Research map / Label-efficient and robust learning / Recognition and long-tail
Positive-Unlabeled Compression on the Cloud
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 |
|---|---|---|---|
| Select training examples from unlabeled data with PU learning instead of manually choosing positives | CIFAR-10 student distillation · CIFAR-10 | accuracy (%): 93.44 -> 93.75 | Helped |
| Use 50,000 randomly selected data, including negative examples, rather than PU-selected positives | CIFAR-10 student distillation · CIFAR-10 | accuracy (%): 87.02 | Hurt |
| Limit training to 50,000 PU-positive examples instead of using all selected positives (ablation) | PU-s1 CIFAR-10 · CIFAR-10 | accuracy (%): 93.75 -> 91.56 | Hurt |
| Use PU-selected Flickr1M data rather than 500k original ImageNet samples | ImageNet student distillation · ImageNet | top-5 accuracy (%): 85.88 -> 85.33 | No clear effect |
| Use PU-selected examples instead of 500k original ImageNet samples | ImageNet student distillation · ImageNet | top-1 accuracy (%): 63.90 -> 61.21 | Hurt |
| Use all PU-selected Flickr1M positives rather than 500k selected samples (ablation) | PU ImageNet · ImageNet | top-5 accuracy (%): 85.33 -> 86.00 | Helped |
| Use one to 20 labeled examples per category with PU augmentation | MNIST student distillation · MNIST | classification accuracy (%): 90.3 -> 98.5 | Helped |
| Use 50,000 randomly selected samples instead of all PU-selected training examples (ablation) | PU-s1 MNIST · MNIST | classification accuracy (%): 98.5 -> 98.3 | Hurt |
| Use the attention-based multi-scale extractor and RKD rather than the compared ResNet-34 and traditional KD configuration (ablation) | PU classifier and student distillation · CIFAR-10, varying class prior | classification accuracy | Helped |
| Vary the assumed class prior away from the true prior (ablation) | PU learning with attention-based multi-scale extractor and RKD · CIFAR-10, varying class prior | classification accuracy | Mixed |
Closest papers by meaning
- Positive-Unlabeled Data Purification in the Wild for Object Detection (CVPR 2021 · similarity 0.84)
- Few Sample Knowledge Distillation for Efficient Network Compression (CVPR 2020 · similarity 0.83)
- Deep k-Means: Re-Training and Parameter Sharing with Harder Cluster Assignments for Compressing Deep Convolutions (ICML 2018 · similarity 0.83)
- Progressive Network Grafting for Few-Shot Knowledge Distillation (AAAI 2021 · similarity 0.83)
- Convolutional Neural Network Compression through Generalized Kronecker Product Decomposition (AAAI 2022 · similarity 0.83)
- EPSD: Early Pruning with Self-Distillation for Efficient Model Compression (AAAI 2024 · 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.