Researchers of CWI’s Life Sciences and Health (LSH) group Tom den Ottelander, Arkadiy Dushatskiy, Marco Virgolin (former CWI), and Peter Bosman (CWI/TU Delft) were awarded a Best Paper award for their paper Local search is a remarkably strong baseline for neural architecture search at the International Conference on Evolutionary Multi-Criterion Optimization (EMO) 2021.
About the paper
Neural Architecture Search (NAS) has recently become a very hot topic in both Deep Learning and Evolutionary Computation communities. However, a profound comparison of search algorithms is often missing, especially in the interesting case where multiple objectives are at play, like accuracy and efficiency. In our paper we showed that multi-objective Local Search is a simple, yet efficient baseline algorithm. Despite its simplicity, our local search algorithm can perform almost as well as complex Evolutionary Algorithms, when network generalization gaps are taken into account. Therefore, we suggest using Local Search as a conventional baseline for novel NAS algorithms. Additionally, we introduced a new challenging NAS benchmark that allows researchers to test their algorithms with little time and computational costs. This benchmark is publicly available.
EMO 2021 is the 11th Edition of International Conference Series on Evolutionary Multi- Criterion Optimization (EMO), aiming to continue the success of previous EMO conferences. We will bring together both the EMO, Multiple Criteria Decision-Making (MCDM) communities, and other related fields and, moreover, focusing on solving real-world problems in government, business and industry.
- CWI’s Life Sciences & Health group
- Paper Local search is a remarkably strong baseline for neural architecture search