Abstract
Repeated exposure to stimuli elicits decreasing sensory neural responses over time (adaptation). However, resulting behavioral responses can either weaken over time (habituation) or remain invariant, indicating that the neural-behavioral link is not fixed. To investigate neural adaptation and its flexible translation into behavioral decision making, we created a mathematical framework hypothesizing (1) that sensory networks optimize the speed and accuracy of encoding exogenous stimuli, and (2) that representations form along two time scales, one embedding immediate information and the other stimulus history. Using experimental recordings of the nematode C. elegans, we validated normative model predictions of this optimal encoding strategy, specifically how neural dynamics and adaptation levels vary with stimulus timing. A parametric Bayesian decoder architecture predicted conditions leading to behavioral habituation or invariance, but also paradoxical inversion, whereby appetitive stimuli elicit aversive responses. Experiments with food odors validated that inversion behavior occurred after several repetitions with a long stimulation time and low odor concentrations. Mechanistically, during sensory neural adaptation, weaker immediate stimulus representations can be compensated by secondaryprocesses through memory effects, with biological origins that remain to be studied.