Explainable Artificial Intelligence (XAI): Promoting the transparency and trust in machine learning models.
Main Article Content
Abstract
The swift proliferation of machine learning and deep learning models into the domain of sensitive areas like healthcare, finance, and the law has created a great worry about their lack of transparency and interpretability. These models tend to run as black box, thus restricting their usage in high stakes situations as the users might not understand, trust or even validate their decisions [10], [13]. Explainable artificial intelligence (XAI) has become a promising paradigm to deal with these two issues by making models explain themselves in human understandable terms and not engaging in predictive performance to a considerable degree [33].
This paper provides an in-depth and organized summary of XAI along with its major concepts in relation to interpretability, transparency, trust, and accountability. It also divides XAI techniques into intrinsically interpretable (ante-hoc) models and post-hoc explanation methods and differentiates between model-specific and model-agnostic [13], [39]. The paper also looks at the association between various kinds of explanations and user confidence in various fields of applications.
It suggests a new conceptual framework that demonstrates the connection between technical attributes of explanations, including fidelity, stability, and completeness, and human-related aspects, including trust, understanding, and fairness. The main contributions made by this work are: a taxonomy of XAI methods are complete aligned with model architectures and data modalities, a more integrated framework with explanation types, their target users, and measures of evaluation, provides deeper insight into the open issues of standardization of evaluation, faithfulness of their explanations and regulatory adherence [38].
In general, the proposed research should inform researchers and practitioners to choose, design, and assess XAI methods to improve the transparency, reliability, and trust in the current AI systems.