Efficient and dynamic neural geometry of value and modality encoding in the primate putamen for value-guided behavior

Abstract

The basal ganglia process diverse reward values from various modalities using limited resources, necessitating efficient processing. The convergence of sensory reward values at the single-neuron level enables the efficient use of limited neural resources. However, this raises a critical question: does such convergence compromise modality-specific information and degrade the overall information quality? Here, we reveal that the population representation of bimodal value neurons in the macaque putamen, which converges value information from tactile and visual inputs, efficiently preserves both value and modality information through shared abstract representations. These population representations generalized across identical modalities and values, establishing and maintaining an efficient low-dimensional representation as the neural geometry dynamically shifted toward value-guided movement within a single trial. Interestingly, a faster transformation of this geometry into a shared-value representation in bimodal value neurons was associated with a cognitive state reflecting well-adapted and well-learned value-guided behavior. In contrast, this relationship was notably absent in unimodal value neurons. Our results indicate that bimodal value neurons in the putamen play a key role in balancing efficiency and information fidelity through shared neural representations, with their dynamic changes facilitating the cognitive states required for value-guided behavior.