Notes — Demis Hassabis (Demis Hassabis, CEO, DeepMind Technologies — The Theory of Everything)
Please find below some notes on Demis Hassabis talk “The Theory of Everything”:
“AI is the science of making machines smart.”
AI: one of the most important technologies humanity will ever invent.
Personal journey:
Two subjects worth studying: physics & neuroscience.
Physics: explaining the External world
Neuroscience: explaining the Internal world
But, mind is more important because it´s with what we interpret the external world
“The mind interprets the world” — Kant
“The ultimate expression of understanding something is recreate it.” : “What I cannot build, I do not truly understand.” — Feynman
Hassabis started with games, chess. Learned chess when he was 4. Chessmaster with 12. Started programming. Loved video games. Created Theme Park, one of the first games to use AI as one of it´s main components. It started a new genre of video games: management simulation games. After 10 years, sold video games company and went back to university to study a PhD in neuroscience. Choses imagination and memory as topics since they are two brain capabilities that we don´t know how to do in AI very well. Wanted inspiration to see how the brain solves this problems. Post docs at MIT & Harvard. New era: start DeepMind in 2010. Idea: create like an Apollo program mission for AI (>100 scientists, top in machine learning and neuroscience fields). It´s neuroscience inspired AI.
DeepMind Mission:
(Very easy to articulate, very hard to do).
- Solve intelligence
- Use it to solve everything else
They want to build the world´s first General Purpose Learning Machine.
Aim:
Learn automatically from raw inputs, not pre-programmed
General — same system can operate across a wide range of tasks
Artificial `General´ Intelligence (AGI) vs. Narrow AI
Flexible, general and adaptive.
We learn about intelligence from a framework called “reinforcement learning”.
System (there is a goal) interacts with the environment (can be real world or virtual).
The system only interacts with the environment in 2 ways:
- Observations through sensory operators (normally vision in DeepMind). They are noisy and incomplete. Unlike chess, the real world is noisy and messy.
- Pick the best possible action that will get her closer to his goal.
From pixels to actions:
To have true cognition the system has to be embedded in a sensorimotor stream
Games are the perfect platform for developing and testing AI algorithms
Atari 2600 testbed: +100 classic 8-bit Atari games from the 80s
- AI system just gets the raw pixels as inputs (30K)
- Goal is simply to maximize score
- Everything learnt from scratch
- One system to play ALL the different games
Concepts and Memory
Inspired by Hassabis and others people work in neuroscience. They are trying to mimmick the hippocampus.
Meta-Solutions
- Information overload: big data (problem is what to do with it once you have a lot of data), genomics, entertainment, all suffer from IO. Personalization technologies are trying to help solver the IO challenge
- System complexity: climate, disease, energy, macroeconomics, physics
Solving AI is potentially the meta-solution to ALL these problems.
DeepMind partnershiped with Google in 2014. Google´s mission statement is “To organize the world´s information and make it universally accessible and useful. DeepMinds mission fits very well with Google´s.
Mission description: EMPOWERING PEOPLE THROUGH KNOWLEDGE. AGI automatically converts information into knowledge.
The Deepest Mysteries of the Universe:
- By building AI and comparing it to the mind we may unlock its mysteries.
- To truly find the Theory of Everything we may have to solve intelligence first.
