Figure 5.

Generalization Disadvantage for Models with Key Features of CGF Elided

(A–C) Cortex; (D–F) thalamus. Each panel contrasts the effects of eliding parameters in two identically sized sections of the CGF (gray rectangles): one corresponding to a CGF feature that appeared to consistently shape input-specific gain, the other a control section where CGF weights were inconsistent or small. Weights in the elided regions were fixed at zero, and the model was re-fit to optimize the remaining model parameters. Histograms show distribution across neurons of differences in cross-validation predictive performance (generalization accuracy) relative to the unelided CGF model; p value indicates significance threshold at which the hypothesis that median change in performance equals or exceeds zero can be rejected (one-tailed sign test, uncorrected; N = 64 in cortex and 101 in thalamus). Scatter plots compare generalization accuracy of the two elided models neuron-by-neuron; p value indicates threshold for rejection of the hypothesis that median difference for feature elision minus control elision equals or exceeds zero (one-tailed sign test, uncorrected). Across the neural population, elision of key CGF features always resulted in poorer generalization accuracy than that achieved by the full (unelided) model. By contrast, control elisions had significantly less impact; the hypothesis that control elisions produced no reduction in predictive performance could not be rejected in any case after correction for multiple comparisons.