New Tool Reveals Why AI Makes the Mistakes It Makes

Researchers have introduced SPICE (Simple Polysemantic Feature Interpretation via Clustering-based Explanation), a generalizable framework for analyzing polysemanticity in deep vision architectures, according to a paper accepted at the European Conference on Computer Vision (ECCV 2026). The core problem: a single neuron is activated by multiple, often unrelated concepts, hindering clear functional understanding. Existing approaches to this problem remain architecture-specific and depend on manual heuristics such as a fixed number of concept clusters (K), limiting their generality and scalability, especially for modern Transformer-based models. SPICE avoids architecture-dependent propagation rules, enabling the first systematic comparison of polysemanticity across both CNNs and Transformers, and automatically determines the number of concept clusters per neuron, eliminating reliance on a preset K and supporting scalable analysis for large models. Using SPICE, the authors conduct a comprehensive investigation into how polysemanticity emerges, varies across depth and architecture, and forms through distinct computational pathways. The paper is by Sehyun Lee, Dahee Kwon, Damin Lee, and Jaesik Choi.