Subject

A major cause of predictive failure of machine learning models is due to disparity between train data and test data. Such samples are known as OOD samples. In this thesis we will explore various visualization tools for identifying OOD samples and explaining them. Our approach will look at different ensembles of OOD Detection methods with grid based visualization. While ensemble based methods have demonstrated promising results, we will look at ways in extracting meaningful representations to separate inlier and outliers. We will also look at various explanation methods and analyze which explanations are powerful and help to distinguish between inlier and outliers.

Kind of work

The student is required to develop a a visualization framework to understand the difference between inlier and outliers. We expect the student to
• Novel Class Discovery: A Dependency Approach
• OODAnalyzer: Interactive Analysis of OOD samples

Number of Students

1 - 2

Expected Student Profile

Prior knowledge in Machine Learning
Prior knowledge in Python and PyTorch