Modularity emerges in neural networks trained to perform context-dependent behavior

Humans and other animals often perform similar actions in different contexts. For instance, we use different criteria when judging whether an apple or avocado are ripe, but use very similar motor actions to move the fruit into our basket if we select it. We hypothesize that this ability to context switch is facilitated by functional modularity within brain regions, such as the specialization of a subpopulation of inferotemporal cortex neurons for face relative to non-face features. However, the link between this functional modularity and context-dependent behavior is not fully understood. Here, we show that functional modularity emerges in feedforward neural networks when they are trained to perform context-dependent tasks, even without any metabolic or anatomical constraints. In this setting, the network always uses the same output dimensions -- analogous to different motor actions -- but must learn to use a unique decision rule for each context, as in the case above. First, we analyze the representational geometry within the modules developed by the network and show that the representations are both abstract (or, disentangled) and flexible. That is, they enable generalization across different conditions even within a single module and they enable novel, nonlinear tasks to be learned with a simple linear readout. Next, we show that these modular representations emerge due to the geometry of the output and the known implicit bias of feedforward networks for shared representational structure. In particular, we show that the modular solution to the tasks has more shared structure than non-modular solutions. This work makes predictions for when and where modular structure will be observed in neural populations. In particular, we predict that modular structure will emerge when the same outputs are used in conflicting behavioral contexts. We are working to connect these results to experimental data that does, and does not, show modularity.