Faqeer ur Rehman's Qualifying Exam -Review of Testing Techniques for Quality Assurance of Machine...
- Monday, January 25, 2021 from 3:00pm to 4:00pm
Review of Testing Techniques for Quality Assurance of Machine Learning Applications
Machine learning applications are becoming very popular in different domains e.g., self-driving cars, bioinformatics, machine translation, image classification, etc. However, due to the complex computation involved in them, it becomes a challenging task to verify their correctness, especially when there is no oracle available or too expensive to apply. Testing machine learning models adds extra challenges. Unlike explicit rules written in traditional software, there are no hard-coded rules written in machine learning models. Instead, these rules are learned from the data used to train the models. Generally, in traditional software, the testing component is the code, whereas, machine learning models may contain a bug either in (i) application code (ii) data, or (3) the framework/library used. It is therefore very important to verify each component from different perspectives. This paper explores five research papers that propose different testing techniques (with special emphasis on Metamorphic testing) to perform the quality assurance of machine learning applications. Among those proposed techniques, we have found Metamorphic testing to be an effective approach in alleviating the oracle problem in testing machine learning applications.
- Gianforte School of Computing