Generative-Embedded-Clustering

Abstract

We present the Continuous Latent Space based Generative Embedded Clustering (GEC), a variant of the Deep Embedded Clustering Model along with Gaussian Mixture Model Parameter Estimation. GEC explores the intercorrelation between Deep Clustering Models and Deep Generative Models. Using the Local Structure Preservation property, we have developed a way to create a continuous latent space such that sampling from any point in this space will yield meaningful images. Our aim is to develop a hierarchical clustering latent space and obtain the distributions of clusters using parameter estimation by expectation-maximization. Our model is capable of generating images belonging to specific cluster distributions which have been clustered in an unsupervised manner. This research explores the inadequately-studied problem of generative clustering models and also introduces a novel approach of generating images without using any Variational Autoencoder or Generative Adversarial Model framework.