The Deleuzian Representation Hypothesis

TL;DR

The paper tackles interpretability of neural representations by replacing sparse autoencoders with a Deleuzian, difference-based concept extractor. It samples activation differences, clusters them with inverse-skewness weighted KMeans, and treats the resulting centroids as interpretable concept directions, framed as unsupervised discriminant analysis. Across five models, five datasets, and three modalities, it demonstrates superior concept quality and consistency relative to SAE variants and competitive with supervised LDA, and shows lossless, causal steering of inner representations in both vision and language models. The approach is simple, scalable, and yields actionable insights for mechanistic interpretability and controlled manipulation of model behavior.

Abstract

We propose an alternative to sparse autoencoders (SAEs) as a simple and effective unsupervised method for extracting interpretable concepts from neural networks. The core idea is to cluster differences in activations, which we formally justify within a discriminant analysis framework. To enhance the diversity of extracted concepts, we refine the approach by weighting the clustering using the skewness of activations. The method aligns with Deleuze's modern view of concepts as differences. We evaluate the approach across five models and three modalities (vision, language, and audio), measuring concept quality, diversity, and consistency. Our results show that the proposed method achieves concept quality surpassing prior unsupervised SAE variants while approaching supervised baselines, and that the extracted concepts enable steering of a model's inner representations, demonstrating their causal influence on downstream behavior.

Figures (8)

• Figure 1: Our method extracts diverse concepts from image, text and audio models.
• Figure 2: Overview of our concept extraction approach. We sample pairwise differences in activation between samples. Then, we use the inverse-skewness of those differences to selected the final concepts, corresponding to vectors in the activation space.
• Figure 3: Steering a painting style in CLIP activations: target is represented by its nearest images. Romanticism is set to zero, while Abstract is steered positively by the same magnitude.
• Figure 4: Steering the concept of countries in a BART model for three sentences (in gray), using $\alpha=+5$ and $\alpha=-5$ in each case
• Figure 5: Performance of our Deleuzian approach using less than 6144 concepts, on CLIP, WikiArt artist task.