Optimized Deep Learning Architectures with Fast Matrix Operation Kernels on Parallel Platform
Ying Zhang, Saizheng Zhang · 2013
In this paper, we introduce an optimized deep learning architecture with flexible layer structures and fast matrix operation kernels on parallel computing platform (e.g. NVIDIA's GPU). Carefully designed layer-wise strategies are conducted to integrate different kinds of deep architectures into a uniform neural training-testing system. Our fast matrix operation kernels are implemented in deep architecture's propagation processes. In our experiment, these kernels save 70% time on average comparing with the kernels in NVIDIA's CUBLAS library (widely used by many other neural network toolkits), and help our parallel deep architecture beats the neural structures using CUBLAS kernels in practical problems.