Research map / Label-efficient and robust learning / Knowledge distillation
QuEST: Quantized Embedding Space for Transferring Knowledge
Himalaya Jain, Spyros Gidaris, Nikos Komodakis, Patrick Pérez, Matthieu Cord
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 |
|---|---|---|---|
| distill on last feature level (Block L) instead of previous stage (Block L-1) (ablation) | QuEST · CIFAR-100 WRN-40-2 -> WRN-16-2 | top-1 accuracy: 74.16 -> 76.10 | Improved |
| distill on last feature level (Block L) instead of previous stage (Block L-1) (ablation) | QuEST · CIFAR-10 WRN-40-1 -> WRN-16-1 | top-1 accuracy: 91.64 -> 91.98 | Improved |
| predict visual word assignments instead of regressing raw features (ablation) | QuEST vs feature regression · CIFAR-100 ResNet50 -> VGG8 | top-1 accuracy: 71.9 -> 75.17 | Improved |
| predict visual word assignments instead of regressing raw features (ablation) | QuEST vs feature regression · CIFAR-100 WRN-40-2 -> WRN-16-2 | top-1 accuracy: 75.53 -> 76.10 | Improved |
| predict visual word assignments instead of regressing raw features, same-architecture teacher (ablation) | QuEST vs feature regression · CIFAR-100 WRN40-2 -> WRN40-2 | top-1 accuracy: 78.26 -> 77.96 | Hurt |
| predict visual word assignments instead of regressing raw features, same-architecture teacher (ablation) | QuEST vs feature regression · CIFAR-100 ResNet32x4 -> ResNet32x4 | top-1 accuracy: 80.57 -> 80.53 | No clear effect |
| use peaky soft assignment instead of hard assignment (tau=0) (ablation) | QuEST · CIFAR-100 | top-1 accuracy | Improved |
| use small temperature tau (peaky soft assignment) rather than large tau (ablation) | QuEST · CIFAR-100 | top-1 accuracy | Improved |
| increase vocabulary size K to at least 2048 (ablation) | QuEST · CIFAR-100 | top-1 accuracy | No clear effect |
| add the KD loss on top of QuEST | QuEST · CIFAR-100 / ImageNet | top-1 accuracy | No clear effect |
| QuEST instead of CRD+KD, ResNet18 student | QuEST · ImageNet val ResNet34 -> ResNet18 | top-1 error: 28.62 -> 28.33 | Improved |
| QuEST instead of OFD, MobileNet student | QuEST · ImageNet val ResNet50 -> MobileNet | top-1 error: 28.75 -> 27.46 | Improved |
Closest papers by meaning
- Compressing Visual-Linguistic Model via Knowledge Distillation (ICCV 2021 · similarity 0.89)
- Contrastive Representation Distillation (ICLR 2020 · similarity 0.89)
- Knowledge Distillation from Internal Representations (AAAI 2020 · similarity 0.89)
- Knowledge Distillation Meets Open-Set Semi-supervised Learning (IJCV 2024 · similarity 0.88)
- Knowledge Distillation for Large Language Models through Residual Learning (ICLR 2026 · similarity 0.88)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · 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.
