Deep Learning State of the Art (2020)

This is the 2020 State of the Art lecture in Lex Fridman's MIT deep learning series, recapping the major advances of 2017-2019. He surveys the maturing of TensorFlow and PyTorch, the transformer explosion (BERT, XLNet, GPT-2, Megatron), reinforcement learning milestones (OpenAI Five in Dota 2, AlphaStar in StarCraft, Pluribus in poker, the Rubik's Cube hand), and the two competing autonomous-vehicle philosophies of Waymo and Tesla. He frames recurring themes of self-play, active learning, common-sense reasoning, and the under-discussed power of recommendation systems. The talk closes with audience Q&A on AGI, machine emotions, ethics, and who will ultimately control AI.

• 2019 was the first year it became cool to highlight the limits of deep learning, with books and press declaring 'the era of deep learning is over.'
  00:09:32
• Reflecting on the GPT-2 staged-release scare, he argues humans turned out to be more dangerous than the AI, but the thought experiment was valuable.
  00:24:43
• He contrasts the two AV approaches: for Waymo deep learning is 'the icing on the cake,' but for Tesla 'deep learning is the cake.'
  00:52:29
• He claims recommendation systems are the most powerful and impactful AI of the coming decades yet are barely discussed publicly.
  01:05:32
• Asked if machines will think and feel, he answers '100 percent yes' because the display of emotion is emotion to him.
  01:15:27
• He predicts the first time a product says 'please don't hurt me' with a straight face is when torturing AI becomes unethical.
  01:17:33
• His real worry is not AI as our masters but owners of large tech companies using AI to control humans.
  01:25:53
• He stresses self-play, where agents learn by competing against incrementally better versions of themselves, as one of deep learning's most exciting ideas.
  00:32:41

• January 1st 2020 marked the end of Python 2 support in TensorFlow and PyTorch.
  00:14:45
• OpenAI Five consumed 800 petaflop/s-days and experienced about 45,000 years of Dota self-play over 10 real-time months.
  00:36:20
• The 2019 OpenAI Five had a 99.9 percent win rate against the 2018 version.
  00:36:52
• AlphaStar reached Grandmaster in StarCraft in 2019 by using a camera and the same constraints humans face.
  00:39:28
• The lottery ticket hypothesis shows small sub-networks inside a large network can match its full accuracy.
  00:45:42
• A best-paper result at ICML 2019 proved disentangled representations are impossible without inductive biases.
  00:47:47
• Waymo logged about 20 million real-world miles and 10 billion simulated miles by that year.
  00:51:58
• Poker pros described Pluribus as extremely hard to read, excelling at thin value bets and mixed strategies humans struggle to execute.
  00:42:38
• Fridman believes more data, bigger networks, and better data selection will take us further than hybrid symbolic approaches.
  01:13:53
• He originally wanted to be a psychiatrist to engineer the human mind, then learned to program in C++ at age 12.
  01:14:24