machine intelligence sweden cto hans salomonsson google’s ... · machine intelligence sweden ab...

24
Machine Intelligence Sweden AB Google’s AlphaGo ...and a career in machine learning Hans Salomonsson CTO Machine Intelligence Sweden

Upload: others

Post on 25-May-2020

10 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Google’s AlphaGo...and a career in machine learning

Hans SalomonssonCTO

Machine Intelligence Sweden

Page 2: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

4 - 1

Page 3: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

GoViewed as the most complex of classical games due to its huge search space

b - no legal moves per positiond - game length (iterations)bd sequences of moves

chess: b = 35 d = 80go: b = 250 d = 150

Page 4: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Search Space

Source: https://thenewstack.io/google-ai-beats-human-champion-complex-game-ever-invented/

Page 5: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Monte Carlo Tree Search

Source: https://en.wikipedia.org/wiki/Monte_Carlo_tree_search

Page 6: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Value Function vs Policy ApproximationDeep Q-learning approximates the action-value function Q(s, a) with an ANN

Page 7: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Value Function vs Policy Approximation

Source: Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484–489 (2016)

Page 8: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Architecture Overview

Source: Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484–489 (2016)

Page 9: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

MCTS in AlphaGo

Source: Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484–489 (2016)

Page 10: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

MCTS in AlphaGo● Do MCTS until time budget is out, then select the action with the maximum

visit count● AlphaGo resigns when its overall evaluation drops below an estimated 10%

probability of winning the game

Page 11: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

HardwareThe final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs.

Source: Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature

529, 484–489 (2016)

Page 12: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

AlphaZeroGoal: Remove all human knowledge, i.e.

● No human features● No human data

In addition to this

● Single neural network● Simpler search (no MCTS)

Page 13: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Source: Silver, D. et al. Mastering the Game of Go without Human Knowledge. Nature 550, 354–359 (2017)

Page 14: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

How to train “Policy Network”

Source: Silver, D. et al. Mastering the Game of Go without Human Knowledge. Nature 550, 354–359 (2017)

Page 15: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

How to train “Value Network”

Source: Silver, D. et al. Mastering the Game of Go without Human Knowledge. Nature 550, 354–359 (2017)

Page 16: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

ResultSource: Silver, D. et al. Mastering the Game of Go without Human Knowledge. Nature 550, 354–359 (2017)

Page 17: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Takeaways ● Superhuman performance at GO with DL and RL● NN flexible tool● Balance trade-off speed and accuracy ● Some things are surprising - test things!● Hardware important

Page 18: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

www.machineintelligence.se

Machine Intelligence Sweden (MIS)an AI garage

member of:

Page 19: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

www.sciencerouter.com

About us

19

Devdatt DubhashiCEO Professor CSE Chalmers UniversityLeads the Machine Learning and Algorithms GroupPh.D. Cornell University, U.S.A.

Hans SalomonssonCTO Has built AI systems since 2010 M.Sc. Complex Adaptive SystemsB.Sc. Engineering PhysicsB.Sc. Industrial and Financial Mgmt

Anton KloekData ScientistM.Sc. Algorithms, Languages and LogicB.Sc. Computer Science and Engg.

autonoma.aiTools for everything autonomous

PhD First LastnameData Scientist(Starts after summer)

Seeded by SSF

Stefan SvahnCEOEntrepreneur (Lumina Adhesives)M.Sc. Entrepreneurship and Inn.B.Sc. Engineering PhysicsB.Sc. International Economics

mission: We do

AI-powered research

and innovations

innovera.ai the platform for the knowledge based economy

Funded by EU and FFI Vinnova

Page 20: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

www.machineintelligence.se

autonoma.aisafer and more robust autonomous

driving systems done cheaper and faster

● Real and simulated data: bridging the gap● Validation of AD: perception and control

Mission:

Realization:

Page 21: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

www.sciencerouter.com

Connecting researchers and industryStandardizing Industry-Academia

interactions

the easy way for researchers and companies to collaborate

infrastructure for the knowledge economy

Page 22: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

www.sciencerouter.com

26 368 universities

7.8 million researchers

> 2.5 million articles / year

77 million companies in China

10 million companies in India

45 508 companies listed in stock

exchanges worldwide

30 million companies in USA

26 million companies in EU

22

Find Initiate Facilitate

Page 23: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

Machine Intelligence Sweden AB

Machine Intelligence Sweden AB

ML Consulting ML Startups ML Research

Page 24: Machine Intelligence Sweden CTO Hans Salomonsson Google’s ... · Machine Intelligence Sweden AB Hardware The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs

www.machineintelligence.se

Do you want to work with ML in Gothenburg?We can offer:

● Permanent positions● M.Sc. thesis● Part time position

Send us an email to

[email protected][email protected]

with CV, PL with ML/tech experience and what application areas you might be interested in working with.