Outstanding seizure detection algorithms have been developed over past two decades. Despite this success, their implementations as part of implantable or wearable devices are still limited. These works are mainly based on heavily handcrafted feature extraction, which is computationally expensive and is shown to be data set specific. These issues greatly limit the applicability of such methods to hardware implementation, including in-silicon implementations such as application specific integrated circuits. In this paper, we propose an integer convolutional neural network (CNN) implementation, Integer-Net, as a memory-efficient unified hardware-friendly CNN framework. The performance of Integer-Net is evaluated with multiple time-series data sets consisting of intracranial and scalp electroencephalogram (EEG) signals. Integer-Net shows a consistent seizure detection performance across three data sets: Freiburg Hospital intracranial EEG data set, Children's Hospital of Boston-MIT scalp EEG data set, and UPenn and Mayo Clinic's seizure detection data set. Our experimental results show that a 4-bit Integer-Net leads to only 2% drop of accuracy compared with a 32-bit real-value resolution CNN model, while offering more than 7 times improvement in memory efficiency. We discuss the structure of the integer convolution to improve the computational gain and reduce the inference time that are crucial for real-time application.
|Number of pages||9|
|Journal||IEEE Journal on Emerging and Selected Topics in Circuits and Systems|
|Publication status||Published - Dec 2018|
- Convolutional neural network
- Integer convolution
- Machine learning
- Seizure detection