Aspects of Artificial Intelligence Conversion from PC to Pumpkin boards

The objective of the paper is to present a way to develop and train a neural network in Matlab, in order to run it on a customized evaluation board. The essence of the development of this route is, on one hand, the fact that the customized Pumpkin board has associated software dedicated to running artificial neural networks on the Graphics Processing Unit - GPU, which gives a high speed of simulating neural networks thanks to the parallel processing feature and on the other hand, the Matlab environment has important resources for the development, characterization and optimization of artificial neural networks - ANN. The considered artificial neural network is of the “deep network” category, with fifteen layers and the data structure is specific to this category, being different from the “shallow network” category. The data presented for training and running the network are of image type with eight-bit greyscale representation and consists of digits from 0 to 9. The classification results obtained by running the neural network in Matlab and on the considered Pumpkin board are compared.