Towards General Artificial Intelligence

31536563883 302bda8bd6 z

Demis Hassabis of DeepMind explained how his company is exploring the frontiers of knowledge via a game called Go.


• Mission:

1) solve intelligence

2) use it to solve everything else


AlphaGo challenge

• game of Go in Asia

• why is Go hard for computers to play?

• more popular than chess, 3,000 years old, 40M players, 10^170 possible configurations

• not to solve with brute force computation? - more possibilities (more board positions) as atoms in the universe

• it’s the holy grail for computer science

• 2 main challenges:

1) search space is really huge (branching factor of 200)

2) "impossible" to write evaluation function to determine who is winning

• Go is primarily a game about intuition rather than calculation like chess

• two networks: policy and value nets - system predictions who is winning the game

• policy network (probability Distribution over moves)

• value network ( real numbers: 0 white - 1 black)

• system to learn it (value), not only programming it


Legendary game

• March 2016, legend for the game Go Lee Sedol, against the predictions, they won

• „A decade before its time“ - before it was predicted to be solved technically

• AlphaGo match: 280M views, 35K press articles, 10x board sales (sold out for months)


Exploring the limits

• fix knowledge gaps and optimize performance

• understand the representations AlphaGo is using, interpretability

• learn from scratch with no bootstrapping from human games

• find the limit of the self-improving process - is there a limit

• How far of optimal is professional play?

• give back to Go community - help to improve human play


New era of Go

• „To Create PERFECTION“

• Go books are rewritten, schools improve/strategies are improved because of AlphaGo

• Ke Jie, Chinese/world No. 1: "AI has shown us, we have not yet even scratched the surface." (after playing Go for thousands of years)

• chess programs are better tactically, AlphaGo is better strategically, used by the best human players

• the power of man and machine - humans and machines collaborating together will achieve amazing things

• the ultimate tool = Hubble telescope, to explore the Universe

• many other domains, domains with „combinatorial explosion“, material design, drug discovery - 2 areas they are looking at

• data center optimization - 40% more efficiency for cooling system = great for the environment, we could use it for national power suppliers

• meta-solutions

• information overload and system complexity are huge challenges


Mentioned in this live blog

Demishassabisquadr
Demis Hassabis
DeepMind
Founder & CEO
DeepMind
http://dld-conference.com/users/demis-hassabis
London
http://s3.amazonaws.com/dldwebsite-production/black_and_white_avatars/demis-hassabis/large/DemisHassabisquadr.jpg

All DLD17 Blogposts