A transformer-based model reveals sparse, stimulus-dependent orientation readout from macaque V1 population activity

Abstract

Orientation information in primary visual cortex (V1) is represented by large populations of neurons with overlapping tuning preferences, yet how this information is selectively read out remains unclear. Here we proposed a transformer-based model to reconstruct oriented Gabor stimuli from two-photon calcium responses of more than 1000 simultaneously recorded macaque V1 neurons. The model reconstructed stimulus orientation with high precision and revealed, through its self-attention maps, a sparse and stimulus-dependent readout structure. For each stimulus orientation, reconstruction was dominated by a small number of highly weighted readout neurons that were tuned near the presented orientation and showed enhanced effective orientation signals after self-attention modulation. After removal of these neurons and retraining, reconstruction recovered through recruitment of substitute neurons with similar response properties, indicating that sparse readout can be flexibly supported by redundant population encoding. Decoder comparisons showed that a simple linear decoder and a multilayer perceptron recovered orientation less precisely than the transformer, suggesting that the model's advantage was not explained simply by generic nonlinear decoding capacity. Together, these findings suggest a population-level principle in which redundant orientation encoding supports sparse, stimulus-dependent, and flexible readout.