Research map / Label-efficient and robust learning / Knowledge distillation
Scaled Decoupled Distillation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Scale Decoupled Distillation (SDD), which splits the global logit map into multi-scale local logits and distills consistent and complementary parts, improving logit-based KD on CIFAR-100, ImageNet, and CUB200.
Problem
Logit-based knowledge distillation underperforms feature-based methods because the global logit output mixes multiple semantic classes, transferring ambiguous knowledge to the student.
Approach
Apply multi-scale average pooling to the teacher and student logit maps before the classifier-output stage, distill each scale's local logits (with the standard KD or DKD/NKD loss), and up-weight the local logits whose predicted class differs from the global prediction (complementary terms) to emphasize ambiguous samples.
Key results
- SD-DKD improves DKD on CIFAR-100 ResNet50-ShuffleNetV1 from 77.01 to 78.11 top-1 accuracy.
- On ImageNet (ResNet34 teacher, ResNet18 student), SD-KD raises KD top-1 from 69.75 to 71.44 (+0.84 is reported for SD-KD vs KD row).
- On CUB200, SD-DKD improves DKD by up to 6.41 points (VGG13-MobileNetV2 pair, 58.45 to 64.86).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add scale-decoupled distillation to DKD | DKD · CIFAR-100 | accuracy: 68.98 -> 70.08 | Helped |
| Add scale-decoupled distillation to DKD | DKD · ImageNet validation, ResNet-34 → ResNet18 | Top-1 accuracy (%): 71.70 -> 72.02 | Helped |
| Add scale-decoupled distillation to DKD | DKD · CUB200 | accuracy: 59.94 -> 62.97 | Helped |
| Add consistent decoupled logits (ablation) | Conventional KD, ResNet32x4 → ResNet8x4 · CIFAR-100 | accuracy: 73.33 -> 75.14 | Helped |
| Add complementary decoupled logits (ablation) | Conventional KD, ResNet32x4 → ResNet8x4 · CIFAR-100 | accuracy: 73.33 -> 74.79 | Helped |
| Combine consistent and complementary decoupled logits (ablation) | Conventional KD, ResNet32x4 → ResNet8x4 · CIFAR-100 | accuracy: 73.33 -> 76.63 | Helped |
| Use scales M={1,2} instead of global-only M={1} (ablation) | SDD, ResNet32x4 → ResNet8x4 · CIFAR-100 | accuracy: 73.33 -> 76.63 | Helped |
| Use scales M={1,2,4} instead of M={1,2} (ablation) | SDD, ResNet32x4 → ResNet8x4 · CIFAR-100 | accuracy: 76.63 -> 75.74 | Hurt |
| Use scales M={1,2,4} instead of global-only M={1} (ablation) | SDD, ResNet32x4 → ShuffleNetV1 · CIFAR-100 | accuracy: 74.07 -> 76.30 | Helped |
| Increase complementary-loss weight to β=2 (ablation) | SDD, ResNet32x4 → ShuffleNetV1 · CIFAR-100 | accuracy: 75.62 -> 76.30 | Helped |
| Increase complementary-loss weight from β=2 to β=4 (ablation) | SDD, ResNet32x4 → ShuffleNetV1 · CIFAR-100 | accuracy: 76.30 -> 75.99 | Hurt |
| Use scale-decoupled pooling in KD | KD, ResNet32x4 → ResNet8x4 · CIFAR-100 | training time (per batch): 11 -> 11 | No clear effect |
Surprising findings
- SDD's gains do not come from better imitating the teacher's global logit output; the global-logit difference map of SD-KD resembles that of KD.
Limitations
- Multi-scale hyperparameters (scale set M, alpha, beta) must be tuned per teacher-student pair; scale set differs between homogeneous and heterogeneous pairs.
- Improvements over DKD/NKD are typically about 1 point on standard benchmarks, with smaller gains on some pairs (e.g. +0.45 on WRN40-2 to MobileNetV2 for SD-DKD).
Experimental setup
Datasets: CIFAR-100 ImageNet CUB200
Benchmarks: CIFAR-100 ImageNet CUB200
Metrics: top-1 accuracy top-5 accuracy
Baselines: KD FitNet SP CRD SemCKD ReviewKD MGD DKD NKD WCoRD PKT RKD
Builds on: Knowledge distillation (Hinton) DKD NKD CRD
Closest papers by meaning
- Maintaining Fairness in Logit-based Knowledge Distillation for Class-Incremental Learning (AAAI 2025 · similarity 0.89)
- Logit Standardization in Knowledge Distillation (CVPR 2024 · similarity 0.89)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.88)
- DeepKD: A Deeply Decoupled and Denoised Knowledge Distillation Trainer (NeurIPS 2025 · similarity 0.88)
- Debiased Distillation for Consistency Regularization (AAAI 2025 · similarity 0.88)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.87)
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.