Efficient implementation of convolutional neural networks on embedded devices

In the field of artificial intelligence, deep convolutional neural network models are very popular because they can yield results close to those of humans. Depending on the application, these deep learning models can be very simple and small, but also very complex and large. Hence, the performance of an embedded systems that implement these models may be poor and infeasible. Through the use of various methods, this thesis aims to improve deep convolutional neural architecture efficiency without a significant loss of in the performance. For this purpose, we first utilize feature dimension reductions in layer activations. We use methods such as Principal Component Analysis and Select K-Best functions for feature dimension reduction. In the following, in order to make a quantization-aware trained binary deep convolutional neural network model more efficient, we also utilize the "Regular Positive and Negative Inference" algorithm by replacing the fully connected layers of the deep learning model as a decision-making mechanism. The ultimate aim of this thesis is to observe if these methods would make our models efficient without a significant loss of performance, and if we can further increase the efficiency of a binary quantized deep convolutional neural network.