Excitatory-Inhibitory Dynamics in Adaptive Decision-Making

Veronica Chelu, Doina Precup

Abstract: Neural circuits rely on excitatory/inhibitory (E/I) interactions to support adaptive learning and decision-making. Here, we investigate how these dynamics contribute to flexible behaviour across three modelling levels. First, using a mean-field model of two-choice decision-making, we examine the computational role of recurrent excitation and feedforward inhibition in stabilizing or amplifying competition between alternative choices. Building on these insights, we then simulate how adjustments in choice distribution resulting from E/I perturbations modulate behavioural adaptation in reversal learning using a Bayesian model of decision-making. Finally, to assess scalability we analyze the learning dynamics of decision-making agents with E/I-constrained recurrent neural networks (RNNs) trained in partially-observable environments. We identify optimization challenges arising from naive E/I partitioning of neurons in RNNs and link them to inhibitory motifs. We also find that an architecture with strict feedforward inhibition bypasses these issues, enabling stable learning dynamics. Overall, these results highlight E/I interactions as a computational mechanism for flexible decision-making, and emphasize the importance of architectural constraints for effective optimization.

External IDs:doi:10.1007/978-981-95-4381-6_5