Deep models, lighter footprint compressing, explaining, and transferring Bayesian neural networks

Despite their widespread adoption and impressive empirical performance, modern neural networks often suffer from overparameterization, limited interpretability, and difficulty in knowledge transfer. These limitations hinder their deployment in settings where computational efficiency, transparency, and adaptability are essential. This thesis addresses these challenges through a unifying principle: building \textit{Bayesian deep learning models with a lighter footprint}—models that retain predictive power while being more compact, interpretable, and transferable. The first contribution presents a structured compression framework for Bayesian neural networks (BNNs), where model sparsity is achieved by placing spike-and-slab priors over weights and inferring posterior inclusion probabilities. A scalable variational inference algorithm is developed that simultaneously learns network weights and their relevance, enabling principled pruning at multiple levels of granularity. The second contribution proposes a sensitivity-based feature selection method, \texttt{SENS}, which quantifies the relevance of input features by analyzing the distribution of local sensitivities under the BNN posterior. This method provides statistically grounded feature importance scores along with confidence measures, allowing for robust interpretability and data-driven selection of relevant inputs. The third contribution introduces a semi-supervised, source-free transfer learning approach using power priors. By generating pseudo-labeled auxiliary data from a black-box model and combining it with a small labeled target set, we derive a principled posterior update that enables adaptive transfer without requiring source data or access to pretrained model internals. Each of these contributions is grounded in Bayesian methodology, equipped with theoretical guarantees, and empirically validated across a wide range of synthetic and real-world datasets. Taken together, they demonstrate that deep models need not be large, opaque, or rigid. Instead, through principled Bayesian design, we can construct neural networks that are efficient, interpretable, and resilient—qualities that are increasingly critical as machine learning systems are deployed in high-stakes, resource-constrained, or regulatory-sensitive environments.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Statistics

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Saha, Diptarka

Contributors dc:contributor

• Liang, Feng
• Wang, Yuexi
• Wang, Shulei
• Chen, Yuguo

• Bayesian Neural Networks
• Model Compression
• Feature Selection
• Variational Inference
• Transfer Learning

Statement dc:rights

• Copyright 2025 Diptarka Saha

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129468

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129468