Decoding Orientation Serial Dependence from V1/V2 Population Responses Using a Cross-Attention Transformer

Abstract

Perception of visual orientation is shaped by not only the current stimulus but also the recent history, showing repulsive biases or attractive serial dependence biases. Although repulsion is often assumed to arise in the early visual cortex, the neural origin of attraction remains debated: Does it reflect top–down modulation embedded in early cortical activity, or does it emerge only from downstream circuits that reverse repulsion? Here we combined two-photon calcium imaging in awake, fixating male macaques with a cross-attention transformer decoder to examine whether V1/V2 population activity contains signals that can support separate readouts of repulsive and attractive biases. We simultaneously recorded large samples of V1/V2 neurons responding to sequential oriented stimuli embedded in white noise. Despite average responses showing only repulsion, the cross-attention transformer, which was designed to model interactions between preceding and current stimulus responses, successfully decoded both repulsive and attractive biases from the same population activity. Attention-map analyses further revealed distinct readout patterns for the two biases, and a constrained shared-representation control with only final-stage output transformation performed substantially worse than the independently trained decoders. These findings indicate that V1/V2 population activity can support different readouts of history-dependent orientation bias.