back to list

Project: VAE-Driven Boundary Sampling for more informative Decision Boundary Visualization

Description

Decision Boundary Maps (DBMs) are visual representations of classifires that rely upon dimensionality reduction layouts to represent classifier decision boundaries. Existing techniques use the concept of inverse projection to decide the color of the pixels in a visualization according to the classifier. This has limitations and the results are usually visual representations that do not faithfully reflect the decision boundaries. In this project we will investigate the use of counter factuals to improve DBMs by generating extra high-dimensional data instances that are used to approximate such boundaries.

Details

Student

IB

Imke Bloemen

Supervisor

Fernando Paulovich