A Comparative Study of Bayesian Neural Networks and Machine Learning Based on COVID-19 Image Classification

Deep neural network methods have become popular tools for data-driven problems in many applications. However, classical neural networks usually suffer overfitting, unsatisfactory calibration and uncertainty quantification. To tackle such issues, Bayesian neural networks and associated sampling and approximation-based inference approaches have been proposed, including stochastic gradient MCMC, variational, and Gaussian approximation. Under a convolutional neural network architecture, we implement several Bayesian deep learning methods and compare their performance with the classical stochastic gradient descent neural network and conventional machine learning methods. The human lung X-ray images with COVID-19 and other lung conditions are used to examine the binary and multi-class classification accuracy and the quality of model uncertainty under in-distribution and out-of-distribution data. Our experiments show that having a set of estimated parameters that captures various high-performing models is crucial for better accuracy and calibration via Bayesian model averaging. Ensemble methods that identify multiple regions of attraction outperform other methods. Furthermore, fully connected networks often underperform compared to those using dropout or subnetwork inference, which enhances regularization and uncertainty estimation. In more straightforward tasks like binary classification, Gaussian processes and ensemble-based machine learning methods such as gradient boosting and random forests remain competitive with more complex neural networks. Supplementary materials for this article are available online.