Connectionists: 2 New Papers

[Jerome Feldman]

      There are two new ICSI Technical Reports on which we invite
      comment. The first one, on the binding problem, is quite
      specialized and will be submitted for publication in some form at
      some time. There is an accompanying web site with a demo
      simulation system that you can play with and the code is also
      available. The second TR is a general discussion of the problem of
      concept learning from a connectionist perspective. It will
      (eventually) appear in an edited book on Fodor's Platonist
      argument for the innateness of concepts.


      ~~~~~~~~~~~~~~~~~~


      "A (Somewhat) New Solution to the Binding Problem"


        L. Barrett, J. Feldman, L. Mac Dermed


        ICSI Technical Report TR-06-001

May 2006

PDF <http://www.icsi.berkeley.edu/pubs/techreports/tr-06-001.pdf>
Overview:


      To perform automatic, unconscious inference, the human brain must
      solve the "binding problem" by correctly grouping properties with
      objects. We propose a structured connectionist model that uses
      short signatures, rather than temporal synchrony or other means,
      to achieve this binding. The proposed system models our ability to
      perform unification and also to handle multiple instantiations of
      logical terms, among other things. To verify its feasibility, we
      simulate our prototype with a computer program modeling simple
      neuron-like elements.


      The web site with demos, etc is:


      http://fluents.barrettnexus.com/


      ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~


      "It’s the Body, Stupid: Concept Learning According to Cognitive
      Science"


        B. Bergen and J. Feldman, ICSI Technical Report TR-06-002, April
        2006


        PDF <http://www.icsi.berkeley.edu/pubs/techreports/tr-06-002.pdf>


          Overview:

The classical question "How do people learn new concepts?" is answered 
by Unified Cognitive Science. Converging evidence from several 
disciplines suggests that:
1) Our core concepts are based on the neural embodiment of all our 
sensory, motor, planning, emotional, social, etc. abilities, most of 
which we share with other primates.
2) We can only be aware of or talk about a limited range of parameters 
over these abilities and human languages are based on these 
parameterizations, plus composition. Composition can give rise to 
additional abilities and parameters.
3) The meanings of all new words and concepts are formed by compositions 
of previously known concepts. We use a wide range of compositional 
operations including conjunction, causal links, abstraction, analogy, 
metaphor, etc.
4) Domain relations, particularly conceptual metaphors, are the central 
compositional operations that allow us to learn technical and other 
abstract concepts.
5) We understand language by mapping it to our accumulated experience 
and imagining (simulating) the consequences.