Deepak Venugopal, Ph.D.

Department of Computer Science, The University of Memphis

Combining symbolic representations with neural representations is a topic that has obtained significant attention to improve interpretability, scalability and generalization in deep learning models. Kautz recently proposed a taxonomy of different possible Neuro-Symbolic integrations. In this talk, first, I will present an overview of such integrations and then summarize recent research results from our group in this direction. Specifically, I will present a model called Hybrid Markov Logic Networks (HMLNs) to combine embeddings learned from a Deep Neural Network (DNN) with symbolic relational knowledge that can express domain constraints. In particular, for this work, we developed approaches to address covariate shifts in neural embeddings that occur due to variability in deep learning. Finally, I will present an overview of applications in education that we are developing in collaboration with Carnegie Learning using symbolic knowledge in conjunction with DNNs.