Semiparametric pairwise model outperforms other models.

2017-09-19T17:24:52Z (GMT) by Jan Humplik Gašper Tkačik

A) Out-of-sample log-likelihood improvement relative to the pairwise model per sample per neuron averaged over subnetworks. Error bars denote variation over subnetworks (1 SD, no errorbars for N = 160 since there is only one subpopulation of that size in the entire dataset). The error in likelihood estimation is much smaller than the displayed error bars. B) The same as in A) but for single populations from two different experiments–one in which the population is stimulated with a random checkerboard stimulus, and the other where the population responds to a full-field flickering. C) The test set error rate averaged over neurons for predicting the response of a neuron from the activities of other neurons in 5 different subpopulations of 100 neurons. D) Average (across neurons) error rate decrease achieved by using a semiparametric pairwise model instead of a K-pairwise model for subpopulations of various sizes. Error bars denote 1 SD variation over subnetworks.