Fig. 7: Model generates sensorimotor mismatch prediction errors in line with experimental observations.
From: Self-supervised predictive learning accounts for cortical layer-specificity
a Illustration of visuomotor task used by Jordan and Keller19 in which mice learned to associate visual flow with locomotion. b Sample of training data from experiments (top) and our synthetic dataset (bottom). As in the experimental setup, we randomly halt the visual flow (flat green line) to generate visuomotor mismatches. c When the visual flow is halted, a sample neuron in L2/3 of the mouse visual cortex shows depolarization (magenta), while a sample neuron in L5 shows hyperpolarization (blue; left). In our model, in line with the data, L2/3 shows a positive mismatch error, while L5 shows a negative mismatch error (right). Shaded areas represent the standard deviation over five runs. d The mismatch error signals in L2/3 of the model are correlated with the modeled locomotion speed when the visual flow is halted (right), in line with experimental observations19,52 (left). e The model generates a distribution of mismatch errors which are biased towards positive errors in L2/3 and negative errors in L5 (right), in line with mismatch responses observed in primary visual cortex19 (left). Panels (a, c–e) were partially reprinted from19, with permission from Elsevier. Statistical tests are two-sided, and no adjustments were made for multiple comparisons. Error bars represent the standard error of the mean over five different initial conditions.
