheses, etc. These promising materials, devices, and integrated technologies will greatly promote the advancement of AI systems toward wide applications in human life. Once the physical circuits are upgraded by advanced functional or smart materials, AI techniques will largely promote the developments and applications of all disciplines.

AI in geoscience

Ai technologies involved in a large range of geoscience fields.

Momentous challenges threatening current society require solutions to problems that belong to geoscience, such as evaluating the effects of climate change, assessing air quality, forecasting the effects of disaster incidences on infrastructure, by calculating the incoming consumption and availability of food, water, and soil resources, and identifying factors that are indicators for potential volcanic eruptions, tsunamis, floods, and earthquakes. 82 , 83 It has become possible, with the emergence of advanced technology products (e.g., deep sea drilling vessels and remote sensing satellites), for enhancements in computational infrastructure that allow for processing large-scale, wide-range simulations of multiple models in geoscience, and internet-based data analysis that facilitates collection, processing, and storage of data in distributed and crowd-sourced environments. 84 The growing availability of massive geoscience data provides unlimited possibilities for AI—which has popularized all aspects of our daily life (e.g., entertainment, transportation, and commerce)—to significantly contribute to geoscience problems of great societal relevance. As geoscience enters the era of massive data, AI, which has been extensively successful in different fields, offers immense opportunities for settling a series of problems in Earth systems. 85 , 86 Accompanied by diversified data, AI-enabled technologies, such as smart sensors, image visualization, and intelligent inversion, are being actively examined in a large range of geoscience fields, such as marine geoscience, rock physics, geology, ecology, seismicity, environment, hydrology, remote sensing, Arc GIS, and planetary science. 87

Multiple challenges in the development of geoscience

There are some traits of geoscience development that restrict the applicability of fundamental algorithms for knowledge discovery: (1) inherent challenges of geoscience processes, (2) limitation of geoscience data collection, and (3) uncertainty in samples and ground truth. 88 , 89 , 90 Amorphous boundaries generally exist in geoscience objects between space and time that are not as well defined as objects in other fields. Geoscience phenomena are also significantly multivariate, obey nonlinear relationships, and exhibit spatiotemporal structure and non-stationary characteristics. Except for the inherent challenges of geoscience observations, the massive data at multiple dimensions of time and space, with different levels of incompleteness, noise, and uncertainties, disturb processes in geoscience. For supervised learning approaches, there are other difficulties owing to the lack of gold standard ground truth and the “small size” of samples (e.g., a small amount of historical data with sufficient observations) in geoscience applications.

Usage of AI technologies as efficient approaches to promote the geoscience processes

Geoscientists continually make every effort to develop better techniques for simulating the present status of the Earth system (e.g., how much greenhouse gases are released into the atmosphere), and the connections between and within its subsystems (e.g., how does the elevated temperature influence the ocean ecosystem). Viewed from the perspective of geoscience, newly emerging approaches, with the aid of AI, are a perfect combination for these issues in the application of geoscience: (1) characterizing objects and events 91 ; (2) estimating geoscience variables from observations 92 ; (3) forecasting geoscience variables according to long-term observations 85 ; (4) exploring geoscience data relationships 93 ; and (5) causal discovery and causal attribution. 94 While characterizing geoscience objects and events using traditional methods are primarily rooted in hand-coded features, algorithms can automatically detect the data by improving the performance with pattern-mining techniques. However, due to spatiotemporal targets with vague boundaries and the related uncertainties, it can be necessary to advance pattern-mining methods that can explain the temporal and spatial characteristics of geoscience data when characterizing different events and objects. To address the non-stationary issue of geoscience data, AI-aided algorithms have been expanded to integrate the holistic results of professional predictors and engender robust estimations of climate variables (e.g., humidity and temperature). Furthermore, forecasting long-term trends of the current situation in the Earth system using AI-enabled technologies can simulate future scenarios and formulate early resource planning and adaptation policies. Mining geoscience data relationships can help us seize vital signs of the Earth system and promote our understanding of geoscience developments. Of great interest is the advancement of AI-decision methodology with uncertain prediction probabilities, engendering vague risks with poorly resolved tails, signifying the most extreme, transient, and rare events formulated by model sets, which supports various cases to improve accuracy and effectiveness.

AI technologies for optimizing the resource management in geoscience

Currently, AI can perform better than humans in some well-defined tasks. For example, AI techniques have been used in urban water resource planning, mainly due to their remarkable capacity for modeling, flexibility, reasoning, and forecasting the water demand and capacity. Design and application of an Adaptive Intelligent Dynamic Water Resource Planning system, the subset of AI for sustainable water resource management in urban regions, largely prompted the optimization of water resource allocation, will finally minimize the operation costs and improve the sustainability of environmental management 95 ( Figure 6 ). Also, meteorology requires collecting tremendous amounts of data on many different variables, such as humidity, altitude, and temperature; however, dealing with such a huge dataset is a big challenge. 96 An AI-based technique is being utilized to analyze shallow-water reef images, recognize the coral color—to track the effects of climate change, and to collect humidity, temperature, and CO 2 data—to grasp the health of our ecological environment. 97 Beyond AI's capabilities for meteorology, it can also play a critical role in decreasing greenhouse gas emissions originating from the electric-power sector. Comprised of production, transportation, allocation, and consumption of electricity, many opportunities exist in the electric-power sector for Al applications, including speeding up the development of new clean energy, enhancing system optimization and management, improving electricity-demand forecasts and distribution, and advancing system monitoring. 98 New materials may even be found, with the auxiliary of AI, for batteries to store energy or materials and absorb CO 2 from the atmosphere. 99 Although traditional fossil fuel operations have been widely used for thousands of years, AI techniques are being used to help explore the development of more potential sustainable energy sources for the development (e.g., fusion technology). 100

Applications of AI in hydraulic resource management

In addition to the adjustment of energy structures due to climate change (a core part of geoscience systems), a second, less-obvious step could also be taken to reduce greenhouse gas emission: using AI to target inefficiencies. A related statistical report by the Lawrence Livermore National Laboratory pointed out that around 68% of energy produced in the US could be better used for purposeful activities, such as electricity generation or transportation, but is instead contributing to environmental burdens. 101 AI is primed to reduce these inefficiencies of current nuclear power plants and fossil fuel operations, as well as improve the efficiency of renewable grid resources. 102 For example, AI can be instrumental in the operation and optimization of solar and wind farms to make these utility-scale renewable-energy systems far more efficient in the production of electricity. 103 AI can also assist in reducing energy losses in electricity transportation and allocation. 104 A distribution system operator in Europe used AI to analyze load, voltage, and network distribution data, to help “operators assess available capacity on the system and plan for future needs.” 105 AI allowed the distribution system operator to employ existing and new resources to make the distribution of energy assets more readily available and flexible. The International Energy Agency has proposed that energy efficiency is core to the reform of energy systems and will play a key role in reducing the growth of global energy demand to one-third of the current level by 2040.

