Wing K. H. Chu^1,2 (), Hsin-Hung Li³, Thomas C. Sprague⁴, Kartik K. Sreenivasan², Clayton E. Curtis¹; ¹New York University, ²New York University Abu Dhabi, ³The Ohio State University, ⁴UC Santa Barbara

Despite ample evidence that working memory content can be decoded from human primary visual cortex (V1), single-neuron activity in macaque V1 rarely persists above pretrial baseline. Persistent activity is the most established neural mechanism supporting working memory; thus, its absence in V1 has led many to conclude that V1 is not involved in working memory storage. Here, we address this contradiction by carefully estimating delay period activity in human V1 from three fMRI studies that successfully decoded spatial working memory content from V1. Leveraging population receptive field (pRF) mapping, we estimated BOLD responses in V1 voxels whose pRFs contained the memory target location (in-RF) and compared these to voxels that did not (out-RF). Across all three datasets, the amplitude of BOLD activity in the in-RF voxels was greater than in the out-RF voxels throughout the memory retention interval, demonstrating that working memory content modulated V1 activity. The BOLD time courses were revealing: following a transient increase in BOLD activity time-locked to the visual target onset, the in-RF response returned to baseline while the out-RF response dipped below baseline. Indeed, only V1 voxels whose pRF overlapped the target location returned to baseline, whereas voxels tuned elsewhere exhibited widespread suppression. Therefore, human V1 activity during working memory delays did not persist above pretrial baseline in the classical sense; instead, spatially-defined differences across subpopulations were sustained. Furthermore, the in- vs. out-RF difference in neural responses predicted decoding performance, suggesting that decoding succeeds precisely because in-RF responses remain distinguishable from suppressed responses elsewhere. Finally, we used a hierarchical neural circuit with recurrence, normalization and interareal feedback to test which computations are necessary to reproduce V1’s empirical dynamics. Together these results reveal how selective V1 activity supports robust decoding of working memory features without persistent firing, explaining V1’s seemingly paradoxical role in working memory.

Acknowledgements: NIH R01 EY-016407 and R01 EY-033925 (CEC), NYUAD Research Institute Grant CG012 (KKS), Global PhD Fellowship (WKHC)