Heterogeneous Accelerated Deep Learning on Spark

Finally. My paper report for my Statistical Machine Learning class. main


As a branch of machine learning, Deep Learning is already becoming a promising approach for developing more sophisticated models for different intelligent applications. Recent deep learning activities such as image recognition, speech categorization, and automatic machine translation have shown significant accomplishments in training the models. But the enormous computation and data input are still a huge burden to a single machine. Recent deep learning framework Tensorflow released a distributed version, but it still cannot provide quality of service (QoS) or utilize integrated accelerators efficiently such as GPU and FPGA to accelerate the computation.

In this work, we designed and implemented a distributed Deep Learning library based on the big data framework Spark, which is widely adopted in the distribution system area. We also want to integrate the accelerators such as GPU and FPGA to show how complex neural networks applications can be speed up. In the evaluation part, we show that the distributed deep learning library improve the performance 1.97 times with two nodes and 2.93 times with three nodes.