Fig. 3: Neocortical circuitry jointly enables self-supervised learning.
From: Self-supervised predictive learning accounts for cortical layer-specificity
a Schematic of the model with individual components knocked out (colored crosses) within the neocortical microcircuit architecture. b Connections from L2/3 to L5 are necessary for L2/3 to learn a predictive representation of the input. c Impact of L4-mediated delay in self-supervised learning (dashed lines represent the optimal prediction). d Summary of decoding accuracy of the current input for L2/3 and L5 when specific connections are knocked out. The x-axis indicates the specific ablation, while the y-axis indicates the decoding accuracy for the current input (xt). e Similar to (d), but for the past input (xt−1). Knockout components in (d, e) are color-coded as in (a). Horizontal dashed lines in d, e represent chance decoding accuracy. Error bars represent the standard error of the mean over five different initial conditions.
