Explanations in Answer Set Programming and Machine Learning

Abstract

In artificial intelligence (AI) systems, generating explanations or justifications for the solution or recommendations provided by these systems is crucial and essential for building user trust and enhancing system transparency. This dissertation falls within the scope of Explainable Artificial Intelligence (XAI) and introduces novel algorithms designed to enhance explainability in two distinct methodologies: the data-centric approach using deep neural networks (DNNs) and the rule-based approach using answer set programming (ASP).

The proposed algorithms for constructing explanation generation systems introduce the following key innovations:

1. Faithful explanations for true and false atoms in ASP: The algorithm generates explanation graphs for true and false atoms with respect to a given answer set of a logic program, ensuring that the explanations are faithful to the input logic program.

2. Handling language extensions in ASP: The approach leverages declarative meta-encoding programming to handle language extensions and improve the system's performance.

3. Revealing deep neural network decision processes with a logic program under answer set semantic: Given a trained neural network and its training dataset, a logic program is extracted under answer set semantics. Ideally, this program faithfully represents the neural network, with each answer set corresponding one-to-one to input-output pairs of the network. This approach preserves a high level of predictive accuracy while also offering valuable insights into the model's inner workings, such as the feature importance and the influence of hidden nodes on predictions. Such insights can inform strategies to simplify the network, for instance, by reducing the number of hidden nodes, ultimately leading to more efficient model architectures.

Through both theoretical analysis and practical examples, this dissertation demonstrates that the proposed algorithms are capable of effectively generating explanations or justifications for the outcomes produced by ASP and DNN.