Neural Networks with Differentiable Belief-Function Layers

A new deep evidential pipeline incorporates differentiable belief-function layers using Yager's 1987 rule of combination, avoiding gradient divergence issues found in Dempster's normaliser. The network treats every mass in the chain as a valid Dempster–Shafer mass function and uses the output mass on the ignorance set as an uncertainty signal, ordering human disagreement in visual and linguistic modalities.