Optimization of rotationally invariant object recognition in a neural network

✍ Scribed by Stavros Busenberg; Louis Rossi

• Publisher: Elsevier Science
• Year: 1988
• Tongue: English
• Weight: 28 KB
• Volume: 1
• Category: Article
• ISSN: 0893-6080
• DOI: 10.1016/0893-6080(88)90505-9

✦ Synopsis

A constrained optimization procedure is developed for obtaining a layered neural network which can recognize objects even when subjected to rotations. The procedure is based on the development and mathematical analysis of both continuous and discrete neural network models and is implemented as either an external optimization process or as a learning rule within the network itself. In the analysis of object recognition with rotational invariance it was found useful to model the network as a continuum of neurons in a multidimensional spatial domain. A software simulation of this process is developed and used to compare its performance with that of the widely used back propagation method. Related mathematical results concerning memory capacity, stability and domain of attraction of steady states are also obtained supporting the experimental evidence based on the simulation.