Multi-objective single-shot neural architecture search via efficient convolutional filters

This paper presents a novel approach for fast neural architecture search (NAS) in Convolutional Neural Networks (CNNs) for end-to-end License Plate Recognition (LPR). The authors propose a one-shot schema that considers the efficiency of different convolutional filters to create a search space for more efficient architectures on vector processing cores. The authors' approach utilizes a super-network for LPR using Connectionist-Temporal-Cost (CTC) and ranks the importance of filters to generate a fine-grain list of architectures. These architectures are evaluated in a multi-objective manner, resulting in several Pareto-optimal architectures with different computational costs and validation errors. Rather than using a single complicated building block for all layers, the authors' method allows each stage to select a custom building block with fewer or more operations. The authors show that their super-network is flexible to calculate filters of any required size and stride in each stage while keeping it efficient by the structural pruning. The authors' experiments, which were performed on Iranian LPR, demonstrate that this method produces a variety of fast and efficient CNNs. Furthermore, the authors discuss the potential of this method for use in other areas of CNN application.