Disentangling Polysemantic Channels in Convolutional Neural Networks

2025-04-17

AI Summary

Convolutional neural network (CNN) channels often exhibit polysemy—i.e., a single channel encodes multiple semantic concepts—severely undermining interpretability. To address this, we propose a weight reconstruction algorithm that achieves the first structural disentanglement of polysemous channels. Our method clusters activation patterns from preceding layers to identify heterogeneous response modes, decomposes each original channel into multiple semantically specialized sub-channels, and performs reparameterization via convolutional kernel remapping and feature response analysis. Crucially, the approach preserves the original network architecture and enables editable, mechanism-level explanations. Evaluated on an ImageNet subset, it increases per-channel semantic purity by 63% on average, substantially improving feature visualization quality, concept localization accuracy, and attribution reliability. This work establishes a novel paradigm for channel-level interpretability in CNNs.

Technology Category

• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Feature Construction/Reformulation
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Semantics and Knowledge: Methods, algorithms and applications for the development of semantic models, knowledge graphs and other forms of structured data models with machine-interpretable semantics
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Mechanistic interpretability is concerned with analyzing individual components in a (convolutional) neural network (CNN) and how they form larger circuits representing decision mechanisms. These investigations are challenging since CNNs frequently learn polysemantic channels that encode distinct concepts, making them hard to interpret. To address this, we propose an algorithm to disentangle a specific kind of polysemantic channel into multiple channels, each responding to a single concept. Our approach restructures weights in a CNN, utilizing that different concepts within the same channel exhibit distinct activation patterns in the previous layer. By disentangling these polysemantic features, we enhance the interpretability of CNNs, ultimately improving explanatory techniques such as feature visualizations.

Problem

• Disentangling polysemantic channels in CNNs for clarity
• Separating mixed concepts in neural network channels
• Enhancing CNN interpretability via feature disentanglement

Innovation

• Algorithm disentangles polysemantic CNN channels
• Restructures weights using distinct activation patterns
• Enhances interpretability via single-concept channels