剑桥大学：2019年度AI全景报告-2019.6-135页_3mb

State of AI Report Summary (June 28, 2019)

Authors

• Nathan Benaich: Founder of Air Street Capital and Venture Partner at Point Nine Capital. Background in biology and cancer research.
• Ian Hogarth: Angel investor and co-founder of Songkick. Focus on applied machine learning and has a background in engineering.

Report Overview

This report provides a snapshot of AI advancements over the past 12 months, highlighting key research, talent, industry, and political developments. It aims to stimulate informed discussions on AI's future implications.

Key Dimensions Analyzed

• Research: Breakthroughs in AI technology.
• Talent: Supply, demand, and concentration of AI professionals.
• Industry: Applications of AI in large platforms and commercial sectors.
• China: A special section due to its unique internet ecosystem.
• Politics: Public opinion, economic impact, and geopolitical dynamics around AI.

Research and Technical Breakthroughs

Reinforcement Learning (RL)

• Montezuma's Revenge: OpenAI used Random Network Distillation (RND) to incentivize exploration, achieving superhuman performance.
• StarCraft II: AlphaStar beat a top player 5-0 using a combination of supervised learning and multi-agent training.
• Quake III Arena: Multiple RL agents learned to play Capture the Flag with human-like behaviors.
• OpenAI Five: Achieved a 99.4% win rate against over 15,000 live players in Dota 2.
• Compute Efficiency: RL progress was driven by increased compute resources, with OpenAI Five consuming 800 petaflop/s-days and 45,000 years of self-play.

Future Directions in RL

• Play-Driven Learning: Robots can learn complex tasks through play, without task-specific training.
• Curiosity-Driven Exploration: Agents are rewarded for novel experiences, aiding in solving tasks with sparse or no rewards.
• Dynamics Models for Online Planning: PlaNet learns environment dynamics from images, reducing data needs by up to 50x compared to A3C and D4PG.
• Production Deployment: Facebook released Horizon, an open-source RL platform for large-scale systems.

Natural Language Processing (NLP)

• Pretrained Language Models: Models like BERT, ELMo, and GPT-2 significantly improved NLP performance, marking an "ImageNet moment" for NLP.
• GLUE Benchmark: A new benchmark for evaluating NLP tasks, showing rapid progress with scores rising from 69 to 88 in 13 months.
• Machine Translation: Facebook demonstrated translating without bitexts using monolingual data and backtranslation techniques.

Deep Learning in Medicine

• Eye Disease Diagnosis: A two-stage deep learning approach achieved expert-level performance in diagnosing eye conditions.
• Cardiac Arrhythmia Detection: An end-to-end model trained on 54k ECGs achieved 0.97 ROC and higher sensitivity than cardiologists.
• Speech Reconstruction from Brain Waves: Researchers reconstructed speech from neural activity with 75% accuracy, improving intelligibility by 65% over linear regression.
• Limb Control Restoration: Neural networks enabled long-term reanimation of a tetraplegic patient's forearm through electrical stimulation.

Emerging Trends and Challenges

• Federated Learning (FL): Google released TensorFlow Federated, enabling decentralized training for real-world applications like healthcare.
• Data Privacy: Libraries like TensorFlow Privacy and TF Encrypted aim to protect user data during model training.
• Data Issues in Medicine: Large datasets improve model performance, but label definitions are shallow and extraction is error-prone. Dataset diversity remains a challenge.

Conclusion

The report highlights the rapid progress in AI across multiple domains, emphasizing the role of reinforcement learning, NLP advancements, and deep learning in medicine. It also notes the growing importance of data privacy and the potential of AI to transform industries and healthcare.