Yuqi You¹, Shurui Li², Yuanning Li², Hsin-Hung Li³, Ru-Yuan Zhang¹; ¹Shanghai Jiao Tong University, ²ShanghaiTech University, ³The Ohio State University

How does the brain utilize attention to enhance the efficiency of information processing? The Bayesian brain hypothesis posits that the brain combines noisy sensory information with prior knowledge to construct a posterior distribution over stimulus value and uses the posterior to guide actions. Although attentional enhancements on visual behaviour are well-established, it remains unclear how attention improves probabilistic neural representation of visual stimuli. Here, we analyzed an fMRI dataset (3T; N = 8) where face stimuli appeared randomly at one of sixteen locations tiling the visual space (4 x 4 grid, 2° spacing) while human participants performed two tasks in different runs: attending to centrally presented digits (digit task) or attending to the face stimuli (face task). We investigated how stimulus position is represented probabilistically in trial-by-trial multivariate cortical responses and how attention modulates these representations. By building a probabilistic voxel-encoding model linking stimulus position and multivariate cortical responses, we decoded the posterior distribution over position from fMRI responses in single trials. We found that, compared to the digit task, the face task markedly reduced both spatial bias and uncertainty according to the decoded posterior in several visual regions, with the strongest effects in hV4. These findings provide direct evidence for attentional enhancement on probabilistic population codes. By parametrically manipulating generative and inference (decoding) models, we further showed that changes in both voxel tuning and covariance are sufficient and necessary to account for the observed reductions in bias and uncertainty. Finally, trial-by-trial fluctuations of decoded uncertainty and its attentional modulation in hV4 correlated with activity in distributed attention-orienting and self-monitoring networks (e.g., parietal cortex and anterior cingulate cortex). Together, our findings bridge attentional control and probabilistic inference, two canonical cortical computations in the brain, and provide new evidence for how top-down modulation optimizes stimulus coding.

Acknowledgements: Supported by the National Natural Science Foundation of China (32441102 to R.-Y. Z., 32371154 to Y.L.), Shanghai Municipal Education Commission (2024AIZD014) to R.-Y.Z. and Shanghai Rising-Star Program (24QA2705500) to Y.L..