Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Sparse Visual Representations via Spatial-Semantic Factorization
Theodore Zhao, Sid Kiblawi, Jianwei Yang, Naoto Usuyama, Reuben Tan, Noel C Codella, Tristan Naumann, Hoifung Poon, Mu Wei
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 |
|---|---|---|---|
| add low-rank reconstruction loss (ablation) | STELLAR ViT-B · IN1K / ADE20K | accuracy / mIoU: 73.26 vs 72.44 (-0.82); ADE 31.33 vs 29.94 (-1.39) | Improved |
| add clustering and set alignment (ablation) | STELLAR ViT-B · IN1K | accuracy: 52.07 (-21.19) | Improved |
| add set alignment and CLS alignment (keep clustering) (ablation) | STELLAR ViT-B · IN1K | accuracy: 2.73 (-70.53) | Improved |
| add clustering loss (alignment present) (ablation) | STELLAR ViT-B · IN1K | accuracy: 42.14 (-31.12) | Improved |
| add CLS token alignment (ablation) | STELLAR ViT-B · IN1K | accuracy: 70.79 (-2.47) | Improved |
| add CLS token alignment (ablation) | STELLAR ViT-B · ADE20K | mIoU: 30.20 (-1.12) | No clear effect |
| add KoLeo regularization (ablation) | STELLAR ViT-B · IN1K | accuracy: 72.05 (-1.21) | Improved |
| use MAE prior instead of DINO prior (ablation) | STELLAR ViT-B · IN1K | accuracy: 66.32 -> 73.26 (+6.9) for MAE; 76.46 -> 73.31 (-3.2) for DINO | Mixed |
| initialize from DINO prior (ablation) | STELLAR ViT-B · ADE20K | mIoU: 26.87 -> 28.17 (+1.3) | Improved |
| train from random prior instead of MAE prior (ablation) | STELLAR ViT-B · IN1K | accuracy: 73.26 -> 65.28 | Hurt |
| increase number of sparse tokens r from 8 to 24 (ablation) | STELLAR · IN1K linear / FID | accuracy / FID | Mixed |
| use 16 sparse tokens instead of 196 dense tokens | STELLAR ViT-B · IN1K | accuracy: 72.21 -> 73.26 | Improved |
Closest papers by meaning
- Both Semantics and Reconstruction Matter: Making Representation Encoders Ready for Text-to-Image Generation and Editing (ICML 2026 · similarity 0.87)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.85)
- Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation (NeurIPS 2026 · similarity 0.85)
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · similarity 0.85)
- dRAE: Representation Autoencoder with Hyper-Spherical Codes (NeurIPS 2026 · similarity 0.85)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.85)
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.
