Personal tools

An Energy Efficient Brain-Computer Interface using Mr.Wolf

From iis-projects

Revision as of 11:09, 21 July 2020 by Herschmi (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search


A brain–computer interface is a device that enables communication and control without movement. The device aims to recognize the human’s intentions from spatiotemporal neural activity typically recorded by a large set of electroencephalogram (EEG) electrodes. What makes it particularly challenging, however, is its susceptibility to errors in the recognition of human intentions. To cope with high variability of EEG data, today’s BCI require highly sophisticated classification algorithms. Traditional MI-BCI architectures are divided into feature extraction and a subsequent classifier. EEG signals are typically pre-processed using tunable spectral and spatial filters followed by log-energy feature extraction, with Riemannian covariance matrices [1] achieving the highest classification accuracy of 75.47% on the BCI competition IV-2a dataset. Alternatively, the feature extractor and classifier can be combined and trained simultaneously with a convolutional neural network (CNN). While being successful in image classification, CNNs are gaining attention in MI-BCIs as well [2]. One of the most popular and compact CNNs is EEGNet [3] achieving an accuracy of 72%.

This thesis analyses the feasibility of operating both EEGNet and Riemannian approaches on Mr.Wolf [4], an embedded platform featuring an 8-core fixed-point cluster for efficient computing. After a prior quanitzation analysis of weights and features from floating-point to fixed-point representation, the classifiers are ported to Mr.Wolf.

Status: In progress

Tibor Schneider

Supervision: Michael Hersche, Xiaying Wang, Lukas Cavigelli


  • Deep Learning
  • Python (preferably Pytorch)
  • Embedded C


20% Theory
80% Software Development


Luca Benini

↑ top


  • [1] Hersche et al., Fast and Accurate Multiclass Inference for MI-BCIs Using Large Multiscale Temporal and Spectral Features
  • [2] Schirrmeister et al., Deep learning with convolutional neural networks for EEG decoding and visualization
  • [3] Lawhern et al., EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces
  • [4] Pullini et al., Mr. Wolf: A 1 GFLOP/s Energy-Proportional Parallel Ultra Low Power SoC for IOT Edge Processing