Enhanced VQA : numerical quantification
Loading...
Authors
Wang, Max
Issue Date
2019-05
Type
Electronic thesis
Thesis
Thesis
Language
ENG
Keywords
Computer science
Alternative Title
Abstract
While it is plausible to hold that artificial intelligence, AI, has made steady progress since its modern inception in 1956, it seems that over the last 10 years, AI has innovated at a particularly rapid pace. One longstanding part of AI is Question Answering (QA), a capability for intelligent computer systems that consists in their ability to accurately answer textual questions regarding certain content. But recently, as part of the explosive innovation to which I refer, visual question answering (VQA) was established for computer systems, which enables them to answer questions regarding visual information, a capacity that promises to be useful for AI needed for example in autonomous vehicles, many of which must deal with visual scenes. But as impressive as today's VQA technology is, there is much room for enhancement over and above today's state of the art. In the context of a long-term vision (shared herein) for such enhancement, this thesis outlines work that focuses on a particular enhancement that is part of this vision: viz.,~using numerical quantifiers in conjunction with automated-theorem-proving in order to secure a more sophisticated form of VQA. This particular enhancement, which is in large measure now implemented, specifically enhances VQA by allowing quantificationally complex questions to be answered, and by enabling an AI to justify these answers with supporting proofs. As such, the thesis marks progress on what can be called `VQ+AJ,' where the superscript indicates that the questions can be more expressive and complicated because they permit use of quantifiers far beyond those seen in standard VQA, and the `J' indicates that justifications of returned answers are provided as well.
Description
May 2019
School of Science
School of Science
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY