Research map / Label-efficient and robust learning / Recognition and long-tail
Labels, Information, and Computation: Efficient Learning Using Sufficient Labels
Shiyu Duan, Spencer Chang, Jose C. Principe
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 |
|---|---|---|---|
| double sufficient pairs from 120k to 240k (ablation) | sufficient-label learning, hinge, 10 full labels · MNIST | test accuracy (%): 98.97 -> 99.12 | Improved |
| double sufficient pairs from 100k to 200k (ablation) | sufficient-label learning, hinge, 10 full labels · CIFAR-10 | test accuracy (%): 90.99 -> 93.56 | Improved |
| use online sufficient pair sampling instead of fixed 200k pairs (ablation) | sufficient-label learning, hinge · CIFAR-10 | test accuracy (%): 93.56 -> 94.24 | Improved |
| double sufficient pairs from 146,514 to 293,028 (ablation) | sufficient-label learning, hinge · SVHN | test accuracy (%): 95.77 -> 96.11 | Improved |
| double sufficient pairs from 400k to 800k (ablation) | sufficient-label learning, hinge, 100 full labels · CIFAR-100 | test accuracy (%): 70.45 -> 72.67 | Improved |
| replace 50k full labels with online sufficient labels plus 10 full labels | LeNet-5 hinge · MNIST | test accuracy (%): 99.22 -> 99.23 | No clear effect |
| replace 73,257 full labels with online sufficient labels plus 10 full labels | ResNet-18 hinge · SVHN | test accuracy (%): 96.49 -> 96.66 | Improved |
| replace 50k full labels with online sufficient labels plus 10 full labels | ResNet-18 hinge · CIFAR-10 | test accuracy (%): 94.09 -> 94.24 | Improved |
| replace 60k full labels with online sufficient labels plus 10 full labels | ResNet-18 cross-entropy · Fashion-MNIST | test accuracy (%): 95.10 -> 95.03 | Hurt |
| replace 50k full labels with online sufficient labels plus 100 full labels | ResNet-18 cross-entropy · CIFAR-100 | test accuracy (%): 74.18 -> 73.98 | Hurt |
| replace 50k full labels with online sufficient labels plus 100 full labels | ResNet-18 hinge · CIFAR-100 | test accuracy (%): 73.22 -> 73.79 | Improved |
| use fixed 120k sufficient pairs plus 10 full labels instead of 60k full labels | LeNet-5 cross-entropy · MNIST | test accuracy (%): 99.32 -> 98.98 | Hurt |
Closest papers by meaning
- Label Efficient Learning by Exploiting Multi-Class Output Codes (AAAI 2017 · similarity 0.89)
- Losses over Labels: Weakly Supervised Learning via Direct Loss Construction (AAAI 2023 · similarity 0.89)
- More Supervision, Less Computation: Statistical-Computational Tradeoffs in Weakly Supervised Learning (NeurIPS 2016 · similarity 0.88)
- A Unified Approach to Count-Based Weakly Supervised Learning (NeurIPS 2023 · similarity 0.88)
- Training Subset Selection for Weak Supervision (NeurIPS 2022 · similarity 0.88)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.88)
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.
