Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Multiencoder VAE for cross-subject alignment of brain responses

Angeliki Papathanasiou¹, Jascha Achterberg¹, Ian Cone¹, Thomas E. Nichols¹, Rui Ponte Costa¹; ¹University of Oxford

Presenter: Angeliki Papathanasiou

Neural responses to identical stimuli vary considerably across individuals despite similar behavioral outcomes. Recent research demonstrates preserved latent neural dynamics in motor cortical populations across monkeys performing identical motor tasks. Inspired by these observations we introduce a multiencoder variational autoencoder (VAE) to model visual cortex responses. Our approach transforms subject-specific fMRI responses from natural scene viewing into a common latent space while predicting artificial neural network (ANN) activations elicited by identical stimuli. Using the Natural Scenes Dataset (NSD), our method outperforms traditional alignment techniques by capturing cross-subject representational similarities. The VAE architecture implements subject-specific encoders which project occipitotemporal cortex responses into a shared latent manifold that preserves semantic organization while accommodating neuroanatomical variability. Simultaneously, the decoder establishes a computational correspondence between this latent representation and ResNet-50 activations. This approach creates a framework for investigating shared neural representations across individuals while quantifying systematic relationships between biological and artificial NNs.

Topic Area: Methods & Computational Tools

Extended Abstract: Full Text PDF