Research map / Architectures and efficiency / Pruning and efficient tuning
Composable Sparse Subnetworks via Maximum-Entropy Principle
Francesco Caso, Samuele Fonio, Simone Monaco, Nicola Saccomanno, Fabrizio Silvestri
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 ME loss instead of Quasi-MaxEnt loss | Shallow MLP, logit merge · Yeast pairwise merge |R|=2 | rewarded accuracy: 0.474 -> 0.565 | Improved |
| use ME loss instead of cross-entropy | CNN, logit merge · FMNIST pairwise merge |R|=1 | rewarded accuracy: 0.633 -> 0.983 | Improved |
| use ME loss instead of cross-entropy | Shallow MLP, weight merge · FMNIST pairwise merge |R|=5 | rewarded accuracy: 0.434 -> 0.827 | Improved |
| merge in logit space instead of weight space | Deep MLP with ME · HAR pairwise merge |R|=2 | rewarded accuracy: 0.870 -> 0.942 | Improved |
| add IMP before weight-space merging (ablation) | Deep MLP with ME · MNIST and FMNIST pairwise merge | rewarded accuracy | Improved |
| add IMP before weight-space merging (ablation) | Shallow MLP and CNN with ME · MNIST and FMNIST pairwise merge | rewarded accuracy | Improved |
| add IMP to ME training of single-class modules (ablation) | Shallow MLP · MNIST single-class submodule | rewarded accuracy: 0.998 -> 0.999 | No clear effect |
| add IMP (ablation) | ResNet18 with ME · FMNIST and MNIST pairwise merge | rewarded accuracy | No clear effect |
| re-estimate BatchNorm statistics on union of rewarded classes (ablation) | VGG11 weight merge with IMP · FMNIST pairwise merge | rewarded accuracy | Improved |
| prune to 99% sparsity instead of 90% or 99.9% (ablation) | Shallow MLP · MNIST complete merge | rewarded accuracy and entropy | Improved |
| use pruning ratio 0.6 instead of other ratios (ablation) | CNN · MNIST and FMNIST pairwise merge | rewarded accuracy | Improved |
| use softmax temperature t=5 instead of 1, 3 or 7 (ablation) | Shallow MLP with ME · MNIST complete merge | entropy and rewarded accuracy | Improved |
Closest papers by meaning
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.88)
- Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks (NeurIPS 2025 · similarity 0.87)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.87)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 2024 · similarity 0.87)
- Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning (NeurIPS 2024 · similarity 0.86)
- Position: A Theory of Deep Learning Must Include Compositional Sparsity (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.
