Investigating evolvable hardware classification for the BioSleeve electromyographic interface

We investigate the applicability of an evolvable hardware classifier architecture for electromyography (EMG) data from the BioSleeve wearable human-machine interface, with the goal of having embedded training and classification. We investigate classification accuracy for datasets with 17 and 11 gestures and compare to results of Support Vector Machines (SVM) and Random Forest classifiers. Classification accuracies are 91.5% for 17 gestures and 94.4% for 11 gestures. Initial results for a field programmable array (FPGA) implementation of the classifier architecture are reported, showing that the classifier architecture fits in a Xilinx XC6SLX45 FPGA. We also investigate a bagging-inspired approach for training the individual components of the classifier with a subset of the full training data. While showing some improvement in classification accuracy, it also proves useful for reducing the number of training instances and thus reducing the training time for the classifier.