An evolvable block-based neural network architecture for embedded hardware

Abstract

Evolvable neural networks are a more recent architecture, and differs from the conventional artificial neural networks (ANN) in the sense that it allows changes in the structure and design to cope with dynamic operating environments. Blockbased neural networks (BbNN) provide a more unified solution to the two fundamental problems of ANNs, which include simultaneous optimization of structure, and viable implementation in reconfigurable embedded hardware such as field programmable gate arrays (FPGAs) due to its modular structure. However, BbNNs still have several outstanding issues to be resolved for an effective implementation. An efficient hardware design can only be obtained with proper design consideration. To date, there has been no previous work reported on BbNNs configured in recurrent mode for complex case studies, even though it is theoretically possible. Existing BbNN models do not explicitly specify or model the latency of the system, determine how it affects the system, nor how it can be optimized. Also, current methods of training BbNNs using genetic algorithm (GA) are slow, especially with large training datasets. This thesis presents an improved BbNN model, proposes a state-of-the-art simulation and co-design environment for it, and implements it on a hardware platform for improved speed and performance. It has a novel architecture with deterministic outputs that can evolve and operate in both feedforward and recurrent modes. The BbNN is redesigned for optimal system latency to achieve higher performance, and supports onchip training for multi-objective optimization using a multi-population parallel genetic algorithm. All the algorithms proposed led to an efficient and scalable hardware implementation. The viability of the resulting BbNN system-on-chip (SoC) is proven with real-time performance analysis of real-world case studies, where performance improvements of up to 410� are observed. The hardware logic utilization is minimized with the help of theoretical analysis and design considerations. A case study requiring the use of recurrent mode BbNN is also presented. All case studies tested with the BbNN give equivalent or better classification accuracies compared to those provided in previous works, but with optimized latency values. As an example, the proposed BbNN solution achieves a classification accuracy of 99.41% for the heart arrhythmia case study, which is an improvement over previous work. The validity of the proposed BbNN model is thus verified.