AI as a building block to promote development in geoscience

The Earth’s system is of significant scientific interest, and affects all aspects of life. 106 The challenges, problems, and promising directions provided by AI are definitely not exhaustive, but rather, serve to illustrate that there is great potential for future AI research in this important field. Prosperity, development, and popularization of AI approaches in the geosciences is commonly driven by a posed scientific question, and the best way to succeed is that AI researchers work closely with geoscientists at all stages of research. That is because the geoscientists can better understand which scientific question is important and novel, which sample collection process can reasonably exhibit the inherent strengths, which datasets and parameters can be used to answer that question, and which pre-processing operations are conducted, such as removing seasonal cycles or smoothing. Similarly, AI researchers are better suited to decide which data analysis approaches are appropriate and available for the data, the advantages and disadvantages of these approaches, and what the approaches actually acquire. Interpretability is also an important goal in geoscience because, if we can understand the basic reasoning behind the models, patterns, or relationships extracted from the data, they can be used as building blocks in scientific knowledge discovery. Hence, frequent communication between the researchers avoids long detours and ensures that analysis results are indeed beneficial to both geoscientists and AI researchers.

AI in the life sciences

The developments of AI and the life sciences are intertwined. The ultimate goal of AI is to achieve human-like intelligence, as the human brain is capable of multi-tasking, learning with minimal supervision, and generalizing learned skills, all accomplished with high efficiency and low energy cost. 107

Mutual inspiration between AI and neuroscience

In the past decades, neuroscience concepts have been introduced into ML algorithms and played critical roles in triggering several important advances in AI. For example, the origins of DL methods lie directly in neuroscience, 5 which further stimulated the emergence of the field of RL. 108 The current state-of-the-art CNNs incorporate several hallmarks of neural computation, including nonlinear transduction, divisive normalization, and maximum-based pooling of inputs, 109 which were directly inspired by the unique processing of visual input in the mammalian visual cortex. 110 By introducing the brain's attentional mechanisms, a novel network has been shown to produce enhanced accuracy and computational efficiency at difficult multi-object recognition tasks than conventional CNNs. 111 Other neuroscience findings, including the mechanisms underlying working memory, episodic memory, and neural plasticity, have inspired the development of AI algorithms that address several challenges in deep networks. 108 These algorithms can be directly implemented in the design and refinement of the brain-machine interface and neuroprostheses.

On the other hand, insights from AI research have the potential to offer new perspectives on the basics of intelligence in the brains of humans and other species. Unlike traditional neuroscientists, AI researchers can formalize the concepts of neural mechanisms in a quantitative language to extract their necessity and sufficiency for intelligent behavior. An important illustration of such exchange is the development of the temporal-difference (TD) methods in RL models and the resemblance of TD-form learning in the brain. 112 Therefore, the China Brain Project covers both basic research on cognition and translational research for brain disease and brain-inspired intelligence technology. 113

AI for omics big data analysis

Currently, AI can perform better than humans in some well-defined tasks, such as omics data analysis and smart agriculture. In the big data era, 114 there are many types of data (variety), the volume of data is big, and the generation of data (velocity) is fast. The high variety, big volume, and fast velocity of data makes having it a matter of big value, but also makes it difficult to analyze the data. Unlike traditional statistics-based methods, AI can easily handle big data and reveal hidden associations.

In genetics studies, there are many successful applications of AI. 115 One of the key questions is to determine whether a single amino acid polymorphism is deleterious. 116 There have been sequence conservation-based SIFT 117 and network-based SySAP, 118 but all these methods have met bottlenecks and cannot be further improved. Sundaram et al. developed PrimateAI, which can predict the clinical outcome of mutation based on DNN. 119 Another problem is how to call copy-number variations, which play important roles in various cancers. 120 , 121 Glessner et al. proposed a DL-based tool DeepCNV, in which the area under the receiver operating characteristic (ROC) curve was 0.909, much higher than other ML methods. 122 In epigenetic studies, m6A modification is one of the most important mechanisms. 123 Zhang et al. developed an ensemble DL predictor (EDLm6APred) for mRNA m6A site prediction. 124 The area under the ROC curve of EDLm6APred was 86.6%, higher than existing m6A methylation site prediction models. There are many other DL-based omics tools, such as DeepCpG 125 for methylation, DeepPep 126 for proteomics, AtacWorks 127 for assay for transposase-accessible chromatin with high-throughput sequencing, and deepTCR 128 for T cell receptor sequencing.

Another emerging application is DL for single-cell sequencing data. Unlike bulk data, in which the sample size is usually much smaller than the number of features, the sample size of cells in single-cell data could also be big compared with the number of genes. That makes the DL algorithm applicable for most single-cell data. Since the single-cell data are sparse and have many unmeasured missing values, DeepImpute can accurately impute these missing values in the big gene × cell matrix. 129 During the quality control of single-cell data, it is important to remove the doublet solo embedded cells, using autoencoder, and then build a feedforward neural network to identify the doublet. 130 Potential energy underlying single-cell gradients used generative modeling to learn the underlying differentiation landscape from time series single-cell RNA sequencing data. 131

In protein structure prediction, the DL-based AIphaFold2 can accurately predict the 3D structures of 98.5% of human proteins, and will predict the structures of 130 million proteins of other organisms in the next few months. 132 It is even considered to be the second-largest breakthrough in life sciences after the human genome project 133 and will facilitate drug development among other things.

AI makes modern agriculture smart

Agriculture is entering a fourth revolution, termed agriculture 4.0 or smart agriculture, benefiting from the arrival of the big data era as well as the rapid progress of lots of advanced technologies, in particular ML, modern information, and communication technologies. 134 , 135 Applications of DL, information, and sensing technologies in agriculture cover the whole stages of agricultural production, including breeding, cultivation, and harvesting.

Traditional breeding usually exploits genetic variations by searching natural variation or artificial mutagenesis. However, it is hard for either method to expose the whole mutation spectrum. Using DL models trained on the existing variants, predictions can be made on multiple unidentified gene loci. 136 For example, an ML method, multi-criteria rice reproductive gene predictor, was developed and applied to predict coding and lincRNA genes associated with reproductive processes in rice. 137 Moreover, models trained in species with well-studied genomic data (such as Arabidopsis and rice) can also be applied to other species with limited genome information (such as wild strawberry and soybean). 138 In most cases, the links between genotypes and phenotypes are more complicated than we expected. One gene can usually respond to multiple phenotypes, and one trait is generally the product of the synergism between multi-genes and multi-development. For this reason, multi-traits DL models were developed and enabled genomic editing in plant breeding. 139 , 140

It is well known that dynamic and accurate monitoring of crops during the whole growth period is vitally important to precision agriculture. In the new stage of agriculture, both remote sensing and DL play indispensable roles. Specifically, remote sensing (including proximal sensing) could produce agricultural big data from ground, air-borne, to space-borne platforms, which have a unique potential to offer an economical approach for non-destructive, timely, objective, synoptic, long-term, and multi-scale information for crop monitoring and management, thereby greatly assisting in precision decisions regarding irrigation, nutrients, disease, pests, and yield. 141 , 142 DL makes it possible to simply, efficiently, and accurately discover knowledge from massive and complicated data, especially for remote sensing big data that are characterized with multiple spatial-temporal-spectral information, owing to its strong capability for feature representation and superiority in capturing the essential relation between observation data and agronomy parameters or crop traits. 135 , 143 Integration of DL and big data for agriculture has demonstrated the most disruptive force, as big as the green revolution. As shown in Figure 7 , for possible application a scenario of smart agriculture, multi-source satellite remote sensing data with various geo- and radio-metric information, as well as abundance of spectral information from UV, visible, and shortwave infrared to microwave regions, can be collected. In addition, advanced aircraft systems, such as unmanned aerial vehicles with multi/hyper-spectral cameras on board, and smartphone-based portable devices, will be used to obtain multi/hyper-spectral data in specific fields. All types of data can be integrated by DL-based fusion techniques for different purposes, and then shared for all users for cloud computing. On the cloud computing platform, different agriculture remote sensing models developed by a combination of data-driven ML methods and physical models, will be deployed and applied to acquire a range of biophysical and biochemical parameters of crops, which will be further analyzed by a decision-making and prediction system to obtain the current water/nutrient stress, growth status, and to predict future development. As a result, an automatic or interactive user service platform can be accessible to make the correct decisions for appropriate actions through an integrated irrigation and fertilization system.

