Investigating the Application of Transfer Learning to Neural Time Series Classification

Abstract

Current approaches to EEG time series classification depend heavily on feature engineering to support the training of classifiers based on generalized linear models, decision trees, neural networks, or other machine learning techniques. This feature engineering demands considerable competence in mathematics, digital signal processing, statistics, linear algebra, etc. Researchers generating these time series often have clinical backgrounds, and may not be in a position to design and extract these features by hand. However, they are likely to be familiar with rudimentary - but fundamental - data visualisation methods. The objective of this paper is to investigate whether the application of transfer learning to such a classification problem can facilitate the replacement of involved feature engineering with straightforward data visualisation. While a classification accuracy of over 80% is achieved, the trained neural network exhibits the hallmarks of overfitting. We suggest alternative data visualisation techniques and modifications to the transfer learning method employed that may yield better results for multichannel neural time series data.