Research map / Architectures and efficiency / Pruning and efficient tuning
Revisiting a kNN-based Image Classification System with High-capacity Storage
Kengo Nakata, Youyang Ng, Daisuke Miyashita, Asuka Maki, Yu-Chieh Lin, Jun Deguchi
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 |
|---|---|---|---|
| store training features and labels and use kNN instead of zero-shot classifier | CLIP ViT-L/14 · ImageNet-1k | top-1 accuracy: 75.3 -> 79.7 | Improved |
| store training features and labels and use kNN instead of zero-shot classifier | CLIP ResNet-50 · CIFAR-100 | top-1 accuracy: 41.6 -> 55.7 | Improved |
| store training features and labels and use kNN instead of zero-shot classifier | CLIP ViT-L/14 · STL-10 | accuracy: 99.3 -> 99.6 | No clear effect |
| use CLIP-pretrained encoder instead of ImageNet-supervised encoder (ablation) | kNN system · CIFAR-100 | accuracy: 64.7 -> 74.3 | Improved |
| use CLIP-pretrained encoder instead of MAE encoder (ablation) | kNN system · ImageNet-1k | top-1 accuracy: 26.7 -> 74.0 | Improved |
| use ImageNet-supervised encoder instead of CLIP on its own training set (ablation) | kNN system · ImageNet-1k | top-1 accuracy: 74.0 -> 74.9 | Mixed |
| increase number of stored pairs (12.8k to 1.28M) (ablation) | kNN system · ImageNet-1k | top-1 accuracy | Improved |
| use larger encoder model (ablation) | kNN system · ImageNet-1k | top-1 accuracy: 65.0 -> 79.7 | Improved |
| add horizontal-flip augmentation of support set (ablation) | kNN system · CIFAR-100 and ImageNet-1k | top-1 accuracy: +1.1 for CIFAR-100 ResNet-50 and +0.5 for ImageNet-1k ResNet-101 | Improved |
| add horizontal-flip augmentation of support set (ablation) | kNN system · CIFAR-10 and STL-10 | accuracy: -0.1 to +0.8 | No clear effect |
| use kNN over stored features instead of fine-tuning methods in task-incremental learning | kNN system · Split CIFAR-100 (ResNet-50) | average accuracy: 80.9 -> 90.8 | Improved |
| use kNN over stored features instead of fine-tuning methods in class-incremental learning | kNN system · Split CIFAR-100 (ResNet-50) | average accuracy: 56.6 -> 67.9 | Improved |
Closest papers by meaning
- Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (TPAMI 2022 · similarity 0.86)
- Towards flexible perception with visual memory (ICML 2025 · similarity 0.86)
- Few-Shot Image Recognition With Knowledge Transfer (ICCV 2019 · similarity 0.86)
- Network of Experts for Large-Scale Image Categorization (ECCV 2016 · similarity 0.86)
- Deeply-Supervised Knowledge Synergy (CVPR 2019 · similarity 0.85)
- VL2Lite: Task-Specific Knowledge Distillation from Large Vision-Language Models to Lightweight Networks (CVPR 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.
