Research map / Label-efficient and robust learning / Self-supervised and contrastive
Vector Quantization With Self-Attention for Quality-Independent Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 VQ codebook with concat fusion (vs. fine-tuning without codebook) (ablation) | ResNet50 on ImageNet-C · ImageNet-C validation | mCE: 53.7 -> 45.7 | Helped |
| concatenate quantized feature with original instead of replacing it (ablation) | VQ codebook features · ImageNet-C validation | mCE: 50.1 (replace) -> 45.7 (concat) | Helped |
| add self-attention module after concat fusion (ablation) | VQ codebook with concat fusion · ImageNet-C validation | mCE: 45.7 -> 43.1 | Helped |
| add self-attention module after concat fusion (ablation) | VQ codebook with concat fusion · ImageNet-C validation | clean top-1: 76.2 -> 76.6 | Helped |
| VQ+SA method vs. QualNet on ResNet50 | ResNet50 on ImageNet-C · ImageNet-C validation | mCE: 50.3 -> 43.1 | Helped |
| VQ+SA method on ResNeXt101 vs. QualNet | ResNeXt101 on ImageNet-C · ImageNet-C validation | mCE: 42.6 -> 37.9 | Helped |
| VQ+SA on unseen corruptions | ResNet50 · ImageNet-C unseen corruptions (speckle noise) | top-1 accuracy: 63.1 -> 68.4 | Helped |
| VQ+SA added to baseline augmentation training | ResNet50 on ImageNet-R · ImageNet-R | top-1 accuracy: 36.2 -> 38.6 | Helped |
| VQ+SA added to baseline augmentation training | ResNet50 on ImageNet-A · ImageNet-A | top-1 accuracy: 0.0 -> 3.7 | Helped |
| VQ+SA added to DeepAugment+AugMix on ImageNet-C | DeepAugment+AugMix ResNet50 · ImageNet-C validation | mCE: 50.8 (DAT) -> 48.7 (DAu+AM+Ours) | Helped |
| VQ+SA on ImageNet-A vs. DAT adversarial training | ResNet50 on ImageNet-A · ImageNet-A | top-1 accuracy: 3.7 vs DAT 6.8 | Hurt |
| select top-k similar codebook items instead of the single nearest item (ablation) | VQ codebook quantization · ImageNet-C | mCE | No clear effect |
Closest papers by meaning
- Implicit Feature Decoupling With Depthwise Quantization (CVPR 2022 · similarity 0.87)
- Measuring Self-Supervised Representation Quality for Downstream Classification Using Discriminative Features (AAAI 2024 · similarity 0.86)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.86)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.86)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.86)
- Channel-wise Vector Quantization (NeurIPS 2026 · 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.