[ CogSci Summaries home | UP | email ]

Croft, D. & Thagard, P. (2000). Dynamic imagery: A computational model of motion and visual analogy. Unpublished manuscript.

  author = 	 {},
  title = 	 {},
  journal = 	 {},
  year = 	 {},
  OPTkey = 	 {},
  OPTvolume = 	 {},
  OPTnumber = 	 {},
  OPTpages = 	 {},
  OPTmonth = 	 {},
  OPTnote = 	 {},
  OPTannote = 	 {}

Author of the summary: Jim R. Davies, 2000, jim@jimdavies.org

Cite this paper for:

This paper deals with analogies between dynamic systems: ones that move and have interacting pieces.

The representational structure used is called a `scene graph.' [p1] They claim humans have structures with some of the same computational powers.

Scene graphs are hierarchically organized. They employ the use of a working memory visual buffer, but what is there is not a bitmap, but has structure. [Figure 1, p3; p4].

The system has a seperate visual and semantic memory [p3] (Summary Author note: Barsalou 1999 would disagree with this theory on these grounds.)

"Skeptics about mental imagery such as Pylyshyn (1984) maintain that no special pictorial representations need to be postulated, but there is much evidence that human thinking employs visual as well as verbal representations (e.g. Simon & Larkin 1989; Kosslyn 1994)" [p4] (Summary author note: Pylyshyn is against bitmap images only, Kosslyn is against Pylyshyn on this point. From the rest of this paper, it sounds like Croft and Thagard agreeing with Pylyshyn, who believes in structured descriptions (Pylyshyn 1978)).

"It is not helpful to think of them as collections of pixels like the display on a computer screen, because a pixel representation does not support high-level processing of objects involving operations such as combination and motion.'' [p4]

Nodes in the scene graph can be associated with behavoirs. That's how the system is dynamic. (Summary Author note: It doesn't appear that it would be easy to show interaction, like the electron revolving around the nucleus. Same for coordinated motion, like a body walking.)

"Behaviors in scene graphs enable the production of any number of different movies." [p10] (Summary Author note: if it makes movies then it makes bitmap representations.)

The system is DIVA, written in Java and Java 3D: ( http://java.sun.com/products/java-media/3D/)

It simulates perceptual input with VRML ( http://www.vrml.org/). It was used because it's standard on the internet. (Summary author note: which is also a structured description language, not depictive. It's hardly computer perception.) Java 3D does the translation from VRML to scene graphs. [p13]

Node properties:

  1. transform: translation, rotation, scaling (applies to everything below node)
  2. group
  3. shape: made up of circles, squares, and meshes of triangles.
  4. material: color, texture
  5. behavior: movement
Behaviors use x, y, and z coordinates. The analogical mapping is done by ACME [p18]

To identify convergence in the fortress/tumor problem, the system looks at each behavior node as well as the common point of convergence among two or more pathways. [p30] This is a convergence abstraction that is inferred. The static object that happens to be at the center has nothing to do with it (later it does, though). [p31]

Right now the system has no physics.[p33]

Summary author's notes:

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
JimDavies (jim@jimdavies.org)
Last modified: Thu Apr 15 11:07:19 EDT 1999