Iterative Models for Machine Learning in the Physical Sciences

Barger, Maria

Etd

Iterative Models for Machine Learning in the Physical Sciences

Öffentlich Deposited

Machine learning models are often applied to problems in the physical sciences yet such problems are often restricted by physical constraints or suffer from a lack of large data traditionally needed to train accurate machine learning models. In this dissertation, we propose to make use of iterative autoencoders, a family of autoencoders designed to produce both reconstructions and predictions from its input and to be iterated over multiple steps. The iteration process is designed to produce better predictions than classical models which perform only a single prediction/reconstruction step. Our models are inspired by the family of "stable" diffusion models recently made prominent in the machine learning community by Stability-AI. Unlike the models produced by Stability-AI, our models can generalize to many applications rather than simply to image processing. Here, we demonstrate the versatility of our models as well as their ability to perform well when used for real-world applications with only a small amount of training data. In particular, we propose methods to cope with such issues and demonstrate the effectiveness of our proposed methods on real-world problems in the fields of electromagnetics and unmanned aerial vehicles. Additionally, the final portion of our work explores a more theoretical application of our iterative neural networks. The theoretical portion consists of using toy datasets to research natural questions which arise when employing iterative neural networks, primarily focusing on the ability of our neural networks to successfully extrapolate to test data beyond the scope of the training data.

Creator