Interpretable Time Series Representation Learning via Disentanglement and Domain Priors

点击翻译按钮获取中文摘要

The World Health Organization projects a shortage of 9.9 million medical providers by 2030. This shortage is especially alarming considering the simultaneous increase in the world population above the age of 65, with elderly Canadians estimated to comprise 23% of the national population by 2050. Compounded by the onslaught of climate change with increased extreme weather events, droughts, flooding, and fire seasons, leading to the loss of lives and livelihoods across the globe, healthcare and climate change pose some of the greatest challenges that humans have ever faced during their time on the planet. Although different at the surface, both healthcare data (e.g., patient medical history) and climate change data (e.g., greenhouse gas emission data) comprise information (data samples) evolving over time, formally known as "time-series data". Consequently, the analysis of patterns in time-series data is critical to the detection, monitoring and prediction of events to improve healthcare outcomes and develop strategies to mitigate climate change. My research program will build automated artificial intelligence (AI) models with temporal reasoning capabilities to tackle critical challenges in these application areas, advancing the state-of-the-art in time-series analysis. My research program will help scale healthcare services by developing AI-based automated assistive healthcare solutions for training, monitoring, diagnostics, and decision-making, facilitating the provision of quality healthcare to remote and marginalized communities. In the context of climate change, my program will help industries meet emission goals by developing automated monitoring and forecasting techniques that can detect and analyze greenhouse gas emissions, one of the primary causes of climate change. These solutions will also be used to model and forecast the dynamics of extreme weather events and forest fire spread that can be used to minimize damage to biodiversity and loss of life. The program will address these pressing real-world problems by developing bio- and physics-inspired models to provide meaningful ways to extract patterns from time-series. The conceptual understanding of time-series data developed as part of this research program will make fundamental advances to enable reliable application of AI in such critical areas via added interpretability (and ultimately trustworthiness) and fairness. Additionally, my program will train students to develop expertise in machine learning for real-world applications, producing graduates with high-demand skills in AI and time-series analysis. These students will develop in-depth understandings of both the theoretical underpinnings and the practical implications of AI to drive the next wave of responsible innovation in Canada via research, entrepreneurship, and engineering.