Sparse Concept Anchoring for Interpretable and Controllable Neural Representations

2025-12-13

AI Summary

This work addresses the interpretability and controllability challenges in neural representations by proposing a geometric concept anchoring method under extremely low supervision—requiring fewer than 0.1% labeled samples per anchored concept. The approach leverages activation normalization, concept separation regularization, anchor/subspace attraction regularization, and a structured autoencoder to precisely embed target concepts along specific directions or axis-aligned subspaces in the latent space, while allowing other concepts to self-organize orthogonally. It is the first to enable both reversible behavioral steering and permanent deletion of sparse concepts via intervention. Theoretically, it achieves reconstruction error approaching the information-theoretic lower bound. Experiments demonstrate selective attenuation or complete removal of target concepts while preserving orthogonal features intact, attaining reconstruction error at the theoretical optimum.

Technology Category

• Machine Learning: Representation Learning
• Computer Vision: Representation Learning for Vision
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

We introduce Sparse Concept Anchoring, a method that biases latent space to position a targeted subset of concepts while allowing others to self-organize, using only minimal supervision (labels for <0.1% of examples per anchored concept). Training combines activation normalization, a separation regularizer, and anchor or subspace regularizers that attract rare labeled examples to predefined directions or axis-aligned subspaces. The anchored geometry enables two practical interventions: reversible behavioral steering that projects out a concept's latent component at inference, and permanent removal via targeted weight ablation of anchored dimensions. Experiments on structured autoencoders show selective attenuation of targeted concepts with negligible impact on orthogonal features, and complete elimination with reconstruction error approaching theoretical bounds. Sparse Concept Anchoring therefore provides a practical pathway to interpretable, steerable behavior in learned representations.

Problem

• Biases latent space to position targeted concepts with minimal supervision
• Enables reversible behavioral steering and permanent concept removal
• Achieves selective attenuation with negligible impact on orthogonal features

Innovation

• Sparse Concept Anchoring biases latent space with minimal supervision
• Combines activation normalization and regularizers for concept separation
• Enables reversible steering and permanent removal via weight ablation