Research map / Architectures and efficiency / Neural architecture search
Analyzing and Mitigating Interference in Neural Architecture Search
Jin Xu, Xu Tan, Kaitao Song, Renqian Luo, Yichong Leng, Tao Qin, Tie-Yan Liu, Jian Li
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 |
|---|---|---|---|
| Gradually substitute one operator between adjacent training steps | SPOS training in the BERT search space · BERT search space, 60 sampled child models | Kendall's Tau: 0.36 -> 0.49 | Improved |
| Add anchor-based input/output alignment | SPOS training in the BERT search space · BERT search space, 60 sampled child models | Kendall's Tau: 0.36 -> 0.44 | Improved |
| Combine gradual topology changes with anchor alignment | SPOS training in the BERT search space · BERT search space, 60 sampled child models | Kendall's Tau: 0.36 -> 0.57 | Improved |
| Use MAGIC-AT rather than SPOS to discover an MLM-pretrained architecture | BERT-base-scale pre-training · GLUE dev | AVG: 84.4 -> 85.7 | Improved |
| Use MAGIC-AT rather than SPOS to discover an MLM-pretrained architecture | BERT-base-scale pre-training · GLUE test | AVG: 82.4 -> 83.4 | Improved |
| Use MAGIC-AT rather than SPOS to discover a compressed architecture | 60M distilled BERT · GLUE dev | AVG: 83.1 -> 84.8 | Improved |
| Use MAGIC-AT rather than SPOS to discover a compressed architecture | 60M distilled BERT · GLUE test | AVG: 81.4 -> 82.0 | Improved |
| Use MAGIC-AT rather than PC-NAS for architecture search | MobileNetV2 search space · ImageNet | Top1 test error: 23.9 -> 23.2 | Improved |
| Increase the number of operators changed at each MAGIC-T step (ablation) | MAGIC-T super-net training · BERT search space | Rank correlation | Hurt |
| Choose a poorly performing child model as the MAGIC-A anchor (ablation) | MAGIC-A super-net training · BERT search space | Rank correlation | Hurt |
| Use an anchor model ranking within the top 30% rather than requiring the top-ranked model (ablation) | MAGIC-A super-net training · BERT search space | Rank correlation | No clear effect |
| Align child-model inputs and outputs to the average across sampled paths (ablation) | Shared operators in the analyzed hybrid BERT search space · Gradient similarity analysis | Gradient cosine similarity | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ELECTRA_base — ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators (ICLR 2020) · better on 4 of 4 shared results
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
- Bridging the Gap between Sample-based and One-shot Neural Architecture Search with BONAS (NeurIPS 2020 · similarity 0.89)
- Distribution Consistent Neural Architecture Search (CVPR 2022 · similarity 0.89)
- Can Weight Sharing Outperform Random Architecture Search? An Investigation With TuNAS (CVPR 2020 · similarity 0.88)
- Landmark Regularization: Ranking Guided Super-Net Training in Neural Architecture Search (CVPR 2021 · similarity 0.88)
- Evaluating The Search Phase of Neural Architecture Search (ICLR 2020 · similarity 0.88)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.88)
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.
