Research map / Label-efficient and robust learning / Recognition and long-tail
The Deleuzian Representation Hypothesis
Clément Cornet, Romaric Besançon, Hervé Le Borgne
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| cluster activation differences instead of raw activations (ablation) | Deleuzian KMeans · WikiArt artist (CLIP) / CoNLL NER (DeBERTa) | probe loss: 0.0133 -> 0.0119 (CLIP); 0.1184 -> 0.0665 (DeBERTa) | Improved |
| replace KMeans with Tk-SAE on activation differences (ablation) | Deleuzian · WikiArt artist (CLIP) / CoNLL NER (DeBERTa) | probe loss: 0.0119 -> 0.0134 (CLIP); 0.0665 -> 0.1093 (DeBERTa) | Hurt |
| add inverse-skewness weighting to KMeans (ablation) | Deleuzian · WikiArt artist (CLIP) / CoNLL NER (DeBERTa) | probe loss: 0.0128 -> 0.0119 (CLIP); 0.0841 -> 0.0665 (DeBERTa) | Improved |
| add inverse-skewness weighting to KMeans (ablation) | Deleuzian · WikiArt artist (CLIP) / CoNLL NER (DeBERTa) | effective rank: 17.9 -> 124.4 (CLIP); 5.65 -> 182.0 (DeBERTa) | Improved |
| add inverse-skewness weighting to KMeans (ablation) | Deleuzian · WikiArt artist (CLIP) / CoNLL NER (DeBERTa) | max. pairwise cosine: 0.6504 -> 0.5677 (CLIP); 0.8357 -> 0.3908 (DeBERTa) | Improved |
| Deleuzian vs Tk-SAE effective rank on differences (ablation) | Deleuzian · WikiArt artist (CLIP) / CoNLL NER (DeBERTa) | effective rank: Tk-SAE on diff 340.5 / 109.2 vs Deleuzian 124.4 / 182.0 | Mixed |
| use Deleuzian concepts instead of best SAE (Tk-SAE) | unsupervised concept extraction · 20 probe-loss tasks | probe loss: lowest probe loss on 13 of 20 tasks; avg rank 1.65 vs 2.65 | Improved |
| use Deleuzian concepts instead of Van-SAE | concept extraction · MPPC consistency across seeds | MPPC: 0.821 vs 0.840 (CLIP ImageNet); 0.789 vs 0.603 (DinoV2 ImageNet) | Mixed |
| replace linear with quadratic Deleuzian extension (ablation) | Deleuzian · WikiArt (CLIP) | probe loss: 0.0119 -> 0.0124 (artist); 0.1230 -> 0.1305 (genre) | Hurt |
| use only 2000 concepts instead of 6144 (ablation) | Deleuzian · WikiArt artist (CLIP) | probe loss | No clear effect |
| Deleuzian vs supervised LDA on BART CoNLL-2003 | Deleuzian · CoNLL-2003 NER (BART) | probe loss: 0.6326 (LDA) -> 0.2148 | Improved |
Closest papers by meaning
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.91)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.89)
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders (ICML 2025 · similarity 0.89)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.89)
- Mechanistic Interpretability with Sparse Autoencoder Neural Operators (NeurIPS 2026 · similarity 0.88)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
