TU Dresden
A joint project with Uni Leipzig DE EN

RelAI: Comperative methods for Reliable AI

Mentored by Ronald Tetzlaff, Guido Montúfar
at Chair of Fundamentals of Electrical Engineering, Institute of Circuits and Systems, TU Dresden or Max Planck Institute for Mathematics in the Sciences

The societal acceptance of modern AI methods is dependent on understandable and reliable network decisions. In traditional neural network based AI methods, these decisions are usually learned by layer-like structures that are optimized through gradient decent. This mathematical procedure does not allow any human understandable conclusion about the learnings of the network. This is a major shortcoming when AI methods and human operators are supposed to cooperate. To keep the human in the loop, we aim to develop novel methods that foster human trust by enabling interpretability of network decisions by enabling local interpretability of complex network models.

The PhD topic investigates new methods for visualizing neural network based decisions and contributes active research on evaluating faithfulness and robustness of the developed method. The research challenges that are crucial in this context are for example:

  • Can local explanations be used to visualize global network decisions?
  • Is it possible to visualize global knowledge of a network about a detected class and if yes does this apply to different datatypes like text or image data?
  • Are statistical visualization methods like t-SNE applicable and what type of knowledge do they provide in real world scenarios?
  • How can explanation methods be compared efficiently in terms of e.g. faithfulness?

Work environment

You will be working in the research group of Prof. Dr. Ronald Tetzlaff at TU Dresden providing access to a high performance computing lab and in-depth expertise in the generation and benchmarking of XAI methods. You will be working with passionate colleges providing active research on the application of your developed explainability methods in a production setting with human operators.

Additionally, you will be supervied by Dr.Guido Montúfar who will provide his insights in mathematical modeling and optimization to foster the joint development of the novel explainable AI methods.

Prerequisites

To conduct this research, you should hold a very good university degree (MSc or an equivalent) in electrical engineering or related technical disciplines (such as computer science, physics or mathematics).

Excellent analytical skills and practical experience in one or more of the research areas, Explainable AI and mathematical optimization for ML is beneficial. You are able to collaborate well in an interdisciplinary and international environment.