FACTORIZED VARIATIONAL AUTOENCODERS

Number: 20190026631
Published: 2019-01-24
Filed: 2017-07-19
Assignee: Disney Enterprises, Inc.
Inventors: CARR; G. Peter K. et al.
CPC: G06F17/16; G06N3/045; G06N3/0455; G06N3/047; G06N3/0475; G06N3/0499; G06N3/084; G06N3/088; G06N3/0895
Verdict: Set aside factorized variational autoencoders, generic ML research
Source: Google Patents · FreePatentsOnline

Abstract

The disclosure provides an approach for learning latent representations of data using factorized variational autoencoders (FVAEs). The FVAE framework builds a hierarchical Bayesian matrix factorization model on top of a variational autoencoder (VAE) by learning a VAE that has a factorized representation so as to compress the embedding space and enhance generalization and interpretability. In one embodiment, an FVAE application takes as input training data comprising observations of objects, and the FVAE application learns a latent representation of such data. In order to learn the latent representation, the FVAE application is configured to use a probabilistic VAE to jointly learn a latent representation of each of the objects and a corresponding factorization across time and identity.

Background

BACKGROUNDField of the Invention

Embodiments of the disclosure presented herein relate to unsupervised machine learning and, more specifically, to factorized variational autoencoders.Description of the Related Art

In unsupervised machine learning, inferences are made from input data that are not labeled. Matrix and tensor factorization techniques have been used in unsupervised machine learning to find underlying patterns from noisy data. In particular, such factorization techniques can be applied to identify an underlying low-dimensional representation of latent factors in raw data. However, when the relationship between the latent representation to be learned and the raw data is complex, such as the behavior of a forest of trees in response to wind in a physical simulation or the reactions of an audience viewing a movie, the data may not linearly decompose into a set of underlying factors as required by conventional matrix and tensor factorization techniques. As a result, such techniques may not achieve good reconstructions of training data or produce interpretable latent factors in complex domains.SUMMARY

One embodiment provides a computer-implemented method for determining latent representations from data. The method generally includes receiving data associated with observations of one or more objects. The method further includes learning, from the received data, a variational autoencoder, the learning being subject to at least a constraint that outputs o

Claims

1. A computer-implemented method for determining latent representations from data, comprising: receiving data associated with observations of one or more objects; and learning, from the received data, a variational autoencoder, the learning being subject to at least a constraint that outputs of an encoder of the variational autoencoder are able to be factorized. 14. A non-transitory computer-readable storage medium storing a program, which, when executed by a processor performs operations for determining latent representations from data, the operations comprising: receiving data associated with observations of one or more objects; and learning, from the received data, a variational autoencoder, the learning being subject to at least a constraint that outputs of an encoder of the variational autoencoder are able to be factorized. 25. A system, comprising: a processor; and a memory, wherein the memory includes a program for determining latent representations from data, the program being configured to perform operations comprising: receiving data associated with observations of one or more objects, and learning, from the received data, a variational autoencoder, the learning being subject to at least a constraint that outputs of an encoder of the variational autoencoder are able to be factorized.