Mutual information minimization enhances noise robustness and induces fault containment in functionally differentiated recurrent neural networks.

Functional differentiation, the emergence of specialized neural populations, is a hallmark of biological brains and has been proposed to provide robustness, metabolic efficiency, and evolvability. In our previous work, we showed that minimizing mutual information (MI) between predefined subgroups of a recurrent neural network (RNN), using mutual information neural estimation (MINE), promotes the emergence of functionally specialized modules. However, whether such information-theoretically induced differentiation translates into functional benefits has remained unclear. Here we examine the robustness of MI-minimized RNNs trained on a chaotic signal separation task in which a superimposed input from the Lorenz and Rössler systems must be demixed into separate output channels. We find that MI minimization, while preserving task accuracy, consistently, and significantly enhances tolerance to input noise and produces qualitatively distinct responses to neuronal ablation: damage to the subgroup an output relies on impairs that output sharply, whereas damage to the other subgroup leaves it largely intact. This selectivity is essentially absent in control networks trained without the MI constraint. These results indicate that minimizing statistical dependence between neural populations not only induces functional specialization but also confers a form of fault containment, isolating damage to the affected output while sparing the others, supporting the view that modular organization is an adaptive consequence of information-theoretic constraints on neural representation.