Encoding and Decoding Representations with Sum- and Max-Product Networks

Antonio Vergari, Robert Peharz, Nicola Di Mauro, Floriana Esposito

Abstract: Sum-Product Networks (SPNs) are deep density estimators allowing exact and tractable inference. While up to now SPNs have been employed as black-box inference machines, we exploit them as feature extractors for unsupervised Representation Learning. Representations learned by SPNs are rich probabilistic and hierarchical part-based features. SPNs converted into Max-Product Networks (MPNs) provide a way to decode these representations back to the original input space. In extensive experiments, SPN and MPN encoding and decoding schemes prove highly competitive for Multi-Label Classification tasks.