Research map / Architectures and efficiency / Neural architecture search
A Study on Encodings for Neural Architecture Search
Colin White, Willie Neiswanger, Sam Nolen, Yash Savani
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 adjacency encoding instead of path encoding in neural predictor (ablation) | neural predictor · NASBench-101 outside search space | test error top 10 avg.: 6.616 -> 6.454 | Improved |
| use categorical adjacency instead of adjacency encoding (ablation) | neural predictor · NASBench-101 outside search space | test error top 10 avg.: 6.454 -> 8.155 | Hurt |
| truncate path encoding (ablation) | neural predictor · NASBench-101 outside search space | test error top 10 avg.: 6.616 -> 6.712 | Hurt |
| make path encoding categorical (ablation) | neural predictor · NASBench-101 outside search space | test error top 10 avg.: 6.616 -> 6.939 | Hurt |
| truncate path encoding from 364 to 40 paths (ablation) | regularized evolution · NASBench-101 | test error | Improved |
| truncate path encoding from 364 to 40 paths (ablation) | other NAS algorithms with path encoding · NASBench-101 | test error | No clear effect |
| use path encodings instead of adjacency matrix encodings for the predictor model (ablation) | BO and neural predictor · NASBench-101 | predictor error / search performance | Improved |
| use adjacency matrix encodings instead of path encodings for random sampling and perturbation (ablation) | Random Search, regularized evolution, local search · NASBench-101 | test error | Improved |
| use path encoding instead of GCN or VAE (ablation) | neural predictor · NASBench-101 | mean absolute error | Improved |
| combine best encodings for each subroutine (ablation) | BANANAS · NASBench-101 | test error | Improved |
| truncate one-hot adjacency encoding from 31 bits to 0 bits versus truncating path encoding (ablation) | regularized evolution · NASBench-101 | test error | Mixed |
| group architectures by path encoding equivalence class (ablation) | path encoding · NASBench-101 | std of accuracy (%): 5.55 -> 0.353 | Improved |
Closest papers by meaning
- Understanding and Accelerating Neural Architecture Search With Training-Free and Theory-Grounded Metrics (TPAMI 2023 · similarity 0.89)
- Understanding Architectures Learnt by Cell-based Neural Architecture Search (ICLR 2020 · similarity 0.89)
- Encodings for Prediction-based Neural Architecture Search (ICML 2024 · similarity 0.89)
- Insights from the Use of Previously Unseen Neural Architecture Search Datasets (CVPR 2024 · similarity 0.89)
- Evaluating The Search Phase of Neural Architecture Search (ICLR 2020 · similarity 0.88)
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.87)
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.
