Decoding orientation serial dependence from V1/V2 population responses using a cross-attention transformer.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Visual Perception

Background:

• Visual perception is influenced by both current stimuli and recent history, leading to repulsive adaptation and attractive serial dependence biases.
• The neural basis of attractive biases is debated, with possibilities including top-down modulation in early visual cortex or downstream circuit reversal of repulsion.

Purpose of the Study:

• To investigate whether V1/V2 population activity contains signals supporting separate readouts of repulsive and attractive visual orientation biases.
• To determine the neural origin of attractive biases in visual perception.

Main Methods:

• Two-photon calcium imaging in awake, fixating male macaques.
• Utilizing a cross-attention transformer decoder to analyze V1/V2 neural population activity.
• Recording responses of V1/V2 neurons to sequential oriented stimuli in white noise.

Main Results:

• Despite average V1/V2 responses showing only repulsion, a cross-attention transformer successfully decoded both repulsive and attractive biases from the same population activity.
• Attention-map analyses revealed distinct readout patterns for repulsive and attractive biases.
• Constrained models with shared representations performed worse than independently trained decoders, suggesting separate readout mechanisms.

Conclusions:

• V1/V2 population activity can support distinct readouts for history-dependent orientation biases.
• These findings suggest that both repulsive and attractive biases may originate from or be represented within early visual cortical circuits.