Integration of AI and remote sensing in smart agriculture

Furthermore, DL presents unique advantages in specific agricultural applications, such as for dense scenes, that increase the difficulty of artificial planting and harvesting. It is reported that CNNs and Autoencoder models, trained with image data, are being used increasingly for phenotyping and yield estimation, 144 such as counting fruits in orchards, grain recognition and classification, disease diagnosis, etc. 145 , 146 , 147 Consequently, this may greatly liberate the labor force.

The application of DL in agriculture is just beginning. There are still many problems and challenges for the future development of DL technology. We believe, with the continuous acquisition of massive data and the optimization of algorithms, DL will have a better prospect in agricultural production.

AI in physics

The scale of modern physics ranges from the size of a neutron to the size of the Universe ( Figure 8 ). According to the scale, physics can be divided into four categories: particle physics on the scale of neutrons, nuclear physics on the scale of atoms, condensed matter physics on the scale of molecules, and cosmic physics on the scale of the Universe. AI, also called ML, plays an important role in all physics in different scales, since the use of the AI algorithm will be the main trend in data analyses, such as the reconstruction and analysis of images.

Scale of the physics

Speeding up simulations and identifications of particles with AI

There are many applications or explorations of applications of AI in particle physics. We cannot cover all of them here, but only use lattice quantum chromodynamics (LQCD) and the experiments on the Beijing spectrometer (BES) and the large hadron collider (LHC) to illustrate the power of ML in both theoretical and experimental particle physics.

LQCD studies the nonperturbative properties of QCD by using Monte Carlo simulations on supercomputers to help us understand the strong interaction that binds quarks together to form nucleons. Markov chain Monte Carlo simulations commonly used in LQCD suffer from topological freezing and critical slowing down as the simulations approach the real situation of the actual world. New algorithms with the help of DL are being proposed and tested to overcome those difficulties. 148 , 149 Physical observables are extracted from LQCD data, whose signal-to-noise ratio deteriorates exponentially. For non-Abelian gauge theories, such as QCD, complicated contour deformations can be optimized by using ML to reduce the variance of LQCD data. Proof-of-principle applications in two dimensions have been studied. 150 ML can also be used to reduce the time cost of generating LQCD data. 151

On the experimental side, particle identification (PID) plays an important role. Recently, a few PID algorithms on BES-III were developed, and the ANN 152 is one of them. Also, extreme gradient boosting has been used for multi-dimensional distribution reweighting, muon identification, and cluster reconstruction, and can improve the muon identification. U-Net is a convolutional network for pixel-level semantic segmentation, which is widely used in CV. It has been applied on BES-III to solve the problem of multi-turn curling track finding for the main drift chamber. The average efficiency and purity for the first turn's hits is about 91%, at the threshold of 0.85. Current (and future) particle physics experiments are producing a huge amount of data. Machine leaning can be used to discriminate between signal and overwhelming background events. Examples of data analyses on LHC, using supervised ML, can be found in a 2018 collaboration. 153 To take the potential advantage of quantum computers forward, quantum ML methods are also being investigated, see, for example, Wu et al., 154 and references therein, for proof-of-concept studies.

AI makes nuclear physics powerful

Cosmic ray muon tomography (Muography) 155 is an imaging graphe technology using natural cosmic ray muon radiation rather than artificial radiation to reduce the dangers. As an advantage, this technology can detect high-Z materials without destruction, as muon is sensitive to high-Z materials. The Classification Model Algorithm (CMA) algorithm is based on the classification in the supervised learning and gray system theory, and generates a binary classifier designing and decision function with the input of the muon track, and the output indicates whether the material exists at the location. The AI helps the user to improve the efficiency of the scanning time with muons.

AIso, for nuclear detection, the Cs 2 LiYCl 6 :Ce (CLYC) signal can react to both electrons and neutrons to create a pulse signal, and can therefore be applied to detect both neutrons and electrons, 156 but needs identification of the two particles by analyzing the shapes of the waves, that is n-γ ID. The traditional method has been the PSD (pulse shape discrimination) method, which is used to separate the waves of two particles by analyzing the distribution of the pulse information—such as amplitude, width, raise time, fall time, and the two particles that can be separated when the distribution has two separated Gaussian distributions. The traditional PSD can only analyze single-pulse waves, rather than multipulse waves, when two particles react with CLYC closely. But it can be solved by using an ANN method for classification of the six categories (n,γ,n + n,n + γ,γ + n,γ). Also, there are several parameters that could be used by AI to improve the reconstruction algorithm with high efficiency and less error.

AI-aided condensed matter physics

AI opens up a new avenue for physical science, especially when a trove of data is available. Recent works demonstrate that ML provides useful insights to improve the density functional theory (DFT), in which the single-electron picture of the Kohn-Sham scheme has the difficulty of taking care of the exchange and correlation effects of many-body systems. Yu et al. proposed a Bayesian optimization algorithm to fit the Hubbard U parameter, and the new method can find the optimal Hubbard U through a self-consistent process with good efficiency compared with the linear response method, 157 and boost the accuracy to the near-hybrid-functional-level. Snyder et al. developed an ML density functional for a 1D non-interacting non-spin-polarized fermion system to obtain significantly improved kinetic energy. This method enabled a direct approximation of the kinetic energy of a quantum system and can be utilized in orbital-free DFT modeling, and can even bypass the solving of the Kohn-Sham equation—while maintaining the precision to the quantum chemical level when a strong correlation term is included. Recently, FermiNet showed that the many-body quantum mechanics equations can be solved via AI. AI models also show advantages of capturing the interatom force field. In 2010, the Gaussian approximation potential (GAP) 158 was introduced as a powerful interatomic force field to describe the interactions between atoms. GAP uses kernel regression and invariant many-body representations, and performs quite well. For instance, it can simulate crystallization of amorphous crystals under high pressure fairly accurately. By employing the smooth overlap of the atomic position kernel (SOAP), 159 the accuracy of the potential can be further enhanced and, therefore, the SOAP-GAP can be viewed as a field-leading method for AI molecular dynamic simulation. There are also several other well-developed AI interatomic potentials out there, e.g., crystal graph CNNs provide a widely applicable way of vectorizing crystalline materials; SchNet embeds the continuous-filter convolutional layers into its DNNs for easing molecular dynamic as the potentials are space continuous; DimeNet constructs the directional message passing neural network by adding not only the bond length between atoms but also the bond angle, the dihedral angle, and the interactions between unconnected atoms into the model to obtain good accuracy.

