Research map / Architectures and efficiency / CNN design
Band-limited Training and Inference for Convolutional Neural Networks
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 |
|---|---|---|---|
| Apply fixed coefficient-based compression across convolutional layers (ablation) | ResNet-18 · CIFAR-10 | accuracy: 92.32 vs 83.37 | Helped |
| Use adaptive energy-based compression rather than fixed compression (ablation) | ResNet-18 · CIFAR-10 | accuracy: 83.37% | Hurt |
| Keep compression fixed instead of changing it dynamically during training (ablation) | CNN training · CIFAR-10 | accuracy: 69.47%, 83.37% and 88.99% | Helped |
| Match inference compression to the compression used during training (ablation) | ResNet-18, DenseNet-121, and UCR FCN · CIFAR-10, CIFAR-100, and UCR time-series | accuracy | Helped |
| Increase frequency compression (ablation) | ResNet-18 and DenseNet-121 · CIFAR-10 and CIFAR-100 | accuracy: 93.69 -> 92.32; 75.30 -> 71.18 | Hurt |
| Use 50% band-limited convolution instead of FP16 arithmetic | ResNet-18 · CIFAR-10 | accuracy: 91.53 -> 92.32 | Helped |
| Increase band-limiting compression (ablation) | ResNet-18 · Gaussian-noise perturbed CIFAR-10 | accuracy under noise | Helped |
| Use band-limited models rather than reduced-precision models | ResNet-18 · CIFAR-10 with Gaussian noise | accuracy under noise | Helped |
| Apply 90% energy-preserving band-limiting | 3-layer FCN · UCR representative time-series dataset | Avg. GPU mem usage (MB): 118 -> 25 | Helped |
| Apply 90% energy-preserving band-limiting | 3-layer FCN · UCR representative time-series dataset | accuracy: 64.40 -> 63.52 | Hurt |
| Apply compression beyond 50% using energy-based rather than fixed coefficient compression (ablation) | DenseNet-121 · CIFAR-100 | accuracy | No clear effect |
| Train with frequency compression rather than full-spectrum convolution (ablation) | FCN · UCR archive time-series datasets | accuracy | Mixed |
Closest papers by meaning
- CNNpack: Packing Convolutional Neural Networks in the Frequency Domain (NeurIPS 2016 · similarity 0.88)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.86)
- High-Frequency Component Helps Explain the Generalization of Convolutional Neural Networks (CVPR 2020 · similarity 0.85)
- Task-Driven Wavelets using Constrained Empirical Risk Minimization (CVPR 2024 · similarity 0.85)
- Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning (NeurIPS 2024 · similarity 0.84)
- Which Frequencies do CNNs Need? Emergent Bottleneck Structure in Feature Learning (ICML 2024 · 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.