A Pre-Search Evaluation Framework for Assessing Search Space Complexity
Keywords: Neural Architecture Search, benchmarking, optimization
TL;DR: Assessing search spaces complexity in NAS
Abstract: The search space represents the most important component of Neural Architecture Search. It defines the ranges of performance, encapsulating the potential for discovering optimal architectures. This paper presents a framework for evaluating these spaces based on size, performance diversity, architecture diversity, and multi-task ability. Our comparative analysis across seven established benchmarks highlights their complexity and adaptability to target multiple tasks, offering a comprehensive tool for NAS strategy assessment.
Submission Number: 131
Loading