Research map / Architectures and efficiency / Neural architecture search
UNAS: Differentiable Architecture Search Meets Reinforcement Learning
Check 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 |
|---|---|---|---|
| Use the generalization-gap objective instead of validation loss during search (ablation) | UNAS with Gumbel-Softmax · CIFAR-10 | mean test error (%): 2.79{ $\pm$0.10} -> 2.81{ $\pm$0.01} | Helped |
| Use the generalization-gap objective instead of validation loss during search (ablation) | UNAS with Gumbel-Softmax · CIFAR-100 | mean test error (%): 17.11{ $\pm$0.38} -> 16.98{ $\pm$0.34} | Helped |
| Use the generalization-gap objective instead of validation loss during search (ablation) | UNAS with Gumbel-Softmax · ImageNet | mean top-1 error (%): 26.06{ $\pm$0.51} -> 24.64{ $\pm$0.13} | Helped |
| Use REBAR instead of Gumbel-Softmax for gradient estimation (ablation) | UNAS with L_gen · CIFAR-10 | mean test error (%): 2.81{ $\pm$0.01} -> 2.65{ $\pm$0.07} | Helped |
| Use REBAR instead of Gumbel-Softmax for gradient estimation (ablation) | UNAS with L_gen · CIFAR-100 | mean test error (%): 16.98{ $\pm$0.34} -> 16.72{ $\pm$0.76} | Helped |
| Use REBAR instead of Gumbel-Softmax for gradient estimation (ablation) | UNAS with L_gen · ImageNet | mean top-1 error (%): 24.64{ $\pm$0.13} -> 24.60{ $\pm$0.06} | Helped |
| Remove the absolute value from the generalization-gap term (ablation) | UNAS search objective · Experiments in the paper | performance | Hurt |
| Use one architecture sample rather than a batch-size number of samples (ablation) | UNAS search · CIFAR-10 | classification error (%): 2.83\% -> 3.06\% | Hurt |
| Apply dropout to the skip connection during latency-based search (ablation) | UNAS in the ProxylessNAS search space · ImageNet latency search | skip-operation selection | Helped |
| Use UNAS with L_gen and REBAR instead of DARTS | DARTS-space architecture search · CIFAR-100 | mean test error (%): 27.83{ $\pm$8.47} -> 16.72{ $\pm$0.76} | Helped |
| Search with UNAS in the ProxylessNAS search space rather than ProxylessNAS-GPU | ImageNet mobile architecture search · ImageNet | GPU latency (ms): 10.1 ms -> 9.8 ms | Helped |
| Search with UNAS in the ProxylessNAS search space rather than ProxylessNAS-GPU | ImageNet mobile architecture search · ImageNet | top-1 error (%): 24.9\% -> 24.7\% | Helped |
Closest papers by meaning
- BaLeNAS: Differentiable Architecture Search via the Bayesian Learning Rule (CVPR 2022 · similarity 0.92)
- Rethinking Architecture Selection in Differentiable NAS (ICLR 2021 · similarity 0.91)
- Improving Differentiable Neural Architecture Search by Encouraging Transferability (ICLR 2023 · similarity 0.90)
- iDARTS: Differentiable Architecture Search with Stochastic Implicit Gradients (ICML 2021 · similarity 0.89)
- EG-NAS: Neural Architecture Search with Fast Evolutionary Exploration (AAAI 2024 · similarity 0.89)
- Zero-Cost Operation Scoring in Differentiable Architecture Search (AAAI 2023 · similarity 0.89)
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.