Shared Hierarchical Representations Explain Temporal Correspondence Between Brain Activity and Deep Neural Networks

The preprint operates within the broader paradigm of cognitive neuroscience, specifically focusing on visual object recognition. Within this field, the dominant paradigm involves hierarchical processing of visual information in both biological and artificial systems.

This preprint investigates the temporal correspondence between brain activity (measured via EEG) and activations in deep neural networks during visual object recognition. The study aligns strongly with "normal science" as defined by Kuhn. The researchers are operating within an established paradigm — the hierarchical processing of visual information — and are seeking to refine and extend our understanding of it. They are not challenging the fundamental assumptions of the paradigm, but rather are testing specific hypotheses related to the temporal dynamics of this hierarchical processing and its consistency across different ANN architectures. Specifically, they found that early EEG responses align with low-level visual features and early layers of the artificial networks and later EEG signals match higher-level features. This isn't revolutionary; it's an incremental increase in knowledge within the current paradigmatic understanding. Moreover, the preprint explores the impact of stimulus duration, and tests these relations over a collection of established ANN architectures. There is no evidence to suggest that the findings will cause a crisis within the existing model, instead confirming its basic elements with additional data and analysis.