DeepMind (DeepMind Technologies)

Contribution to AI

The founding bet was that reinforcement learning, which had spent decades on toy problems, would scale. DQN was the first real evidence: one network and one set of hyperparameters learning to play dozens of Atari games from raw pixels and a score, with no game-specific engineering. Nothing before it had learned that many different tasks from the same starting point. AlphaGo made the case publicly. Go had resisted the search techniques that beat chess because the branching factor is far too large to enumerate, so the programme paired tree search with networks that learned to judge positions and propose moves, restoring search to a game where search alone was hopeless. AlphaZero then discarded the human games entirely and reached the same standard from self-play and the rules. AlphaFold is the clearest instance of the field paying something back. Predicting a protein's folded shape from its sequence had been open for fifty years; by 2020 the predictions were accurate enough for many experimental purposes, and the structures were released as a public database rather than sold as a product. The lasting argument is that deep learning is not only a perception technique — it can supply the judgement that search and planning need.