Computing That Serves

DRE: A Differentiable Reasoning Engine for Common-sense Knowledge Inference

Daniel Ricks: MS Thesis Defense
Tuesday, September 12, 2:00PM
3350 TMCB
Advisor: David Wingate


Common-sense reasoning is an essential component for an autonomous agent's interactions with the real world, but previous approaches to capturing and providing this reasoning are complex and expensive. Basic knowledge is necessary for true understanding in machines, which becomes more important as natural language systems become more prevalent in society. Recent advances in state-of-the-art deep learning algorithms lack this common-sense understanding of the world, but it may be possible to imbue these algorithms with that common-sense knowledge such that they can respond with answers to simple queries when interacted with. Common-sense knowledge can be scraped from plain text and projected into high-dimensional space by algorithms like word2vec. We propose the creation of a system comprised of a learning component in conjunction with word2vec's embedding space, which will learn and make this knowledge available to autonomous agents.