DS MS Thesis Defense | Dennis Hofmann | Tuesday, April 23, 2024 @ Noon, Gordon Library
12:00 pm to 1:00 pm
DATA SCIENCE
MS Thesis Defense
Dennis Hofmann
Tuesday, April 23, 2024 | 12:00PM - 1:00PM
Location: Gordon Library, 303 Conference Room
Thesis Committee:
Advisor: Elke Rundensteiner
Reader: Frank Zou
Title: Agree to Disagree: Robust Anomaly Detection with Noisy Labels
Abstract:
Anomaly detection is extremely challenging due to the scarcity of reliable anomaly labels. Recent techniques thus rely on learning from generated lower-quality labels employing either clean sample selection or label refurbishment to correct the noisy labels. Both these approaches struggle for anomaly detection as a result of conflating anomalous samples with noisy labeled samples. For sample selection, the class imbalance of anomaly detection combined with the higher noise rate of anomalies (driven by their high diversity) leads selection techniques to unintentionally discard crucial anomaly samples. On the other hand, label refurbishment methods rely on anomalies having distinct properties from inliers, such as higher prediction variance. This can lead to incorrect refurbishment, especially for marginal clean samples which exhibit similar characteristics. To overcome these limitations, we introduce Unity, a new learning-from-noisy-labels approach for anomaly detection that elegantly leverages the merits of both sample selection and label refurbishment. Unity leverages two deep anomaly classifiers to collaboratively select easy samples with clean labels based on prediction agreement and marginal samples with clean labels via disagreement resolution. Instead of discarding samples that may have noisy labels, Unity introduces a feature-space-based metric called ContrastCorr to refurbish the remaining labels. The set of selected and refurbished clean samples are then combined to robustly update the anomaly classifiers in an iterative label cleaning process. Our experimental study on a rich variety of anomaly detection benchmark datasets demonstrates that Unity consistently outperforms state-of-the-art techniques for learning from noisy labels.