An Evidential Deep Network Based on Dempster-Shafer Theory for Large Dataset
Résumé
We introduce a novel deep neural network architecture based on Dempster-Shafer theory capable of handling large image datasets with numerous classes, such as ImageNet. Our approach involves analyzing images through multiple experts, composed of convolutional deep neural networks that predict mass functions. These experts are then merged using Dempster's rule, thereby returning a set of potential classes by selecting the best expected utility based on the previously computed mass functions. Our innovative algorithm can identify the best set of classes among the 2 K possible sets for K classes while maintaining a complexity of O(K log(K)). To illustrate our approach, we apply it to an out-of-distribution example search problem, demonstrating its efficiency.
Origine : Fichiers produits par l'(les) auteur(s)