AASS Seminar - Addressing Shortcomings of Explainable Machine Learning Methods

The research centre AASS arranges a seminar with Emmanuel Amr Alkhatib, KTH.

Remote attendance: https://oru-se.zoom.us/j/65096623293

Abstract

Machine learning algorithms increasingly achieve near-human performance across a wide range of real-world applications. However, these models often lack interpretability, making it difficult for users to understand the reasoning behind predictions, a critical requirement for building trust in the predictive models and, more importantly, addressing legal and ethical concerns. As a result, post-hoc explanation techniques have gained attention as a way to achieve transparency without sacrificing predictive performance. However, such techniques are constrained in their ability to provide a comprehensive and faithful insight into the prediction process.  We propose several solutions that address the limitations of current explanation methods. Additionally, we explore the use of the conformal prediction framework to ensure validity in the explanations provided. Furthermore, we introduce explainable machine learning methods that eliminate the need for post-hoc explanation techniques, offering a more integrated and transparent approach to machine learning.

Bio

Amr Akhatib is a final-year doctoral student at KTH Royal Institute of Technology, specializing in Trustworthy Machine Learning. His research spans explainability, Conformal Prediction, and Graph Neural Networks. He has published nine research papers during his doctoral studies, with more submissions currently under review. Amr's expertise also extends to Natural Language Processing, with hands-on experience in Language Modeling and Transformers. He is a recipient of the WASP Ph.D. scholarship and was awarded the Alexey Chervonenkis Best Student Paper at COPA 2023.  Amr will present the research findings from his doctoral studies.