Research map / Architectures and efficiency / CNN design
LST-Net: Learning a Convolutional Neural Network with a Learnable Sparse Transform
Lida Li, Kun Wang, Shuai Li, Xiangchu Feng, Lei Zhang
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 |
|---|---|---|---|
| initialize spatial and channel transforms with 2D-DCT instead of random (ablation) | LST-Net · CIFAR-100 | top-1 error: 31.12 -> 28.21 (20 layers); 28.94 -> 21.94 (164 layers) | Improved |
| use soft thresholding with tau=1e-4 instead of tau=0 (plain ReLU) (ablation) | LST-Net-20 · CIFAR-100 | top-1 error: 28.92 -> 28.21 | Improved |
| use LST-II bottleneck instead of LST-I (ablation) | LST-Net · CIFAR-100 | top-1 error: 27.64 -> 28.21 (20 layers, LST-II worse); 23.15 -> 21.94 (164 layers) | Improved |
| use ReLU-ST instead of ReLU (ablation) | LST-Net · CIFAR-100 | top-1 error: 28.92 -> 28.21 (20 layers); 22.86 -> 21.94 (164 layers) | Improved |
| use ReLU-ST instead of ELU (ablation) | LST-Net-164 · CIFAR-100 | top-1 error: 29.94 -> 21.94 | Improved |
| replace Conv2d bottleneck with LST (ResNet-110) | ResNet-110 · CIFAR-100 | top-1 error: 25.2 -> 22.7 | Improved |
| replace Conv2d bottleneck with LST (ResNet-18) | ResNet-18 · ImageNet | top-1 error: 30.24 -> 26.55 | Improved |
| add soft thresholding to LST-Net activation (ablation) | LST-Net-18 · ImageNet-C | mCE: 80.34 -> 79.89 | Improved |
| add soft thresholding to LST-Net activation (ablation) | LST-Net-50 · ImageNet-C | mCE: 70.85 -> 70.54 | Improved |
| replace Conv2d with LST vs adding SE/CBAM attention | ResNet-18 · ImageNet-C | mCE: 83.97 -> 79.89 | Improved |
| increase WRN-28 width multiplier from 10 to 12 | WRN · CIFAR-10 | error rate: 4.17 -> 4.33 | Mixed |
| vary tau across 1e-1 to 1e-6 (ablation) | LST-Net-20 · CIFAR-100 | top-1 error: 28.32, 28.28, 28.86, 28.21, 28.43, 28.70 | Mixed |
Closest papers by meaning
- Sparse Deep Transfer Learning for Convolutional Neural Network (AAAI 2017 · similarity 0.88)
- Convolutional Neural Networks Analyzed via Convolutional Sparse Coding (JMLR 2017 · similarity 0.87)
- Holistic CNN Compression via Low-Rank Decomposition with Knowledge Transfer (TPAMI 2018 · similarity 0.87)
- Inference, Learning and Attention Mechanisms that Exploit and Preserve Sparsity in CNNs (IJCV 2020 · similarity 0.87)
- Sequencer: Deep LSTM for Image Classification (NeurIPS 2022 · similarity 0.87)
- DeltaCNN: End-to-End CNN Inference of Sparse Frame Differences in Videos (CVPR 2022 · 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.
