Scalable Prototype Learning Using GPUs

Prototype learning is widely used in character recognition field. Unfortunately, current learning algorithms require intensive computation burden for large category applications, such as Japanese/Chinese character recognition. To resolve this challenge, a principled parallel method is proposed on GPUs instead of CPUs. We have implemented the method in mini-batch manner as well as stochastic gradient descent (SGD) manner. Our evaluations on a Chinese character database show that our method posses a high scalability while preserving its performance precision. Up to 194X speedup can be achieved in the case of mini-batch. Even to the more difficult SGD occasion, a more than 30-fold speedup is observed.