Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Learning by Predicting Noise
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 continuous noise targets instead of discrete canonical-basis targets (ablation) | NAT · ImageNet transfer classification | accuracy: 19% -> 33.5% | Helped |
| Use unit-normalized features with square loss instead of softmax (ablation) | AlexNet supervised classifier · ImageNet supervised classification | accuracy: 59.2 -> 58.4 | No clear effect |
| Use high-pass-filtered images instead of clean images (ablation) | AlexNet supervised classifier · ImageNet supervised classification | acc@1: 59.7 -> 58.5 | Hurt |
| Use Sobel-filtered images instead of clean images (ablation) | AlexNet supervised classifier · ImageNet supervised classification | acc@1: 59.7 -> 57.4 | Hurt |
| Use grayscale image gradients as NAT input (ablation) | NAT · ImageNet | transfer feature quality | Helped |
| Update assignments every 3 epochs rather than at a high frequency (ablation) | NAT · ImageNet transfer classification | accuracy | Helped |
| Train NAT for longer (ablation) | NAT · ImageNet transfer classification | accuracy | Helped |
| Use NAT features rather than BiGAN features | AlexNet transfer features · ImageNet classification | Acc@1: 32.2 -> 36.0 | Helped |
| Use NAT features rather than BiGAN features | AlexNet transfer features · Pascal VOC 2007 classification (fc6-8) | mean average precision: 52.3 -> 56.7 | Helped |
| Use NAT features rather than BiGAN features | AlexNet transfer features · Pascal VOC 2007 classification (all layers) | mean average precision: 60.1 -> 65.3 | Helped |
| Use NAT features rather than BiGAN features | AlexNet transfer features · Pascal VOC 2007 detection | mean average precision: 46.9 -> 49.4 | Helped |
| Use NAT features instead of SIFT+FV | ImageNet transfer classifier · ImageNet classification | Acc@1: 55.6 -> 36.0 | Hurt |
Closest papers by meaning
- Unsupervised Representation Learning by Predicting Image Rotations (ICLR 2018 · similarity 0.87)
- Unsupervised Deep Learning via Affinity Diffusion (AAAI 2020 · similarity 0.86)
- Augmenting Supervised Neural Networks with Unsupervised Objectives for Large-scale Image Classification (ICML 2016 · similarity 0.86)
- NC-TTT: A Noise Constrastive Approach for Test-Time Training (CVPR 2024 · similarity 0.85)
- Noise Matters: Optimizing Matching Noise for Diffusion Classifiers (NeurIPS 2025 · similarity 0.84)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · similarity 0.84)
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.