Techniques for Compressing Deep Convolutional Neural Network

摘要

Deep Convolutional Neural Network (CNNs) has evolved to be a progressive technique for both artificial intelligence and computer vision applications. However, these models are also known for being large, computationally expensive and requires lot of memory therefore its hard to implement them on various embedded systems or mobile phones as they have limited hardware resources, low power budget and strict latency requirements. This has motivated researchers to develop various techniques to achieve the goal of model compression of deep CNNs and also to simultaneously optimize its performance. In this paper, recent bench marking advancements in the techniques of pruning, quantization and coding are systematically explored which are mostly implemented on the top of convolutional neural networks with convolutional layer and fully connected layer to provide a compressed network. Also the challenges and future directions are discussed to reduce the model size.