1. Flexible Shaping: How learning in small steps helps Hierarchical Organization of Behavior, NIPS 2007 Kai Krueger and Peter Dayan Gatsby Computational Neuroscience Unit

2. Outline Introduction learning may require external guidance shaping as a concept: Set-Up 12-AX task, LSTM network, shaping procedure Results simple shaping when does shaping help most? flexibility to adapt to variations Conclusions, issues and future work rules and habits

3. Introduction Learning essential for flexibility trial and error external guidance: one shot teaching by verbal explanation of abstract rules imitation shaping Guidance critical for complex behavior branching, working memory, rapid changes

4. Shaping a method of successive approximations (Skinner 1938) Key features : external alteration of reward contingencies withdrawal of intermittent rewards Creates behavioral units e.g. lever pressing of a rat Separate time scales / branching points by providing separate stages in shaping Ubiquitously (and implicitly) in animal experiments

5. 12-AX task Demo

6. LSTM network Long Short-Term memory (Hochreiter and Schmidthuber 1997) 3-layer recurrent neural network Provides built-in mechanisms for: working memory gating (input, output and forget) Abstract over-simplified model of PFC basis to motivate PBWM (OReilly et al.)

7. Shaping procedure Teach 12-AX as successive approximations Separate WM timescales: long: (1 / 2 ) short: (AX/BY) Learning in 7 stages last stage: full 12-AX Resource allocation currently done by hand each stage learned into a new block all other memory blocks disabled provides separation / No interference

8. Simple shaping Improvement in learning times: 8 fold decrease (only final stage) significantly better (including complete training) median: 13 epochs, min: 8 epochs Need the 4 stages of shaping 1 and 2 High variance in shaping times

9. What makes shaping work Robustness to additional structure: irrelevant experience related and unrelated tasks / inputs Resource allocation: interference between tasks => no benefits

10. Shaping: when is it useful? Can shaping prevent scaling of learning time with task complexity? One aspect of complexity: Temporal credit assignment increase the outer loop length => higher temporal complexity Results: training time still increases, but scales much slower. increasing complexity => shaping more important

11. Rule abstraction Rule abstraction: flexibility to cope with change in statistics Train on the base 12-AX task (loop length 4) Test with variations loop lengths 4, 5, 12, 40 disable learning Should perform perfectly abstract rules have not changed

12. Generalisation Generalisation: cope with yet unseen data (inner loop combinations) Train 12-AX task without AZ and CX Test performance on full task Only 7 combinations one valid generalisation only? Mixed results: differences in emphasis (1- back / 0-back) overall shaping still better

13. Reversal learning Reverse stimulus – rule association shape all components needed Repeatedly reverse (after 500 epochs) learning of reversals. Identify flexibility to perform reversals unshaped: mostly fails shaped: succeeds more often

14. Conclusions Shaping works Reduces learning times Helps learning long time delays separating time scales of actions recombine behavioral units into sequences Improves abstraction and separation Increases flexibility to reversals Take home message: need to take sequential and transfer learning more into account when looking at learning architectures. Still issues to solve though

15. Limitations Resource allocation prime computational issue done by hand (Homunculus) ideas to automate: compute responsibilities Mosaic Experimental data no published data on learning 12-AX interesting manipulations: loop length, target frequency, … natural grouping of alphabet

16. Future Work Still based on habitual learning => no instant reprogramming Need additional mechanisms: more explicit rules variable substitution Bilinear rules framework: (Dayan 2007) recall match execute Close interaction between habitual and rule based learning rules supervise habit learning habits form basis of rule execution Results in a task grammar?

17. Questions? Thank you

18. Shaping in computational work Surprisingly few implementations: instead: elaborate learning framework to learn in one go Learning grammars: simple examples first (Elman 1993) self simplification by network limitations Robotics learning in a grid world (Singh1992) (Dorigo and Colombetti 1998) robot vs animal shaping (Savage 1998)

19. Shaping procedure Remembering 1 and 2 anything following a 1 => L / 2 => R progressively increase loop length 1-back memory for A / B unconditional CPT-AX 2 Separate stages for A and B

20. LSTM network

21. Examples of learning in LSTM

22. Shaping at work

23. Rule abstraction

24. Reversal learning continued

25. Alternative shaping paths Not all shaping paths have equal properties Trivial shaping path: each stage full 12-AX rules increase loop length from 1 through 4 Similar training time, but bad generalisation

26. Rules individual rules: recall match execute