Proposal of an architecture for integrating explainability in deep learning models for spatio-temporal problems

Artificial Intelligence (AI) is transforming society, from industrial processes to interaction with the services we consume on a daily basis. The relevance of AI has given rise to concerns about the transparency and fairness associated with it. The field of Explainable Artificial Intelligence (XAI) was developed to address these concerns by providing transparency in AI decision-making, alleviating the effect of biases, and providing confidence. However, the application of XAI is a complex problem. Techniques for the application of explainability not only have complexity at the technical level but are different based on the architecture being used. Thus, an explainability technique that is applied to machine vision may require modifications to be applicable to other fields, and may not be as effective because it was not designed for the new problem. This thesis addresses the problem of the application of XAI to spatio-temporal problems, which is a challenge due to the limited number of techniques focused on this type of problem, most of these techniques being developed for other problems and subsequently adapted. In this way, explainability techniques for spatio-temporal problems are a challenge, both due to the limited number of existing options and to the complexity of the design of methods adapted to the characteristics of the problem. Thus, in this thesis the integration of explainability in a deep learning model for spatiotemporal problems has been sought, minimizing the impact on the accuracy of the model while increasing the explainability. To achieve this goal, the models that represent the state of the art in the prediction of spatio-temporal problems and the different existing techniques for integrating explainability in these models have been explored. Based on these fundamentals, a proprietary architecture has been developed to optimally integrate explainability into a deep learning model for spatio-temporal problems.

Author

• García-Sigüenza, Javier

• Deep learning
• Explicabilidad
• XAI
• Inteligencia artificial explicable
• Inteligencia artificial
• Problemas espaciotemporales

Identifier

hdl:10045/164239

OAI identifier oai:identifier

oai:rua.ua.es:10045/164239