Research map / Architectures and efficiency / RNNs and normalization
Matryoshka Representation Learning
Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, Ali Farhadi
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 |
|---|---|---|---|
| train nested multi-dimension losses (MRL) instead of independent fixed-feature models | ResNet50 · ImageNet-1K linear probe | top-1 accuracy | No clear effect |
| train with MRL instead of fixed-feature models | ResNet50 · ImageNet-1K 1-NN | accuracy: up to +2% | Improved |
| train with MRL instead of fixed-feature models | ResNet50 · ImageNet-1K retrieval | mAP@10: up to +3% | Improved |
| use weight-tied MRL-E instead of separate classifiers (ablation) | ResNet50 MRL · ImageNet-1K linear probe | top-1 accuracy: within 1% from 16-dim | Hurt |
| use MRL instead of SVD, random features or slimmable nets | ResNet50 · ImageNet-1K | accuracy at small dims | Improved |
| add adaptive classification cascade over nested dimensions | MRL ResNet50 · ImageNet-1K | representation size at equal accuracy: ~37 vs 512 dims for 76.3% | Improved |
| use adaptive retrieval (shortlist at 16-dim, rerank at 2048) instead of single-shot | MRL ResNet50 · ImageNet-1K | retrieval cost: 128x theoretical, 14x wall-clock | Improved |
| train with MRL instead of fixed features | ResNet50 · ImageNet-A | accuracy: +0.6% | Improved |
| train with MRL instead of fixed features | ResNet50 · FLUID long-tail | novel-class accuracy: up to +2% | Improved |
| train with MRL instead of fixed features | ResNet50 · few-shot nearest class mean | accuracy | No clear effect |
| logarithmic instead of uniform granularity spacing (ablation) | MRL · ImageNet-1K | accuracy | Improved |
Closest papers by meaning
- Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation (ICML 2025 · similarity 0.89)
- Matryoshka Query Transformer for Large Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- MMRL: Multi-Modal Representation Learning for Vision-Language Models (CVPR 2025 · similarity 0.85)
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders (ICML 2025 · similarity 0.85)
- Matryoshka Multimodal Models (ICLR 2025 · similarity 0.85)
- Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements (CVPR 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.
