Presenter
Roland Gunics
Eszterházy Károly Catholic University
Authors
Roland Gunics, Gergely Kovásznai, Ádám Kovács, Tibor Tajti
Abstract
Neural network ensembles with soft voting improve accuracy and stability by aggregating multiple models; however, their reliability under individual model failure remains a critical concern. This paper addresses the robustness of soft-voting ensembles in safety-critical settings by combining empirical analysis and formal verification. We evaluate the impact of single-model failures on ensemble performance and find that soft voting yields graceful degradation, with only minimal loss in accuracy when one component model is removed or corrupted. In parallel, we develop a formal verification framework to investigate whether the ensemble's final prediction remains unchanged under any single-model failure scenario. The results demonstrate that soft-voting ensembles can maintain reliable outputs despite individual model failures, providing both empirical evidence and provable guarantees of fault tolerance in neural network ensembles.
Slides
TBA