Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Generalization via Scalable Neighborhood Component Analysis
Zhirong Wu, Alexei A. Efros, Stella X. Yu
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 |
|---|---|---|---|
| train with NCA and augmented memory instead of parametric softmax, kNN k=1 | ResNet18 · ImageNet val | top-1 accuracy: 62.91 -> 67.39 | Improved |
| train with NCA instead of softmax, kNN k=30 | ResNet50 · ImageNet val | top-1 accuracy: 75.09 -> 76.67 | Improved |
| compare NCA kNN with softmax classifier itself | ResNet50 · ImageNet val | top-1 accuracy: 76.01 -> 76.67 | Improved |
| project baseline feature to 128-d with PCA | softmax baseline ResNet18 · ImageNet val k=1 | top-1 accuracy: 62.91 -> 60.43 | Hurt |
| reduce embedding size from 128 to 64 (ablation) | SNCA ResNet18 · ImageNet val k=1 | top-1 accuracy: 67.39 -> 65.32 | Hurt |
| reduce embedding size from 128 to 32 (ablation) | SNCA ResNet18 · ImageNet val k=30 | top-1 accuracy: 70.59 -> 68.01 | Hurt |
| increase embedding size from 128 to 256 (ablation) | SNCA ResNet18 · ImageNet val k=1 | top-1 accuracy: 67.39 -> 67.54 | No clear effect |
| lower temperature from 0.1 to 0.05 (ablation) | SNCA ResNet18 · ImageNet val k=1 | top-1 accuracy: 63.87 -> 67.39 | Improved |
| lower temperature from 0.05 to 0.03 (ablation) | SNCA ResNet18 · ImageNet val k=1 | top-1 accuracy: 67.39 -> 66.98 | No clear effect |
| use temperature 0.02 (ablation) | SNCA ResNet18 · ImageNet val | top-1 accuracy: N/A | Hurt |
| use NCA instead of softmax for coarse-to-fine transfer | ResNet · CIFAR100 from CIFAR20 training | top-1 accuracy: 54.17 -> 62.32 | Improved |
| use NCA instead of softmax for coarse-to-fine transfer | ResNet · ImageNet 1000 from ImageNet127 training | top-1 accuracy: 48.07 -> 52.75 | Improved |
Closest papers by meaning
- Memory-Based Neighbourhood Embedding for Visual Recognition (ICCV 2019 · similarity 0.89)
- Large Scale Visual Recognition through Adaptation using Joint Representation and Multiple Instance Learning (JMLR 2016 · similarity 0.88)
- Visual Recognition with Deep Nearest Centroids (ICLR 2023 · similarity 0.88)
- Unsupervised Deep Learning by Neighbourhood Discovery (ICML 2019 · similarity 0.88)
- Is Second-Order Information Helpful for Large-Scale Visual Recognition? (ICCV 2017 · 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.
