Abstract Mounting evidence indicates that information processing in the visual hierarchy progresses sequentially, from low-level perceptual to high-level conceptual features during visual perception, and in reverse order during memory retrieval. However, the nature of this processing hierarchy and its modulation by selective attention remain unclear. By combining a deep neural network and the drift-diffusion model, we identified parallel processing of perceptual and conceptual features during encoding. The slower reaction times for conceptual tasks compared to perceptual tasks were primarily due to differences in decision boundary and non-decision time, which could not be compensated by the faster evidence accumulation rates of conceptual features. Using single-trial multivariate decoding of MEG (Magnetoencephalography) data, we tracked the timing of feature representation during visual perceptual and mnemonic tasks. During perception, selective attention reversed the onset times of perceptual and conceptual features in the occipital and parietal lobes, leading to earlier detection of conceptual features. Stronger theta oscillation interactions between occipital and temporal regions during task preparation correlated with earlier onset times of target features in the occipital lobe. During retrieval, selective attention resulted in earlier peak times for perceptual features compared to conceptual features in the frontal lobe, leading to perceptual features being reconstructed with the highest fidelity before conceptual features. These findings provide novel insights into the nature of the hierarchical processing during perceptual and retrieval, highlighting the critical role of selective attention in modulating information accumulation speed.

Selective attention modulates feature accumulation speed during perception and memory retrieval