Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Companion Learning: Enhancing Generalization Through Historical Consistency
Ruizhao Zhu, Venkatesh Saligrama
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 |
|---|---|---|---|
| add DCL (companion consistency regularizer) to cross-entropy training | ResNet18 · CIFAR-100 | accuracy: 78.3 -> 82.4 | Improved |
| add DCL to cross-entropy training | Swin-T · ImageNet-1K | accuracy: 81.3 -> 81.7 | Improved |
| add DCL to cross-entropy training | ResNet50 · ImageNet-1K | accuracy: 76.0 -> 77.1 | Improved |
| train from scratch with DCL instead of ImageNet pre-training then fine-tuning | ResNet18 · CIFAR-100 | accuracy: 80.5 -> 82.4 | Improved |
| add DCL on top of FixMatch with 40 labels | FixMatch WRN-28-2 · CIFAR-10 (40 labels) | accuracy: 86.2 -> 88.7 | Improved |
| add DCL to MAE pre-training and fine-tune | MAE ViT-B · Stanford Car | accuracy: 82.5 -> 87.2 | Improved |
| add DCL to knowledge distillation (basic KD) | ResNet8x4 student · CIFAR-100 | accuracy: 74.0 -> 75.2 | Improved |
| use MSE instead of KL-Divergence, InfoNCE or L1 as consistency distance (ablation) | DCL · CIFAR-100 | accuracy: 81.5 -> 82.4 | Improved |
| set EMA alpha to 0.6 instead of extremes (0 or 0.99) (ablation) | DCL companion update · CIFAR-100 | accuracy: 80.0 -> 82.4 | Improved |
| use fixed alpha instead of time-varying alpha_t=(t-1)/t (ablation) | DCL · CIFAR-100 | accuracy: 81.5 -> 82.4 | Improved |
| use a smaller companion model (ResNet10-xs) instead of ResNet18 (ablation) | DCL · CIFAR-100 | accuracy: 82.4 -> 80.2 | Hurt |
| train companion on 10% of data instead of 100% (ablation) | DCL · CIFAR-100 | accuracy: 82.4 -> 80.7 | Hurt |
Closest papers by meaning
- Reasoning-Driven Multimodal LLM for Domain Generalization (ICLR 2026 · similarity 0.84)
- Asymmetric Duos: Sidekicks Improve Uncertainty (NeurIPS 2025 · similarity 0.84)
- MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · similarity 0.83)
- Learning from Teaching Regularization: Generalizable Correlations Should be Easy to Imitate (NeurIPS 2024 · similarity 0.83)
- Enhancing Evolving Domain Generalization through Dynamic Latent Representations (AAAI 2024 · similarity 0.83)
- Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification (CVPR 2024 · similarity 0.83)
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.
