剑桥大学：2019年度AI全景报告-2019.6-135页_7__3mb

State of AI Report Summary (June 28, 2019)

Authors

• Nathan Benaich: Founder of Air Street Capital and Venture Partner at Point Nine Capital. He has a background in biology and cancer research.
• Ian Hogarth: Angel investor in over 50 startups, co-founder and former CEO of Songkick. He has a background in engineering and computer vision.

Core Content

This report provides an overview of the exponential progress in artificial intelligence (AI) over the past 12 months, focusing on research, talent, industry applications, and geopolitical implications.

Key Dimensions Analyzed

1. Research: Technological breakthroughs and their capabilities.
2. Talent: Supply, demand, and concentration of AI professionals.
3. Industry: Major applications and developments in AI-driven innovation.
4. China: A separate section due to its distinct internet ecosystem.
5. Politics: Public perception, economic implications, and geopolitical dynamics of AI.

Main Technical Breakthroughs in AI

Reinforcement Learning (RL) Advancements

• Montezuma's Revenge: OpenAI used Random Network Distillation (RND) to achieve superhuman performance by incentivizing curiosity.
• StarCraft II: DeepMind's AlphaStar beat a world-class player 5-0 using a combination of supervised learning and multi-agent training.
• Quake III Arena: Multiple RL agents learned to cooperate and compete, achieving human-like behaviors through rich internal representations.
• OpenAI Five: Achieved a 99.4% win rate against 15,019 players in Dota 2. Training required 800 petaflop/s-days and 45,000 years of self-play.
• Play-Driven Learning for Robots: Robots trained through play showed improved dexterity and generalization across 18 tasks, outperforming models trained on expert demonstrations.
• Curiosity-Driven Exploration: RL agents were rewarded for exploring new states, improving their ability to solve tasks with sparse or no explicit rewards.
• Learning Dynamics Models: PlaNet, an RL agent, learned environmental dynamics from images, achieving 50x less interaction than state-of-the-art methods.
• Moving Research to Production: Facebook released Horizon, an open-source RL platform for optimizing large-scale systems like Messenger and video streaming.

Natural Language Processing (NLP) Breakthroughs

• Pretrained Language Models: Models like BERT, ELMo, and GPT-2 significantly improved NLP performance, marking an "ImageNet moment" for the field.
• GLUE Benchmark: A new benchmark for evaluating NLP systems across multiple tasks, with performance increasing from 69 to 88 over 13 months.
• SUPERglue: Introduced to address the rapid progress in NLP, offering a more comprehensive evaluation framework.
• Machine Translation: Google and Facebook demonstrated techniques to improve translation using monolingual data and back-translation, achieving better results than previous methods.
• Common Sense Reasoning: Researchers explored whether text alone could enable common sense reasoning, using a dataset of over 300k everyday events to train models.

Deep Learning in Medicine

• Eye Disease Diagnosis: A two-stage deep learning approach achieved expert-level performance in diagnosing eye conditions using 3D U-Net for segmentation and a classification network for severity prediction.
• Cardiac Arrhythmia Detection: An end-to-end deep learning model trained on 54k patients outperformed cardiologists in detecting rhythm classes, with an average ROC of 0.97.
• Speech Reconstruction from Brain Waves: Researchers used invasive electrocorticography to synthesize speech from brain activity, achieving 75% accuracy in digit recognition.
• Neural Networks for Limb Control: A neural network decoder enabled long-term reanimation of a tetraplegic patient's forearm through electrical stimulation, with high accuracy over a year.

Key Challenges and Considerations

• Data Privacy and Security: ML systems are vulnerable to attacks, prompting the development of privacy-preserving technologies like TensorFlow Privacy and TF Encrypted.
• Generalization and Diversity: Large datasets improve model performance, but issues like shallow label definitions and lack of diversity in data (e.g., 70% of scans from 30% of patients) hinder generalization.
• Federated Learning (FL): Google introduced TensorFlow Federated, enabling decentralized training while preserving user privacy, particularly in healthcare.

Conclusion

The 2019 State of AI Report highlights significant progress in AI, especially in reinforcement learning and natural language processing, with promising applications in medicine. It also underscores the importance of data privacy, generalization, and the potential of AI to revolutionize various industries. The report serves as a comprehensive overview of the current state of AI and its future implications.