Visual perception in brains and models

From familiarity to emotion

Recognizing a familiar face and interpreting a changing expression both involve visual information unfolding over time. I use neural networks as computational comparisons: by changing what a model learns, we can ask which aspects of human neural responses it captures, where, and when.

During my MSc, supervised by Prof. Karim Jerbi and Prof. Shahab Bakhtiari, I compared source-localized MEG with seven convolutional-network architectures trained for face recognition, object recognition, or both. Time-resolved representational analyses connect the structure of model activations with the structure of human brain responses.

What the familiarity work suggests

Our 2026 bioRxiv preprint reports earlier brain–model alignment for familiar faces in lateral occipital cortex, and stronger alignment later in fusiform cortex. Face-recognition training was most consistently associated with the earlier timing effect; broader visual learning also captured the later fusiform representations. The comparison suggests that learning objectives matter differently across processing stages.

Continuing with emotion perception

I am also involved in collaborative work comparing CNNs with human responses during dynamic facial-expression perception. This asks how pretraining and model specialization shape alignment with EEG as emotional expressions evolve. The work is not yet published.

The common thread is to use differences between models as a way to investigate perception, while keeping predictive performance and similarity to human neural processing as distinct questions.