Invariance testing for machine learning models

Thesis

Machine learning models are expected to produce consistent results when the input data objects are slightly modified whilst the ground truth remains unchanged. Such characteristics are also known as invariance qualities. However, previous studies assessing ML model invariance have primarily relied on one or a few accuracy metrics, which provide limited information on how models behave under different transformations. This thesis addresses this gap by proposing a novel, systematic invariance testing framework that moves beyond aggregated accuracy scores and offers more detailed visual information to enable more complex and explainable analysis of model performance.

Our framework introduces a generic and extendable methodology for evaluating invariance across various attributes, including simple transformations (e.g., object rotation) as well as complex environmental variations (e.g., background changes). It is designed to be fully automated, enabling extensive and reproducible testing while ensuring consistent and explainable assessments of models' invariance qualities. A key challenge in invariance testing is ordering and structuring test data, especially when testing an entire input space is infeasible. To address this, we develop a novel approach to sampling large information spaces, leveraging context-sensitive training data to construct context-free models. This method ensures that our framework can effectively assess invariance properties without requiring exhaustive test cases.

By providing visualized patterns of model behavior, our approach enhances diagnostic capabilities, making it a valuable tool for both researchers and practitioners. Our findings demonstrate that the evaluation of ML models' invariance qualities based on the visualized patterns can be: more informative, automatic, reliable and consistent. This thesis contributes to the broader effort of improving ML model evaluation, ensuring that claims of invariance are substantiated with comprehensive and reproducible evidence.

Authors: Liao, Z

• DOI: 10.5287/ora-9epgd4voy
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: model evaluation; background invariance testing; image similarity; invariance testing
• Subjects: Ensemble learning (Machine learning); Machine Learning Testing