Research map / Architectures and efficiency / Neural architecture search
BlockQNN: Efficient Block-Wise Neural Network Architecture Generation
Zhao Zhong, Zichen Yang, Boyang Deng, Junjie Yan, Wei Wu, Jing Shao, Cheng‐Lin Liu
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 Q-learning guided search instead of random search (ablation) | BlockQNN · CIFAR-100 early-stop accuracy | accuracy: by over 1% | Improved |
| add Q-learning guided search instead of random search with limited parameters (ablation) | BlockQNN-L · CIFAR-100 early-stop accuracy | accuracy: by over 1% | Improved |
| add epsilon-greedy exploitation phase after random exploration (ablation) | BlockQNN · CIFAR-100 early-stop accuracy | mean accuracy: 56% -> 65% | Improved |
| search the combined PCC cell instead of separate ReLU, Conv and BN (ablation) | BlockQNN search · CIFAR-100 Q-learning | accuracy | Improved |
| add optimal block connection style instead of sequential stacking (ablation) | Block-QNN-S · CIFAR-10 | error rate: 3.30 -> 3.01 | Improved |
| add more filters to Block-QNN-Connection (ablation) | Block-QNN-Connection · CIFAR-10 | error rate: 3.01 -> 2.35 | Improved |
| add more filters per conv layer (ablation) | Block-QNN-S · CIFAR-10 | error rate: 3.30 -> 2.80 | Improved |
| add more filters per conv layer (ablation) | Block-QNN-Faster · CIFAR-10 | error rate: 3.57 -> 3.15 | Improved |
| replace conv with depthwise-separable cell | BlockQNN · CIFAR-10 | error rate: 3.30 -> 2.58 | Improved |
| use performance predictor instead of training sampled networks | Faster BlockQNN · CIFAR-10 | error rate and cost: 2.80 -> 3.15; 3 days on 32 GPUs -> 0.8 day on 1 GPU | Mixed |
| train on 224 and test on 320 instead of 224 | Block-QNN-Depthwise · ImageNet | top-1 error: 19.00 -> 18.00 | Improved |
| use Block-QNN-Depthwise instead of NASNet-A (6 @ 4032) at same 224 training | Block-QNN-Depthwise · ImageNet | top-1 error: 19.90 -> 19.00 | Improved |
Later papers that compared with this method
1 later accepted paper reports BlockQNN as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Data-Free Neural Architecture Search via Recursive Label Calibration (ECCV 2022 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Practical Block-Wise Neural Network Architecture Generation (CVPR 2018 · similarity 0.99)
- BlockSwap: Fisher-guided Block Substitution for Network Compression on a Budget (ICLR 2020 · similarity 0.86)
- Blockout: Dynamic Model Selection for Hierarchical Deep Networks (CVPR 2016 · similarity 0.86)
- Designing Neural Network Architectures using Reinforcement Learning (ICLR 2017 · similarity 0.85)
- Efficient ConvBN Blocks for Transfer Learning and Beyond (ICLR 2024 · similarity 0.84)
- Compute Better Spent: Replacing Dense Layers with Structured Matrices (ICML 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.
