Generative Inferences Based on Learned Relations

Abstract

A key property of relational representations is their generativity: From partial descriptions of relations between entities, additional inferences can be drawn about other entities. A major theoretical challenge is to demonstrate how the capacity to make generative inferences could arise as a result of learning relations from non‐relational inputs. In the present paper, we show that a bottom‐up model of relation learning, initially developed to discriminate between positive and negative examples of comparative relations (e.g., deciding whether a sheep is larger than a rabbit), can be extended to make generative inferences. The model is able to make quasi‐deductive transitive inferences (e.g., “If A is larger than B and B is larger than C, then A is larger than C”) and to qualitatively account for human responses to generative questions such as “What is an animal that is smaller than a dog?” These results provide evidence that relational models based on bottom‐up learning mechanisms are capable of supporting generative inferences.

Document Details

Document Type
Pub Defense Publication
Publication Date
Nov 17, 2016
Source ID
10.1111/cogs.12455

Entities

People

  • Dawn Chen
  • Hongjing Lu
  • Keith Holyoak

Organizations

  • National Science Foundation
  • Office of Naval Research
  • University of California, Berkeley
  • University of California, Los Angeles

Tags

Fields of Study

  • Computer science

Readers

  • Artificial Intelligence
  • Brain and Cognitive Science; Experimental Psychology; Cognitive Neuroscience
  • Distributed Systems and Data Platform Development

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks