2. Dana Ballard - University of Rochester1 Distributed Synchrony: a model for cortical communication Madhur Ambastha Jonathan Shaw Zuohua Zhang Dana H. Ballard Department of Computer Science University of Rochester Rochester, NY

3. Summary 1. There is a computational hierarchy. 2. At the bottom of the hierarchy is the need to calibrate. 3. To communicate throughout cortex quickly, calibration uses the  band

6. ContextSelect a set of active behaviors ~10s ResourceMap active behaviors onto motor system ~.3s Routinesupdate state information~100ms Calibrationrepresent sensory/motor/reward ~20ms Computational quanta ~2ms 1. Computational Timescales

7. 2. How can the Cortical Memory Self-Calibrate? Olshausen and Field 97 Rao and Ballard 99

8. Code Input I with synapses U and output r Coding cost of residual error Coding Cost of model Min E(U,r)= |I-Ur| 2 + F(r) + G(U)

9. Synapses are Trained with Natural Images 1. Apply Image 2. Change firing 3. Change Synapses

10. An Example: LGN-V1Circuit r - + U r est I U T e = I - Ur LGN Cortex

11. Hierarchical Memory Organization Fellerman and Van Essen 85

12. A Slice Through The Cortex - + r - + r - + r LGNV1V2 X

13. Rao and Ballard, Nature Neuroscience 1999 RF Endstopping

14. 3. Can Predictive Coding work with individual spikes?

15. Spike Timing Model _ + r Loop delay - 20 milliseconds

16. LGN-V1 Circuit using Spikes r - + U r est I U T e - + U r est I- U T e

17. Spike Models Spike is probabilistic Deterministic spike has area

18. inputfeedback prediction error LGN ON LGN OFF IUrI-Ur

19. Receptive Fields Orientation Distribution Coding Cells

20. Responses are Random and Phasic

21. Projection Pursuit Iu1u1 u2u2 r1r1 r2r2 r 1 = I u 1 r 2 = ( I - r 1 u 1 ) u 2

22. Microcircuit Details 1 I I I I I  r 1 u 1 u 2 r 1 = I u 1 r 2 = I u 2 - r 1 u 1 u 2 2

24. Summary 1: Distributed Synchrony is motivated by four principle constraints 1. Fast, reliable intercortical communication 2. The ‘need’ for a cell to multiplex 3. Need to poll the input 4.The need to reproduce observed cell responses

25. Summary 2: Isolating Computations = The Binding problem Solutions: 1. There is no binding problem - 2. Fast weight changes at synapses - 3.Synchrony encodes the stimulus - 4.Synchrony encodes the answer - 5.Synchrony encodes the process - Solutions: 1. There is no binding problem - Movshon 2. Fast weight changes at synapses - von der Malsburg 3.Synchrony encodes the stimulus - Singer 4.Synchrony encodes the answer - Koch and others 5.Synchrony encodes the process - Distributed Synchrony

26. Thanks !

27. Handling the Error Term with Predictive Coding I r1r1 r2r2 LGN Cortex

28. Roelfsema et al PNAS 2003

29. Diesmann, Gewaltig,Aertsen Nature 402, p529 1999 Synchronous Spikes Can Propagate

30. Max M P(M|D)= Max M [P(D|M)P(M)/P(D)] Minimum Description Length - Bayesian Version Can neglect P(D) and take logs… Max M [log P(D|M)+ log P(M)] Or equivalently minimize negative logs… Min M [ - log P(D|M) - log P(M)] If we use exponentiated probability distributions, log cancels negated exponent so… Coding cost of residual error Coding cost of model

31. Singer group, J Neuroscience 1997

32. Cortical Inhibitory Cells Can Oscillate at 20-50 Hz Beierlein, Gibson, Connors Nature Neuroscience 3 p904 2000

33. Temporal Rate Coding: A Strategy that cannot possibly work

35. Reconstruction as a function of Coding Cost low high inputfeedbackerror LGN ON LGN OFF LGN ON LGN OFF

36. Spectral software supplied by Daeyeol Lee

37. Distributed Synchrony

38. Coding Cost as a function of Signaling Strategy

39. Axonal Propagation Speeds: Evidence? 2-6 cm/s 0.1 - 0.4 cm/s

40. Visual Routine

42. Reverse Correlation  + + +

43. Spatio-temporal behavior of LGN Cells Experiment (Reid & Usrey) Model Time - milliseconds 30 507090 Using Reverse Correlation