A GPGPU-Based Acceleration of Fault-Tolerant MLP Learnings

A method to speed-up the learning processes proposed in [1] using the GPGPU technology is proposed. The method in [1] is called the deep learning method (Deep LM), and a multilayer perceptron (MLP) which has successfully trained using the Deep LM is tolerant to multiple weight and neuron faults where the weight faults are between the hidden and output layers, and the neuron faults are in the hidden layer. The core part of the Deep LM is the traditional back-propagation (BP) algorithm. However, it takes so much time to learn by the Deep LM because of its additional processes to realize the fault tolerance. Furthermore, there is an additional process such that SECDED code is used to detect or correct the neuron faults in the output layer. To cope with this, the hot-spots in using the Deep LM and the outline of their CUDA C source codes, including the auto-tuning process in the hot-spots, are shown. Next, on using the Deep LM, the computing speeds using a GPU are compared with those using a CPU for the concrete examples of character recognitions. Then, it is shown that the former is extremely accelerated.