AI helps explore the Universe

AI is one of the newest technologies, while astronomy is one of the oldest sciences. When the two meet, new opportunities for scientific breakthroughs are often triggered. Observations and data analysis play a central role in astronomy. The amount of data collected by modern telescopes has reached unprecedented levels, even the most basic task of constructing a catalog has become challenging with traditional source-finding tools. 160 Astronomers have developed automated and intelligent source-finding tools based on DL, which not only offer significant advantages in operational speed but also facilitate a comprehensive understanding of the Universe by identifying particular forms of objects that cannot be detected by traditional software and visual inspection. 160 , 161

More than a decade ago, a citizen science project called “Galaxy Zoo” was proposed to help label one million images of galaxies collected by the Sloan Digital Sky Survey (SDSS) by posting images online and recruiting volunteers. 162 Larger optical telescopes, in operation or under construction, produce data several orders of magnitude higher than SDSS. Even with volunteers involved, there is no way to analyze the vast amount of data received. The advantages of ML are not limited to source-finding and galaxy classification. In fact, it has a much wider range of applications. For example, CNN plays an important role in detecting and decoding gravitational wave signals in real time, reconstructing all parameters within 2 ms, while traditional algorithms take several days to accomplish the same task. 163 Such DL systems have also been used to automatically generate alerts for transients and track asteroids and other fast-moving near-Earth objects, improving detection efficiency by several orders of magnitude. In addition, astrophysicists are exploring the use of neural networks to measure galaxy clusters and study the evolution of the Universe.

In addition to the amazing speed, neural networks seem to have a deeper understanding of the data than expected and can recognize more complex patterns, indicating that the “machine” is evolving rather than just learning the characteristics of the input data.

AI in chemistry

Chemistry plays an important “central” role in other sciences 164 because it is the investigation of the structure and properties of matter, and identifies the chemical reactions that convert substances into to other substances. Accordingly, chemistry is a data-rich branch of science containing complex information resulting from centuries of experiments and, more recently, decades of computational analysis. This vast treasure trove of data is most apparent within the Chemical Abstract Services, which has collected more than 183 million unique organic and inorganic substances, including alloys, coordination compounds, minerals, mixtures, polymers, and salts, and is expanding by addition of thousands of additional new substances daily. 165 The unlimited complexity in the variety of material compounds explains why chemistry research is still a labor-intensive task. The level of complexity and vast amounts of data within chemistry provides a prime opportunity to achieve significant breakthroughs with the application of AI. First, the type of molecules that can be constructed from atoms are almost unlimited, which leads to unlimited chemical space 166 ; the interconnection of these molecules with all possible combinations of factors, such as temperature, substrates, and solvents, are overwhelmingly large, giving rise to unlimited reaction space. 167 Exploration of the unlimited chemical space and reaction space, and navigating to the optimum ones with the desired properties, is thus practically impossible solely from human efforts. Secondly, in chemistry, the huge assortment of molecules and the interplay of them with the external environments brings a new level of complexity, which cannot be simply predicted using physical laws. While many concepts, rules, and theories have been generalized from centuries of experience from studying trivial (i.e., single component) systems, nontrivial complexities are more likely as we discover that “more is different” in the words of Philip Warren Anderson, American physicist and Nobel Laureate. 168 Nontrivial complexities will occur when the scale changes, and the breaking of symmetry in larger, increasingly complex systems, and the rules will shift from quantitative to qualitative. Due to lack of systematic and analytical theory toward the structures, properties, and transformations of macroscopic substances, chemistry research is thus, incorrectly, guided by heuristics and fragmental rules accumulated over the previous centuries, yielding progress that only proceeds through trial and error. ML will recognize patterns from large amounts of data; thereby offering an unprecedented way of dealing with complexity, and reshaping chemistry research by revolutionizing the way in which data are used. Every sub-field of chemistry, currently, has utilized some form of AI, including tools for chemistry research and data generation, such as analytical chemistry and computational chemistry, as well as application to organic chemistry, catalysis, and medical chemistry, which we discuss herein.

AI breaks the limitations of manual feature selection methods

In analytical chemistry, the extraction of information has traditionally relied heavily on the feature selection techniques, which are based on prior human experiences. Unfortunately, this approach is inefficient, incomplete, and often biased. Automated data analysis based on AI will break the limitations of manual variable selection methods by learning from large amounts of data. Feature selection through DL algorithms enables information extraction from the datasets in NMR, chromatography, spectroscopy, and other analytical tools, 169 thereby improving the model prediction accuracy for analysis. These ML approaches will greatly accelerate the analysis of materials, leading to the rapid discovery of new molecules or materials. Raman scattering, for instance, since its discovery in the 1920s, has been widely employed as a powerful vibrational spectroscopy technology, capable of providing vibrational fingerprints intrinsic to analytes, thus enabling identification of molecules. 170 Recently, ML methods have been trained to recognize features in Raman (or SERS) spectra for the identity of an analyte by applying DL networks, including ANN, CNN, and fully convolutional network for feature engineering. 171 For example, Leong et al. designed a machine-learning-driven “SERS taster” to simultaneously harness useful vibrational information from multiple receptors for enhanced multiplex profiling of five wine flavor molecules at ppm levels. Principal-component analysis is employed for the discrimination of alcohols with varying degrees of substitution, and supported with vector machine discriminant analysis, is used to quantitatively classify all flavors with 100% accuracy. 172 Overall, AI techniques provide the first glimmer of hope for a universal method for spectral data analysis, which is fast, accurate, objective and definitive and with attractive advantages in a wide range of applications.

AI improves the accuracy and efficiency for various levels of computational theory

Complementary to analytical tools, computational chemistry has proven a powerful approach for using simulations to understand chemical properties; however, it is faced with an accuracy-versus-efficiency dilemma. This dilemma greatly limits the application of computational chemistry to real-world chemistry problems. To overcome this dilemma, ML and other AI methods are being applied to improve the accuracy and efficiency for various levels of theory used to describe the effects arising at different time and length scales, in the multi-scaling of chemical reactions. 173 Many of the open challenges in computational chemistry can be solved by ML approaches, for example, solving Schrödinger's equation, 174 developing atomistic 175 or coarse graining 176 potentials, constructing reaction coordinates, 177 developing reaction kinetics models, 178 and identifying key descriptors for computable properties. 179 In addition to analytical chemistry and computational chemistry, several disciplines of chemistry have incorporated AI technology to chemical problems. We discuss the areas of organic chemistry, catalysis, and medical chemistry as examples of where ML has made a significant impact. Many examples exist in literature for other subfields of chemistry and AI will continue to demonstrate breakthroughs in a wide range of chemical applications.

