Intro to DeepMind

DeepMind’s initial aim was a programme to achieve Artificial General Intelligence (AGI)
- Step 1: solve intelligence.
- Step 2: solve everything else.
  - “To advance science and benefit humanity”
Two approaches to build AI
- Expert systems
  - Limited to what they have already seen
  - Relies on hardcoded knowledge
- Learning systems
  - Can generalise and learn new tasks
  - Inspired by neuroscience
Merge deep learning and reinforcement learning
- Deep reinforcement learning systems can learn new knowledge through a process of trial and error
- Feedback loop: Observations —> Agent —> Actions —> Environment —>

AlphaGo

The Game of Go
- The search space is huge. Number of configurations of the board is more than the number of atoms in the universe
  - All the computers in the world working for 1 million years couldn’t calculate all the possible moves.
- Impossible to write evaluation functions unlike in Chess
- Go players rely on intuition rather than calculation “this felt right”
How do we do it (Alpha Go Zero)? - V1 neural network (NN) plays against itself 100k times. It picks moves at random. This generates a dataset - V2 neural net is trained on V1 - V2 NN plays V1 NN 100 times - Best NN (55% win rate) then plays against itself another 100k times - Cycle repeats
- Combine the neural network with Monte Carlo search for efficient move choice.
Beat Lee Sedol (18 time world champion) 4-1.
Changed the way human beings view the game of Go.
- Move 37 will go down in history.
Generalisation
- AlphaZero generalised AlphaGo to 3 games: Chess, Go and Shogi
  - New style of chess: Mobility over Materiality
  - It will sacrifice pieces to get more mobility for it’s remaining ones; The Immortal Zugzwang Game
  - Human Chess Grandmaster searches 100 moves per decision, Stockfish: 100K, AlphaZero: 10K

Scientific Discovery

What makes for a suitable problem?
- Massive combinatorial search space
- Clear objective function
- Lots of training data or an accurate simulator to generate data.
Protein folding problem
- Can you predict the 3D form of a protein based only on its Amino Acid sequence?
- Proteins are essential to life. Their 3D structure determines their function
- It takes one PhD student (4 years) per protein structure.
- Levinthal’s paradox: 10^300 possible confirmations for an average protein. Yet nature solves this spontaneously.
- It was a long road to getting fully into protein folding. 90s intro at Uni full start 2016.
Alpha Fold 2
- AlphaFold 2 achieved atomic accuracy for protein structure prediction in 2020 CASP.
- Needed 32 component algorithms. 60 pages of suppl. info. Every part was a requirement for Alpha Fold 2 based on ablation studies
- Made the system completely end to end, rather than relying on an intermediate step.
- Architecture: attention based neural net.
- Included biological and physical constraints that didn’t impact the learning.
- Start with domain knowledge. Generality is secondary.
- 2 weeks to train. Prediction now takes minutes.
- Predicted 20K proteins encoded by the human genome.
Roadmap
- Prioritise neglected tropical disease for proteins structure prediction.
- Plan to solve 100m proteins over the next year. All proteins currently known to humans
- New era: digital biology
  - Can you derive the laws of motion of a cell? Can you learn them?
- Demis has been shipping. Ticking off childhood dream projects.
- Many products you use will probably have interacted with code from DeepMind.
  - Data centre and grid optimisation.
  - YouTube bit rates
Ethics:
- Has been central to their mission from the beginning.
- We should not move fast and break things.
- Use the scientific method: deliberation and foresight. Control tests. Update based on empirical data. Aim to get a better understanding.