Designing Efficient NoC-Based Neural Network Architectures for Identification of Epileptic Seizure

Artificial Neural Networks (ANNs) mirror the analytical functions of human neural networks. The performance of smart healthcare systems has been limited to the increasing size and intricacy of information. Several ANN architectures help in the analysis of EEG signals for the identification of epileptic seizures. However, real-time performance needs to be accurate and very quick. Consequently, it is important to design efficient ANN models without compromising the feasibility of hardware realization. Since, CPUs and GPUs are based on conventional bus-system architectures, processing large complex datasets decreases the efficiency, scalability and versatility of the systems. To counter the bottlenecks of the bus-based architectures, Network-on-Chip has been efficient for complex computations. In this paper, we develop NoC-based feed-forward neural network and convolutional neural network models for the identification of epileptic seizure by analysis of continuously monitored EEG signal. The trained neural network models are mapped onto the Network-on-Chip to increase the throughput, power efficiency, parallelism and scalability of the architecture. The performance of all models is thoroughly explored in terms of throughput, energy, latency and identification accuracy of an epileptic seizure.