Neural Representations of face recognition in biological and artificial systems: Insights from MEG and CNNs

Poster Session F - Tuesday, April 16, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Hamza Abdelhedi^1,2,4 (hamza.abdelhedi@umontreal.ca), Shahab Bakhtiari^1,3, Karim Jerbi^1,2,3,4; ¹Université de Montréal, ²Computer Science Department, Université de Montréal, ³Psychology Department, Université de Montréal, ⁴Mila - Quebec AI Institute

Artificial neural networks, inspired by brain structure and function, have surpassed human performance in various tasks, but the link between Artificial Intelligence and neuroscience is still underexplored. Combining these fields has offered mutual reinforcement, especially in the field of Neuro-AI, where comparing artificial and biological systems in cognitive tasks, such as visual categorization, has yielded promising insights. Face-recognition, however, is less explored in this context. Do Convolutional Neural Networks (CNNs) trained for face-recognition mimic neural dynamics of face-recognition in brain circuits? A question addressed only by a handful of studies, which in non-human primate mainly focus on the IT-cortex, and in humans, largely rely on fMRI/behavioral-data. Here we compare human brain activity collected using Magnetoencephalography (MEG) during a face-recognition task to activations across seven CNNs. Compared to previous work, we leverage the high temporal resolution of MEG and source reconstruction techniques to compare these models to the brain across time, frequency, and space. Out of the tested models, FaceNet emerged as the most brain-like model during face-recognition. Crucially, training on face-recognition, rather than on object-recognition or both simultaneously, was necessary and sufficient for high model-brain similarity. In terms of temporal alignment, peak similarities were observed around 170ms which corresponds to the M170-component linked with face perception. Examining the Fusiform-Face-Area (FFA), we observed that, compared to an untrained model, the similarity to FaceNet trained on face recognition significantly increased in certain FFA-regions. Our study provides novel insights into the spatio-temporal similarity patterns between artificial and biological neural responses associated with face-recognition.

Topic Area: PERCEPTION & ACTION: Vision