Abstract

Content Addressable Networks are a family of neural networks designed for efficient implementation in optoelectronics and VLSI. Three CAN systems have been constructed for pattern classification, employing the supervised, self-organized, and tutored algorithm variations. The experimental systems use planes of parallel binary optical computations for both learning and recall. Experimental results are presented to demonstrate the fault tolerance of the supervised CAN algorithms. The supervised CAN network was able to learn around optical errors resulting from noise, stuck pixels, and failing pixels, even when these errors caused imperfect application of the learning algorithm.

© 1993 Optical Society of America