Towards Uncertainty Estimation in Microwave Imaging using Latent Space Analysis

We present two methods using the latent space encoding of a trained data-to-image microwave imaging deep learning network to asses the certainty of network output. In Method 1, Kernel Density Estimation is applied to the latent vectors corresponding to the training set to obtain a probability density function. The value of the density function for a given test example is shown to have a relatively strong correlation with target reconstruction accuracy. In Method 2, the true absolute error in training predictions from the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{K}$</tex> nearest latent vectors in the training set to the latent vector of the test example are combined to produce an uncertainty map for the test prediction. Results demonstrate that the distribution of training examples in the latent space can provide a measure of confidence in imaging results.