AI enables robotics capable of automating the synthesis of molecules

Organic chemistry studies the structure, property, and reaction of carbon-based molecules. The complexity of the chemical and reaction space, for a given property, presents an unlimited number of potential molecules that can be synthesized by chemists. Further complications are added when faced with the problems of how to synthesize a particular molecule, given that the process relies much on heuristics and laborious testing. Challenges have been addressed by researchers using AI. Given enough data, any properties of interest of a molecule can be predicted by mapping the molecular structure to the corresponding property using supervised learning, without resorting to physical laws. In addition to known molecules, new molecules can be designed by sampling the chemical space 180 using methods, such as autoencoders and CNNs, with the molecules coded as sequences or graphs. Retrosynthesis, the planning of synthetic routes, which was once considered an art, has now become much simpler with the help of ML algorithms. The Chemetica system, 181 for instance, is now capable of autonomous planning of synthetic routes that are subsequently proven to work in the laboratory. Once target molecules and the route of synthesis are determined, suitable reaction conditions can be predicted or optimized using ML techniques. 182

The integration of these AI-based approaches with robotics has enabled fully AI-guided robotics capable of automating the synthesis of small organic molecules without human intervention Figure 9 . 183 , 184

A closed loop workflow to enable automatic and intelligent design, synthesis, and assay of molecules in organic chemistry by AI

AI helps to search through vast catalyst design spaces

Catalytic chemistry originates from catalyst technologies in the chemical industry for efficient and sustainable production of chemicals and fuels. Thus far, it is still a challenging endeavor to make novel heterogeneous catalysts with good performance (i.e., stable, active, and selective) because a catalyst's performance depends on many properties: composition, support, surface termination, particle size, particle morphology, atomic coordination environment, porous structure, and reactor during the reaction. The inherent complexity of catalysis makes discovering and developing catalysts with desired properties more dependent on intuition and experiment, which is costly and time consuming. AI technologies, such as ML, when combined with experimental and in silico high-throughput screening of combinatorial catalyst libraries, can aid catalyst discovery by helping to search through vast design spaces. With a well-defined structure and standardized data, including reaction results and in situ characterization results, the complex association between catalytic structure and catalytic performance will be revealed by AI. 185 , 186 An accurate descriptor of the effect of molecules, molecular aggregation states, and molecular transport, on catalysts, could also be predicted. With this approach, researchers can build virtual laboratories to develop new catalysts and catalytic processes.

AI enables screening of chemicals in toxicology with minimum ethical concerns

A more complicated sub-field of chemistry is medical chemistry, which is a challenging field due to the complex interactions between the exotic substances and the inherent chemistry within a living system. Toxicology, for instance, as a broad field, seeks to predict and eliminate substances (e.g., pharmaceuticals, natural products, food products, and environmental substances), which may cause harm to a living organism. Living organisms are already complex, nearly any known substance can cause toxicity at a high enough exposure because of the already inherent complexity within living organisms. Moreover, toxicity is dependent on an array of other factors, including organism size, species, age, sex, genetics, diet, combination with other chemicals, overall health, and/or environmental context. Given the scale and complexity of toxicity problems, AI is likely to be the only realistic approach to meet regulatory body requirements for screening, prioritization, and risk assessment of chemicals (including mixtures), therefore revolutionizing the landscape in toxicology. 187 In summary, AI is turning chemistry from a labor-intensive branch of science to a highly intelligent, standardized, and automated field, and much more can be achieved compared with the limitation of human labor. Underlying knowledge with new concepts, rules, and theories is expected to advance with the application of AI algorithms. A large portion of new chemistry knowledge leading to significant breakthroughs is expected to be generated from AI-based chemistry research in the decades to come.

Conclusions

This paper carries out a comprehensive survey on the development and application of AI across a broad range of fundamental sciences, including information science, mathematics, medical science, materials science, geoscience, life science, physics, and chemistry. Despite the fact that AI has been pervasively used in a wide range of applications, there still exist ML security risks on data and ML models as attack targets during both training and execution phases. Firstly, since the performance of an ML system is highly dependent on the data used to train it, these input data are crucial for the security of the ML system. For instance, adversarial example attacks 188 providing malicious input data often lead the ML system into making false judgments (predictions or categorizations) with small perturbations that are imperceptible to humans; data poisoning by intentionally manipulating raw, training, or testing data can result in a decrease in model accuracy or lead to other error-specific attack purposes. Secondly, ML model attacks include backdoor attacks on DL, CNN, and federated learning that manipulate the model's parameters directly, as well as model stealing attack, model inversion attack, and membership inference attack, which can steal the model parameters or leak the sensitive training data. While a number of defense techniques against these security threats have been proposed, new attack models that target ML systems are constantly emerging. Thus, it is necessary to address the problem of ML security and develop robust ML systems that remain effective under malicious attacks.

Due to the data-driven character of the ML method, features of the training and testing data must be drawn from the same distribution, which is difficult to guarantee in practice. This is because, in practical application, the data source might be different from that in the training dataset. In addition, the data feature distribution may drift over time, which leads to a decline of the performance of the model. Moreover, if the model is trained with only new data, it will lead to catastrophic “forgetting” of the model, which means the model only remembers the new features and forgets the previously learned features. To solve this problem, more and more scholars pay attention on how to make the model have the ability of lifelong learning, that is, a change in the computing paradigm from “offline learning + online reasoning” to “online continuous learning,” and thus give the model have the ability of lifelong learning, just like a human being.

Acknowledgments

This work was partially supported by the National Key R&D Program of China (2018YFA0404603, 2019YFA0704900, 2020YFC1807000, and 2020YFB1313700), the Youth Innovation Promotion Association CAS (2011225, 2012006, 2013002, 2015316, 2016275, 2017017, 2017086, 2017120, 2017204, 2017300, 2017399, 2018356, 2020111, 2020179, Y201664, Y201822, and Y201911), NSFC (nos. 11971466, 12075253, 52173241, and 61902376), the Foundation of State Key Laboratory of Particle Detection and Electronics (SKLPDE-ZZ-201902), the Program of Science & Technology Service Network of CAS (KFJ-STS-QYZX-050), the Fundamental Science Center of the National Nature Science Foundation of China (nos. 52088101 and 11971466), the Scientific Instrument Developing Project of CAS (ZDKYYQ20210003), the Strategic Priority Research Program (B) of CAS (XDB33000000), the National Science Foundation of Fujian Province for Distinguished Young Scholars (2019J06023), the Key Research Program of Frontier Sciences, CAS (nos. ZDBS-LY-7022 and ZDBS-LY-DQC012), the CAS Project for Young Scientists in Basic Research (no. YSBR-005). The study is dedicated to the 10th anniversary of the Youth Innovation Promotion Association of the Chinese Academy of Sciences.

Author contributions

Y.X., Q.W., Z.A., Fei W., C.L., Z.C., J.M.T., and J.Z. conceived and designed the research. Z.A., Q.W., Fei W., Libo.Z., Y.W., F.D., and C.W.-Q. wrote the “ AI in information science ” section. Xin.L. wrote the “ AI in mathematics ” section. J.Q., K.H., W.S., J.W., H.X., Y.H., and X.C. wrote the “ AI in medical science ” section. E.L., C.F., Z.Y., and M.L. wrote the “ AI in materials science ” section. Fang W., R.R., S.D., M.V., and F.K. wrote the “ AI in geoscience ” section. C.H., Z.Z., L.Z., T.Z., J.D., J.Y., L.L., M.L., and T.H. wrote the “ AI in life sciences ” section. Z.L., S.Q., and T.A. wrote the “ AI in physics ” section. X.L., B.Z., X.H., S.C., X.L., W.Z., and J.P.L. wrote the “ AI in chemistry ” section. Y.X., Q.W., and Z.A. wrote the “Abstract,” “ introduction ,” “ history of AI ,” and “ conclusions ” sections.

Declaration of interests

The authors declare no competing interests.

