Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Astrocyte-Inspired Modulation for Robust Explanations in Vision Transformers

Nicolas Echevarrieta-Catalan¹, Odelia Schwartz¹, Vanessa Aguiar-Pulido¹; ¹University of Miami

Presenter: Nicolas Echevarrieta-Catalan

A central goal at the intersection of AI and neuroscience is not only to build high-performing models, but also to understand the internal mechanisms by which those models transform inputs into outputs. Post-hoc explanation methods are often used for that purpose in deep vision models, yet the resulting saliency maps can be highly unstable under image corruption, limiting their usefulness as explanations. Motivated by evidence that astrocytes integrate neural activity over time and modulate local circuit excitability, here we test whether an astrocyte-inspired modulation can make transformer explanations more robust. Explanations from vision transformers with and without astrocytic modulation are evaluated on Gaussian-noise-corrupted images using Normalized Scanpath Saliency (NSS) to measure alignment with human ground truth. This framework tests a specific astrocyte-inspired hypothesis: that slow, local, activity-dependent modulation can stabilize model explanations under noisy inputs. The project therefore uses explainability robustness as a quantitative testbed for neuroscience-inspired architectural ideas.

Topic Area: Methods, Tools, Theory & Neural Coding