Researcher, TU Berlin | Hubject GmbH
Mahmoud is a researcher at the Technische Universität Berlin. He studies data-driven models to explain and predict the stability of nonlinear systems and their dynamics. Parallel to his research, Mahmoud has developed a master's degree course for the Computer Science and Electrical Engineering programs that focuses on applied machine learning methods and optimization techniques in the energy domain. Besides that, Mahmoud leads at Hubject GmbH a number of national and European research projects that narrow the gap between research and industry.
This workshop will cover the use of post-hoc XAI methods and tools such as saliency, class activation and gradient weighted maps, PALM, LIME, SHAP and permutation importance. Also, will put hands on some advanced methods that are still being explored such as information bottleneck theory. The workshop will focus on putting most of these tools into practice while providing intuitions and an overview of the theory driving each method used to help you better understand them.
By the end of the workshop, you will:
learn about the key trends driving the development of XAI;
build and train machine learning models for different modalities;
be able to apply XAI methods to explain and interpret the decisions made by these models;
be able to decide on the suitability of the XAI method for the problem you are addressing; understand the strengths and weaknesses of each XAI method presented.
Brief introduction to XAI
Overview of the datasets and problems used
Use of SHAP, LIME, saliency, class activation and gradient weighted maps for image modality
Implementation of SHAP and LIME on tabular data -Brief overview of information bottleneck (IB) theory
Application of IB theory to MINST datasets