Cosine convolutional neural network and its application for seizure detection.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Signal Processing

Background:

• Traditional Convolutional Neural Networks (CNNs) face challenges with high memory consumption and kernel redundancy.
• These limitations hinder real-time, low-power applications like seizure detection systems.
• Overfitting is a common issue in conventional CNNs, impacting performance in resource-constrained environments.

Purpose of the Study:

• To introduce a novel Cosine Convolutional Neural Network (CosCNN) for efficient seizure detection.
• To develop a post-training quantization algorithm for hardware implementation of CosCNN.
• To demonstrate the real-time and low-power capabilities of the proposed seizure detection system.

Main Methods:

• Developed Cosine Convolutional Neural Network (CosCNN) using robust cosine kernels with two learnable factors.
• Proposed a post-training quantization algorithm utilizing a cosine lookup table and KL-divergence.
• Designed a configurable cosine convolution accelerator on Field Programmable Gate Array (FPGA).
• Deployed the quantized CosCNN on Zedboard for real-time seizure detection.

Main Results:

• CosCNN achieved a nearly 75% reduction in memory cost post-quantization with minimal accuracy loss.
• The FPGA-based CosCNN accelerator demonstrated real-time and low-power operation.
• Evaluations on Bonn and CHB-MIT epileptic EEG databases confirmed CosCNN's superior performance.
• CosCNN outperformed traditional CNNs and other existing seizure detection methods.

Conclusions:

• The proposed CosCNN offers a highly efficient and accurate solution for seizure detection.
• Quantization and hardware acceleration enable real-time, low-power deployment on edge devices.
• CosCNN represents a significant advancement for wearable and accessible epilepsy monitoring systems.