Published Online: October 28, 2021

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• NATURE INDEX
• 12 October 2022

Growth in AI and robotics research accelerates

It may not be unusual for burgeoning areas of science, especially those related to rapid technological changes in society, to take off quickly, but even by these standards the rise of artificial intelligence (AI) has been impressive. Together with robotics, AI is representing an increasingly significant portion of research volume at various levels, as these charts show.

Across the field

The number of AI and robotics papers published in the 82 high-quality science journals in the Nature Index (Count) has been rising year-on-year — so rapidly that it resembles an exponential growth curve. A similar increase is also happening more generally in journals and proceedings not included in the Nature Index, as is shown by data from the Dimensions database of research publications.

Source: Nature Index, Dimensions. Data analysis by Catherine Cheung; infographic by Simon Baker, Tanner Maxwell and Benjamin Plackett

Leading countries

Five countries — the United States, China, the United Kingdom, Germany and France — had the highest AI and robotics Share in the Nature Index from 2015 to 2021, with the United States leading the pack. China has seen the largest percentage change (1,174%) in annual Share over the period among the five nations.

AI and robotics infiltration

As the field of AI and robotics research grows in its own right, leading institutions such as Harvard University in the United States have increased their Share in this area since 2015. But such leading institutions have also seen an expansion in the proportion of their overall index Share represented by research in AI and robotics. One possible explanation for this is that AI and robotics is expanding into other fields, creating interdisciplinary AI and robotics research.

Nature 610 , S9 (2022)

doi: https://doi.org/10.1038/d41586-022-03210-9

This article is part of Nature Index 2022 AI and robotics , an editorially independent supplement. Advertisers have no influence over the content.

Related Articles

Partner content: AI helps computers to see and hear more efficiently

Partner content: Canada's welcoming artificial intelligence research ecosystem

Partner content: TINY robots inspired by insects

Partner content: Pioneering a new era of drug development

Partner content: New tool promises smarter approach to big data and AI

Partner content: Intelligent robots offer service with a smile

Partner content: Hong Kong’s next era fuelled by innovation

Partner content: Getting a grip on mass-produced artificial muscles with control engineering tools

Partner content: A blueprint for AI-powered smart speech technology

Partner content: All in the mind’s AI

Partner content: How artificial intelligence could turn thoughts into actions

Partner content: AI-powered start-up puts protein discovery on the fast track

Partner content: Intelligent tech takes on drone safety

• Computer science
• Mathematics and computing

AI now beats humans at basic tasks — new benchmarks are needed, says major report

News 15 APR 24

High-threshold and low-overhead fault-tolerant quantum memory

Article 27 MAR 24

Three reasons why AI doesn’t model human language

Correspondence 19 MAR 24

NATO is boosting AI and climate research as scientific diplomacy remains on ice

News Explainer 25 APR 24

Are robots the solution to the crisis in older-person care?

Outlook 25 APR 24

Lethal AI weapons are here: how can we control them?

News Feature 23 APR 24

AI’s keen diagnostic eye

Outlook 18 APR 24

Use game theory for climate models that really help reach net zero goals

Correspondence 16 APR 24

Junior Group Leader

The Imagine Institute is a leading European research centre dedicated to genetic diseases, with the primary objective to better understand and trea...

Paris, Ile-de-France (FR)

Imagine Institute

Director of the Czech Advanced Technology and Research Institute of Palacký University Olomouc

The Rector of Palacký University Olomouc announces a Call for the Position of Director of the Czech Advanced Technology and Research Institute of P...

Czech Republic (CZ)

Palacký University Olomouc

Course lecturer for INFH 5000

The HKUST(GZ) Information Hub is recruiting course lecturer for INFH 5000: Information Science and Technology: Essentials and Trends.

Guangzhou, Guangdong, China

The Hong Kong University of Science and Technology (Guangzhou)

Suzhou Institute of Systems Medicine Seeking High-level Talents

Full Professor, Associate Professor, Assistant Professor

Suzhou, Jiangsu, China

Suzhou Institute of Systems Medicine (ISM)

Postdoctoral Fellowships: Early Diagnosis and Precision Oncology of Gastrointestinal Cancers

We currently have multiple postdoctoral fellowship positions within the multidisciplinary research team headed by Dr. Ajay Goel, professor and foun...

Monrovia, California

Beckman Research Institute, City of Hope, Goel Lab

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

What’s next with AI?

Consumers are already putting ai tools to use for both productivity and creativity. but what’s next.

By Verge Press Room , Andrew Melnizek , and Edwin Wong

Share this story

Artificial intelligence is undeniably the story of the year. 

To better understand the rapid rise and adoption of generative AI tools, in late June 2023, The Verge and Vox Media partnered to conduct a representative study of how adult Americans are using and thinking about AI. The Verge published its initial findings at the time on The Verge and in Vox Media’s press room . This initial AI report was a follow-up to our major consumer tech trust surveys conducted in 2017 , 2020 , and 2021 .

Now, nearly one year later, we’re sharing our full, updated findings and building on our initial AI report. This is the next wave of our AI consumer survey.

With AI being top of mind for The Verge and Vox Media’s partners, we also wanted to share our overall key findings for brands looking to understand and leverage AI tools. Here are some of those overall takeaways from the survey, with a more comprehensive look at the data available in the report above.

One year later: where we are after a year with AI.

• Interest continues, but usage is growing at a much smaller rate.
• New adoption is slowing; however, users are evolving into superusers.
• OpenAI took a stronger lead in market share, but AI features are becoming a marketing asset for technology brands, creating strong brand awareness.
• The first meaningful disruption in search in 20 years is coming into full view.

Creativity and productivity: how AI will scale differently — namely, by age — when it comes to creativity and productivity. 

• Consumers are putting AI tools to use for both productivity and creativity, but productivity is pulling ahead, especially for older users.
• The fastest-growing use case for AI is email. 
• While most people think AI can improve their own skills across mediums, they don’t think it’s beating the professionals. 
• As adoption grows, norms in work and school are changing to become more open to AI tools.

Trusting the data: understanding how consumers think about the data and whether they trust it.

• As AI search grows, consumers are putting a lot of trust in the quality of information.
• Consumers want to see government regulation but expect that companies will have to take the lead on transparency, energy efficiency, and misinformation.

Coexisting with AI: for all the discussion about AI replacing things, it’s really about augmentation and incrementally coexisting with AI.

• People immediately grasp the potential for AI to simplify and streamline their daily lives but still prefer humans for more personal tasks like dating.
• Data privacy is still a concern, but local AI and personal apps show promise.
• The race is now on to build general — not just generative — AI.

This next phase in AI is about proving AI, where hype will transition into reality, practicality, and usage in our everyday lives.

How you can join the conversation about AI Here at Vox Media and The Verge , we’re obsessed with quality storytelling and journalism. As artificial intelligence continues to infiltrate our lives, our editorial brands will help guide our audience of over 100 million people here in the US through this next wave of innovation, discerning what is hype and what will have a real impact on humanity.

Methodology In December 2023, the Vox Media Insights and Research team surveyed more than 2,000 consumers with our partner, The Circus. The Circus is an insight and data storytelling consultancy that specializes in original trend research, thought leadership, and strategic brand positioning rooted in a human-data centric approach.

