CAIMS 2023

Short Talks

Addressing Overfitting and Uncertainty in Neural Network with Nonlinear Sparse Bayesian learning

Deep neural networks suffer from two significant limitations in supervised learning tasks. First, their large number of weight parameters make them prone to overfitting. Second, the inability to comprehensively account for uncertainty in the training data, can result in overconfident predictions. While regularization schemes like early stopping, weight decay, and dropout have been proposed to address the overfitting issue, they fail to addres the uncertainty quantification problem. Conversely, Bayesian neural networks (BNNs) introduce uncertainties in the network’s weights to address the uncertainty issue, but overfitting remains a significant challenge.

To address both these problems simultaneously, the current study proposes a sparse Bayesian neural network (SBNN). The proposed algorithm aims to address the practical and computational issues associated with BNNs by incorporating a sparsity-inducing prior based on the automatic relevance determination (ARD) concept. While the use of a data-informed prior increases the level of hierarchy in the Bayesian analysis, it is computationally manageable within the efficient semi-analytical framework of NSBL. The SBNN algorithm identifies the optimal sparse neural network by removing redundant weight parameters, resulting in a tractable treatment for overfitting through an evidence optimization algorithm. To demonstrate the benefits of the SBNN algorithm, the study presents an illustrative regression problem and compares the results with the standard Bayesian neural network (BNN) and the hierarchical Bayesian formulation of the BNN.

 Overview  Program