Research map / Architectures and efficiency / Neural architecture search
BATS: Binary ArchitecTure Search
Adrian Bulat, Brais Martinez, Georgios Tzimiropoulos
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 |
|---|---|---|---|
| use binary-oriented grouped-convolution search space instead of DARTS search space (ablation) | BATS · CIFAR-10 | Top-1 accuracy: 51.0±7.5 -> 93.7±0.6 | Improved |
| lower softmax temperature from 1 to 0.2 (ablation) | BATS search · CIFAR-10 | Top-1 accuracy: 86.0 -> 93.7 | Improved |
| lower temperature from 0.5 to 0.2 (ablation) | BATS search · CIFAR-10 | Top-1 accuracy: 92.1 -> 93.7 | Improved |
| lower temperature from 0.2 to 0.05 (ablation) | BATS search · CIFAR-10 | Top-1 accuracy: 93.7 -> 93.3 | Hurt |
| search with binary activations and real weights instead of real search then binarize (ablation) | BATS search · CIFAR-10 | Top-1 accuracy: 88.4±0.8 -> 93.7±0.6 | Improved |
| binarize both weights and activations during search instead of only activations (ablation) | BATS search · CIFAR-10 | Top-1 accuracy: 93.7±0.6 -> 84.6±1.7 | Hurt |
| keep activations real and binarize weights during search instead of binarizing activations (ablation) | BATS search · CIFAR-10 | Top-1 accuracy: 93.7±0.6 -> 91.1±1.7 | Hurt |
| add autoaugment (ablation) | BATS · CIFAR-100 | Test accuracy: 70.7 -> 71.3 | Improved |
| add autoaugment (ablation) | BATS (large) · CIFAR-10 | Test accuracy: 95.5 -> 96.1 | Improved |
| scale up width and groups of the cells to large (ablation) | BATS · CIFAR-100 | Test accuracy: 70.7 -> 75.7 | Improved |
| widen the network 2x (ablation) | BATS · ImageNet | Top-1 accuracy: 60.4 -> 66.1 | Improved |
| use BATS search space and method for real-valued networks (ablation) | BATS · CIFAR-10 real-valued | Test error: 2.40 vs 2.50 (P-DARTS) | Improved |
Later papers that compared with this method
2 later accepted papers report BATS (Small) as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- DCP–NAS: Discrepant Child–Parent Neural Architecture Search for 1-bit CNNs (IJCV 2023 · better on 4 of 4 shared results)
- High-Capacity Expert Binary Networks (ICLR 2021 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- XNOR-Net++ — XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks (ECCV 2016) · better on 2 of 2 shared results
- Bi-Real Net — Bi-Real Net: Enhancing the Performance of 1-bit CNNs with Improved Representational Capability and Advanced Training Algorithm (ECCV 2018) · 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
- Binarized Neural Architecture Search (AAAI 2020 · similarity 0.89)
- Binarized Neural Architecture Search for Efficient Object Recognition (IJCV 2020 · similarity 0.88)
- BN-NAS: Neural Architecture Search With Batch Normalization (ICCV 2021 · similarity 0.88)
- Learning Architectures for Binary Networks (ECCV 2020 · similarity 0.88)
- ISTA-NAS: Efficient and Consistent Neural Architecture Search by Sparse Coding (NeurIPS 2020 · similarity 0.87)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · 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.
