Code for my Master's Thesis at the Institute of Medical Informatics, Universität zu Lübeck.
I carried out studies on Multi-Task Learning using three approaches (Hard Parameter Sharing, Regularised Soft Parameter Sharing and Cross-Stitch Networks) on two time-series datasets with sensor-based data (OPPORTUNITY for Human Activity Recognition and DEAP for Emotion Recognition from EEG).
For a gentle introduction into Multi-Task Learning, I recommend this excellent paper by S. Ruder.
Here is the abstract from my thesis:
Multi-Task Learning (MTL) in the domain of Deep Neural Networks (DNNs) is an idea where a network performs multiple tasks at once to produce latent representations of the input that are more plausible than what is generated using classical Single-Task Learning (STL). In this way, many approaches to this concept have demonstrated results that surpass those using STL. However, the use of MTL approaches on sensor- based time-series data has not received much attention so far. This is unfortunate, as time-series data processing has an enormous range of potential applications, especially in the medical domain.
In this thesis, three very different approaches to MTL—Hard Parameter Sharing (HPS), Soft Parameter Sharing (SPS) and Cross-Stitch Networks (CSNs)—are applied to two different datasets from the domain of Human Activity Recognition (HAR) and Emotion Recognition—OPPORTUNITY and DEAP. We demonstrate that not every approach is equally beneficial. In particular, benefits were observed using HPS on both datasets and CSNs on only one dataset. We could not demonstrate that the chosen approach to SPS works for time-series. To aid further research in this field, our source code is made public.
The thesis is available in this repository.