A multichip, programmable analog neural network for realtime dynamic computations is described. The network's interconnection structure, the neuron characteristics, synaptic connections and synaptic time constants are modifiable. The chips are designed to allow a modular and expandable gross architecture that can be adjusted to the complexity of the task. The network operates fully analog in real time. Howeve. a digital host is used to set the network parameters and monitor the neuron outputs.A prototype neural computer consisting of 72 neurons has been assembled and tested. The network has been successfully configured for several applications and found to have a performance that is equivalent to a digital machine of 10" FLOPS.Neural networks consist of many massively parallel, highly interconnected processing elements. This unique architecture which is based on the structure of the brain has proven to be successful for solving problems which are in general hard to solve with traditional computers. It is not suprising that these kind of problems can be categorized as human-like tasks which involve sensory processing -seeing, hearing -as well as classification, completion and associative recall of patterns often from incomplete and noisy data, generalization, optimization and learning. It is the distributed nature and internal organization of the information that provides the network with its unique characteristics and emergent properties'.A common feature of real-world problems is the rimeaspect. In speech, vision, motion and other real-life problems, time is an essential variable. Biological neural networks deal with time utilizing built-in mechanisms for temporal integration. However, the majority of artificial neural networks to date, have ignored this aspect and process incoming signals at discrete time intervals. The network described in this paper is different in this respect and is intended to deal with problems where time-domain computations are required. Time is used as an explicit and continuous variable. This is accomplished by working in the analog domain and introducing adjustable time constants.The system has several other unique features. It is a multichip system in which the chips contain arrays of modifiable neurons, synapses, synaptic time constants and analog crosspoint switches. The network is modular and can be expanded to any size by adding additional modules. Each chip is designed in such a fashion that it can be interchanged with other modules or added to the system. Each chip has only connections to the next neighbor's, except for a couple of global control lines, resulting in simple overall packaging and board design. This approach not only allows a programmable interconnection structure, but also an expandable gross architecture. This is important in order to be able to adjust the network to the complexity of the task as well as to explore different network architectures for a wide range of applications. OvervieEThe network architecture is shown in Fig. 1. It consists of different modul...
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.