Future trend of deep learning frameworks — From the perspective of big data analytics and HPC
Takuya Araki, Yuichi Nakamura · 2017
Deep learning is becoming more and more important in many areas including signal processing, because of its high recognition accuracy and interesting applications built on top of the models. There is no doubt that deep learning frameworks played an important role for its popularization; they made it easy to utilize deep learning by hiding the detailed implementation of the algorithms including utilization of accelerators like GPU. On the other hand, there have been Big Data analytics frameworks for classical machine learning like Hadoop and Spark. In addition, there also have been many researches in the field of HPC middleware, whose workload is similar to that of deep learning. In this paper, we discuss such frameworks and middleware including their targeted workloads and tradeoffs, and also discuss the future trend of deep learning frameworks including our preliminary proposal.