Re-training and parameter sharing with the Hash trick for compressing convolutional neural networks

Abstract

As an ubiquitous technology for improving machine intelligence, deep learning has largely taken the dominant position among nowadays most advanced computer vision systems. To achieve superior performance on large-scale datasets, convolutional neural networks (CNNs) are often designed as complex models with millions of parameters. This limits the deployment of CNNs in embedded intelligent computer vision systems, such as intelligent robots that are resource-constrained with real-time computing requirement. This paper proposes a simple and effective model compression scheme to improve the real-time sensing of the surrounding objects. In the proposed framework, the Hash trick is first applied to a modified convolutional layer, and the compression of the convolutional layer is realized via weight sharing. Subsequently, the Hash index matrix is introduced to represent the Hash function, and its relaxation regularization is introduced into the fine-tuned loss function. Through the dynamic retraining of the index matrix, the Hash function can be updated. We evaluate our method using several state-of-the-art CNNs. Experimental results showed that the proposed method can reduce the number of parameters in AlexNet by 24× with no accuracy loss. In addition, the compressed VGG16 and ResNet50 can achieve a more than 60× increased speed, which is significant.