Program for Tuning Hyper Parameters of Neural Networks for Image Recognition on Mobile Devices

Neural network had become a very popular technology lately. With extended usage of this technology a new challenge had come adapting neural network for mobile devices. It became a challenge because, despite the fact that computational power of mobile devices constantly grows, computational power required for tasks of machine learning requires even more power. I happens because most of the tasks are computer vision and image processing tasks and resolution of pictures and photos taken by those phone are constantly improving. And, with single image, image processing can be done I cloud. However with real-time video processing is not possible, because of internet speed. So, it is required to process those images locally, with a minimum speed of 30 images per second. Lottery ticket hypothesis is a hypothesis that tells us that every neural network has a sub net inside of it that has the same accuracy that the base neural network. We can determine architecture of this sub network, however, if we decide to train this network from ground up, this neural network would not yield the same results as the base neural network. This is most likely happening due to random nature of neural network during initial initialization. However, the problem remains: how should we optimize neural networks, making them smaller to reduce required computing power? Neural network pruning is a an activity that disposes of spare neurons in neural network. It helps in optimization of neural network, however traditional pruning techniques are not optimal. Spare neurons that are being deleted from the final network are still contributing to the output, therefore final network won’t be as accurate as the base network and won’t be trainable further. In this work we present an algorithm that can create accurate and trainable neural networks. For this task we used a genetic algorithm and achieved the mentioned result. Keywords: neural network; pruning; genetic algorithm;