Combining an Autoencoder and a Variational Autoencoder for Explaining the Machine Learning Model Predictions

A method for explaining a deep learning model prediction is proposed. It uses a combination of the standard autoencoder and the variational autoencoder. The standard autoencoder is exploited to reconstruct original images and to produce hidden representation vectors. The variational autoencoder is trained to transform the deep learning model outputs (embedding vectors) into the hidden representation vectors of the standard autoencoder. In explaining or testing phase, the variational autoencoder produces a set of vectors based on the explained image embedding. Then the trained decoder part of the standard autoencoder reconstructs a set of images which form a heatmap explaining the original explained image. In fact, the variational autoencoder plays a role of the perturbation technique of images. Numerical experiments with the well-known datasets MNIST and CIFAR10 illustrate the propose method.