Fluid Concepts and Creative Analogies

The book Fluid Concepts and Creative Analogies describes a cognitive model that is very compelling and fun to study.

When you run an implementation of this cognitive model, the cognitive model seems to actively pursue an agenda in an incredible, human like way. Some aspects of its behaviour are reminiscent of ant colony simulators, with codelets as the ants. Each codelet is attracted to a "salient" ("pheromone scented") part of the workspace and does one little piece of work - proposing or building a structure, changing the salience of a new part of the workspace, changing the activation and slippability of one of the concepts in the slipnet, and so on. As the codelets run, the system seems to be pursuing an agenda, an agenda that is different every time it runs, and that is emergent rather than explicitly programmed in.

Trying to apply this cognitive model to other domains is a mind bending exercise. By way of contrast, when applying a genetic algorithm to a problem for example, you just need to define a few recombination operators for generating new proposed solutions from old ones, and a fitness function for measuring a proposed solution. When this program runs, its style is simply a parallel prioritized search. The fluid concepts cognitive model on the other hand has a much more compelling, beautiful style as it works toward a solution. Whether it is more or less efficient is not as important to me as its behaviour.

When applying this cognitive model to a new domain, you need examine the domain in more detail than with other AI techniques. You need to choose the concepts involved, how they are connected to each other and how these concepts manifest themselves through the "top-down" codelets they can generate and how "bottom-up" codelets should be influenced by the activations of the concepts. When you see this program running, you see a story unfold about a thought process rather than just a dry search. The system notices features, which may activate concepts, which may direct the system to search for features related to that activated concept. The activation of a concept may also strengthen the bond between two other concepts, causing a slippage of activation. For example, thinking about the letter "a" and the concept of successorship may lead you to start thinking about the letter "b". Its this richness and depth of domain knowledge that gives systems based on this cognitive model such a compelling, beautiful character.

As well as Fluid Concepts and Creative Analogies, I've been reading the following two books written by the co-authors:

Analogy-Making as Perception - A Computer Model by Melanie Mitchell

The Subtlety of Sameness by Robert French

Also its helpful when studying this to read An evaluation of the Fluid Concept Architecture, a good examination of the cognitive model.

Links:

The book

A review and description of the book by Daniel Dennett

Fluid Analogies Research Group

A java implementation of Copycat by Scott Bolland.
Update: Scott Bolland appears to have disappeared from the internet. A copy of the Copycat program along with the tutorial pages appears on the Internet Archive.

Synthbench v002

tech notes:

  Synthbench is an audio synthesis tool inspired by the "Reactable" instrument.
 
  It features:
    - Generators and LFO controllers with waveforms editable in real time
      either interactively with the mouse or through lua commands.
    - Generators and LFO controllers that use standard waveforms.
    - High pass, low pass and band pass filter effect nodes
    - Delay / phase shift effect node
    - A tracker, accessible using tracker controller nodes that can be
      attached to a generator like any other controller node
    - The two parameters of a generator node can be separately controlled by
      a controller node

interface notes:

  To adjust the tempo, the center node (small purple circle) needs to be
  selected and the controls window needs to be visible ("Controls" menu
  item toggles it).
 
  The "mf" and "ma" controller nodes need to be attached to a generator
  to enable it to be controlled by the tracker. Sliding the "Track"
  parameter selects which track controls the controller node.
 
  If you find that a controller node is controlling the pitch instead of
  the amplitude or vice versa, try moving the controller around the
  generator until what it controls changes.
 
  When placing a high or low pass filter, you may want to increase the Q
  parameter slightly to make the effect more dramatic.
 
  The generator and controller nodes that allow you to modify the
  waveform are the "lua" nodes. The "WaveformEdit" menu item toggles the
  waveform editor.
 
  The "Waveform", "Tracker" and "Synthbench" menu items toggles the
  visibility of the respective panes.

Download