DNN pruning and mapping on NoC-Based communication infrastructure

Machine learning algorithm-based applications have been deployed for supporting the intemet of things (IoT) and web search engines without losing accuracy in order to satisfy human requests. Developments in deep learning-based applications and complexity of machine learning algorithms increase the depth of artificial neural networks (ANN). Increasing depth of neural network (NN) is challenging regarding the delay, energy consumption, learning, and inference speed up. We train a deep neural network (DNN) gradient descent-based method based on two Booth and Matyas standard generating functions. We also propose a method for pruning weights, neurons, and layers of DNNs based on minimal distance error before and after pruning in a range of safety margin error. This paper employs a new elastic dataflow and DNN mapping on the mesh topology for decreasing delay and energy consumption. Simulation results show reducing the delay and energy consumption of training and inference phases by approximately 22.56%-77% and 65.94%-88.54% compared with not employing a DNN pruning.