Running Convolutional Neural Network on tiny devices

Abstract

The introduction of deep neural network architectures has resulted in an interesting development in artificial intelligence. These architectures demonstrate how useful they are in a variety of fields, such as the classification of images, the detection of objects, and the recognition of speech. However, this efficiency is a result of sophisticated networks with very high requirements for computational power, memory space, and energy, which makes it difficult to use such models on devices with less powerful hardware. Within the scope of this study, we enable offline CNN (Convolutional Neural Network) inference on smartphones operating in real time. The following is the procedure for deployment: 1) The development of an architecture for a convolutional neural network 2) Teaching the model to perform on a reference dataset from the ground up. 3) Reduce the size of this model as well as the total amount of computation it requires while still striving to keep at least the same level of precision as the first version, since it is expected that the accuracy after optimization become lower than the baseline. 4) Deploy the improved architecture and carry out the inference process in real time on a mobile device, preferably a smartphone. The models VGG16 and ResNet that were trained on the Cat vs Dogs dataset are the ones that have been implemented.