Abstract

Semantic hashing is a technique to represent high-dimensional data using similarity-preserving binary codes for efficient indexing and search. Recently, variational autoencoders with Bernoulli latent representations achieved remarkable success in learning such codes in supervised and unsupervised scenarios, outperforming traditional methods thanks to their ability to handle the binary constraints architecturally.