SINEX, a local posthoc SIamese Network Explainer

Date: January 25, 2024

Presenter: Andrea Fedele

Location: Officine Garibaldi - il Cantiere delle Idee, Via Vincenzo Gioberti, 39, 56124 Pisa PI, Italy

Abstract: SINEX is a local, data-agnostic explanation method tailored for Siamese Networks within the context of C-way one-shot learning. The explainer aims to unveil the decisive features that empower the model to learn and generalize, mirroring human cognition when predicting instances of previously unseen objects. Employing a perturbation-based approach, SINEX calculates the contribution of each feature through a segment-weighted-average evaluation. The results can be visually represented as heatmaps, offering an intuitive insight into the behavior of Siamese Networks.