A MIC-based acceleration model of Deep Learning

Abstract

In the era of Computational Intelligence developing rapidly, Deep Learning (DL) has gradually won acceptance from the world of Artificial Intelligence (AI) and it has been widely applied to the industry. However, the training of the network requires a considerable amount of time. For instances, the training of Convolution Neural Network (CNN) and Deep Belief Network (DBN) may take one week or even longer. Therefore, a new challenge has been put forward to the world of Artificial Intelligence, which demands decrease on the training time of Deep Learning algorithm effectively. And in this paper, we proposed a Deep Learning acceleration model based on MIC, which can reduce the training time significantly by using Restricted Boltzmann Machine (RBM) and Logistic Regression (LR). First, it conducts vectorization on the program, and then accelerates it by using the model we proposed in this paper. And the paper mainly consists of the design of the parallel model, which comprising data parallelism, model parallelism, a hybrid of data and model parallelism and so on. And experiments showed that the MIC-based acceleration model can reduce the training time to 1/10 of the original.