Chihye Han¹ (), Michael Bonner¹; ¹Johns Hopkins University

A central question in visual neuroscience is how neural population activity supports perceptual behavior. Traditional approaches have focused on interpretable dimensions that explain large portions of neural variance, such as those corresponding to object categories or semantic properties (Huth et al., 2012; Tarhan & Konkle, 2020; Hebart et al., 2020). Here, we tested whether perceptual structure is concentrated in these leading dimensions or extends throughout the full spectrum of neural population activity. Using fMRI responses to natural scenes from the midventral visual stream (Natural Scenes Dataset; Allen et al., 2022), we applied PCA to identify latent dimensions of neural population activity. For each dimension, we sampled images with extreme values to create target clusters and paired them with randomly sampled foil images. We confirmed that early dimensions (e.g., PCs 1-5) contained interpretable categories like faces, while later dimensions (e.g., PCs beyond rank 10) showed no obvious semantic interpretation. We then asked participants (N=80) to view target and foil image clusters sampled from different dimension ranges (1-10, 10-100, 100-1000) and judge which cluster was more similar to a reference image sampled from the respective ranges. We found that behavioral relevance was just as strong for dimensions 10-100 as for dimensions 1-10. To test whether this extended to individual dimensions and a more challenging task paradigm, we asked participants (N=74) to judge which of two image clusters sampled from individual dimensions was more coherent, without presenting a reference image. Dimensions showing significant behavioral relevance (p<0.05, FDR-corrected) were distributed throughout the entire range tested. These findings demonstrate that behaviorally relevant information extends well beyond interpretable, high-variance dimensions. The striking dissociation between variance and behavioral relevance across individual dimensions suggests that the neural code utilizes information distributed across the full spectrum of representational dimensions, not just a subset of leading components.