Low redundancy cell-based Neural Architecture Search for large convolutional neural networks

The cell-based search space is one of the main paradigms in Neural Architecture Search (NAS). However, the current research on this search space tends to optimize on small-size models, and the performance improvement of NAS might be stuck in a bottleneck. This situation has led to a growing performance gap between NAS and hand-designed models in recent years. In this paper, we focus on how to effectively expand the cell-based search space and proposes Low redundancy Cell-based Neural Architecture Search for Large Convolutional neural networks ($L{C}^{2}NAS$), a gradient-based NAS method to search large-scale convolutional models with better performance based on low redundant cell search space. Specifically, a cell-based search space with low redundancy and large kernel is designed. Then train and sample a super network under computational constraints. Finally the network structure is optimized by gradient-based search. Experimental results show that the performance of the proposed search method is comparable to the popular hand-designed models in recent years at different scales. Moreover, LC-NASNet-B achieves an 83.7% classification accuracy on the ImageNet-1k dataset with 86.2M parameters, surpassing previous NAS methods and comparable to the most prominent hand-designed models.