Neural Network Implementation and Analysis on Low-Power FPGA-based Devices

Abstract

With the evolution of powerful processing units, memory-intensive algorithms are able to provide better accuracy and superfast computing. Machine learning and AI-based platforms are creating new algorithms each passing day for better accuracy and superfast computing. However, on the other side in the event of increase number of IoT driven demand for mobile platforms have also increased manifold. Thus, on one side AI based neural network algorithms are in demand and on the other hand mobile computing hardware tools are exponentially evolving for especially IoT applications. This has led to a trade-off in the various application such as pattern recognition and classification. Therefore, algorithms are being developed so as to accommodate neural networks in low resource modules. A major challenge is trade-off between speed, accuracy, and memory usage. In the first stage, this dissertation aims for various edge detection algorithms on hardware platform. For this a Xilinx-based FPGA board was chosen for the purpose of experimentation. Filters such as Sobel, Prewitt, and Canny was implemented and various standard images were tested on image parameters on the behavior of the hardware and comparison with CPU. Secondly, the thesis also discusses various approaches that can be deployed for optimizing various CNN and QNN models with additionally available tools. The work is performed on the Xilinx Zynq 7000 series, Pynq Z2 board, which serves as FPGA-based low-power IoT device. The MNIST and CIFAR-10 databases are considered for simulation and experimentation. In this work, CNN, QNN and BNN-based pattern recognition techniques are implemented and analyzed on an FPGA. The FPGA hardware acts as an IoT device due to the connectivity with the cloud, and QNN and BNN are considered to offer better performance in terms of low power and low resource use on hardware platforms. This dissertation also implements and examines medical image processing on deep neural network. Here, attempt to perform medical image segmentation on a mobile platform based hardware is considered as an IoT device. We use two sets of medical data sets, HIS2828 and ISIC2017. We perform deep neural network (DNN), convolutional neural network (CNN), state vector machine (SVM), and neural network (NN). We perform simulation on CPU and also implement on FPGA using PYNQ-Z2, Zynq 7000 board. The primary aim is to detect any type of cancer or tumor in the medical images. Finally, an unsupervised learning methodology has been implemented and analyzed for medical image processing, i.e., fuzzy c means (FCM) for medical image processing. FCM is one of better algorithm in terms of other unsupervised algorithm. The work is done on brain image segmentation and uses clustering method for classification diagnosis.