Bespoke Neural Network Architectures for Rapid Multivariate Time Series Classification and Representation

Moore, Alexander

Etd

Bespoke Neural Network Architectures for Rapid Multivariate Time Series Classification and Representation

Public Deposited

Accurate chemical sensors are vital in medical, military, and home safety applications. Training machine learning models to be accurate on real world chemical sensor data requires performing many diverse, costly experiments in controlled laboratory settings to create a data set. In practice even expensive, large data sets may be insufficient for generalization of a trained model to a real-world testing distribution. This dissertation is concerned with the application of modern machine learning and deep learning techniques to a real-world, low-data chemical sensing task. In order to mitigate the challenges of an application with costly data, we develop algorithms in adversarial learning and data synthesis, regularize models with multitask and multi-loss learning, and transfer knowledge between multiple domains such that the ultimate goal of chemical detection is improved. We include novel research on data sets within the chemical sensing as well as natural image and molecular representation literature. Machine learning and deep learning models have been adapted with novel architectures from tabular, time series, and natural image domains which ultimately improve downstream classifier performance.

Creator