Research map / Architectures and efficiency / CNN design
Towards Learning Convolutions from Scratch
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 |
|---|---|---|---|
| train S-FC with beta-lasso (beta=50) instead of SGD | S-FC · CIFAR-10 | accuracy: 78.63 -> 85.19 | Helped |
| train S-FC with beta-lasso (beta=50) instead of SGD | S-FC · CIFAR-100 | accuracy: 51.43 -> 59.56 | Helped |
| train S-FC with beta-lasso (beta=50) instead of SGD | S-FC · SVHN | accuracy: 91.80 -> 94.07 | Helped |
| raise beta from 0 to 50 in beta-lasso (more aggressive soft thresholding) (ablation) | S-FC beta-lasso · CIFAR-10 | accuracy: 82.45 -> 85.19 | Helped |
| raise beta from 1 to 50 in beta-lasso (ablation) | S-FC beta-lasso · CIFAR-100 | accuracy: 55.96 -> 59.56 | Helped |
| use local connectivity (locally connected) instead of fully connected, shallow (ablation) | S-FC · CIFAR-10 (400 epochs) | accuracy: 72.77 -> 81.52 | Helped |
| add weight sharing (conv vs local), shallow net (ablation) | S-local · CIFAR-10 (400 epochs) | accuracy: 81.52 -> 84.14 | Helped |
| train deep fully-connected D-FC instead of shallow S-FC (ablation) | FC networks · CIFAR-10 (400 epochs) | accuracy: 72.77 -> 64.78 | Hurt |
| train shallow architectures for 4000 instead of 400 epochs (ablation) | S-FC · CIFAR-10 | accuracy: 72.77 -> 78.63 | Helped |
| train deep D-conv for 4000 instead of 400 epochs (ablation) | D-conv · CIFAR-100 | accuracy: 63.73 -> 62.26 | No clear effect |
| use beta-lasso instead of SGD to train ResNet18 with different kernel sizes | ResNet18 · CIFAR-10/100, SVHN | accuracy | Helped |
| use S-FC with wide first layer instead of 3-FC at equal parameters (ablation) | fully-connected net · CIFAR-10 | accuracy: 75.12 -> 78.63 | Helped |
Closest papers by meaning
- Revisiting Sparse Convolutional Model for Visual Recognition (NeurIPS 2022 · similarity 0.88)
- Refining Architectures of Deep Convolutional Neural Networks (CVPR 2016 · similarity 0.87)
- Linear CNNs Discover the Statistical Structure of the Dataset Using Only the Most Dominant Frequencies (ICML 2023 · similarity 0.87)
- Structured Convolutions for Efficient Neural Network Design (NeurIPS 2020 · similarity 0.86)
- RePr: Improved Training of Convolutional Filters (CVPR 2019 · similarity 0.86)
- Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling (ICLR 2023 · 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.