Multi-layer multi-feature map architecture for situational analysis
Jakubowicz · 1989
A neural network architecture is described that can recognize and later reconstruct spatially related groupings of various objects. The recognition and reconstruction properties are invariant under input patterns that are translated, distorted, noisy, incomplete, and rotated by approximately 30 degrees with respect to the training patterns. The system utilizes massive redundancy and localized coagulation of spatial information in a manner related to K. Fukushima's Neocognitron for visual image recognition. This system is the first to use multilayered multitopologically ordered T. Kohonen (1984) feature maps in a Neocognitron-related architecture. The model is described, applications are discussed, and results exhibiting an example run are presented.>