Dissertation Presentation

 

This dissertation explores several ways to relieve the data shortage problem and improve the performance of deep learning models on signal data recorded in the form of time series. First, we develop a transfer learning methodology to transfer the general features learned from bigger time-series datasets to new domain smaller datasets. We have explored supervised, unsupervised, and self-supervised methods to train such feature extractors. Our approach also solves the problem of configuration incompatibilities between datasets. Second, we leverage generative adversarial networks (GANs) to generate synthetic signal data so as to expand the size of original datasets. We have introduced two transformer-based GAN models, TTS-GAN, and TTS-CGAN, which can generate multi-dimensional and arbitrary-length time-series data. We use several qualitative and quantitative evaluation metrics to demonstrate the fidelity of our GAN model-generated data, and compare their performance with many other state-of-the-art time-series GANs.

 

To conclude the work of this dissertation, we will continue to explore GAN solutions to improve the performance of deep learning applications in this field. We will adopt the idea of Style Transfer GAN, which is booming in the image processing field, to solve the data imbalance issue that commonly exists in health-related signal datasets. Furthermore, we will tackle the problem of information loss during the process of signal noise filtering. We will develop super-resolution GAN models to artificially increase the sampling rate of a signal, and generate a high-resolution signal, from low-resolution input samples, that maintains the properties of the original signals.