۱- Introduction

۲- Methods

۳- Datasets

۴- Results

۵- Discussion

۶- Conclusions

References

Abstract

We introduce two Convolutional Neural Network (CNN) classifiers optimized for inferring brain states from magnetoencephalographic (MEG) measurements. Network design follows a generative model of the electromagnetic (EEG and MEG) brain signals allowing explorative analysis of neural sources informing classification. The proposed networks outperform traditional classifiers as well as more complex neural networks when decoding evoked and induced responses to different stimuli across subjects. Importantly, these models can successfully generalize to new subjects in real-time classification enabling more efficient brain–computer interfaces (BCI).

Introduction

Deep Neural Networks have revolutionized many domains such as image recognition and natural language processing. To date, their application in the analysis of electro- and magnetoencephalographic (EEG and MEG) data has been limited by several domain-specific factors. First of all, electromagnetic brain signals are characterized by very low signal-to-noise ratio (SNR). Here, the term “noise” is understood widely and includes external interference, physiological (e.g. cardiac or oculomotor) artifacts as well as background brain activity unrelated to the studied phenomena. SNR in single-trial EEG and MEG measurements is typically assumed to be < 1 for evoked responses and
۱ for oscillatory activity, which puts these data to stark contrast with those in traditional applications of deep learning. Typical EEG/MEG analysis employs a wide range of techniques to increase the SNR, e.g. spatial and temporal filtering, averaging a large number of observations, sourceseparation algorithms, and other complex feature extraction methods (e.g. wavelet transform). Thus, efficient noise suppression is required for high-accuracy classification of EEG/MEG signals. Second, these data have a complex, high-dimensional spatiotemporal structure. Modern EEG and MEG systems comprise several hundreds of sensors capable of sampling brain activity with sub-millisecond temporal resolution. In the case of MEG, these sensors may also measure different components of the neuromagnetic field. On one hand, this multitude of data points enables sophisticated analysis methods to extract finer details of brain function. On the other hand, manual analysis and interpretation of these data become increasingly complex and time-consuming. Machine-learning algorithms can be of great help in such tasks but the mere classification result is often not sufficient; ideally, the experimenter should understand why the algorithm is able to classify the data, i.e., the learned model should be interpretable in neurophysiological terms. A model able to reliably identify those neural sources that contribute to the discrimination between given experimental conditions could enable efficient exploitative analysis of these complex data sets and ultimately allow more complex experimental designs. Finally, deep-learning models require large numbers of training samples to perform optimally. In a typical EEG/MEG or BCI experiment, however, time constraints and data acquisition costs limit the sample sizes severely. Open datasets usually comprise data from a large number of individuals, providing a promising strategy to overcome this limitation, but in this case, the classifier needs to be robust to high interindividual variability stemming from differences in cortical anatomy, mappings of function to structure and physiological state. Taken together, these factors may easily lead to over-fitting (especially in more complex models) and poor interpretability of the findings. To address these challenges, we propose a Convolutional Neural Network (CNN) whose architecture is based on a generative model of non-invasive electromagnetic measurements of the brain activity (Daunizeau and Friston, 2007).