Research map / Architectures and efficiency / Pruning and efficient tuning
Structured Bayesian Pruning via Log-Normal Multiplicative Noise
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 |
|---|---|---|---|
| use structured Bayesian pruning instead of the original network | LeNet-500-300 · MNIST | FLOPs speedup: 1.00× -> 11.23× | Helped |
| use structured Bayesian pruning instead of the original network | LeNet-500-300 · MNIST | Error (%): 1.54 -> 1.55 | No clear effect |
| use structured Bayesian pruning instead of the original network | LeNet5-Caffe · MNIST | FLOPs speedup: 1.00× -> 10.49× | Helped |
| use structured Bayesian pruning instead of the original network | LeNet5-Caffe · MNIST | Error (%): 0.80 -> 0.86 | No clear effect |
| use structured Bayesian pruning instead of SparseVD | VGG-like · CIFAR-10, k=1.0 | FLOPs speedup: 2.27× -> 2.30× | No clear effect |
| scale the KL term by layer computational complexity (ablation) | StructuredBP · CIFAR-10, VGG-like k=1.0 | Error (%): 7.5 -> 9.0 | Hurt |
| scale the KL term by layer computational complexity (ablation) | StructuredBP · CIFAR-10, VGG-like k=1.0 | FLOPs speedup: 2.30× -> 3.16× | Helped |
| scale the KL term by layer computational complexity (ablation) | StructuredBP · CIFAR-10, VGG-like k=1.5 | Error (%): 7.2 -> 7.8 | Hurt |
| scale the KL term by layer computational complexity (ablation) | StructuredBP · CIFAR-10, VGG-like k=1.5 | FLOPs speedup: 3.32× -> 3.93× | Helped |
| train both the noise mean and variance instead of fixing the mean (ablation) | SBP · MNIST | model quality and sparsity | Helped |
| use Bayesian ensembling at test time rather than expected noise values (ablation) | SBP · classification experiments | accuracy | No clear effect |
| apply SBP to random-label data instead of Binary Dropout | LeNet5 on MNIST and VGG-like on CIFAR-10 · random-label classification | overfitting | Helped |
Closest papers by meaning
- Structured Dropout Variational Inference for Bayesian Neural Networks (NeurIPS 2021 · similarity 0.90)
- BMRS: Bayesian Model Reduction for Structured Pruning (NeurIPS 2024 · similarity 0.90)
- Dropout as a Structured Shrinkage Prior (ICML 2019 · similarity 0.90)
- Joint Inference for Neural Network Depth and Dropout Regularization (NeurIPS 2021 · similarity 0.88)
- Dynamic Sparse Training with Structured Sparsity (ICLR 2024 · similarity 0.86)
- Stochastic Modified Equations and Dynamics of Dropout Algorithm (ICLR 2024 · similarity 0.85)
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.