Autoencoder-Augmented Machine-Learning-Based Uncertainty Quantification for Electromagnetic Imaging

Uncertainty quantification of machine learning (ML) predictions is of key importance for the widespread adoption of ML-enabled electromagnetic imaging. As ML inference is a predictive process, providing a best (most likely) guess, supplementing that prediction with quantitative uncertainty can help avoid costly errors when interpreting the output of a network. In this work, we present a novel two-output-branch neural network architecture that combines the Monte-Carlo Dropout Bayesian convolutional neural network (BCNN) with an autoencoder (AE) to solve the data-to-image inverse problem. The inclusion of the AE branch complements the predicted uncertainty image from the BCNN with a reconstruction of the network input. The data reconstruction (AE) path provides the user with additional information on the quality of the reconstruction, as a failed data reconstruction may be indicative of an out-of-range input, warranting further investigation.