Deep learning (DL) has transformed the field of data analysis by dramatically improving the state of the art in various classification and prediction tasks. Especially in the area of computer vision and speech processing, DL has recently demonstrated better performance and generalisation properties, compared to classical machine learning approaches, which are based on the extraction of hand-crafted model-based features followed by classification. Hand gestures and speech constitute two of the most important modalities in human-to-human communication and man-machine interaction. In biomedical engineering, a lot of new work is directed towards electromyography-based gesture recognition. In this paper, we present a brief overview of DL methods for electromyography-based hand gesture recognition and then we select from literature a simple model based on Convolutional Neural Networks that we consider as the baseline model. The proposed modifications to the baseline model yield a 3% classification improvement. In the current paper, we concentrate on the explanatory analysis of this performance improvement. An ablation study identifies which modifications are the most important ones, and label smoothing is investigated to verify if the results can be improved by reducing a priori bias. The analysis helps in understanding the limitations of the model and exploring new ways to improve the performance.
|Title of host publication||Holzinger A., Pope A., Plácido da Silva H. (eds) Physiological Computing Systems. PhyCS 2016, PhyCS 2017, PhyCS 2018. Lecture Notes in Computer Science|
|Publication status||Published - 14 Aug 2019|