V-125 | Temporal correspondence of visual processing of rapid stimuli between the human brain and artificial neural network architectures

Theoretical and Computational Neuroscience
Author: Eric Lützow Holm | Email: elholm90@gmail.com

Eric Lützow Holm^1°2°, Giovanni Franco Gabriel Marraffini,Enzo Tagliazucchi^1°2°3°

^1° Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Godoy Cruz 2290, CABA 1425, Argentina.
^2° Instituto de Física Interdisciplinaria y Aplicada, Departamento de Física, Universidad de Buenos Aires, Pabellón 1, Ciudad Universitaria, CABA 1425, Argentina.
^3° Latin American Brain Health (BrainLat), Universidad Adolfo Ibáñez, Av. Diag. Las Torres 2640, Peñalolén 7941169, Santiago Región Metropolitana, Chile

The visual cortex processes information hierarchically, from low-level features to complex patterns that enable object categorization. Similarly, in the most successful artificial models for object recognition, images are processed through multiple layers of artificial neural networks trained to determine the corresponding class. The goal of this work is to compare these models in terms of their temporal correspondence with EEG data recorded during visual perception tasks to determine whether the similarity between both systems depends on generic aspects or if it is influenced by the specific computations of each model. Using public EEG data from tasks involving rapid visual stimuli and the architectures AlexNet, ResNet, MoCo, VGG19, and ViT, we found that the initial layers better correlate with activity evoked in early stages and low-level luminance features, while the later layers correlate better with late components and semantic information. Finally, we verified that it is possible to transition from one representation to another using simple linear transformations. These results suggest a universal parallelism between human processing and that of artificial systems for the recognition of rapid visual stimuli.