Comparative Performances of CNN Implementation for EEG Signal Analysis on Embedded CPU Boards

The aim of this study is to give elements to predict time performances and power consumption performances in CNN implementation on embedded CPU boards. CNNs are effective in classifying EEG signals but are computationally intensive, which leads to increased energy consumption. To achieve this objective, we compare the implementation of a trained convolutional neural network (CNN) for epilepsy detection, analyzing EEG signals on some resource-constrained embedded platforms commonly used in IoT applications. We evaluate the ease of deployment and the time and energy performance metrics across four different CPU embedded boards to guide and anticipate future implementations of CNNs on embedded solutions. Depending on the board, (ARM based, 32 bits / 64 bits architecture / Multicore) the execution time performances vary from 10 time more to 10 time less compare to a standard Laptop. Obviously, the power consumption varies accordingly. Multicore platforms require also specific tools and we evaluate the benefits of it compared to a standard approach.