Evidence-Based Reliability Estimation in Deep Neural Networks using Dempster-Shafer Theory
Conference
Waseem, A, Pons, AP, Nadeem, M et al. (2026). Evidence-Based Reliability Estimation in Deep Neural Networks using Dempster-Shafer Theory
. 2824-2833. 10.1109/COMPSAC69091.2026.00426
Waseem, A, Pons, AP, Nadeem, M et al. (2026). Evidence-Based Reliability Estimation in Deep Neural Networks using Dempster-Shafer Theory
. 2824-2833. 10.1109/COMPSAC69091.2026.00426
Deep neural networks have good predictive power and are non-transparent in regard to internal consistency and credibility. The current interpretability and uncertainty techniques are centred on the input-output behaviour or predictive confidence, and provide minimal information on the reliability of internal network elements. This paper shows a novel technique, which is an evidence-based estimation of internal reliability based on DS theory. The neurons of the hidden layer and the groups of weights are modelled as separate sources of evidence, which are either supportive or conflicting with the proper model behaviour. Neuron activations and weight perturbation responses are used to determine belief, plausibility, and conflict measures, which are then fused with the Dempster rule of combination. Synthetic data experiments, MNIST, and MRI-based CNN pruning experiments have shown that the neuron activations have substantial internal conflict, whereas weight structures show great stability to perturbation. The suggested framework offers a theoretical basis for the internal trust measurement and pruning of a deep neural network based on evidence.
