Hierarchical Bayesian Neural Networks (HBNNs) for Large Data Classification

Abstract

Abstract Views: 0

The incorporation of hierarchical structures into the Bayesian framework gave rise to Hierarchical Bayesian Neural Networks (HBNNs), which further extend existing Bayesian neural networks due to their modelling uncertainty capabilities in complex data. This systematic literature review focuses on the development, application, and performance of HBNNs trained for large-scale classification tasks across various fields. Specific aims include uncovering existing modelling frameworks, analysing real-word dataset implementations, reviewing the underlying inference methods and architectures, and assessing relative performance against benchmark Bayesian and non-Bayesian models. The review also highlights the gaps and issues of scaling HBNNs with the intent for big data classification. This study analyses 42 peer-reviewed publications dated between 2005 and 2025 and uncovers that HBNNs, in comparison to traditional models, consistently estimate uncertainty and reason probabilistically at better rates, but the lack of refinement in their parallel implementation due to high computational demand and convergence problems leads to a suboptimal solution for expansion bottlenecks. There is optimistic outlook with increasing focus on hybrid models, variational inference approaches, and spatially-aligned hierarchical priors for high-dimensional frameworks. This review provides the basic building blocks for researchers and practitioners seeking to harness the full potential of HBNNs in large-scale classification problems.