Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Strong and Simple Deep Learning Baseline for BCI MI Decoding

Yassine El Ouahidi, Vincent Gripon, Bastien Pasdeloup, Ghaith Bouallegue, Nicolas Farrugia, Giulia Lioi

TL;DR

The paper introduces EEG-SimpleConv, a lightweight yet strong baseline for motor imagery decoding using a simple 1D CNN with standard ingredients. It demonstrates competitive or superior performance across four open MI datasets, with particularly strong cross-subject transfer when augmented with Fine-Tuning, and shows favorable inference times that support online applicability. Through extensive ablations, Mixup and a Batch Normalization trick emerge as key contributors to robustness, while minimal preprocessing and EA-based normalization maintain data fidelity across subjects. The work argues for using off-the-shelf components to improve reproducibility and adoption of deep learning in BCI, and it releases code to enable benchmarking and further development.

Abstract

We propose EEG-SimpleConv, a straightforward 1D convolutional neural network for Motor Imagery decoding in BCI. Our main motivation is to propose a simple and performing baseline to compare to, using only very standard ingredients from the literature. We evaluate its performance on four EEG Motor Imagery datasets, including simulated online setups, and compare it to recent Deep Learning and Machine Learning approaches. EEG-SimpleConv is at least as good or far more efficient than other approaches, showing strong knowledge-transfer capabilities across subjects, at the cost of a low inference time. We advocate that using off-the-shelf ingredients rather than coming with ad-hoc solutions can significantly help the adoption of Deep Learning approaches for BCI. We make the code of the models and the experiments accessible.

A Strong and Simple Deep Learning Baseline for BCI MI Decoding

TL;DR

The paper introduces EEG-SimpleConv, a lightweight yet strong baseline for motor imagery decoding using a simple 1D CNN with standard ingredients. It demonstrates competitive or superior performance across four open MI datasets, with particularly strong cross-subject transfer when augmented with Fine-Tuning, and shows favorable inference times that support online applicability. Through extensive ablations, Mixup and a Batch Normalization trick emerge as key contributors to robustness, while minimal preprocessing and EA-based normalization maintain data fidelity across subjects. The work argues for using off-the-shelf components to improve reproducibility and adoption of deep learning in BCI, and it releases code to enable benchmarking and further development.

Abstract

We propose EEG-SimpleConv, a straightforward 1D convolutional neural network for Motor Imagery decoding in BCI. Our main motivation is to propose a simple and performing baseline to compare to, using only very standard ingredients from the literature. We evaluate its performance on four EEG Motor Imagery datasets, including simulated online setups, and compare it to recent Deep Learning and Machine Learning approaches. EEG-SimpleConv is at least as good or far more efficient than other approaches, showing strong knowledge-transfer capabilities across subjects, at the cost of a low inference time. We advocate that using off-the-shelf ingredients rather than coming with ad-hoc solutions can significantly help the adoption of Deep Learning approaches for BCI. We make the code of the models and the experiments accessible.
Paper Structure (40 sections, 7 figures, 9 tables)

This paper contains 40 sections, 7 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Convolutionals Neural Networks (CNN) architectures
  4. 1D CNN
  5. 2D CNN
  6. Other types of Neural Networks
  7. Recurrent Neural Networks and Transformers
  8. Riemannian geometry based Neural Networks
  9. More complex architectures
  10. contributions
  11. EEG-SimpleConv architecture
  12. Key ingredients of the training routine
  13. EEG Signal Preprocessing
  14. Normalisation
  15. Regularization
  16. ...and 25 more sections

Figures (7)

  • Figure 1: EEG-SimpleConv architecture
  • Figure 2: Evaluation paradigms under a LOSO approach
  • Figure 3: TSNE Embeddings of the Sujbect 2 of BNCI
  • Figure 4: Impact of the number Fine-Tuning epochs on the classification performance on BNCI using the C.S.+F.T. paradigm
  • Figure 5: TSNE Embeddings. From top to bottom: (a,b) the "hardest", (c,d) the "average" and (e,f) the "easiest" BNCI subject to classify
  • ...and 2 more figures