Deep Learning for Time-Series Classification Enhanced by Transfer Learning Based on Sensor Modality Discrimination

Progress in hardware development has caused wearable devices to become pervasive in our daily lives. Their ability to passively collect time-series data has led to an increasing overlap between Ubiquitous computing (Ubicomp) and machine learning, making it common to translate an Ubicomp application into a classification problem. This thesis focuses on time-series classification via two main axes: feature extraction and deep transfer learning.

Feature extraction is nowadays mainly divided into two categories: feature engineering and feature extraction based on deep learning. The thesis firstly attempts to verify whether deep feature learning convincingly outperforms feature engineering like for image classification. Transfer learning refers to the transfer of knowledge from a source to a target domain to improve classification performances on the latter. It has shown to consistently enhance deep feature learning for image classification, but remains under investigation for time-series. The thesis secondly proposes a new deep transfer learning approach transferring features learned by sensor modality classification on a source domain containing diverse types of time-series data.

Experiments carried out for various Ubicomp applications (human activity, emotion and pain recognition) show that deep feature learning is not always the best option for time-series feature extraction, and that the proposed deep transfer learning approach can consistently enhance deep feature learning.