Research map / Label-efficient and robust learning / Recognition and long-tail
Parametric Instance Classification for Unsupervised Visual Feature 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 |
|---|---|---|---|
| Use a cosine soft-max loss with stronger augmentation and a two-layer projection head (ablation) | PIC · ImageNet-1K linear evaluation, 200 epochs | Top-1 accuracy: 62.3 -> 66.2 | Helped |
| Add cosine soft-max and stronger augmentation while omitting the two-layer head (ablation) | PIC · ImageNet-1K linear evaluation, 200 epochs | Top-1 accuracy: 46.5 -> 60.5 | Helped |
| Add weight update correction for sampled negative classes (ablation) | PIC with 2^9 negatives · ImageNet-1K linear evaluation, 200 epochs | Top-1 accuracy: 56.8 -> 65.5 | Helped |
| Use 2^16 sampled negatives with weight correction instead of all negative classes (ablation) | PIC · ImageNet-1K linear evaluation, 200 epochs | Top-1 accuracy: 66.2 -> 66.2 | No clear effect |
| Use sliding-window scheduling rather than epoch-based scheduling (ablation) | PIC · ImageNet-1K, 50 epochs | Top-1 accuracy: 53.3 -> 60.4 | Helped |
| Use sliding-window scheduling rather than epoch-based scheduling (ablation) | PIC · ImageNet-1K, 1600 epochs | Top-1 accuracy: 70.4 -> 70.8 | Helped |
| Use sliding-window scheduling rather than epoch-based scheduling (ablation) | PIC · ImageNet-11K, 25 epochs | Top-1 accuracy: 22.38 -> 30.73 | Helped |
| Use sliding-window scheduling rather than epoch-based scheduling (ablation) | PIC · ImageNet-11K, 100 epochs | Top-1 accuracy: 32.02 -> 32.94 | Helped |
| Use one augmentation view per iteration and train for twice as many epochs (ablation) | PIC · ImageNet-1K, matched 200 augmentation views | Top-1 accuracy: 65.0 -> 67.3 | Helped |
| Use one augmentation view per iteration and train for twice as many epochs (ablation) | PIC · ImageNet-1K, matched 400 augmentation views | Top-1 accuracy: 67.6 -> 69.0 | Helped |
| Increase sliding-window size from 32768 to 131072 at stride 16384 (ablation) | PIC · ImageNet-1K linear evaluation, 200 epochs | Top-1 accuracy: 66.2 -> 67.3 | Helped |
| Increase sliding-window stride from 4096 to 16384 at window size 131072 (ablation) | PIC · ImageNet-1K linear evaluation, 200 epochs | Top-1 accuracy: 66.5 -> 67.3 | Helped |
Closest papers by meaning
- Train a One-Million-Way Instance Classifier for Unsupervised Visual Representation Learning (AAAI 2021 · similarity 0.91)
- Unsupervised Feature Learning via Non-Parametric Instance Discrimination (CVPR 2018 · similarity 0.89)
- Unsupervised Representation Learning by Predicting Image Rotations (ICLR 2018 · similarity 0.85)
- Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference (AAAI 2024 · similarity 0.84)
- Learning from Memory: Non-Parametric Memory Augmented Self-Supervised Learning of Visual Features (ICML 2024 · similarity 0.83)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.82)
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.