Vector Space Logic: A Unified Representation for Neural and Symbolic Knowledge
Abstract
The central goal of this research is to develop a representation language for AI that combines the best features of deep networks and symbolic formalisms. The key properties of symbolic representations are formal semantics, compositionality, and flexible inference. The key properties of neural representations are continuity, distributed encoding, and graded inference.We believe that a representation with all these properties is essential for the progress of AI. A closely-tied goal is to develop powerful and efficient inference and learning algorithms for it. Because inference in logic and inference in the more flexible types of deep architecture are both intractable, designing tractable approximations is key. Our learning algorithms will be able to take advantage of pre-existing knowledge, rather than start tabula rasa each time. Similar tohumans, learning can then become cumulative, successively adding to the system???s knowledge based on experience, extraction and generalization from new sources, and transfer to new situations. We will pursue these goals by a combination of theoretical, algorithmic, and experimental approaches. We will develop and study the formal properties of vector space logic, which willhave first-order logic and tensor algebra as special cases. We will design a full range of inference and learning algorithms, including exact inference, approximate inference, parameter learning, structure learning, and others. We will evaluate the proposed framework in answering questions based on textual and visual sources. If successful, this research will make several contributions with potential impact on DoD capabilities. Autonomous systems must be highly adaptable and be able to robustly interpret and respond to noisy input. Deep learning excels at these. At the same time, AI systems that workcollaboratively with humans need to produce output that is interpretable to humans, and they need human-level models of the world including objects, classes, and relations. Symbolic AI excels at these. In addition, intelligent agents need to efficiently build and continuously maintain a well-integrated model of their environment, and predict its future behavior, from large streamsof multimodal data, which requires both strong statistical and symbolic elements. By building the foundation for a successful unification of the two, including representation, inference, and learning, this project will enable a major leap forward in the degree of integration of all these capabilities, and as a result in the performance of defense systems and the warfighters usingthem.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 19, 2018
- Source ID
- N000141812826
Entities
People
- Pedro Domingos
Organizations
- Office of Naval Research
- United States Navy
- University of Washington