Deep Learning State of the Art (2020)
Lex Fridman
Summary
The deep learning landscape in 2020 saw significant maturation in frameworks, AI achieving superhuman performance in complex games and robotic manipulation, a growing awareness of its limitations in common sense reasoning, and an increasing need for ethical transparency, especially in recommendation systems.
Key Takeaways
- Framework & NLP Evolution: Deep learning frameworks like TensorFlow 2.0 and PyTorch 1.3 matured, converging on best features, while Python 2 support ended. In NLP, transformer models (e.g., BERT, XLNet, GPT-2) dominated benchmarks, and tools like Hugging Face emerged to simplify their use. 12:53 17:53 19:06
- Reinforcement Learning Triumphs: AI agents demonstrated world-class performance in complex multi-agent games: OpenAI Five beat Dota 2 champions in 2019 after massive compute, DeepMind's AlphaStar achieved StarCraft Grandmaster level, and CMU's Pluribus beat professional players in six-player No-Limit Texas Hold'em. 35:54 39:30 42:16
- Robotic Manipulation Success: OpenAI's robotic hand learned to solve the Rubik's Cube through reinforcement learning, employing automated domain randomization (ADR) to train across progressively more difficult environments and encourage the learning of general manipulation concepts. 43:06 43:22
- Autonomous Vehicle Approaches: Two primary development paths exist: Level 2 (human-supervised, e.g., Tesla Autopilot), which is camera-based and treats deep learning as the core "cake" requiring vast amounts of data, and Level 4 (AI-responsible, e.g., Waymo), which leverages LIDAR and high-definition maps for higher reliability with less data, treating deep learning as "icing on the cake." 53:09 53:26 55:34
- Importance of Online/Active Learning: Real-world machine learning deployment, especially in autonomous vehicles, fundamentally relies on an "active learning" or "data engine" process: a continuous loop of a neural network performing a task, discovering edge cases, searching for similar scenarios, annotating them, and retraining to iteratively improve. 53:43 54:14
- Deep Learning Limitations and Future Research: In 2020, discussions highlighted deep learning's limitations in common sense reasoning and general knowledge accumulation. Future research should focus on reasoning, active/lifelong learning, open-domain dialogue, multimodal/multitask learning, and interdisciplinary collaboration to overcome these challenges. 9:46 11:26 11:42
- Ethical Scrutiny of Recommendation Systems: Recommendation systems are identified as the most impactful AI systems for the coming decades due to their influence on information consumption and societal thought. There is a critical need for greater transparency from tech companies and more informed public and governmental discourse regarding their ethical implications and control mechanisms. 1:05:40 1:06:18
