Research map / Architectures and efficiency / Pruning and efficient tuning
Integral 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 |
|---|---|---|---|
| Initialize the integral network from a pre-trained discrete network instead of training it from scratch (ablation) | INN · Image classification and super-resolution benchmarks in Table 1 | accuracy / PSNR: 91.8 -> 92.5; 89.4 -> 91.6; 93.1 -> 95.3; 68.5 -> 72.4; 66.5 -> 70.0; 71.2 -> 74.1; 32.6 -> 32.9; 32.2 -> 32.4; 29.0 -> | Helped |
| Use an integral network initialized from the discrete model rather than the discrete model | INN · Image classification and super-resolution benchmarks in Table 1 | accuracy / PSNR: 92.3 -> 92.5; 95.3 -> 95.3; 72.3 -> 72.4; 74.1 -> 74.1; 32.9 -> 32.9; 32.4 -> 32.4; 29.4 -> 29.4; 28.7 -> 28.7; 26.8 -> | No clear effect |
| Add the filter/channel permutation when converting a pre-trained network to an INN (ablation) | INN conversion with partition tuning · Cifar10, ResNet-18 at 40% compression | accuracy: 91.3 -> 93.0 | Helped |
| Add the filter/channel permutation when converting a pre-trained network to an INN (ablation) | INN conversion with partition tuning · Cifar10, NIN at 40% compression | accuracy: 84.71 -> 89.4 | Helped |
| Use a non-uniform trainable partition instead of uniform sampling (ablation) | INN pruning and re-discretization · Image classification pruning | accuracy | Helped |
| Train the INN from scratch rather than initialize it from a converted discrete network (ablation) | INN · Classification and super-resolution benchmarks in Table 1 | accuracy / PSNR: 92.5 -> 91.8; 91.6 -> 89.4; 95.3 -> 93.1; 72.4 -> 68.5; 70.0 -> 66.5; 74.1 -> 71.2; 32.9 -> 32.6; 32.4 -> 32.2; 29.4 -> | Hurt |
| Prune an INN by changing its discretization without fine-tuning | Integral Neural Networks · ImageNet, ResNet-18 | accuracy: 2% accuracy loss; up to 30% | Hurt |
| Use INN-based pruning rather than conventional pruning methods under the same conditions | ResNet-18 on ImageNet · ImageNet | accuracy: 2% accuracy loss; 65% accuracy loss | Helped |
Closest papers by meaning
- Accelerate CNNs from Three Dimensions: A Comprehensive Pruning Framework (ICML 2021 · similarity 0.85)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.84)
- Pruning from Scratch (AAAI 2020 · similarity 0.84)
- ChipNet: Budget-Aware Pruning with Heaviside Continuous Approximations (ICLR 2021 · similarity 0.84)
- Pruning Neural Networks at Initialization: Why Are We Missing the Mark? (ICLR 2021 · similarity 0.84)
- Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.83)
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.