Heteroassociative Mapping with Self-Organizing Maps for Probabilistic Multi-output Prediction

In recent years, Representation Learning (RepL) has experienced a surge, especially for cross-modal applications with mixed-type data. Most current cross-modal RepL approaches rely mainly on supervised deep learning models, with unsupervised models playing auxiliary roles. In this paper, we present a fully unsupervised approach to demonstrate the capabilities of such models for cross-modal RepL. Our method jointly learns representations into topologically coherent cross-modal heteroassociative mappings. We apply the proposed framework to a yet to be solved multi-output prediction problem in petroleum geoscience: generating a complete set of regular petrophysical well logs from acoustic borehole images in highly heterogeneous Brazilian pre-salt reservoirs.