Fig. 4

Sensory contexts cannot fully explain behavioral variability across fish. We encode the behavior of each fish in transition matrices built with an increasing number of coarse-grained states , and use regularized logistic regression to classify each fish to their respective sensory contexts. (A) Accuracy of the classification task (fraction of correct classifications) for the Train (black) and Test (yellow) sets plotted as a function of the number of coarse-grained states (SI Appendix): the Train accuracy grows continuously as a function of and rapidly reaches while the test accuracy reaches only , far from perfect. Minimal improvements are seen above above which the classifier overfits. states correspond to Left?Right variations of slow-fast cruising?wandering motor strategies (Right, example trajectories). (B) Confusion matrix of the classifier for in which each row reflects the proportion of individuals assigned from each sensory context to all conditions. A strong diagonal component indicates that a large portion of fish are correctly assigned, while off-diagonal components point to fish misclassified into a different sensory context. Sensory contexts with visual stimuli are labeled with squares while prey exposure conditions are labeled with a diamond.