Compact analog/digital electronic data processors of a novel type are based on the concept of the optimization cellular neural network (OCNN) and are implemented in very-large-scale integrated (VLSI) circuitry. These VLSI neural processors are being developed to perform a variety of computation-intensive tasks in maximizing the return of useful scientific information from future multiple-sensor microspacecraft, given such considerations as synergy among multiple sensors, limits on the power and bandwidth available for transmission of data, and the need to recognize significant data in the absence of prior explicit instructions. The OCNN concept also has obvious potential for terrestrial "smart"-sensor and other data-reduction applications.
The OCNN concept is founded on the concept of the cellular neural network (CNN), which is a recursive neural network that comprises a multidimensional array of mainly identical artificial neural cells, wherein (1) each cell is a dynamic subsystem with continuous state variables and (2) each cell is connected to only the few other cells that lie within a specified radius (see Figure 1). Since its original publication in 1988, the concept of the CNN has evolved rapidly and now provides a unified theoretical framework for such computation-intensive applications as signal processing and optimization. The CNN concept provides the architecture for a locally connected, massively parallel computing system with simple synaptic operators - very suitable for VLSI implementation in real-time, high-speed applications. The behavior of a CNN depends on the computing model, network topology, and coefficient template; the behavior in a given application depends primarily on the coefficients of the template (synaptic weights, threshold values, and boundary conditions) and the procedure for applying them.
An OCNN is designed to perform programmable functions for fine-grained processing with annealing control to enhance the quality of the output. Going beyond the basic CNN concept, the OCNN concept incorporates a hardware-based annealing method, according to which the annealing function is embedded in the network. The annealing function is a parallel version of fast mean-field annealing in analog networks, and is highly efficient in finding globally optimal solutions. Whereas synaptic weights in a basic CNN are fixed, those in an OCNN are made digitally programmable to accommodate applications in which predetermined operators are needed.
Other important features of the OCNN concept can be described with reference to the example of a typical intended application, in which a VLSI OCNN processor would receive input from an image-sensing array of photodetectors and would perform such image-data-processing functions as detecting edges or detecting motion. If a basic CNN were to be used in this application, then each pixel would be represented by one neuron. In an OCNN, every pixel could be represented by multiple neurons, that, collectively, would constitute a hyperneuron and would execute the maximum evolution function for selecting among various profiles or exploiting data synergy. For instance, in an OCNN designed to detect motion, every image pixel would be represented by multiple, mutually exclusive neurons in a hyperneuron for velocity selection. Only the winning neuron within each hyperneuron would remain active, while the others would be turned off. The operation of the neural network would be terminated when the energy function of the network reached a minimum. To improve the global interconnections and the input and output of image data with external circuitry beyond those of a basic CNN, an OCNN could be integrated with optical receivers and transmitters (see Figure 2)
VLSI neural processors of this type would likely also prove useful in terrestrial robotic systems and in other systems in which there are requirements for "smart" sensor subsystems and/or high-performance on-board computing for optimization, autonomous control, and/or reduction of data. Preliminary studies have shown that the OCNN concept offers the potential for orders-of-magnitude improvements in performance in onboard computation for autonomous control and reduction of data from multiple sensors.
This work was done by Wai-Chi Fang, Bing J. Sheu, and James Wall of Caltech for NASA's Jet Propulsion Laboratory. For further information, access the Technical Support Package (TSP) free on-line at www.techbriefs.com under the Electronic Systems category, or circle no. 113on the TSP Order card in this issue to receive a copy by mail ($5 charge).
In accordance with Public Law 96-517, the contractor has elected to retain title to this invention. Inquiries concerning rights for its commercial use should be addressed to:
Technology Reporting Office
JPL
Mail Stop 122-116
4800 Oak Grove Drive
Pasadena, CA 91109
(818) 354-2240
Refer to NPO-19989, volume and number of this NASA Tech Briefs issue, and the page number.
NASA Tech Briefs, January 1998