Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unsupervised Representation Learning by Balanced Self Attention Matching
Daniel Shalam, Simon Korman
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 attention balancing (Sinkhorn balanced target matrix) (ablation) | BAM · ImageNet linear probing (ViT-S/16, 150 epochs) | accuracy: collapse -> 73.7 | Improved |
| use global normalization over the full nk x nk matrix (ablation) | BAM · ImageNet linear probing (ViT-S/16, 150 epochs) | accuracy: 60.7 -> 73.7 | Improved |
| add multi-crop (2 global + 6 local views) (ablation) | BAM · ImageNet linear probing (ViT-S/16, 150 epochs) | accuracy: 68.1 -> 73.7 | Improved |
| mask positive (augmentation) pairs in the similarity matrix (ablation) | BAM · ImageNet linear probing (ViT-S/16, 150 epochs) | accuracy: 72.6 -> 73.7 | Improved |
| add EMA student-teacher (ablation) | BAM · ImageNet linear probing (ViT-S/16, 150 epochs) | accuracy: 72.3 -> 73.7 | Improved |
| add ContrastiveCrop adapted to ViT (ablation) | BAM · ImageNet linear probing (ViT-S/16, 150 epochs) | accuracy: 73.5 -> 73.7 | Improved |
| use BAM instead of DINO for ImageNet-1% semi-supervised fine-tuning (ViT-S) | ViT-S/16 · ImageNet 1% labels | top-1 accuracy: 58.7 -> 60.0 | Improved |
| BAM vs DINO under linear probing, ViT-S/16 at 300 epochs | ViT-S/16 · ImageNet linear probing | top-1 accuracy: 75.9 (DINO) vs 75.0 (BAM) | Hurt |
| use CAFormer-M36 backbone instead of ViT-B/16 | BAM linear probing · ImageNet linear probing | accuracy: 77.5% -> 78.9% | Improved |
| use ResNet-50 instead of ViT-S/16 backbone | BAM linear probing · ImageNet linear probing | accuracy: 74.8% vs 75.0% | No clear effect |
| increase video frames T from 7 to 28 in DAVIS-2017 propagation | BAM vs DINO · DAVIS-2017 (J&F)_m | (J&F)_m: 62.4 -> 64.0 (BAM); 62.3 -> 63.1 (DINO) | Improved |
Closest papers by meaning
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.89)
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.88)
- Train a One-Million-Way Instance Classifier for Unsupervised Visual Representation Learning (AAAI 2021 · similarity 0.88)
- Efficient Unsupervised Visual Representation Learning with Explicit Cluster Balancing (ECCV 2024 · similarity 0.87)
- Unsupervised Deep Learning via Affinity Diffusion (AAAI 2020 · similarity 0.87)
- Unsupervised Feature Learning via Non-Parametric Instance Discrimination (CVPR 2018 · similarity 0.87)
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.
