Search Papers | Poster Sessions | All Posters
Poster C129 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Recurrent circuit mechanisms for reward learning in multidimensional environments
Michael Chong Wang1, Alireza Soltani1; 1Dartmouth College
Decision making and learning in naturalistic environments involve choice options with multiple features, whereas usually only a few features and/or their conjunctions are predictive of their associated reward outcomes. It has been shown that humans deploy attention to selectively learn about the predictive values of features and feature conjunctions and generalize those values to similar stimuli/objects. This behavior can be captured by reinforcement learning models with explicit value representations. But how are such representations learned and used for decision making in neural circuits with mixed selectivity, and how does attention modulate these processes? To address these questions, we trained multi-area recurrent neural networks endowed with reward-dependent Hebbian plasticity on a multidimensional reward learning task. After training the networks to perform the task across diverse reward schedules, we tested them on the reward schedule used in a recent human study. The networks exhibited similar attentional biases as those observed experimentally. Despite their distinct topographies, we found that different networks shared an interpretable latent circuit organization that resembled the architecture of attractor network models. Specifically, distributed but orthogonal subspaces were used to encode and communicate information about different features and conjunctions within and across network areas, enabling the simultaneous learning of feature and conjunction values through reward-dependent Hebbian learning. Finally, we discuss how this structure gives rise to value-based selective attention, providing insight into how the underlying mechanisms can be validated in future experiments.
Keywords: recurrent neural networks Hebbian plasticity reinforcement learning selective attention