Machine learning tools for identifying structural artifacts in data
File(s)
Author(s)
Mikhailova, Aleksandra
Type
Thesis
Abstract
In recent years, machine learning methods have advanced to process datasets of unprecedented complexity and volume across a multitude of domains. Often, such datasets are challenging to manually analyse and may contain hidden artifacts that are not perceptible to the human eye. Examples of such hidden artifacts may include anomalies in the data or hidden structures like temporal drift. Being unaware of the presence of such artifacts in the data exposes researchers to risks of reduced model performance or model misspecification. In light of these challenges, the research presented in this thesis is concerned with two key questions. The first is automated detection of such hidden artifacts in datasets. A civil engineering architecture problem with a focus on anomaly detection is demonstrated, and an application of a state-of-the-art unsupervised deep learning technique is introduced. In addition to anomalous events in the data and their automated detection, we discuss another type of hidden artifact in the data – temporal variation – that may impede the model training process if not accounted for. The second challenge is monitoring the joint interaction of the model and the data used to train it. We introduce a new tool called memory maps which is designed to analyse the learning dynamics of machine learning methods and help to identify potential hidden artifacts of different character in the data. Finally, we discuss possibilities of automated processing of memory maps to further aid the analyst.
Version
Open Access
Date Issued
2023-06
Date Awarded
2024-01
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Adams, Niall
Hallsworth, Christopher
Jones, Daniel
Publisher Department
Mathematics
Publisher Institution
Imperial College London
Qualification Level
Masters
Qualification Name
Master of Philosophy (MPhil)