[Comp-neuro] New Ph.D. Thesis: Model for Learning & Recognition in the Neocortex

Dileep George dil at stanford.edu
Mon Aug 25 08:26:42 CEST 2008


I am pleased to announce the availability of my Ph.D. dissertation,  
completed this June at the Department of Electrical Engineering at  
Stanford University. The thesis is titled "How the brain might work: A  
Hierarchical and Temporal Model for Learning and Recognition"

This thesis makes the following contributions:

1) Algorithms and networks, collectively called Hierarchical Temporal  
Memory (HTM), used to learn hierarchical models of data. The HTM  
algorithms, when applied to a visual pattern recognition problem,  
exhibit invariant recognition, robustness to noise and generalization.  
Inference in the hierarchy is performed using Bayesian belief  
propagation equations adapted to HTMs.

2) An analysis of learning in hierarchical systems: Computational  
learning theory does not offer any justification for the efficiency of  
hierarchical learning. This dissertation proposes a generative model  
for hierarchical spatio-temporal data and uses this model to analyze  
the efficiency of hierarchical learning.

3) A mathematical model for cortical microcircuits: The microcircuit  
model is derived by combining known anatomical constraints with the  
computational specifications for Bayesian belief propagation in HTMs.  
The proposed model has a laminar and columnar organization that  
matches many known anatomical features. The circuit model is then used  
in the modeling of two well known physiological phenomena including  
the illusory contour effect.

The thesis can be downloaded from http://www.numenta.com/for-developers/education/DileepThesis.pdf
Parts of this thesis are incorporated into the NuPIC software  
developed by Numenta (available for download from www.numenta.com).

Thanks and Best Regards
-- Dileep George
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.neuroinf.org/pipermail/comp-neuro/attachments/20080824/72aaa1d5/attachment.html


More information about the Comp-neuro mailing list