Energy Efficient Convolutional Neural Networks for EEG Artifact Detection

Author/Creator ORCID




Citation of Original Publication


This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Public Domain Mark 1.0
This work was written as part of one of the author's official duties as an Employee of the United States Government and is therefore a work of the United States Government. In accordance with 17 U.S.C. 105, no copyright protection is available for such works under U.S. Law.


This paper proposes an energy e cient Convolutional Neural Network based architecture for detecting different types of artifacts in multi-channel EEG signals. Our method achieves an average artifact detection accuracy of 74% and precision of 92% across seven different artifact types which outperforms existing techniques in terms of classification accuracy as well as the more common ICA based solution in terms of computational complexity and memory requirements. We designed a minimal neural network processor whose Verilog HDL is configurable for implementing 2ⁿ processing engines (PEs). We deployed our CNN on the processor, placed and routed on Artix-7 FPGA and examined different number of PEs at different operating frequencies. Our experiments indicate that utilizing 4 PEs operating at a clock frequency of 11.1 MHz is the optimal configuration for our hardware to yield the least classification energy consumption of 32 mJ accomplished in the maximum allowed prediction time of 1 Sec. We also implemented our CNN on TX2 NVIDIA Jetson platform and, by tweaking the CPU and the GPU frequencies, explored the minimum power and energy configuration. Our FPGA results indicate that the 4-PE implementation outperforms the low power configuration of TX2 by 65x in terms of power, and the low energy configuration of TX2 by 2x in terms of energy per classification. Our CNN-based FPGA implementation method also outperforms the ICA method by 11x in terms of energy consumption per classification.