Subject

The aim of this project is to design a Natural Language Explanations (NLEs) system for diverse tasks, eliminating the limit for specific NLEs datasets. This can be done by re-formulating existing non-NLEs datasets (which are diverse) into the form of NLEs.

Kind of work

Natural Language Explanations (NLEs) is a topic in explainable artificial intelligence, where the aim is to understand the decision-making process and underlying mechanisms of deep neural networks through human-friendly and detailed text. However, current datasets for NLEs are scarce and limited to specific tasks only. On the other hand, there exists an abundant of datasets for other tasks which can be converted to produce NLEs.

Framework of the Thesis

As an example, the datasets of Visual Abductive Reasoning, Visual Story Telling, and other video datasets can be manipulated to achieve datasets of textual explanations. This allows to have a much larger portion of NLEs which can be combined along with the existing NLEs datasets to create one huge dataset, where a model can be trained on it. This creates a strong model with rich knowledge about the world, which can be used to give NLEs for diverse tasks.

Number of Students

1

Expected Student Profile

Good knowledge of Machine Learning, AI and data processing. Good programming skills in Python (particularly PyTorch)