Smart Engineering Systems Architectures

Home Page

Summary

Research in this area focuses on designing new architectures for enhancing capabilities of Smart Engineering Systems. Current research concentrates on the following:

Developing a biologically inspired artificial vision system. The system will consist of two architectures that will operate together in an attempt to emulate many of the initial charactersitics of natural vision.

1. The first architecture will contain artificial neurons and connections that mimic those observed in the retina of the eye. The modeling will concentrate mainly on the connections observed within the fovea (highest acuity). Images sent to the nextwork will be static and gray scale, therefore the color and temporal dynamics of the retina are not incorporated into this architecture. Charactersitics of the retina to be considered include photoreceptor adaption, photoreceptor blurring, and contrast enhancement. Properties of the architecture will include elimination of small noise levels, contrast enhancement, and invariance to illumination levels.
2. The signals leaving the retina model will converge on the second architecture, which attempts to model the interactions between the lateral geniculate nuclues (LGN) and the first layers of the primary visual cortex. A self-organizing network will be developed to model the orientations columns observed in the cortex. Both lateral interactions between cortical neurons, and feedback connections from the cortex to the LGN will play a part in the self organization. The addition of feedback to the LGN will also be instrumental in allowing the architecture to modify and adapt its structure in the presence of scotomas or illusory images.

Recent architectures developed are the Hausdorff-Voronoi Network (or HAVNET) and SIMNET.
The HAVNET neural network consist of three layers, the plastic layer, the Voronoi layer, and the Hausdorff layer. The plastic layer contains neurons with the weights that are trained during the learning process, the Voronoi layer serves to measure the distance between individual points in the input and learned patterns, and the Hausdorff layer uses information from the Voronoi layer to compute the overall level of similarity between the input pattern and the learned pattern.

The node is shown in a configuration for one-dimensional inputs for reasons of clarity. In the actual network, the input pattern, plastic layer, and voronoi layer are all two-dimensional. The learning is employed in HAVNET to adapt the individual nodes to recognize certain classes. The specific details of how learning and recognition are preformed can be found in Ryan Rosandich's dissertation.

Graduate Students Involved in Research

David Enke

Hsi-Chieh Lee

Ryan Rosandich

Last update: December 20, 1995.