Bayesian neural network generalised additive models

• Main Author: Tay, Caleb Wei Hua
• Other Authors: Xiang Liming
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Science::Mathematics::Statistics
• Online Access: https://hdl.handle.net/10356/172098
• Institution: Nanyang Technological University

In recent years, neural networks (NNs) have gained wide and lasting traction as the machine learning architecture of choice in many contexts, due to its flexibility and ability to represent complex functions. However, in the context of a regression task, NNs face difficulties in interpretability and understanding of the effects of each predictor, due to the interactions between each predictor. Additive models, which are simpler models than NNs and lack interaction terms, allow insight into the effects of individual predictors, at the potential cost of model accuracy. More generally, machine learning models may also be ‘overconfident’ in their predictions; in that the model is unable to specify its confidence it is in its prediction. Taking a Bayesian viewpoint allows for machine learning models to represent its confidence (or lack thereof) in its predictions. This paper aims to collect these ideas together to form a new machine learning architecture that is interpretable and Bayesian in nature.