PROJECT TITLE :

Discriminative Mixture Variational Autoencoder for Semisupervised Classification

ABSTRACT:

For the purpose of feature extraction in semisupervised learning, a deep probability model known as the discriminative mixture variational autoencoder (DMVAE) has been developed and is described in this article. Encoding, decoding, and classification modules make up the DMVAE. The encoding module comes first, followed by the decoding module, and then the classification module. The encoder projects the observation onto the latent space while in the encoding module. The encoder then feeds the latent representation into the decoding part, which depicts the generative process from the hidden variable onto the data. In particular, the decoding module that is a part of our DMVAE takes the observed dataset and divides it up into a few clusters using multiple decoders, the number of which is automatically determined by the Dirichlet process (DP), and then it learns a probability distribution for each cluster. In contrast to the variational autoencoder (VAE) that describes all of the data with a single probability function, the DMVAE is able to provide a more accurate description for observations, which in turn improves the ability of the extracted features to be characterized. This is especially true for data that have complex distributions. In addition, in order to acquire a discriminative latent space, the class labels of the labeled data are utilized to restrict the feature learning accomplished by means of a softmax classifier. This is done with the intention of ensuring that the minimum entropy of the predicted labels for the features obtained from unlabeled data is maintained. Last but not least, the joint optimization of the marginal likelihood, label, and entropy constraints causes the DMVAE to have a higher classification confidence for unlabeled data while accurately classifying the labeled data, which ultimately leads to better overall performance. Experiments performed on a number of benchmark datasets in addition to the measured radar echo dataset demonstrate the benefits that our DMVAE-based semisupervised classification has in comparison to other similar methods.