Efficient Evolutionary Architecture Search for CNN Optimization on GTSRB

Abstract

Neural network inference on embedded devices has to meet accuracy and latency requirements under tight resource constraints. The design of suitable network architectures is a challenging and time-consuming task. Therefore, automatic discovery and optimization of neural networks is considered important for continuing the trend of moving classification tasks from cloud to edge computing. This paper presents an evolutionary method to optimize a convolutional neural network (CNN) architecture for classification tasks. The method runs efficiently on a single GPU-workstation and provides simple means to direct the tradeoff between complexity and accuracy of the evolved network. Using this method, we achieved a 11x reduction in the number of multiply-accumulate (MAC) operations of the winning network for the German Traffic Sign Recognition Benchmark (GTSRB) without accuracy reduction. An ensemble of four of our evolved networks competes the winning ensemble with a 0.1% lower accuracy but 70x reduction in MACs and 14x reduction in parameters.