A Priori Quantification of Transfer Learning Performance on Time Series Classification for Cyber-Physical Health Systems

Abstract

Cyber-Physical Systems fully encompass the intelligent system from signal acquisition through to physical computing and computation - it requires consideration of the deep entanglement between computational and physical elements. Human health and performance is increasingly being measured and analyzed using machine learning to identify complex relationships using wearable and pervasive computing. This combination defines the focused area of Cyber-Physical Health Systems. Modern deep learning algorithms, such as one-dimensional convolutional neural networks, have demonstrated excellent performance in classifying time series data because of the ability to identify time-invariant features. A primary challenge of deep learning for time series classification is the large amount of data required for training and many application domains, such as medicine, have challenges obtaining sufficient data. Transfer learning is a deep learning method used to apply feature knowledge from one deep learning model to another; this is a powerful tool when both training datasets are similar and offers smaller datasets the power of more robust larger datasets. This makes it vital that the best source dataset is selected when performing transfer learning and presently there is no a priori metric defined for this purpose. Analyzing time-series data from public human-activity-recognition datasets a neural network autoencoder was used to first transform the source and target datasets into a time-independent feature space. To quantify the suitability of transfer learning datasets the average embedded signal from each dataset was used to calculate the distance between each dataset centroid. Our metric was then applied to predict the success of transfer learning from one dataset to another for the purpose of general time series classification.