Superposition disentanglement of neural representations reveals hidden alignment

2025-10-03

AI Summary

Superposition—where multiple features are encoded via heterogeneous linear combinations in neural representations—distorts cross-model and cross-species representational alignment metrics, causing conventional methods (e.g., linear regression, soft alignment) to underestimate true alignment. Method: The authors identify superposition-induced structural mismatch as a key mechanism behind degraded alignment scores and propose the first application of sparse autoencoders (SAEs) to disentangle superposed representations, thereby recovering latent alignment obscured by overlapping feature encodings. Contribution/Results: Evaluated on synthetic models, vision DNNs, and neural recordings, SAE-based disentanglement consistently yields significantly higher alignment scores, revealing more accurate cross-system representational correspondences. This work establishes a novel theoretical framework for analyzing representational alignment under superposition and provides a scalable, principled methodology—enhancing reliability in model-to-brain and inter-model alignment studies.

Technology Category

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

Application Category

• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

The superposition hypothesis states that a single neuron within a population may participate in the representation of multiple features in order for the population to represent more features than the number of neurons. In neuroscience and AI, representational alignment metrics measure the extent to which different deep neural networks (DNNs) or brains represent similar information. In this work, we explore a critical question: extit{does superposition interact with alignment metrics in any undesirable way?} We hypothesize that models which represent the same features in extit{different superposition arrangements}, i.e., their neurons have different linear combinations of the features, will interfere with predictive mapping metrics (semi-matching, soft-matching, linear regression), producing lower alignment than expected. We first develop a theory for how the strict permutation metrics are dependent on superposition arrangements. This is tested by training sparse autoencoders (SAEs) to disentangle superposition in toy models, where alignment scores are shown to typically increase when a model's base neurons are replaced with its sparse overcomplete latent codes. We find similar increases for DNN( ightarrow)DNN and DNN( ightarrow)brain linear regression alignment in the visual domain. Our results suggest that superposition disentanglement is necessary for mapping metrics to uncover the true representational alignment between neural codes.

Problem

• Investigating superposition's interference with neural alignment metrics
• Testing if different superposition arrangements reduce measured alignment
• Developing disentanglement methods to reveal true representational similarity

Innovation

• Using sparse autoencoders to disentangle superposition
• Replacing base neurons with sparse latent codes
• Increasing alignment scores via superposition disentanglement