Research map / Label-efficient and robust learning / Knowledge distillation
Born Again 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 |
|---|---|---|---|
| Distill from an identical DenseNet teacher | DenseNet BAN · CIFAR-100 | test error: 17.69 -> 16.69 | Helped |
| Distill from an identical Wide-ResNet teacher | Wide-ResNet BAN · CIFAR-100 | test error: 30.05 -> 29.43; 25.32 -> 24.38; 20.88 -> 20.93; 19.08 -> 18.25 | Mixed |
| Train sequential BAN generations | DenseNet-80-80 · CIFAR-100 | test error: 17.16 -> 16.36; 15.5 | Helped |
| Ensemble multiple BAN generations | BAN DenseNet · CIFAR-100 | test error: 14.9 | Helped |
| Add ground-truth label loss to BAN distillation (ablation) | DenseNet-90-60 BAN · CIFAR-100 | test error: 16.69 -> 16.93 | Hurt |
| Replace the teacher distribution with teacher-max-weighted labels (ablation) | DenseNet BAN · CIFAR-100 | test error: 18.25 -> 17.84; 17.69 -> 17.42; 17.16 -> 17.16; 16.87 -> 17.12 | Mixed |
| Permute the teacher's non-argmax outputs (ablation) | DenseNet BAN · CIFAR-100 | test error: 18.25 -> 17.84; 17.69 -> 17.43; 17.16 -> 16.84; 16.87 -> 16.34 | Helped |
| Reduce student depth to half the teacher depth (ablation) | DenseNet-90-60 BAN · CIFAR-100 | test error: 17.69 -> 16.95 | Helped |
| Increase student depth to twice the teacher depth (ablation) | DenseNet-90-60 BAN · CIFAR-100 | test error: 17.69 -> 16.43 | Helped |
| Halve the compression rate (ablation) | DenseNet-90-60 BAN · CIFAR-100 | test error: 17.69 -> 19.83 | Hurt |
| Distill DenseNet-90-60 into a ResNet student | BAN-ResNet · CIFAR-100 | test error: 20.28 -> 18.8; 18.84 -> 17.39; 20.4 -> 19.12; 18.83 -> 17.42; 19.63 -> 17.13; 18.77 -> 17.18 | Helped |
| Distill with teacher outputs and label loss | BAN language models · PTB test | perplexity: 80.05 -> 76.97; 71.87 -> 68.56 | Helped |
Closest papers by meaning
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · similarity 0.90)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.89)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.89)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.88)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.88)
- Debiased Distillation for Consistency Regularization (AAAI 2025 · 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.