Silicon Integration of Learning Algorithms and Other Auto-Adaptive Properties in a Digital Feedback Neural Network

In the past few years, a lot of efforts has been devoted to the integration of neural networks [1,2], with much emphasis on the implementation of the network itself, leaving the burden of learning to a host, possibly parallel computer [3]. However, the idea of implementing training on the chip itself is attractive for two reasons: (i) the learning phase is usually very time-consuming; (ii) on-chip learning makes the network more autonomous and opens the way to building elaborate assemblies of networks. The present paper discusses the capabilities of a neural network chip, fully connected with feedback, using binary neurons with parallel synchronous dynamics, intended to be used as an associative memory [4]; the chip integrates a learning algorithm and also some additional, potentially useful features such as self identification of correct relaxation on a stored vector (a prototype), and to the discussion of the main silicon implementation issues.