Research map / Label-efficient and robust learning / Recognition and long-tail
Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
Xinting Huang, Michael Hahn
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| use NDM subspaces instead of an identity partition | GPT-2 Small subspace partition · GPT-2 Small test suite (average) | Gini coefficient: 0.21 -> 0.71 | Improved |
| use NDM subspaces instead of a random orthogonal partition | GPT-2 Small subspace partition · GPT-2 Small test suite (average) | Gini coefficient: 0.21 -> 0.71 | Improved |
| use NDM instead of PCA-based partition (PCA 1) | GPT-2 Small subspace partition · GPT-2 Small test suite (average) | Gini coefficient: 0.38 -> 0.71 | Improved |
| use NDM instead of clustering SAE features | GPT-2 Small subspace partition · GPT-2 Small test suite (average) | Gini coefficient: 0.56 -> 0.71 | Improved |
| use NDM instead of SAE feature clustering on test 4 | GPT-2 Small subspace partition · GPT-2 Small test suite, test 4 | Gini coefficient: 0.74 -> 0.50 | Hurt |
| use NDM instead of SAE feature clustering on test 3 | GPT-2 Small subspace partition · GPT-2 Small test suite, test 3 | Gini coefficient: 0.73 -> 0.63 | Hurt |
| use NDM instead of PCA 2 on test 1 | GPT-2 Small subspace partition · GPT-2 Small test suite, test 1 | Gini coefficient: 0.66 -> 0.89 | Improved |
| increase group sparsity of toy features from 0.25 to 0.75 | Toy superposition model · Toy superposition model | orthogonality between groups | Hurt |
| lower hidden dimension from 12 to 10 in toy model with four groups | Toy superposition model · Toy superposition model (group=(5,15,5,15)) | FVU: 0.066 -> 0.166 | Hurt |
| raise hidden dimension from 12 to 16 in toy model with four groups | Toy superposition model · Toy superposition model (group=(5,15,5,15)) | FVU: 0.066 -> 0.016 | Improved |
| use Euclidean distance instead of cosine similarity in NDM objective (ablation) | NDM · Toy model | partition quality | Improved |
| search nearest neighbors among a small number of activations (N=4) in toy setting (ablation) | NDM · Toy model | partition quality | Improved |
Closest papers by meaning
- Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching (ICLR 2024 · similarity 0.86)
- Neural networks are more modular than single neurons suggest (NeurIPS 2026 · similarity 0.85)
- URLOST: Unsupervised Representation Learning without Stationarity or Topology (ICLR 2025 · similarity 0.85)
- Latent Functional Maps: a spectral framework for representation alignment (NeurIPS 2024 · similarity 0.84)
- From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · similarity 0.84)
- Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision (AAAI 2024 · similarity 0.84)
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.
