Statistically efficient neural encoding of natural object variability shapes the temporal dynamics of visual processing.

Object perception unfolds dynamically over millisecond timescales, yet the organisational principles that shape the emerging neural responses are not fully understood. Traditional hypothesis-driven approaches risk constraining interpretations by focusing on pre-selected object features. To circumvent this limitation, we applied a data-driven framework to behavioural and neuroimaging data obtained from the THINGS initiative, which provides a systematic sampling of real-world objects. Behaviourally relevant stimulus dimensions were derived from prior large-scale similarity judgements, offering an unbiased, ecologically grounded representation of object space. Using Partial Least Squares Regression (PLSR), we generated neural encoding models to predict time-resolved evoked responses in EEG and MEG from these dimensions. Across both modalities, the PLSR identified a small set of latent components that reliably captured the temporal dynamics of the neural activity. These components were similar across the EEG and MEG datasets and with a prior MRI analysis. The components encoded a diverse range of object features, including visual and semantic properties, yet did not map straightforwardly onto canonical accounts of visual cortical organisation. Instead, our findings suggest that object representations in the brain are structured by principles of statistical efficiency, capturing the co-occurrence of features amongst natural variability in real-world objects to support dynamic visual processing.