Zirui Chen¹ (), Leyla Isik¹, Michael F. Bonner¹; ¹Johns Hopkins University

Human vision excels at processing rapidly changing inputs while maintaining coherent representations of the preceding content. Previous studies have shown that stimulus features can be decoded from the brain for hundreds of milliseconds after presentation even as new visual inputs arrive, but decoders fit at a certain time point do not generalize across the entire decodable range. This suggests that there are dynamic shifts in the representational structure over time. However, the nature of these transformations remains poorly understood. To address this gap, we analyzed the THINGS EEG2 dataset, a large-scale collection of human neural responses to rapidly presented visual objects, and we investigated how representations of behaviorally relevant object features unfold over time. Specifically, we decoded these features via regression at each time point and tracked the temporal evolution of decoder weight vectors, allowing us to characterize how latent axes that represent stimulus features transform over time. Our analyses revealed two key findings. First, the transformations in feature representations are not merely the result of signal suppression by new visual inputs. Instead, these transformations reflect rotations of latent axes that are consistent across trials, enabling the same feature to be encoded along orthogonal axes at different time points. Second, we found that there are common representational transformations over time that are shared by multiple clusters of distinct features. These findings suggest that the human brain employs structured and reliable modes of temporal transformation to encode information from a rapid succession of visual inputs. This work offers promising directions for understanding representational dynamics in biological and artificial systems through time.