Neural geometry efficiently representing abstract form of value and modality in the primate basal ganglia

Abstract

Basal ganglia process diverse values from various modalities using limited resources, necessitating efficient processing. This involves converging tactile and visual information within bimodal value-coding neurons, ensuring efficient processing with limited number of neurons. However, convergence at the single neuron level may compromise modality-specific information, raising the question: Does efficient processing inevitably degrade information quality? We investigated the representational geometry in the putamen of macaque monkeys trained to learn values from tactile and visual inputs. Here, we demonstrated that the population representation of bimodal value-coding neurons in the putamen preserved both value and modality information, and these representations were shared to efficiently maintain quality. Notably, these representations were generalized across identical modalities and values, resulting in an efficient low-dimensional representation. Furthermore, a faster transformation to a generalized value representation within neural geometry reflected greater confidence in value-guided choice behavior--this correlation not observed in conventional decoding. Our results suggest that bimodal value-coding neurons play a key role in balancing efficiency and information fidelity, facilitating cognitive states required for confident decision-making.