SPNS KNOW WHAT THEY DON’T KNOW

Probabilistic Deep Learning using Random Sum-Product Networks

4.1 GENERATIVE LEARNING: RAT-SPNs ARE COMPARABLE TO STATE-OF-THE-ART

4.2 DISCRIMINATIVE LEARNING: RAT-SPNs ARE COMPETITIVE WITH NEURAL NETS

4.3 RAT-SPNs ARE ROBUST UNDER MISSING FEATURES

4.4 RAT-SPNs KNOW WHAT THEY DON’T KNOW

output uncertainties input uncertainties

this clearly highlights the ability of RAT-SPNs to properly calibrate uncertainties when compared to current deep generative models based on neural networks, which fall prey to the “likelihood mirage”.

SPNs are capable connectionist models with additional advantages like calibrated anomaly detection, treatment of missing features, or most importantly, the power of tractable probabilistic inference. Exploring these feature jointly with deep neural networks, e.g. as calibrated loss layers, is the perhaps the most promising avenue for future work.