Interpretable MARL Models a Developmental Transition in Zebrafish Social Behavior

Machine-learning models of development are routine in biology, but relatively few analogues exist in neuroethology. Here we use multi-agent deep reinforcement learning (MARL) to test a candidate mechanism for a developmental transition in larval zebrafish social behavior. Larval zebrafish shift from avoiding conspecifics early in development to greater social proximity later. We train recurrent neural network (RNN) agents in a collective foraging task and model development as a reduction in the cost of close physical contact or collisions. This manipulation is sufficient to reproduce the developmental reduction in nearest-neighbor distance under patchy prey. We then interpret the trained models at three levels: behavior, visual attention, and RNN representations. Agents reproduce the endpoints of the transition behaviorally, as measured by nearest-neighbor distances. Across development, agents reorganize visual allocation, while RNN activity becomes lower dimensional and more strongly shared between nearby agents; inter-agent distance also becomes more linearly decodable from recurrent state. Together, these results show how our interpretable MARL framework can serve as an in silico system for testing mechanistic hypotheses about developmental behavior and generating experimentally testable predictions for biological systems.