Christina Sartzetaki

The paper aligns video models with EEG data to understand dynamic visual processing, finding that mid-level temporally-integrative features best match posterior brain responses.

The paper compares human behavior, fMRI data, and DNN activations to show that the human brain represents locomotive action affordances independently of other visual features, and that standard DNNs do not capture this representation well.

The paper benchmarks video models against human brain activity to understand alignment factors, finding that temporal modeling and task relevance are key for different brain regions.

BrainACTIV uses diffusion models to manipulate images and control neural activity in specific brain regions, identifying visuo-semantic properties that drive cortical selectivity.