Abstract
Deep Learning puts forward some of the most sophisticated machine learning solutions for applications including natural language processing, audio processing, computer vision, etc. In spite of all the recent achievements, we have yet to achieve the true semantic learning required to reason about the data. This lack of reasoning is partially attributed to the discriminative methods that memorize patterns and characteristics from training samples and partially to the lack of synthesis of the underlying, often hidden, semantic relationships. Over the last decade, there has been extensive research on semantic based representation learning that focuses on bridging the semantic gap between low-level features and their high-level contextual meanings. Similar to how humans create latent representations of observed data and ascribe them meaningful labels, the next generation of deep learning methods will also extract meaning from latent representations. In this manuscript, we propose a novel generative framework that uses counterfactual inference based representation learning methods to uncover the hidden relationships in observed data. The proposed generative framework will have the ability to learn the structure of the data at a deeper level rather than just learning hyperplane boundaries. It works on the principle of variational inference and variational autoencoders, a method that has gained a lot of attention as being the models of choice for generative models. Variational autoencoders help us uncover deep representations by mapping the observed data to latent spaces and use the latent relationships and generative modeling to infer hidden information about the data. The results indicate that the proposed framework outperforms several state-of-the-art methods in getting the most accurate performance metrics.