Knowledge-Centric AI for Learning and Reasoning in Space and Time: Applications in Computational Sustainability
Access to this document is restricted. Some items have been embargoed at the request of the author, but will be made publicly available after the "No Access Until" date.
During the embargo period, you may request access to the item by clicking the link to the restricted file(s) and completing the request form. If we have contact information for a Cornell author, we will contact the author and request permission to provide access. If we do not have contact information for a Cornell author, or the author denies or does not respond to our inquiry, we will not be able to provide access. For more information, review our policies for restricted content.
Deep learning holds great potential to help with specialized scientific tasks, such as forecasting crop yields, monitoring vegetation health, and modeling complex Earth systems such as the carbon cycle. However, deep learning often requires massive amounts of labeled data, which is rarely available for these problems. Rather, labels may be scarce, coarse-resolution, or limited to specific variables or regions. In this dissertation, I overcome this weak supervision through knowledge-centric AI, which combines the power of data-driven learning with established scientific knowledge about the problem's structure. This knowledge informs which inductive biases and constraints should be built into the model. For example, in crop yield forecasting, each datapoint (a location in a given year) may have thousands of input features, but only a single numeric label. It is challenging to disentangle the contributions from each time-period and variable within the year, without overfitting to the small number of labels. I thus design architectures leveraging spatial smoothness and locality inductive biases (learning features from temporal neighborhoods) to learn generalizable relationships. In vegetation monitoring from satellite images, our labels are only available at a coarse resolution, but I propose a method to predict at finer resolutions through locality constraints and smoothness regularization. In carbon cycle modeling, I embed physical process-based equations into our architecture, combining prior scientific knowledge with interpretable data-driven learning into ScIReN, a fully-transparent framework that discovers new insights. Overall, these works demonstrate the potential of overcoming weak or limited supervision by injecting prior knowledge into deep learning, expanding the potential of deep learning to address challenging predictive tasks in the sustainability sciences.