james anderson

Professor
Ph.D. Massachusetts Institute of Technology
James_Anderson@Brown.edu

 

Anderson, J.A. 1995. An Introduction to Neural Networks. Cambridge, MA: MIT Press.

Andeson, J.A. Seven times Seven is About Fifty. In D. Scarborough and S. Sternberg (eds) Invitation to Cognitive Science, Volume 4. Cambridge, MA: MIT Press, 1997.

Anderson, J.A. and Rosenfeld, E. Talking Nets: An Oral History of Neural Network Research.  Cambridge, MA: MIT Press, 1998, Paperback edition, 2000.

Anderson, J.A. 2002. Hybrid computation with an attractor neural network. In: Y. Wang, R.H. Johnson and M.R. Smith, Proceedings, First IEEE International Conference on Cognitive Informatics: ICCI 2002, pp. 3-12, Piscataway, NJ:
IEEE Computer Society.

JA Anderson, P Allopenna, GS Guralnik, D Sheinberg, JA Santini, Jr., D Dimitriadis, BB Machta, and BT Merritt (2007). Programming a Parallel Computer: The Ersatz Brain Project. In W Duch, J Mandzuik, and JM Zurada (Eds.), Challenges to Computational Intelligence. Springer: Berlin.

 

We want to develop preliminary hardware designs, programming techniques, and software applications for a brain-like computing system. 

The Ersatz Brain Project:  Our overall project is referred to as the Ersatz Brain Project.  It presently consists of faculty, staff, and students from several departments at Brown, including Cognitive and Linguistic Sciences, Physics, and Neuroscience and from Aptima, Inc. (Woburn, MA).  Examples of potential cognitive applications for the Ersatz Brain include natural language understanding, text processing, conceptually based Internet search, natural human-computer interactions, sensor- fusion and information integration.  Traditional von Neumann computer architecture is poorly suited to these applications. 


 

Proposed Hardware.  The proposed hardware architecture, and resulting software architecture, is based on ideas taken from mammalian neo-cortex.  It is a massively parallel, two-dimensional locally connected array of CPUs. 

Although the brain is very large, the design is made feasible by using an approximation to cortical function called the Network of Networks.  This approximation assumes the basic computing unit in the cortex is not a single neuron but small groups of neurons [modules] working together to form attractor networks.  There are many such “intermediate level” structures in mammalian cerebral cortex, most notably cortical columns, small (~  1/3 mm), tightly interconnected groups of cortical cells. 

 

 

 

 


 

 

 

 

Software.  In general, developing software is harder and slower than building hardware.  Software and applications take an unfamiliar form in this system.  It has both analog and digital aspects, that is, it works using both the continuous and discrete domains and is not logic-based as are traditional computers.  It can take multiple steps (a long time) to move data to distant modules.  Topographic representation of data from multiple sources allows useful filtering and computation to be done simply as a result of the details of array geometry, that is, how the data is arranged on the 2-D array of modules.

 

 

 

 

 


 

Software Applications:  We have worked in some detail with several potential Ersatz applications. 

  • Performance of “abstract” arithmetic operations, for example, counting, and magnitude comparisons.  Once a set of number facts is learned, they can then be generalized to new domains based on “analog” (topographic differential weighting) filters without further learning.  The “number” application can be generalized so it can estimate the number of identical items in a field.
  • The ratio between two values can be represented topographically on the 2-D arrays.
  • Two patterns can be recognized as identical.  Two patterns can be recognized as symmetrical.
  • Computation of “logic” is possible using the identity function, subject to caveats.
  • Active perception of short lists of items (Sternberg experiments) is possible using the identity function.
  • It is possible to build an “anomaly detector” that will respond strongly to a different object that is present in an array of identical objects.  (This phenomenon in humans is called “pop-out”.)
  • A specific topographic data representation can automatically compute the ratios of the formants of vowels, thereby compensating for the different vocal tract sizes in men, women, and children.
  • A preliminary design of a reader for Morse code has been developed as an exercise in cognitive signal processing using the Ersatz approach..