Multilevel explainable artificial intelligence methods and applications

Abstract

Thanks to their astonishing prediction ability, deep neural networks (DNNs) have been deployed in various disciplines, from computer vision to natural language processing. However, their opaque decision-making mechanism makes it challenging to employ them in sensitive areas such as healthcare, legal settings, and autonomous driving. Many works have been proposed in the explainable artificial intelligence (XAI) field to overcome this issue and make DNNs more transparent, trustworthy, and deployable. However, most of these methodologies suffer from several drawbacks.

This thesis explores the current landscape of XAI and identifies critical shortcomings in the field that require urgent attention. Through a thorough examination of these limitations, we reveal key gaps that motivate our contributions. To address these challenges, we propose a novel methodology, multilevel XAI, which generates human-like explanations in the form of linguistic and visual concepts for machine learning and computer vision tasks. Our approach demonstrates that producing multilevel concept-based explanations can be both cost-effective and achieved without significantly compromising model performance.

Building on this, we introduce a novel weakly supervised semantic segmentation framework, semantic proportions-based semantic segmentation (SPSS). This approach facilitates effective semantic segmentation without the need for costly and impractical pixel-wise ground-truth segmentation maps, which are often challenging to obtain in real-world scenarios. By leveraging class proportions as the sole supervision during training, SPSS enables an intuitive and efficient generation of segmentation maps. Furthermore, this framework opens opportunities to integrate the explainability components of multilevel XAI paving the way for future research to achieve semantic segmentation with significantly reduced annotation costs.

We further identified that one of the most significant gaps in the concept-based XAI field---on which this thesis also specifically focuses---is the absence of standardised measures and benchmarks for fair evaluation and selection of the most effective methodologies. To address this, we propose three novel measures and benchmarks to advance the field. We encourage the research community to employ these measures and benchmarks for a fair comparison among concept-based XAI techniques.

Finally, we discuss the limitations of our work and possible future directions that, once realised, could significantly impact the XAI, machine learning, and computer vision communities.

More information

Identifiers

ORCID for Halil Ibrahim Aysel: orcid.org/0000-0002-4981-0827

ORCID for Xiaohao Cai: orcid.org/0000-0003-0924-2834

Catalogue record

Export record