Research map / Architectures and efficiency / RNNs and normalization
Toward the Connection Between Activation Sparsity and Flat Minima
Ze Peng, Jian Zhang, Lei Qi, Yang Gao, Y Shi
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 |
|---|---|---|---|
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | ViT-B · ImageNet-1K pretraining | training sparsity: 0.104 -> 0.046 | Improved |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | ViT-B · ImageNet-1K pretraining | testing sparsity: 0.087 -> 0.055 | Improved |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | ViT-B · ImageNet-1K pretraining | Acc@1: 77.35% -> 76.77% | Hurt |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | SwinTransformer-B · ImageNet-1K pretraining | testing sparsity: 0.057 -> 0.041 | Improved |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | SwinTransformer-B · ImageNet-1K pretraining | Acc@1: 79.75% -> 79.82% | No clear effect |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | ViT-L · Places365 pretraining | testing sparsity: 0.039 -> 0.018 | Improved |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | ViT-B · LAION-400M subset | test loss: 0.984 -> 1.067 | Hurt |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | T5-Base · C4 pretraining | testing sparsity: 0.299 -> 0.180 | Improved |
| add sparsity-aware modifications (zeroth biases, lower-bounded LayerNorm, new activation) | T5-Base · C4 pretraining | test loss: 4.88 -> 4.78 | Improved |
| sparsity-aware LoRA finetuning instead of vanilla finetuning | ViT-Base · ImageNet-1K finetuning | testing sparsity / Acc@1: 0.122 -> 0.102 sparsity; 77.17% -> 74.46% Acc@1 | Mixed |
| sparsity-aware LoRA finetuning instead of vanilla finetuning | T5-Base · C4 finetuning | testing sparsity: 0.304 -> 0.199 | Improved |
| add lower-bounded LayerNorm affine factors and bounded zeroth biases (ablation) | ViT · ImageNet-1K pretraining | training sparsity | Improved |
Closest papers by meaning
- DEEP NEURAL NETWORK INITIALIZATION WITH SPARSITY INDUCING ACTIVATIONS (ICLR 2024 · similarity 0.91)
- Exploring the Benefit of Activation Sparsity in Pre-training (ICML 2024 · similarity 0.90)
- Sparsing Law: Towards Large Language Models with Greater Activation Sparsity (ICML 2025 · similarity 0.88)
- Asymmetric Scaling Laws from Sparse Features (NeurIPS 2026 · similarity 0.88)
- Scaling Laws for Sparsely-Connected Foundation Models (ICLR 2024 · similarity 0.88)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.87)
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.
