OpenTalks.AI - Сергей Шумский, Символьное мышление роботов

Transcript

1. The Symbolic Mind of Robots Sergey Shumsky, 2019
2. Motivation: common architecture of robotic brains Reinforcement learning
3. Motivation: common architecture of robotic brains
4. Agenda I. Problem Deep Reinforcement Learning II. Solution Deep Symbolic Learning III. Prospects Robot Operating System
5. Agenda I. Problem Deep Reinforcement Learning
6. AlphaZero: intuition + calculation ▪ Deep intuition: Yes ▪ Deep planning: No ▪ Complexity of learning ∝ 2 , 3 TFLOPSyears – weights update 500 TFLOPSyears – search options
7. Google DeepMind research program "One way you can think about our research program is: 'Can we build out from our perception, using deep-learning systems and learning from first principles? Can we build out all the way to high-level thinking and symbolic thinking?' " D. Hassabis (Google DeepMind)
8. Agenda I. Problem Deep Reinforcement Learning II. Solution Deep Symbolic Learning
9. Deep symbolic learning ▪ Hierarchy of plans state sequences ▪ Planning and Learning in real time ▪ Complexity of learning ∝ ∝
10. Images and Symbols Coding Decoding Image ~ 106 bits Sensors  States  Actions  Effectors Image ~ 106 bits Symbols ~ 5 bits Planning
11. The trick: symbolic coding of images 2106 ~ , → 1 2 … > ( = 30, k = 7)  symbols  images  = 210 <
12. Symbolic thinking ~ planning ▪ Coding of sequences ~ k variants (words) ~ 30 Symbolic sequences can be remembered symbol +1 +2 ~ 30 ~ 30
13. Symbolic thinking ~ planning символ +1 +2 Image code Sequence code (pattern)
14. Deep symbolic learning Layer L Layer L -1 1 2 …
15. Deep symbolic learning ▪ Learning useful patterns of interaction with the world ▪ At all hierarchical levels ▪ In real time Encoder- Decoder +1 +1 Parser +1 Encoder- Decoder Parser symbols patterns symbols
16. Agenda I. Problem Deep Reinforcement Learning II. Solution Deep Symbolic Learning III. Prospects Robot Operating System
17. Prospects Sensory intelligence Strategic intelligence Robotic intelligence   Goal setting and planning to achieve them Achieving the goal
18. Open AI gym Mountain car (Igor Pivovarov)
19. Interested? Join us! Serge.Shumsky@gmail.com igorpivovarov@yandex.ru