General Learning Algorithms

History / Edit / PDF / EPUB / BIB /
Created: March 18, 2016 / Updated: February 6, 2021 / Status: finished / 2 min read (~366 words)

This is a summary of the notes I've taken while listening to the presentation by Demis Hassabis available at https://www.youtube.com/watch?v=08Cl7ii6viY. It builds on the presentation Systems Neuroscience and AGI but can be watched on its own. This presentation covers most of the content in the presentation The Theory of Everything amd How Deep Learning Can Give Birth to General Artificial Intelligence as well.

  • DeepMind was founded in 2010
  • Acquired by Google in 2014
  • An Apollo Programme for AI (>100 scientists)
  • A new way to organize science

  • Learn automatically from raw inputs - not pre-programmed
  • General - same system can operate across a wide range of tasks

  • Build the best approximate model of the world
  • Update the model every time a new observation comes in
  • Use the model to make plans (simulations) toward a goal
  • Reinforcement learning in the brain through the dopamine system called TD (temporal difference) learning

  • Agents just get the raw pixels as inputs (~30K)
  • Goal is simply to maximize score
  • Everything learnt from scratch
  • One system to play all the different games

  • The key to flexible general intelligence
    • Apply previously learnt knowledge to a new situation
  • Identify the salient features in an environment
  • Re-represent those features as an abstract concept
  • Select and appropriately apply prior knowledge

  • Classical Computer (leads to)
  • Recurrent Neural Network + Memory Store (leads to)
  • Neural Turing Machine

  • Information overload and system complexity
  • Solving AI is potentially the meta-solution to all these problems
  • Empowering people through knowledge