User Tools

Site Tools


public:events:agi-summerschool-2012:ben

«– HUMANOBS.ORG | Main Page | Schedule Day-by-Day



Ben Goertzel: Bridging the Symbolic / Subsymbolic Gap

Invited Lecture III - August 8 2012

Core arguments at a glance

  • Given the nature of human-like general intelligence and the presence of relatively severe computational resource requirements, the most practical (and perhaps only plausible) way to create human-level AGI is to tightly integrate a number of distinct components, each focused on remembering a certain sort of knowledge and doing learning based on its memory store
  • Many critical aspects of human-like general intelligence, such as self and reflective consciousness, arise via synergetic interactions between multiple, complex, tightly coupled components
  • Visual and auditory pattern recognition is best carried out via hierarchical pattern recognition systems that form “compositional spatiotemporal deep learning networks” (CSDLN)
  • More abstract, symbolic and linguistic cognition is best carried out via networks of processing elements interacting via a combinatory topology, centrally concerned with carrying out discrete “symbolic” transformations
  • To integrate visual and auditory pattern recognition with abstract symbolic/linguistic cognition, one robust approach is to use pattern mining algorithms to recognize patterns in CSDLNs, and export these patterns to an abstract cognition module; and in turn enable the latter module to imprint patterns on the CSDLN’s state
  • The above idea may be implemented, for example, via integrating the DeSTIN CSDLN with the OpenCog integrated cognitive architecture (that focuses on uncertain symbolic reasoning, also incorporating other aspects such as evolutionary learning and subsymbolic attention allocation)

Abstract

The central challenge in the AI field, since its beginning in the 1950s, has been the apparent gap between subsymbolic and symbolic intelligence. After outlining the core aspects of the issue, and reviewing historical and current attempts to solve it, I will outline a unique approach to bridging this gap, via using novel pattern recognition technology to interface two existing AI software systems: the DeSTIN (subsymbolic) hierarchical pattern recognition engine, and the OpenCog (largely symbolic) integrated cognitive architecture. I will then discuss currently in-progress work aimed at implementing this design, and to demonstrating its success via using the integrated OpenCog/DeSTIN system to control a Robokind humanoid robot carrying out a variety of preschool-type activities. As well as indicating exciting current and future research directions, this topic provides a concrete and intuitive context in which to explain how OpenCog and DeSTIN work individually; and a substantial portion of the lecture will be occupied with such explanations, including a review of current OpenCog work controlling intelligent video game characters, and DeSTIN work recognizing patterns in images and videos.

Readings For This Presentation

From Our Required Readings List

  • B. Goertzel, “Opencog prime: A cognitive synergy based architecture for embodied artificial general intelligence,“ in Proceedings of ICCI-09, Hong Kong, 2009, PDF
  • Goertzel, Ben (2011). A Novel Strategy for Hybridizing Subsymbolic and Symbolic Learning and Representation. Proceedings of AAAI Symposium on Cognitive Systems, Arlington VA. PDF
  • Goertzel, Ben and Stephan Vladimir Bugaj. AGI Preschool: A Framework for Evaluating Early-Stage Human-like AGIs. Proceedings of the Second Conference on Artificial General Intelligence, Atlantis Press. PDF
  • Goertzel, Ben (2010). What Must a World Be Like That a Human-Like Intelligence May Develop In It, PDF
  • Goertzel, Ben, Joel Pitt, Jared Wigmore, Nil Geisweiller, Zhenhua Cai, Ruiting Lian, Deheng Huang, Gino Yu (2011). Cognitive Synergy between Procedural and Declarative Learning in the Control of Animated and Robotic Agents Using the OpenCogPrime AGI Architecture. Proceedings of AAAI-11 , PDF
public/events/agi-summerschool-2012/ben.txt · Last modified: 2012/08/13 09:14 by thorisson