Research map / Architectures and efficiency / RNNs and normalization
Fast, Differentiable and Sparse Top-k: a Convex Analysis Perspective
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 |
|---|---|---|---|
| Replace hard magnitude top-k pruning with the differentiable or differentiable-a.e. operator | MLP weight pruning · MNIST | training speed | Helped |
| Use differentiable magnitude top-k pruning instead of hard top-k pruning | MLP weight pruning · MNIST | test error | Helped |
| Use the entropy-regularized approximation with a straight-through trick | MLP weight pruning · MNIST | test error: 5.9 % | No clear effect |
| Replace cross-entropy with the smooth sparse top-3 loss | ViT-B/16 fine-tuning · CIFAR-100 | top-k error | Helped |
| Replace hard top-k routing with the differentiable a.e. top-k operator | V-MoE ViT-S/32 · JFT-300M | precision-at-1 | Helped |
| Use p=4/3 regularization instead of p=2 (ablation) | Smooth top-3 loss for ViT-B/16 · CIFAR-100 | runtime (sec/step): 9.7 sec/step -> 10 sec/step | Hurt |
| Use the differentiable a.e. top-k router instead of the hard top-k router | V-MoE ViT-S/32 · JFT-300M | runtime: 10 hours -> 15 hours | Hurt |
Closest papers by meaning
- ST$_k$: A Scalable Module for Solving Top-k Problems (NeurIPS 2024 · similarity 0.87)
- Statistical Perspective of Top-K Sparse Softmax Gating Mixture of Experts (ICLR 2024 · similarity 0.85)
- Searching for Efficient Linear Layers over a Continuous Space of Structured Matrices (NeurIPS 2024 · similarity 0.85)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.85)
- SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization (NeurIPS 2024 · similarity 0.84)
- LapSum - One Method to Differentiate Them All: Ranking, Sorting and Top-k Selection (ICML 2025 · similarity 0.84)
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.