DISSERTATION DEFENSE PRESENTATION
Title: Mitigating the Effects of Label Noise in Time-Series Sensor Data Using Deep Feature Extraction
Presenter: Gentry Atkinson
Advisor: Dr. Metsis
Date: Wednesday, April 5, 2023
Time: 5:00 p.m.
Location: Comal 212
Abstract:
Uncertainty in training datasets can badly affect the ability of machine-learning models to learn solutions to real-world problems. Label noise, a disagreement
between the assigned labels and the correct characterizations of instances of data, can be particularly troublesome in time series data due to the low interpretability of this domain of data. Many real-world systems rely on data gathered from one or more sensors including medical, human activity recognition, weather, and emotion recognition. Better approaches to identifying and relabeling incorrectly labeled instances in sensor datasets will have an immediate and pervasive impact on these fields. This works describes a suite of approaches to processing sensor data with incorrect labels. We show that self-supervised deep feature extraction is learns representations of time series data without being impacted by the uncertainty of incorrect labels. These deep features are used with our novel adaptation of the KNN algorithm along with an estimated label transition matrix to relabel instances of data with high precision. This work will also describe a novel adaptation of generative diffusion for producing new instances of data from examples with noisy labels. This approach to data generation is shown to fit more closely to the distribution of the example data than existing techniques when measured using an adaptation of the Fréchet Inception Distance.
Deadline: April 7, 2023, midnight