剑桥大学_2019年度AI全景报告-2019.6-135页_14__4mb

State of AI Report Summary (June 28, 2019)

Authors

• Nathan Benaich: Founder of Air Street Capital, co-founder of the Research and Applied AI Summit, and writer of the AI newsletter nathan.ai.
• Ian Hogarth: Angel investor in over 50 startups, co-founder and former CEO of Songkick, and Visiting Professor at UCL.

Report Overview

The report provides an analysis of the exponential progress in AI over the past 12 months, highlighting key developments in research, talent, industry applications, and geopolitical implications. It aims to spark informed discussions about AI's current state and future impact.

Key Dimensions Analyzed

1. Research: Breakthroughs in AI technologies.
2. Talent: Supply, demand, and concentration of AI professionals.
3. Industry: Applications of AI in large platforms and emerging sectors.
4. China: Special focus on AI developments within the Chinese context.
5. Politics: Public perception, economic impact, and geopolitical dynamics.

Core Content and Main Points

Reinforcement Learning (RL) Breakthroughs

• Montezuma's Revenge: OpenAI achieved superhuman performance using Random Network Distillation (RND), which encourages exploration by rewarding novel experiences.
• StarCraft II: DeepMind's 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 cooperate and compete, achieving human-like behaviors through a temporally hierarchical representation.
• OpenAI Five: Improved to a 99.4% win rate in Dota2, with significant compute resources (800 petaflop/s-days) and 45,000 years of simulated self-play.
• Curiosity-Driven Exploration: RL agents are incentivized to explore novel states, leading to better performance in complex tasks like 3D maze navigation.
• Learning Dynamics Models: PlaNet, an RL agent, learns environment dynamics from images, achieving 50x less interaction than A3C and D4PG while maintaining similar performance.

Natural Language Processing (NLP) Advancements

• Pretrained Language Models: Models like BERT, ELMo, GPT-2, and others have transformed NLP, similar to the ImageNet moment in computer vision.
• GLUE Benchmark: A new performance benchmark for language understanding tasks, with scores increasing from 69 to 88 in 13 months.
• Machine Translation Without Bitexts: Facebook demonstrated how monolingual data can be used for machine translation, using backtranslation and bilingual dictionaries.

Deep Learning in Medicine

• Eye Disease Diagnosis: A two-stage deep learning system using 3D U-Net and classification networks achieved expert-level performance in diagnosing eye conditions.
• Cardiac Arrhythmia Detection: An end-to-end deep learning model trained on 54k patients outperformed cardiologists in sensitivity for all rhythm classes.
• Speech Reconstruction from Brain Waves: Researchers at Columbia used electrocorticography to reconstruct speech from brain activity, achieving 75% accuracy with deep learning methods.
• Neural Networks for Limb Control: A neural network decoder, combined with electrical stimulation, restored limb control in a tetraplegic patient.

Emerging Trends

• Federated Learning (FL): Google released TensorFlow Federated, enabling decentralized training for real-world applications, especially in healthcare.
• Data Privacy and Security: Libraries like TensorFlow Privacy and TF Encrypted are designed to protect user data and prevent model tampering.
• Transfer Learning: A key area of research that allows knowledge from one task to be applied to another, reducing the need for extensive data and compute.

Key Information

• Research: RL has made significant strides in complex environments, with new methods enabling better exploration and generalization.
• Industry: AI is increasingly applied in real-world systems, with notable examples in gaming, healthcare, and communication.
• China: A major player in AI, with distinct developments and a focus on both research and commercial applications.
• Politics: AI's impact on global competition and trade policies is growing, with the US-China trade war including access to semiconductor companies.

Conclusion

The report highlights the rapid advancements in AI across multiple domains, emphasizing the importance of research, talent, and industry collaboration. It also underscores the growing concerns around data privacy and the need for robust, secure, and ethical AI development.