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

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Learning to cluster neuronal function

Nina S. Nellen¹, Polina Turishcheva¹, Alexander S. Ecker¹; ¹Georg-August Universität Göttingen

Presenter: Nina S. Nellen

Deep predictive models have recently shown great potential to create digital twins to predict neuronal activity in the visual cortex. These models provide per-neuron embeddings, which have been proposed as a basis to identify functional cell types. However, so far no clear clusters have been observed in the mouse visual cortex and the structure of the embedding space is not highly reproducible across independent model fits. To address these problems, we build upon state-of-the-art predictive networks and introduce an explicit inductive bias to enhance cluster separability. If functional cell types exist, such a clustering bias should improve model performance and consistency of clustering. Our approach is based on training a predictive model and adding an auxiliary loss function that encourages the per-neuron embeddings to be distributed according to a $t$ mixture model. We jointly optimize both neuronal feature embeddings and clustering parameters. Our approach improves consistency of clusters and therefore leads to more consistent embedding spaces across models.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF