Σεμινάριο: "A Βayesian concept selection approach towards concept-based reasoning"

ΚΥΚΛΟΣ ΣΕΜΙΝΑΡΙΩΝ ΣΤΑΤΙΣΤΙΚΗΣ 2024-2025

Ομιλητής: Konstantinos Panousis, Post-Doctoral Researcher, Inria

A Βayesian concept selection approach towards concept-based reasoning

ΑΙΘΟΥΣΑ T107

ΠΕΡΙΛΗΨΗ

The increasing deployment of deep neural networks (DNNs) in critical applications has intensified the need for interpretable models, sparking significant research interest in approaches that ensure model transparency and accountability. However, many existing interpretability methods face challenges such as (i) reduced predictive performance and (ii) compromised clarity due to the involvement of too many features or concepts in decision-making processes. In this presentation, we propose a novel Bayesian concept selection approach designed to enhance interpretability while maintaining or improving performance. Our method leverages a Bayesian framework to enforce sparsity by inferring the relevance of features or concepts through a data-driven Bernoulli distribution. This approach leads to more interpretable models by limiting the number of features or concepts contributing to each decision. Our empirical evaluation demonstrates that this method not only outperforms recent interpretability techniques in terms of accuracy but also yields higher sparsity per decision, facilitating a clearer understanding of the model's decision-making process.