Neural Architecture Search and Multi-Objective Evolutionary Algorithms for Anomaly Detection

Chakkrit Termritthikun, Lin Xu, Yemeng Liu, Ivan Lee

Abstract: The processing speed and memory footprint are important factors for applications processing on resource-constrained devices such as IoT devices and embedded systems. Deep learning has been evolving continuously so that it can be used on resource-constrained devices but there are still some limitations in using it because these devices are not designed for processing complicated tasks. Further, the complexity of the Convolutional Neural Network (CNN) model is the barrier to implementation on these devices. In this paper, we have developed Neural Architecture Search (NAS) that uses a Multi-Objective Genetic Algorithm and Firefly Algorithm for creating a less complicated and robust CNN model, focusing on searching the model with faster processing time and minimum storage size. Five image datasets are applied to examine the performance of the proposed techniques, including two crack detection datasets for surface or built infrastructure inspection for industrial applications. Experimental results show that the proposed technique consistently lowers the parameter counts and processing time at comparable classification accuracies.