Optical: systems and elements – Holographic system or element – Using a hologram as an optical element
Patent
1992-06-11
1995-06-27
Ben, Loha
Optical: systems and elements
Holographic system or element
Using a hologram as an optical element
359 1, 359 11, 359561, G02B 532, G03H 112
Patent
active
054284660
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention relates to neural networks and in particular to optical neural networks.
2. Related Art
Neural networks are networks modeled on the interconnections of neurons in animals in which each node of the network, corresponding to a neuron, has an output which depends on the total value of inputs to that node, each input being given a weight value.
One example of the neural network is the single layer neural network in which each neural node sums the weighted values of a set of inputs. In such networks, in particular, the number of interconnections between input and output nodes grows rapidly as the number of output nodes increases. As pointed out by Mitsubishi Electric Corporation in their article in JETRO, Mar. 1989 entitled "Optical Neurochip Developed" the minimum number of nodes for a commercially realisable neural network is of the order of 1000 which requires 1 million interconnections, a number with which existing LSI circuit technologies cannot cope.
The approach of Mitsubishi was to interconnect the input and output nodes by means of optical beams. They constructed an optical neural network in which a single neuron was represented by a line detector which provided a summed output of light inputs as the input for a threshold comparator, a column of line optical sources in the form of light omitting diodes provide inputs to the neuron, and an optical mask between the line sources and line detectors provides a light intensity from a particular diode corresponding to its weight value in the neural net. The remaining output nodes were provided by further rows of optical detectors and associated threshold detectors positioned to receive optical input from the columns of light emitting diodes via the optical mask. Whether a particular input optical beam impinged on a photodetector was determined by switching the light emitting diode on or off.
SUMMARY OF THE INVENTION
According to the present invention, a neural network has at least one node responsive to the weighted sum of a plurality of inputs to provide an output dependant on said weighted sum, the node comprising an optical detecting means and the inputs comprising input optical beams whose intensities determine the weight values of the inputs, the network including a holographic means for generating the input optical beams and an array of optical modulators for controlling which of the input optical beams impinge on the optical detectors.
The use of a holographic means to provide optical beams allows the provision of the required optical interconnections without the need to fabricate an array of light emitting diodes. As the modulators have a lower power consumption than light emitting diodes, this method of forming the optical beams allows the neural network elements to be fabricated in higher densities and be driven at a faster rate than possible before.
The optical detecting means may be a series of optical detectors or a single line detector, for example.
The structure of the present invention also allows the modulators and optical detectors to be formed as fixed, compact devices in a neural network the programing of which can be readily changed by changing the intensity distribution of optical beams generated externally to them. In particular the removal of the need for a mask between the inputs and neural nodes permits the use of the particularly compact structure of modulators and optical detectors described fully in the applicant's co-pending application GB 8926183, namely one in which they are formed on respective sides of a common substrate, for example based on multiple quantum well structures formed by double-sided epitaxy.
The single source of optical power required to illuminate a hologram of the holographic means needs to be as powerful as the sum of the powers of the equivalent number of light emitting diodes but it has the advantage that the optical beams are more readily and simultaneously matched to detectors' sensitive wavelength. The hologram may provide equal in
REFERENCES:
patent: 5004309 (1991-04-01), Caulfield et al.
patent: 5068801 (1991-11-01), Clark et al.
patent: 5121231 (1992-06-01), Jenkins et al.
patent: 5129041 (1992-07-01), Pernick et al.
IEEE International Conference on Neural Networks, 24-27 Jul. 1988, San Diego, Calif., IEEE, (New York, US), J. Singh et al: "Implementation of Neural Networks Using Quantum Well Based Excitonic Devices--Device Requirement Studies."
Optics Letters, vol. 13, No. 3, Mar. 1988, Optical Society of America, (New York, US), J. S. Jang et al: "Optical implementation of the Hopfield model for two-dimensional associative memory:", pp. 248-250.
Scientific American, vol. 256, No. 3, Mar. 1987, (New York, US), Y. S. Abu-Mostafa et al: "Optical neural computers", pp. 66-73.
Healey Peter
Rejman-Greene Marek A. Z.
Scott Edward G.
Webb Roderick P.
Wood David C.
Ben Loha
British Telecommunications public limited company
Collins Darryl J.
LandOfFree
Neural networks does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Neural networks, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Neural networks will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-291177