Answers are grounded in this paper and cite their sources.
Natural image representations are distributed across many dimensions of visual cortex activity, but little is known about how the multidimensional structure of these representations evolves over time following stimulus onset. Herewe examined the temporal dynamics and latent dimensional structure of natural object representations in large-scale EEG and MEG data from male and female human participants. We also compared these datawith leading representational models derived from large-scale human similarity judgments and deep neural networks. Our findings reveal a rapid expansion of stimulus dimensionality in the brain, which peaks within 100 milliseconds and gradually decays over hundreds of milliseconds. The dynamics of these dimensionality changes tracked the decoding accuracy for both behavioral embeddings and neural network features, suggesting that dimensionality may be a general indicator of representational expressivity. Interestingly, the dimensionality of the neural representations could not be fully explained by a leading behavior-based model or deep neural network. Follow-up experiments showed that the remaining neural variance carried additional perceptually relevant information not yet explained by leading models. Together, these findings reveal previously unrecognized complexity in measurements of dynamic human brain responses to natural objects. Significance Statement When we look at an object, our brains rapidly construct a rich internal representation of what we see. Using large-scale, time-resolved brain recordings from people viewing thousands of natural images, we find that the brain rapidly expands its representational complexity within the first 100 milliseconds, sustaining this richness for hundreds of milliseconds thereafter. A leading behavioral model and a large-scale deep neural network do not fully account for this complexity — yet the unexplained brain activity is not noise: people can reliably use these residual neural patterns to judge image similarity. These findings challenge the assumption that such brain recordings are informationally limited and reveal structure in the human visual system that future models will need to explain.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: