Research map / Label-efficient and robust learning / Knowledge distillation
Bag of Instances Aggregation Boosts Self-supervised Distillation
Haohang Xu, Jiemin Fang, XIAOPENG ZHANG, Lingxi Xie, Xinggang Wang, Wenrui Dai, Hongkai Xiong, Qi Tian
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 |
|---|---|---|---|
| Use teacher-guided bags of similar instances with bag-aggregation distillation | ResNet-18 and ResNet-34 students · ImageNet linear evaluation | top-1 accuracy: 52.5 -> 65.5; 57.4 -> 68.9 | Improved |
| Use teacher-guided bags of similar instances | BINGO student representations · ImageNet KNN validation, 10-NN | accuracy: 55.3 -> 61.0; 58.2 -> 64.9 | Improved |
| Use BINGO distillation instead of MoCo-v2 pretraining | ResNet-18 · COCO detection, 1× schedule | APbb: 31.3 -> 32.0 | Improved |
| Use BINGO distillation instead of MoCo-v2 pretraining | ResNet-18 · COCO instance segmentation, 1× schedule | APmk: 28.8 -> 29.6 | Improved |
| Use BINGO distillation instead of DisCo | ResNet-18 student with ResNet-50 teacher · ImageNet semi-supervised, 1% labels | accuracy: 39.2 -> 42.8 | Improved |
| Use BINGO distillation instead of DisCo | ResNet-18 student with ResNet-50 teacher · ImageNet semi-supervised, 10% labels | accuracy: 50.1 -> 57.5 | Improved |
| Use BINGO instead of DisCo | ResNet-18 student · CIFAR-10 linear evaluation | accuracy: 85.3 -> 86.8 | Improved |
| Use BINGO instead of DisCo | ResNet-18 student · CIFAR-100 linear evaluation | accuracy: 63.3 -> 66.5 | Improved |
| Form bags using a supervised-pretrained model rather than a self-supervised pretrained model (ablation) | BINGO · ImageNet linear evaluation, ResNet-18 | top-1 accuracy: 64.0 -> 64.8 | Improved |
| Form bags using ground-truth labels rather than self-supervised teacher features (ablation) | BINGO · ImageNet linear evaluation, ResNet-18 | top-1 accuracy: 64.0 -> 65.8 | Improved |
| Use a randomly initialized model to form bags instead of a pretrained model (ablation) | BINGO · ImageNet linear evaluation, ResNet-18 | top-1 accuracy: 64.0 -> 46.6 | Hurt |
| Use teacher relation rather than online student relation to form bags (ablation) | BINGO without pretrained teacher parameters · ImageNet linear evaluation, ResNet-18 | top-1 accuracy: 57.2 -> 62.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DisCo — DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for Deep Neural Networks (CVPR 2021) · better on 6 of 6 shared results
- SEED — SEED: Self-supervised Distillation For Visual Representation (ICLR 2021) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SEED: Self-supervised Distillation For Visual Representation (ICLR 2021 · similarity 0.87)
- Robust Self-Supervised Multi-Instance Learning with Structure Awareness (AAAI 2023 · similarity 0.87)
- Towards Efficient and Effective Self-Supervised Learning of Visual Representations (ECCV 2022 · similarity 0.87)
- ISD: Self-Supervised Learning by Iterative Similarity Distillation (ICCV 2021 · similarity 0.86)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.86)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 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.
