W6 An improved estimator of Variance Explained in M. the presence of noise RaalofratoryHaefner & Bruce G. Cumming Bethesda MD Lb for Sensorimotor Research, NEI/NIH,
- Traditional estimator strongly biased - Previous improvement accounted for noise in data (Sahani & Linden, 2003) - New estimator also adjusts for overfitting without extra validation data - Additional conditioning term to improve performance for very high noise E x a m p le : Virtually all variance in V1 disparity tuning curves explained by G abor fu n c tio n s Noise power / signal power

B ia s

Noise level

simulation

Unbiased VE
data

RMSE