There’s always lots of interesting new applications of machine learning being released. In this article, we’ll take a step back from the avalanche of news and consider just three exciting results from DeepMind, namely:
- AlphaGo is cool, nuclear fusion is cooler: how DeepMind and the team at the EPFL tokamak used reinforcement learning to control the plasma in the reactor.
- Ithaca: a deep learning model that has been used to restore ancient greek texts and promises to help historians better understand ancient times.
- How my former supervisor, Petar, has built with his colleagues a machine-learning system to guide intuition in pure mathematics and propose conjectures, Ramanujan style.
Deep Reinforcement Learning used to control Nuclear Fusion Reactions
Nuclear fusion can release four times more energy than nuclear fission (i.e. what happens in our nuclear reactors), and doesn’t generate any nuclear waste! The main ingredient: hydrogen, and you can find that in water.
Physicists have been trying to achieve stable fusion reactions since the mid-1960s. The best design for such a reactor is known as a “tokamak”, a machine much like a hollow doughnut, except it’s a magnet. It has to be a doughnut so you can get particles to zoom round…