Learning Probability Distributions over Structured Spaces

Abstract

Project Summary/Abstract This proposal is concerned with the learning of probability distributions over structured spaces. Such distributions arise in a variety of real world applications, including speech recognition, video/image interpretation, configuration problems, diagnostics systems, and when reasoning about user preferences. In each of these applications, one may have a set of logical constraints that define the “physics” of the corresponding situation, therefore limiting the space of valid configurations in the probability space. This space of valid configurations is what we call a structured space, and the interest is in inducing and learning distributions over such structured spaces (i.e., distributions that are guaranteed to assign a zero probability to every configuration that is outside the structured space). The proposal is based on the main observation that such logical constraints are typically ignored by existing learning formalisms, which can lead to a sub-optimal performance of the learned distributions. The proposal is also based on a recent discovery, known as the Probabilistic Sentential Diagram (PSDD), which can be used to systematically induce distributions over structured spaces. The proposal aims to advance the PSDD and mature it into a domain-independent tool for machine learning that can be used by the broad community. This includes the learning of PSDD parameters and structures; the development of normalized methods for generating the constraints that define structured probability spaces; the efficient compilation of these constraints into PSDD structures; and the integration of these constraints with independence constraints that are prevalent in existing machine learning formalisms, such as probabilistic graphical models.

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 12, 2016
Source ID
N000141512339

Entities

People

  • Adnan Darwiche

Organizations

  • Office of Naval Research
  • United States Navy
  • University of California, Los Angeles

Tags

Fields of Study

  • Computer science

Readers

  • Aerospace Engineering.
  • Distributed Systems and Data Platform Development
  • Mathematical Modeling and Probability Theory.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • Space