Year: 2023
Type: dissertation
Abstract: The rapid advancement of Artificial Intelligence, particularly in the realm of convolutional neural networks (CNNs), coupled with the favorable capabilities of Field Programmable Gate Arrays (FPGAs), has motivated this research project. The primary objective is to establish a foundational framework by designing an initial implementation of the Lenet5 neural network. In our pursuit of implementing the neural network, we carefully considered various alternatives. Given the nature of this initial project, we made a deliberate choice to develop the neural network from scratch. This approach enabled us to gain a deeper understanding of the intricate design principles underlying Convolutional Neural Networks (CNNs) and provided us with complete control over the hardware implementation. We designed the Lenet5 Neural Network, specifically designed for handwritten digit recognition, using VHDL and the Vivado tools. Concurrently, we developed a MATLAB simulation of the network to serve as a benchmark for comparing the results obtained from the hardware implementation. The accuracy of the system was found to be excellent, with a minimum confidence of 95.22% in digit recognition. However, further work is required to optimize resource utilization, as we have exceeded the capacity of the Zedboard FPGA, which was used for system simulation, particularly in terms of Look-Up Tables (LUTs) usage. The results obtained in this project provide a strong foundation for its continuation, showcasing the viability of the implemented system. These results also highlight three key areas to focus on in future developments. Firstly, there is a need for optimizing the bit utilization for coefficient and data representation within the network. Exploring more efficient methods of filter allocation outside the on-chip memory is also crucial. Lastly, the project opens up new avenues for implementing improved algorithms that can effectively reduce the latency of the system.
Author: Giménez Fernández-Caro, Andreu
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6982404833