Title: Mitigating Data Shortage in Biomedical Signal Analysis: An Investigation into Transfer Learning and Generative Models
Presenter: Xiaomin Li
Advisor: Vangelis Metsis
Date/Time: Thursday, June 8th, 10 a.m. to 12 p.m.
Location: Comal 212 Zoom link: https://txstate.zoom.us/j/84498592850
Title: Mitigating Data Shortage in Biomedical Signal Analysis: An Investigation into Transfer Learning and Generative Models
Abstract:
This dissertation provides innovative strategies to address data scarcity in deep learning applications, focusing specifically on biomedical signal data. These strategies aim to enhance the effectiveness of deep learning models and consequently contribute to advancements in biomedical signal analysis and synthesis.
We first introduce a versatile transfer learning approach that capitalizes on the generalizable features extracted from large time-series datasets to improve the performance on smaller target datasets. This approach incorporates a range of supervised, unsupervised, and self-supervised methodologies to train feature extractors, thereby mitigating the issues associated with incompatible configurations among various datasets.
Subsequently, we utilize generative adversarial networks (GANs) to synthesize signal data and expand the original datasets. We introduce two novel transformer-based GAN models, TTS-GAN and BioSGAN, which can generate multi-dimensional and arbitrary-length time-series data. The fidelity and performance of these GAN model-generated data are thoroughly evaluated using robust qualitative and quantitative metrics, further contrasting them with the current state-of-the-art time-series GANs.
In addition, we delve into the potential of Denoising Diffusion Probabilistic Models (DDPMs) as an alternative to GANs for generating biomedical signals. We discover that DDPMs produce synthetic signals of higher quality compared to their GAN counterparts. We extend our exploration to conditional DDPMs, revealing their applicability in tackling challenges such as signal denoising, imputation, super-resolution, and individual subject generation.
In summary, this dissertation offers a suite of innovative solutions to overcome the constraint of data shortage and bolster the performance of deep learning models for biomedical signal data. The established transfer learning strategy, the development of transformer-based GANs, and the insightful exploration of DDPMs collectively enrich our understanding and address crucial challenges in biomedical signal processing.
Deadline: July 1, 2023, midnight