Use of corrections in evidential dense-layer deep neural networks in a multi-view fusion approach


In this article, we explore, within a multi-view fusion framework, the use of different contextual corrections in evidential deep neural networks with dense layers (multilayer perceptrons), i.e. networks where the last softmax layer has been replaced by a Dempster-Shafer layer to obtain a belief function instead of a probability, and where a correction layer has also been added subsequently, followed by a fusion layer and finally a decision layer. Initial results obtained from experimental data show the interest of this approach.