Transfer Learning to Address Subject Based Variability in Surface Electromyogram (SEMG) Signals
The emergence of inexpensive and unobtrusive physiological sensors has widened the application of physiological sensing to newer and innovative areas including human-computer interface, proactive human health monitoring, emotion recognition, activity recognition and many more such areas. Wide application of these sensors over multiple subjects poses a great challenge to traditional machine learning algorithms due to high variabilities in physiological signals across subjects. Subject-based variability, inherent in physiological signals such as SEMG signals from the human muscles, make the task of developing classification frameworks that model these signals, a very complex challenge. Variations in physiological signals across multiple subjects give rise to differences in data distribution between the training and test data, thus deteriorating the performance of traditional machine learning algorithms.
As a part of this project, we develop transfer learning or domain adaptation frameworks that address the problem of distribution difference between the training and test data so as to be able to develop classification frameworks for detecting muscle fatigue from SEMG signals from a test subject, using the available data from several other subjects.
The optimization frameworks can be adapted to suit any application having the challenge of distribution difference between the test and training data.