An electroencephalogram (EEG) is a recording of brain activity. The accurate recognition of movement intentions from EEG signals is essential for achieving Motor imagery (MI)-based Brain-computer interfaces (BCIs). 

Among the rapid development of pattern recognition algorithms, various classification methods have been widely used for EEG recognition. Support matrix machine (SMM) showed unique advantage for it can grasp the structural information of feature matrices. 

Researchers from the Shenzhen Institutes of Advanced Technology (SIAT) of the Chinese Academy of Sciences proposed a novel deep architecture that used SMM as its basic module and random shift as the stacking element. The study was published in Computer Methods and Programs in Biomedicine. 

The novel deep architecture was named deep stacked support matrix machine (DSSMM). It was constructed in a layer-by-layer manner using SMM as its basic module, in which the random projections of weak predictions from the previous layers' SMM were combined with the original feature matrices to transform the manifold of single trial EEG data. 

In this way, DSSMM combined the virtue of SMM with the powerful feature representation derived from the deep architecture, making it suitable for EEG classification.

Besides, DSSMM involved an efficient feed-forward instead of parameter fine-tuning using back-propagation, where each layer was a convex optimization problem. 

Experimental results demonstrated that, compared with the other competitive methods, the DSSMM showed superior classification performance on three public EEG datasets and a self-collected EEG dataset.  

This was the first attempt to incorporate a matrix classification model into the deep architecture. The team will devote to improving the generalization capability of DSSMM in scenes with insufficient EEG data in the future.  

Flowchart of DSSMM for MI-based EEG classification (Image by SIAT)