Dynamics of sensorimotor adaptation Sen Cheng, Philip N Sabes University of California, San Francisco Annual Swartz-Sloan Centers Meeting, 26 th July 2005.

Dynamics of sensorimotor adaptation

Sen Cheng, Philip N SabesUniversity of California, San Francisco

Annual Swartz-Sloan Centers Meeting, 26th July 2005

2

A simple sensorimotor task

3

Motivation and outline

trial-by-trial dynamics

What is the learning rule of adaptation?1. What signals drive

learning?

2. Noise in the learning process?

3. Spatial anisotropies?

More powerful correlation between behavior and neural activity.

Steady-state of adaptation Compare average behavior

pre- and post-exposure

block design

4

Virtual reality setup

5

Concurrent test and exposure

6

Model for dynamics of adaptation

),0(~),,0(~

1

RNtrQNtq

ttt

tttt

rxy

qBuAxx

Linear dynamical system (LDS) ut : inputs (?)

xt : internal state, planned/expected reach error

yt : actual reach error

qt : learning noise

rt : motor noise

general state space model

7

1. What signals drive learning?

2. Noise in the learning process?

3. Spatial anisotropies?

Questions

8

Two candidate learning signals

System identification with expectation-maximization (EM) algorithm, Cheng and Sabes, 2005, submitted

Learning equation with two input signals

t : visual error

t : perturbation/ discrepancy betw. vision and proprioception

),0(~),,0(~

1

RNtrQNtq

ttt

tttt

rxy

qBuAxx

ttt

ttt

rx

y

ttttt qHGAxx 1

9

Sample data and vis-model fit

),0(~),,0(~

1

RNtrQNtq

ttt

tttt

rxy

qBAxx

perturbation

reach error

model prediction

10

Portmanteau test for serial autocorrelations

tt yy ˆIs the sequence of residuals a white noise sequence?

010

1

2 tr1

CCCCn

nPm

m

)ˆ)(ˆ(1

1

ttt

n

tt yyyy

nC

Portmanteau statistic (Hosking, 1980)

Residual autocorrelations Portmanteau test for vis-model

11

pert-model fit to sample data

),0(~),,0(~

1

RNtrQNtq

ttt

tttt

rxy

qBAxx

perturbation

reach error

vis-model

pert-model

12

Portmanteau test cannot distinguish models

for vis-modelfor pert-model

13

Likelihood ratio test (LRT) for nested models

M1: no input

M2: pert

M4: pert and vis

M3: vis error

p < 10-4 (n=18)p < 10-4 (n=18)

p=0.006 (n=1), p>0.067 (n=17) p>0.22 (n=18)

)(~)data|(

)data|(log2 2

ijj

i

ML

ML

ttt qAxx 1

tttt qGAxx 1

ttttt qHGAxx 1

tttt qHAxx 1

14

1. What signals drive learning?

2. Noise in the learning process?

3. Spatial anisotropies?

Questions

15

The signal that drives learning

tttt qGAxx 1

ttt qAxx 1

apply to no feedback (noFB) reaches:

Estimated modelspert-model

vis-model

tttt qHAxx 1

pert-model

vis-model

16

1. What signals drive learning? 2. Noise in the learning process?

3. Spatial anisotropies?

Questions

17

Learning noise

stochastic

pert

LRT (n=18)

p < 10-4

noFB

LRT (n=18)p < 0.0003

),0(~),,0(~

1

RNtrQNtq

ttt

tttt

rxy

qBuAxx

x

18

1. What signals drive learning? 2. Noise in the learning process? 3. Spatial anisotropies?

Questions

19

Anisotropy in learning and noise

*

),0(~),,0(~

1

RNtrQNtq

ttt

tttt

rxy

qBAxx

20

Conclusions

LDS are good models for adaptation dynamics

New insights into adaptation1. Visual error drives adaptation predominantly

2. There is learning noise

3. Dynamics are anisotropic

Can now correlate trial-by-trial changes of behavior with neural activity.

supported by the Swartz foundation