Options
2019
Conference Paper
Titel
Applications of Projected Belief Networks (PBN)
Abstract
The projected belief network (PBN) is a layered generative network, with tractable likelihood function (LF) that can be trained by gradient ascent as a probability density function (PDF) estimator and classifier. The PBN is derived from a feed-forward neural network (FF-NN) by finding the generative network that implements the probability distribution with maximum entropy (MaxEnt) consistent with the knowledge of the distribution at the output of the FF-NN. The FF-NN, from which the PBN is derived, is a complementary feature extractor that exactly recovers the PBN's hidden variables. This paper presents a multi-layer PBN and a deterministic PBN that are tested using a subset of MNIST data set. When the deterministic PBN is combined with the dual FF-NN, it forms an auto-encoder that achieves much lower reconstruction error on testing data than the equivalent conventional network and functions significantly better as a classifier.