Exploring Neural Architecture Search Spaces via Visual Analytics [Application Notes]

Deep neural networks represent a significant driving force behind the accelerated advancement of artificial intelligence (AI). However, traditional approaches based on manual model architecture design appear inadequate for satisfying various modern applications due to their inherent inefficiency and scalability limitations. The advent of neural architecture search (NAS) methodologies in recent years has rendered the automated search for optimal model architectures a novel possibility. The emergence of NAS benchmarks has also provided an effective solution to the problem of costly evaluation in the NAS process. However, a significant challenge persists that most existing NAS methods cannot provide explanations for the exceptional model architectures identified in the benchmark search space. In light of the vast deployment of visual analytics methods in the domain of explainable AI, this paper introduces a visual analytics framework to facilitate interactive exploration of the search space illustrated in NAS benchmarks. The framework is designed to enable the extraction of potential relationships between model architectures and performance metrics. Furthermore, this paper introduces an analytical pipeline that integrates a range of data processing techniques to efficiently map model architectures across diverse types of large-scale search spaces into a low-dimensional space, which is then adapted for analysis in our visual analytics framework. The effectiveness of the proposed framework is demonstrated through case studies and a user study.