[UCL WI Talk]: Explainable AI for Interpretable, Robust, Controllable NLP Models

Abstract:
Larger and more complex models have consistently raised the performance bar in NLP. However, their black-box nature raises
significant concerns regarding their trustworthiness, as their scale and complexity hinder our ability to understand and control them. We explore the usage of model explanations to address the three key requirements of (1) interpretability, (2) robustness, and (3) human oversight. Most interestingly, we show that model explanations carry strong signals enabling the explicit and model-agnostic detection of adversarial text attacks. Also, we propose a human-model interaction platform, enabling annotators to influence and control deployed models by editing model explanations and thus providing human feedback.

Bio:
Edoardo Mosca is an XAI and NLP researcher from the Technical University of Munich who just completed his PhD with the thesis "Explainable AI for the Human-Centric Development of NLP Models”. He co-authored ~15 scientific publications in the area of XAI/NLP and led over 20 ML-based projects. He describes himself as a mathematician who loves languages.
Personal website: edoardomosca.github.io