Research map / Label-efficient and robust learning / Noisy labels
Collaborative Learning for Deep Neural Networks
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 |
|---|---|---|---|
| Train multiple classifier heads collaboratively instead of one classifier | ResNet-32, ResNet-110, and DenseNet-40-12 · CIFAR-10 | test error (%): 6.66 ± 0.21 -> 5.86 ± 0.13; 5.56 ± 0.16 -> 4.98 ± 0.12; 5.26 ± 0.08 -> 4.86 ± 0.12 | Helped |
| Increase the number of independent collaborative instances from 2 to 4 | CIFAR-10 classifiers · CIFAR-10 | test error (%): 6.19 ± 0.17 -> 6.16 ± 0.17; 5.21 ± 0.14 -> 5.16 ± 0.13; 5.11 ± 0.15 -> 5.00 ± 0.05 | Helped |
| Share intermediate-level representations across heads | Collaborative CIFAR-10 classifiers · CIFAR-10 | test error (%): 6.16 ± 0.17 -> 5.97 ± 0.07; 5.16 ± 0.13 -> 5.15 ± 0.14; 5.00 ± 0.05 -> 5.04 ± 0.10 | Helped |
| Use simultaneous rather than alternative optimization | Two ResNet-32 instances · CIFAR-100 | test error (%): 28.81 -> 27.48 ± 0.37 | Helped |
| Increase distillation temperature from T=1 to T=2 | Collaborative ResNet-32, head 1 · CIFAR-100 | test error (%): 27.48 ± 0.37 -> 26.36 ± 0.27 | Helped |
| Increase distillation temperature from T=1 to T=2 | Collaborative ResNet-32, head 2 · CIFAR-100 | test error (%): 27.64 ± 0.36 -> 26.32 ± 0.26 | Helped |
| Add backpropagation rescaling rather than using no scaling (ablation) | ResNet-32 with 4 heads and simple ILR sharing · CIFAR-10 | test error (%): 6.04 ± 0.17 -> 5.82 ± 0.08 | Helped |
| Add backpropagation rescaling rather than using loss scaling (ablation) | ResNet-32 with 4 heads and simple ILR sharing · CIFAR-10 | test error (%): 6.09 ± 0.24 -> 5.82 ± 0.08 | Helped |
| Use collaborative learning rather than individual learning | ResNet-50 · ImageNet validation | Top-1 error: 23.47 -> 22.29 | Helped |
| Use collaborative learning rather than individual learning | ResNet-50 · ImageNet validation | Top-5 error: 6.83 -> 6.21 | Helped |
| Train collaboratively with multiple heads | ResNet-32 · CIFAR-10 with label noise | test error | Helped |
| Favor hierarchical ILR sharing over greater prediction diversity at high noise | Four-head collaborative CIFAR-10 classifiers · CIFAR-10 with high label noise | test error | Hurt |
Closest papers by meaning
- Collaborative Group Learning (AAAI 2021 · similarity 0.89)
- DNNs as Layers of Cooperating Classifiers (AAAI 2020 · similarity 0.86)
- A Kernel Perspective on Distillation-based Collaborative Learning (NeurIPS 2024 · similarity 0.85)
- Deep Mutual Learning (CVPR 2018 · similarity 0.85)
- CoDTS: Enhancing Sparsely Supervised Collaborative Perception with a Dual Teacher-Student Framework (AAAI 2025 · similarity 0.84)
- MAT-Agent: Adaptive Multi-Agent Training Optimization (NeurIPS 2025 · 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.