Research map / Architectures and efficiency / CNN design
Residual Distillation: Towards Portable Deep Neural Networks without Shortcuts
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 |
|---|---|---|---|
| add teacher-gradient joint training (JointRD) over naive training of plain-CNN 50 | plain-CNN 50 · ImageNet val | top-1 accuracy: 69.34 -> 76.08 | Helped |
| JointRD vs logits KD | plain-CNN 50 · ImageNet val | top-1 accuracy: 71.47 -> 76.08 | Helped |
| add teacher gradient passing on top of KD(MSE)+Dirac | plain-CNN 50 · CIFAR-100 | accuracy: 50.38 -> 78.16 | Helped |
| add final-layer KD loss to JointRD (ablation) | JointRD plain-CNN 50 · CIFAR-100 | accuracy: 69.67 -> 78.16 | Helped |
| add gradient passing from teacher to Dirac-only / KD-only training (ablation) | plain-CNN 50 · CIFAR-100 | accuracy: 71.12 -> 78.16 | Helped |
| use pretrained fixed ResNet teacher instead of random-init jointly trained (ablation) | JointRD plain-CNN 50 · ImageNet val | top-1 accuracy: 67.45 -> 76.08 | Helped |
| use pretrained fixed ResNet teacher instead of random-init (ablation) | JointRD plain-CNN 50 · CIFAR-100 | accuracy: 78.22 -> 78.16 | No clear effect |
| decay teacher gradient weight with cosine annealing vs constant (ablation) | JointRD plain-CNN 50 · CIFAR-100 | accuracy: 74.23 -> 78.16 | Helped |
| JointRD plain-CNN replaces ResNet-50 in transfer fine-tuning | ResNet-50 · MIT67 | accuracy: 80.34 -> 81.67 | Helped |
| remove shortcuts via JointRD vs ResNet-50 | plain-CNN 50 · CIFAR-10 | accuracy: 95.31 -> 94.40 | Hurt |
| apply KD(logits) on top of JointRD | plain-CNN 34 · CIFAR-100 | accuracy: 78.47 -> 79.23 | Helped |
| prune 60% then JointRD | plain-CNN 34 · CIFAR-100 | accuracy: 78.47 -> 75.38 | Hurt |
Closest papers by meaning
- Towards Understanding the Importance of Shortcut Connections in Residual Networks (NeurIPS 2019 · similarity 0.90)
- Simplified Reversible Residual Networks (NeurIPS 2026 · similarity 0.89)
- Layer Compression of Deep Networks with Straight Flows (AAAI 2024 · similarity 0.88)
- Revisiting Residual Connections: Orthogonal Updates for Stable and Efficient Deep Networks (NeurIPS 2025 · similarity 0.87)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.86)
- LAuReL: Learned Augmented Residual Layer (ICML 2025 · similarity 0.86)
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.