A Parallel Optimization Algorithm Of Deep Learning On Heterogeneous Platform

As the time for the era of big data,deep learning algorithm ushered anoth er development.Nowaday,deep neural networks has gradully become the rease arch focus in the acadimia and industry.The recognition accurancy of the deep neural networks model is far more than that of the tranditional macheine learni ng algorithm.The deep belief networks is one of the well-known deep learning models.The deep belief networks is a deep nerual networks based on unsuperv ised learning concepts,presented by Professor Geffrey Hinton.At present,the deep belief network is widely used in image recognition,character recognition and other fields,and have achive good accuracy.However,because of the High-dimensinal and large amount of traning dat a,traning a DBN in tranditional algorithm always cost a lot of time.With the continuous development of computer hardware technology,the use of heterogen eous computing resources for program algorithm for parallel optimization has b ecome one of the main method in the field of parallel computing.In this pape r,We propose a paralle optimization algorithm based on heterogeneous platform s to improve the DBN training efficiency.Firstly,In order to optimize the DBN parallel optimization algorithm,we proposed a serial optimization algorithm based on single-core CPU platform.A good serial optimization algorithm is the basis of the parallel optimizition algor ithm.In order to facilitate the optimization of algorithm performance,We use C programming language achieve deep belief network.Meanwhile,we use BL AS to optimize matrix manipulation and use loop unrolling+matrix manipulation to optimize loop structure to solve performance bottlenecks.Compared with the unoptimized serial algorithm,deep belief networks serial optimization algorithm achieved 2x speedup.Secondly,In order to solve solve coarse-grained task module parallel algor ithm design difficulties and heterogeneous platform computing resource allocatio n problems.We propose a coarse-grained task decomposition strategy and a su b-task scheduling strategy.Through this strategy,to making optimize subtask m ore convenience and loading balance tasks.Finally,In order to solve the problem of slow deep belief network trainin g training,we propose DBN parallel optimization algorithm based on heterogeneous platform.Based on the task decomposition strategy and the sub-task sched uling strategy,the method of data parallelism and task parallelism.We use MN IST handwritten data set.Compared with the unoptimized serial algorithm,we achieve 10 x speedup;compared with serial algorithm,we achieve 4.5x speedup.