Research map / Architectures and efficiency / Quantization
Model compression via distillation and quantization
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 distillation loss during quantized training instead of using normal loss (ablation) | CIFAR-10 student model 1 · CIFAR-10, 2 bits | accuracy: 67.22 % -> 82.40 % | Helped |
| add distillation loss during quantized training instead of using normal loss (ablation) | CIFAR-10 student model 1 · CIFAR-10, 4 bits | accuracy: 86.01 % -> 88.00 % | Helped |
| add distillation loss during quantized training instead of using normal loss (ablation) | OpenNMT student model 1 · OpenNMT integration test, 4 bits | perplexity: 32.67 ppl -> 25.43 ppl | Helped |
| add distillation loss during quantized training instead of using normal loss (ablation) | OpenNMT student model 1 · OpenNMT integration test, 4 bits | BLEU: 15.03 BLEU -> 15.73 BLEU | Helped |
| add bucketing to post-mortem quantization | CIFAR-10 student model 1 · CIFAR-10, 2 bits | accuracy: 9.30 % -> 10.53 % | Helped |
| add bucketing to post-mortem quantization | CIFAR-10 student model 1 · CIFAR-10, 4 bits | accuracy: 67.99 % -> 87.18 % | Helped |
| add gradient-norm-based bit redistribution (ablation) | Differentiable Quantization · CIFAR-10, 2 bits | accuracy: 19.69 % -> 82.94 % | Helped |
| initialize quantization points with quantiles rather than uniform spacing (ablation) | Differentiable Quantization with automatic bit redistribution · CIFAR-10, 2 bits | accuracy: 78.76 % -> 82.94 % | Helped |
| add distillation loss to differentiable quantization (ablation) | Differentiable Quantization with automatic bit redistribution and quantile initialization · CIFAR-10, 2 bits | accuracy: 83.67 % vs 82.94 % | No clear effect |
| add distillation loss to differentiable quantization (ablation) | Differentiable Quantization with automatic bit redistribution and quantile initialization · CIFAR-10, 4 bits | accuracy: 88.93 % vs 88.80 % | No clear effect |
| use a shallower student instead of the deeper teacher | ImageNet ResNet distillation · ImageNet validation | inference speed: 263 seconds -> 169 seconds | Helped |
Closest papers by meaning
- Few Sample Knowledge Distillation for Efficient Network Compression (CVPR 2020 · similarity 0.90)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.89)
- Knowledge distillation via softmax regression representation learning (ICLR 2021 · similarity 0.89)
- N2N learning: Network to Network Compression via Policy Gradient Reinforcement Learning (ICLR 2018 · similarity 0.89)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · similarity 0.89)
- Layer Compression of Deep Networks with Straight Flows (AAAI 2024 · 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.