Efficient Learning with Human-in-the-Loop in Structured, Noisy and Temporal Domains
Abstract
Machine learning has been successful in transforming large amounts of data to actionable knowledge. Despite this remarkable success, most current approaches make assumptions that are incompatible with the domain of this proposal including: (a) ?at features during data generation, (b) high quality examples, (c) restriction of humans as mere labelers, (d) homogeneous data types during learning, and (e) existence of a natural discretization of time when learning temporal models. For example, consider Electronic Health Records (EHRs) that contain demographic information, lab tests, diagnoses, medical images, genomic information and other forms of patient data. These hybrid data are not recorded at uniform intervals over all patients but only when patients require it. Modeling, thus, requires extensive feature engineering to account for the large amounts of non-uniformity between patients and their histories.We aim to relax these assumptions by - (1) enriching data representation (relational, noisy, heteroge-neous) and (2) incorporating richer forms of human knowledge as expert advice. We consider probabilistic logic models (PLMs) that model multi-relational and noisy data faithfully by combining the powerful for-malisms of logic and probability.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 28, 2018
- Source ID
- FA95501810462
Entities
People
- Sriraam Natarajan
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Texas at Dallas