Xin Wang¹, Shu-Chen Guan¹, Shi-Ming Tang¹, Cong Yu²; ¹Peking University, ²Zhejiang University

The scarcity of V1 neurons selective for low spatial frequencies (<1 cpd) raises the question whether low spatial frequency information is encoded in population activity. We analyzed V1 population data collected using two-photon calcium imaging in multiple FOVs (each containing >1000 neurons) while presenting a parafoveal Gabor in twelve orientations (0°–180°), six SFs (0.25–8 cpd). PCA analysis of population responses revealed that, like high spatial frequencies, low spatial frequencies (0.25, 0.5, & 1 cpd) formed distinct manifolds in representational space, demonstrating robust encoding of low-frequency information in the geometry of V1 population responses. Moreover, as each frequency-specific manifold originate from trials at different orientations, we used manifold volume to estimate orientation discriminability at each frequency. Volumes were minimal at very low frequencies and increased until peaking at 4 cpd before declining at 8 cpd, indicating that orientation discriminability was poor at low frequencies and became greatest at intermediate frequencies. To compare low- and high-spatial-frequency decoding, we trained a transformer-based model to reconstruct the stimuli, achieving >90% accuracy. An ablation experiment examined the dynamics of stimulus reconstruction. Neurons were incrementally added according to their self-attention scores while others were nullified to track how stimulus features were recovered. For higher spatial frequencies (1–4 cpd), reconstruction first recovered orientation before frequency. However, for lower spatial frequencies (0.25–0.5 cpd), reconstruction either recovered frequency first or frequency and orientation simultaneously. These results indicate that spatial frequency and orientation decoding follows a coarse-to-fine principle, recovering coarser orientation first at higher spatial frequencies. However, as lower spatial frequency is a coarser feature, frequency may be recovered either first or simultaneously with orientation. We conclude V1 encodes low spatial frequencies through population-level geometry and supports a coarse-to-fine decoding process reflected in the dynamics of spatial frequency and orientation reconstruction.

Acknowledgements: STI2030-Major Projects grant (2022ZD0204600)