Colloquium Computer Science - Prof. Dr. Thomas Villmann, Hochschule Mittweida, Germany

Title: Challenges for Prototype-based Reliable Classification Learning

Abstract:

Classification learning by Generalized Matrix Vector Quantization (GMLVQ) provides an intuitive and interpretable approach with comfortable additional extension like relevance learning and explanation by counterfactuals. Learning is realized by gradient descent learning (GDL) for the class dependent prototypes as well as the mapping matrix with respect to a cost function approximating the classification error. The mapping matrix finally realizes a data-embedding supporting class separation.

Although GMLVQ constitutes a shallow model, the cost function may realize a complex topological structure in the parameter space (prototypes and mapping matrix). Thus, gradient learning may get stuck at very different close to optimum solutions (COS), which can be interpreted as Rashomon-sets. These COS can offer qualitatively distinguished solutions, which makes a unique interpretation difficult or impossible. This challenge is particularly valid for high-dimensional data.