Colloquium Computer Science - Dr. ing. Sascha Saralajew, NEC Heidelberg

Title: Prototype-based Classification Learning: From traditional versions to modern approaches

Abstract:

Prototype-based classification learning is a powerful machine learning method with the great properties of being interpretable and at the same time provably robust. These properties make the method an interesting approach for various domains (e.g. biomedicine) and especially for domains classified as high-risk in the EU AI Act. In this talk, I will introduce the concept of prototype-based classification learning, starting from the traditional methods (e.g. vector quantization) to modern approaches that combine prototype-based classification concepts with deep neural networks (e.g. ProtoPNet). In this introduction, I will focus on the properties of prototype-based classification methods, such as interpretability, and whether they are preserved in modern approaches and other variants. I will give examples, show possibilities and talk about open questions.