Special thanks to these individuals who helped bring this survey to life: Andrew Melnizek – VP and GM, The Verge and Polygon Edwin Wong – SVP, Insights and Research, Vox Media Sebastian Fernandez – Founder, The Circus Diana Young – Designer, D-Constructed Jacob Kastrenakes – Executive Editor, The Verge Kara Verlaney – Managing Editor, The Verge Nilay Patel – Editor-in-Chief, The Verge Helen Havlak – Publisher, The Verge

This self-transforming Megatron is as badass as it is expensive

I traded in my macbook and now i’m a desktop convert, tesla’s autopilot and full self-driving linked to hundreds of crashes, dozens of deaths, blizzcon 2024 has been canceled, dji might get banned next in the us.

More from Press Room

Join The Verge at the 2024 Chicago Humanities Spring Festival

The Verge Promotes Jake Kastrenakes, Alex Cranz, and Kara Verlaney

The Verge nominated for two 2024 National Magazine Awards

Hot Pod Summit returns at On Air Fest 2024

Our approach

• Responsibility
• Infrastructure
• Try Meta AI

RECOMMENDED READS

• 5 Steps to Getting Started with Llama 2
• The Llama Ecosystem: Past, Present, and Future
• Introducing Code Llama, a state-of-the-art large language model for coding
• Meta and Microsoft Introduce the Next Generation of Llama
• Today, we’re introducing Meta Llama 3, the next generation of our state-of-the-art open source large language model.
• Llama 3 models will soon be available on AWS, Databricks, Google Cloud, Hugging Face, Kaggle, IBM WatsonX, Microsoft Azure, NVIDIA NIM, and Snowflake, and with support from hardware platforms offered by AMD, AWS, Dell, Intel, NVIDIA, and Qualcomm.
• We’re dedicated to developing Llama 3 in a responsible way, and we’re offering various resources to help others use it responsibly as well. This includes introducing new trust and safety tools with Llama Guard 2, Code Shield, and CyberSec Eval 2.
• In the coming months, we expect to introduce new capabilities, longer context windows, additional model sizes, and enhanced performance, and we’ll share the Llama 3 research paper.
• Meta AI, built with Llama 3 technology, is now one of the world’s leading AI assistants that can boost your intelligence and lighten your load—helping you learn, get things done, create content, and connect to make the most out of every moment. You can try Meta AI here .

Today, we’re excited to share the first two models of the next generation of Llama, Meta Llama 3, available for broad use. This release features pretrained and instruction-fine-tuned language models with 8B and 70B parameters that can support a broad range of use cases. This next generation of Llama demonstrates state-of-the-art performance on a wide range of industry benchmarks and offers new capabilities, including improved reasoning. We believe these are the best open source models of their class, period. In support of our longstanding open approach, we’re putting Llama 3 in the hands of the community. We want to kickstart the next wave of innovation in AI across the stack—from applications to developer tools to evals to inference optimizations and more. We can’t wait to see what you build and look forward to your feedback.

Our goals for Llama 3

With Llama 3, we set out to build the best open models that are on par with the best proprietary models available today. We wanted to address developer feedback to increase the overall helpfulness of Llama 3 and are doing so while continuing to play a leading role on responsible use and deployment of LLMs. We are embracing the open source ethos of releasing early and often to enable the community to get access to these models while they are still in development. The text-based models we are releasing today are the first in the Llama 3 collection of models. Our goal in the near future is to make Llama 3 multilingual and multimodal, have longer context, and continue to improve overall performance across core LLM capabilities such as reasoning and coding.

State-of-the-art performance

Our new 8B and 70B parameter Llama 3 models are a major leap over Llama 2 and establish a new state-of-the-art for LLM models at those scales. Thanks to improvements in pretraining and post-training, our pretrained and instruction-fine-tuned models are the best models existing today at the 8B and 70B parameter scale. Improvements in our post-training procedures substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. We also saw greatly improved capabilities like reasoning, code generation, and instruction following making Llama 3 more steerable.

*Please see evaluation details for setting and parameters with which these evaluations are calculated.

In the development of Llama 3, we looked at model performance on standard benchmarks and also sought to optimize for performance for real-world scenarios. To this end, we developed a new high-quality human evaluation set. This evaluation set contains 1,800 prompts that cover 12 key use cases: asking for advice, brainstorming, classification, closed question answering, coding, creative writing, extraction, inhabiting a character/persona, open question answering, reasoning, rewriting, and summarization. To prevent accidental overfitting of our models on this evaluation set, even our own modeling teams do not have access to it. The chart below shows aggregated results of our human evaluations across of these categories and prompts against Claude Sonnet, Mistral Medium, and GPT-3.5.

Preference rankings by human annotators based on this evaluation set highlight the strong performance of our 70B instruction-following model compared to competing models of comparable size in real-world scenarios.

Our pretrained model also establishes a new state-of-the-art for LLM models at those scales.

To develop a great language model, we believe it’s important to innovate, scale, and optimize for simplicity. We adopted this design philosophy throughout the Llama 3 project with a focus on four key ingredients: the model architecture, the pretraining data, scaling up pretraining, and instruction fine-tuning.

Model architecture

In line with our design philosophy, we opted for a relatively standard decoder-only transformer architecture in Llama 3. Compared to Llama 2, we made several key improvements. Llama 3 uses a tokenizer with a vocabulary of 128K tokens that encodes language much more efficiently, which leads to substantially improved model performance. To improve the inference efficiency of Llama 3 models, we’ve adopted grouped query attention (GQA) across both the 8B and 70B sizes. We trained the models on sequences of 8,192 tokens, using a mask to ensure self-attention does not cross document boundaries.

Training data

To train the best language model, the curation of a large, high-quality training dataset is paramount. In line with our design principles, we invested heavily in pretraining data. Llama 3 is pretrained on over 15T tokens that were all collected from publicly available sources. Our training dataset is seven times larger than that used for Llama 2, and it includes four times more code. To prepare for upcoming multilingual use cases, over 5% of the Llama 3 pretraining dataset consists of high-quality non-English data that covers over 30 languages. However, we do not expect the same level of performance in these languages as in English.

To ensure Llama 3 is trained on data of the highest quality, we developed a series of data-filtering pipelines. These pipelines include using heuristic filters, NSFW filters, semantic deduplication approaches, and text classifiers to predict data quality. We found that previous generations of Llama are surprisingly good at identifying high-quality data, hence we used Llama 2 to generate the training data for the text-quality classifiers that are powering Llama 3.

We also performed extensive experiments to evaluate the best ways of mixing data from different sources in our final pretraining dataset. These experiments enabled us to select a data mix that ensures that Llama 3 performs well across use cases including trivia questions, STEM, coding, historical knowledge, etc.

Scaling up pretraining

To effectively leverage our pretraining data in Llama 3 models, we put substantial effort into scaling up pretraining. Specifically, we have developed a series of detailed scaling laws for downstream benchmark evaluations. These scaling laws enable us to select an optimal data mix and to make informed decisions on how to best use our training compute. Importantly, scaling laws allow us to predict the performance of our largest models on key tasks (for example, code generation as evaluated on the HumanEval benchmark—see above) before we actually train the models. This helps us ensure strong performance of our final models across a variety of use cases and capabilities.

We made several new observations on scaling behavior during the development of Llama 3. For example, while the Chinchilla-optimal amount of training compute for an 8B parameter model corresponds to ~200B tokens, we found that model performance continues to improve even after the model is trained on two orders of magnitude more data. Both our 8B and 70B parameter models continued to improve log-linearly after we trained them on up to 15T tokens. Larger models can match the performance of these smaller models with less training compute, but smaller models are generally preferred because they are much more efficient during inference.

