Skip to main content

IEOR Seminar by Dr. Dinesh Garg

Title: Quantum Embedding of Knowledge for Machine Reasoning

Speaker: Dr. Dinesh Garg, AI Reasoning, IBM Research, Bengaluru

Time and Date: 3:30 PM to 4.30 PM, on 28th November 2019 (Thursday)

Venue: IEOR Seminar Room (2nd Floor, IEOR Building)

Abstract: Knowledge representation is a field of AI dedicated to inventing ways of representing worldly knowledge in a form that computing machines can easily store, access, and utilise it to solve complex reasoning tasks such as Question Answering, Dialogue, Debate, etc. The classical approach towards knowledge representation and reasoning include symbolic methods which are slow and sometime brittle despite being accurate. A more modern approach include Statistical Relational Learning (SRL) methods to generate distributional representations of the symbolic Knowledge Bases (KBs). These methods are fast and robust, but approximate. SRL methods embed any given KB into a vector space by exploiting statistical similarities among its entities and predicates but without any guarantee of preserving the underlying logical structure of the KB. This, in turn, limits the performance of the logical reasoning tasks which make use of such distributional representations.
 
In this talk, we will present an alternate way of representing knowledge in vector spaces. We call this representation as Quantum Embedding because our approach is inspired by the theory of Quantum Logic. The beauty of Quantum Embedding lies in the fact that it preserves the original "logical structure" of the knowledge post its embedding into the vector space. Furthermore, such an embedding allows one to perform Boolean logical operations directly in the vector space in a manner similar to how one would perform on the symbolic form of the same knowledge. We have demonstrated this claim by showing an impressive accuracy improvements over popular SRL baselines while answering the membership based complex logical reasoning queries.

This work will be appearing at NeurIPS’19 conference as a research paper titled "Quantum Embedding of Knowledge for Reasoning".
 
Speaker Bio: Dinesh Garg is currently a senior researcher in the Machine Reasoning department of IBM India Research Lab, Bangalore. In the past, he has served as a faculty member in the CSE department of IIT Gandhinagar and as a researcher at Yahoo! Labs, Bangalore. His current research is focusing in the areas of machine reasoning, automated question answering, knowledge embedding, explainable question answering, and representation learning for NLP tasks. He completed his master’s and PhD thesis from Computer Science department of IISc, Bangalore. He is a recipient of the best master’s thesis award, the best PhD thesis awards, and INAE Young Engineer Award. He is a senior member of IEEE and also serves on the program committees of key AI/ML conferences on regular basis.