Fig. 8: Population activity analyses of contrast discrimination learning in monkey V4.
a,b, Contrast discrimination training significantly enhanced stimulus information at the population level (for decoding accuracy: one-sided paired t-test, t(5) = −6.03, P < 0.001, one-sided 95% CI −∞ to −3.6 × 10−2, Cohen’s d = −3.10 (a); for aLFI: one-sided paired t-test, t(5) = −2.21, P = 0.039, one-sided 95% CI −∞ to −52, Cohen’s d = −0.76 (b)). c–f, Consistent with VPL in the DCNNs and the human brain, training monkeys on a contrast discrimination task reduced Fano factors (c; one-sided paired t-test, t(5) = 7.28, P < 0.001, one-sided 95% CI 8.8 × 10−2 to ∞, Cohen’s d = 3.43), noise correlations (d; one-sided paired t-test, t(5) = 7.46, P < 0.001, one-sided 95% CI 2.6 × 10−2 to ∞, Cohen’s d = 5.80) and response variance (f; one-sided paired t-test, t(5) = 13.24, P < 0.001, one-sided 95% CI 1.6 × 10−1 to ∞, Cohen’s d = 8.70) but had no significant effect on signal separation (e; one-sided paired t-test, t(5) = −1.957, P = 0.054, one-sided 95% CI −3.7 × 10−1 to ∞, Cohen’s d = −0.30, BF10 2.41). g–i, We also found evidence for signal rotation (g) and manifold warping (h for PC variance and i for PC rotation). j, The stepwise information analyses also show the similar pattern of the four mechanisms. The unit ‘spk/s’ indicates the number of spikes per second (that is, firing rate). We calculate aLFI and information gain using stimulus contrast as decimal values (that is, 0.29), so they have arbitrary units. Each point is averaged over the two monkeys. See plots for individual monkeys in Supplementary Fig. 4. Data are presented as mean ± s.e.m., with error bars indicating the s.e.m. across the six conditions (n = 6).