To train our largest Llama 3 models, we combined three types of parallelization: data parallelization, model parallelization, and pipeline parallelization. Our most efficient implementation achieves a compute utilization of over 400 TFLOPS per GPU when trained on 16K GPUs simultaneously. We performed training runs on two custom-built 24K GPU clusters . To maximize GPU uptime, we developed an advanced new training stack that automates error detection, handling, and maintenance. We also greatly improved our hardware reliability and detection mechanisms for silent data corruption, and we developed new scalable storage systems that reduce overheads of checkpointing and rollback. Those improvements resulted in an overall effective training time of more than 95%. Combined, these improvements increased the efficiency of Llama 3 training by ~three times compared to Llama 2.

Instruction fine-tuning

To fully unlock the potential of our pretrained models in chat use cases, we innovated on our approach to instruction-tuning as well. Our approach to post-training is a combination of supervised fine-tuning (SFT), rejection sampling, proximal policy optimization (PPO), and direct preference optimization (DPO). The quality of the prompts that are used in SFT and the preference rankings that are used in PPO and DPO has an outsized influence on the performance of aligned models. Some of our biggest improvements in model quality came from carefully curating this data and performing multiple rounds of quality assurance on annotations provided by human annotators.

Learning from preference rankings via PPO and DPO also greatly improved the performance of Llama 3 on reasoning and coding tasks. We found that if you ask a model a reasoning question that it struggles to answer, the model will sometimes produce the right reasoning trace: The model knows how to produce the right answer, but it does not know how to select it. Training on preference rankings enables the model to learn how to select it.

Building with Llama 3

Our vision is to enable developers to customize Llama 3 to support relevant use cases and to make it easier to adopt best practices and improve the open ecosystem. With this release, we’re providing new trust and safety tools including updated components with both Llama Guard 2 and Cybersec Eval 2, and the introduction of Code Shield—an inference time guardrail for filtering insecure code produced by LLMs.

We’ve also co-developed Llama 3 with torchtune , the new PyTorch-native library for easily authoring, fine-tuning, and experimenting with LLMs. torchtune provides memory efficient and hackable training recipes written entirely in PyTorch. The library is integrated with popular platforms such as Hugging Face, Weights & Biases, and EleutherAI and even supports Executorch for enabling efficient inference to be run on a wide variety of mobile and edge devices. For everything from prompt engineering to using Llama 3 with LangChain we have a comprehensive getting started guide and takes you from downloading Llama 3 all the way to deployment at scale within your generative AI application.

A system-level approach to responsibility

We have designed Llama 3 models to be maximally helpful while ensuring an industry leading approach to responsibly deploying them. To achieve this, we have adopted a new, system-level approach to the responsible development and deployment of Llama. We envision Llama models as part of a broader system that puts the developer in the driver’s seat. Llama models will serve as a foundational piece of a system that developers design with their unique end goals in mind.

Instruction fine-tuning also plays a major role in ensuring the safety of our models. Our instruction-fine-tuned models have been red-teamed (tested) for safety through internal and external efforts. ​​Our red teaming approach leverages human experts and automation methods to generate adversarial prompts that try to elicit problematic responses. For instance, we apply comprehensive testing to assess risks of misuse related to Chemical, Biological, Cyber Security, and other risk areas. All of these efforts are iterative and used to inform safety fine-tuning of the models being released. You can read more about our efforts in the model card .

Llama Guard models are meant to be a foundation for prompt and response safety and can easily be fine-tuned to create a new taxonomy depending on application needs. As a starting point, the new Llama Guard 2 uses the recently announced MLCommons taxonomy, in an effort to support the emergence of industry standards in this important area. Additionally, CyberSecEval 2 expands on its predecessor by adding measures of an LLM’s propensity to allow for abuse of its code interpreter, offensive cybersecurity capabilities, and susceptibility to prompt injection attacks (learn more in our technical paper ). Finally, we’re introducing Code Shield which adds support for inference-time filtering of insecure code produced by LLMs. This offers mitigation of risks around insecure code suggestions, code interpreter abuse prevention, and secure command execution.

With the speed at which the generative AI space is moving, we believe an open approach is an important way to bring the ecosystem together and mitigate these potential harms. As part of that, we’re updating our Responsible Use Guide (RUG) that provides a comprehensive guide to responsible development with LLMs. As we outlined in the RUG, we recommend that all inputs and outputs be checked and filtered in accordance with content guidelines appropriate to the application. Additionally, many cloud service providers offer content moderation APIs and other tools for responsible deployment, and we encourage developers to also consider using these options.

Deploying Llama 3 at scale

Llama 3 will soon be available on all major platforms including cloud providers, model API providers, and much more. Llama 3 will be everywhere .

Our benchmarks show the tokenizer offers improved token efficiency, yielding up to 15% fewer tokens compared to Llama 2. Also, Group Query Attention (GQA) now has been added to Llama 3 8B as well. As a result, we observed that despite the model having 1B more parameters compared to Llama 2 7B, the improved tokenizer efficiency and GQA contribute to maintaining the inference efficiency on par with Llama 2 7B.

For examples of how to leverage all of these capabilities, check out Llama Recipes which contains all of our open source code that can be leveraged for everything from fine-tuning to deployment to model evaluation.

What’s next for Llama 3?

The Llama 3 8B and 70B models mark the beginning of what we plan to release for Llama 3. And there’s a lot more to come.

Our largest models are over 400B parameters and, while these models are still training, our team is excited about how they’re trending. Over the coming months, we’ll release multiple models with new capabilities including multimodality, the ability to converse in multiple languages, a much longer context window, and stronger overall capabilities. We will also publish a detailed research paper once we are done training Llama 3.

To give you a sneak preview for where these models are today as they continue training, we thought we could share some snapshots of how our largest LLM model is trending. Please note that this data is based on an early checkpoint of Llama 3 that is still training and these capabilities are not supported as part of the models released today.

We’re committed to the continued growth and development of an open AI ecosystem for releasing our models responsibly. We have long believed that openness leads to better, safer products, faster innovation, and a healthier overall market. This is good for Meta, and it is good for society. We’re taking a community-first approach with Llama 3, and starting today, these models are available on the leading cloud, hosting, and hardware platforms with many more to come.

Try Meta Llama 3 today

We’ve integrated our latest models into Meta AI, which we believe is the world’s leading AI assistant. It’s now built with Llama 3 technology and it’s available in more countries across our apps.

You can use Meta AI on Facebook, Instagram, WhatsApp, Messenger, and the web to get things done, learn, create, and connect with the things that matter to you. You can read more about the Meta AI experience here .

Visit the Llama 3 website to download the models and reference the Getting Started Guide for the latest list of all available platforms.

You’ll also soon be able to test multimodal Meta AI on our Ray-Ban Meta smart glasses.

As always, we look forward to seeing all the amazing products and experiences you will build with Meta Llama 3.

Our latest updates delivered to your inbox

Subscribe to our newsletter to keep up with Meta AI news, events, research breakthroughs, and more.

Join us in the pursuit of what’s possible with AI.

Product experiences

Foundational models

Latest news

Meta © 2024