Evolutionary Multi-Objective Model Compression for Deep Neural Networks

While deep neural networks (DNNs) deliver state-of-the-art accuracy on various applications from face recognition to language translation, it comes at the cost of high computational and space complexity, hindering their deployment edge devices. To enable efficient processing DNNs in inference, a novel approach, called Evolutionary Multi-Objective Model Compression (EMOMC), is proposed optimize energy efficiency (or model size) simultaneously. Specifically, network pruning quantization are explored exploited by using architecture population evolution. Furthermore, taking advantage orthogonality between quantization, two-stage co-optimization strategy developed, which considerably reduces time search. Lastly, different dataflow designs parameter coding schemes considered optimization process since they have significant impact consumption size. Owing cooperation evolution architectures population, set compact that offer trade-offs objectives (e.g., accuracy, can be obtained single run. Unlike most existing approaches designed reduce size weight parameters with no loss method aims achieve trade-off desirable objectives, for meeting requirements Experimental results demonstrate approach obtain diverse suitable broad range memory usage requirements. Under negligible loss, EMOMC improves compression rate VGG-16 CIFAR-10 factor more than 8 9. X 2.4 X, respectively.