Evaluando la confianza en las redes convolucionales aplicando LIME

Abstract

[eng] In recent years, there has been an explosion of artificial intelligence thanks to neural networks. Neural networks have enabled significant advances in computer vision and natural language processing, including the use of large language models (LLMs) and chatbots developed by companies such as OpenAI and Google. This widespread adoption has brought artificial intelligence to millions of people. However, a major challenge with these models is their lack of explainability, as they are often seen as black boxes with unclear decision-making processes. To address this issue, the field of Explainable AI (XAI) has emerged. This paper focuses on investigating the confidence levels generated by LIME (Local Interpretable Model-Agnostic Explanations) explanations in convolutional network algorithms. The study involved a questionnaire administered to a sample of 43 participants who were asked to discriminate between predictions made by a trained and an untrained network. The results showed that the participants were able to correctly identify predictions in 82.91 % of the 20 images included in the survey.