high performance computing on graphics processing units: hgpu.org

The goal of this thesis is to design and implement a hyper neural network that has a topology with limited number inputs of individual neurons and uses genetic programming as the learning algorithm. Parallelization of this neural network is done with use of OpenCL standard which allows running it on wide range of devices. From the learning algorithms two methods are presented and implemented – a method that performs genetic evolution of trees of simple action and a method implements the idea of search in the compressed space and uses a matrices of coefficients of discrete cosine transformation. At the end the performance of the block neural – cornerstone of this network – is tested.