Patent
1993-05-10
1995-02-14
MacDonald, Allen R.
395 24, 395 25, G06F 1518
Patent
active
053902858
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND
This application is related to co-pending commonly assigned application Ser. No. 07/859,698, filed Jun. 11, 1992 naming Rejman-Greene et al as inventors.
I. Field of the Invention
This invention relates to a method of training a neural network having an input for inputting input vectors, an output for outputting output vectors and an adjustable response determining means for determining which output vector is output by the network in response to the inputting of a given input vector.
II. Related Art and Other Considerations
A neural network can, in general terms, be regarded as a series of nodes, each node providing an output which is some function of the outputs of other nodes to which the node is coupled. For example, a particular node might output a signal at a first level if the weighted sum of the outputs of those other nodes exceeds some set threshold and a signal at a second level if it doesn't. Different nodes may receive the outputs from different sets of other nodes, with weightings and threshold values particular to that node. An input vector is coupled to a set of the nodes via the input of the network and an output vector is generated at the output of the network. The response of the neural network to the input vector is determined in this case by the values of the weightings and thresholds which collectively form the response determining means for this type of network.
Other implementations of neural networks may employ techniques different to the weighting and threshold technique described above but will nevertheless have some response determining means which determines the output vectors provided by the particular network in response to input vectors.
An example of an optical implementation of a neural network is described in an article by N. M. Barnes, P. Healey, P. McKee, A. W. O'Neill, M. A. Z. Rejman-Greene, E. G. Scott, R. P. Webb and D. Wood entitled "High Speed Opto-Electronic Neural Network", Electronics Letters 19th July 1990, Vol 26, No. 15, pp 1110-1112. This exploits the possibilities of parallel processing and optical interconnections to provide a high throughput. That is, the neural network has a rapid response to an applied input vector, which in the reported configuration can be clocked at rates in excess of 10 Mbits/s. The network is shown at FIG. 1 of this application. It is a two-layer perception network which can recognise exclusive-or (EXOR) combinations in a pair of input stream A and B.
Its basic operation is to perform a matrix-vector multiplication in which each row of a 4.times.4 optical detector array 80 provides an electrical signal corresponding to the sum of the light intensities impinging on the row.
A computer generated hologram 100 and lenses 102 and 104 generate intensity coded beams (not shown) directly from a single laser source 106 which impinge on the individual modulators of a modulator array 82. The relative intensities of the beams are shown by the numerals on the modulators. These different intensity beams provide the different weightings to be applied to the outputs of the modulators in their connection to the nodes formed by the rows of detectors.
A signal is passed to the next level via pre-amplifiers 84 or not, depending on whether this weighted sum is above or below some set threshold either a fixed bias from a bias supply 86 or relative to one of the other weighted sums.
There are several ways in which the number of intensity coded beams could be generated including, for example, using a fixed mask to filter the intensity of individual beams from an array of sources.
A novel method of determining the weights devised by the applicant (not published at the time of filing this application) for an optical neural computer is based on the fact that the modulation depth that can be achieved at a given wavelength for a multiple quantum well (MQW) modulator is determined not only by the voltage swing of the digital drive but also by the applied bias across the modulators. See FIG. 3 of N. M. Barnes, P. Healey, M. A. Z. Rejman-Greene, E. G.
REFERENCES:
patent: 4914603 (1990-04-01), Wood
patent: 4918618 (1990-04-01), Tomlinson, Jr.
patent: 5004309 (1991-04-01), Caulfield et al.
patent: 5068801 (1991-11-01), Clark et al.
patent: 5099434 (1992-03-01), Stoll et al.
Vector Quantization Codebook Generation Using Simulated Annealing, J. Kelly Flanagan, Darryl R. Morrell . . . IEEE/ 23-26 May 1989.
Image Vector Quantization Using Neural Networks and Simulated Annealing, M. Lech and Y. Hua, IEEE/1992 7-9 Apr. 1992.
Optics Letters, vol. 13, No. 3, Mar. 1988, "Optical Implementation of the Hopfield Model for two-Dimensional Associative Memory," pp. 248-250.
Optical Neural Computers, "Can computers be built to solve problems, such as recognizing patterns, that entail memorizing all possible solutions? The key may be to arrange optical elements in the same way as neurons are arranged in the brain," by Yaser S. Abu-Mostafa et al., pp. 66-73.
IEEE International Conference on Neural Networks, Jul. 24-27, 1988, "Implementation of Neural Networks Using Quantum Well Based Excitonic Devices--Device Requirement Studies" by Singh et al., pp. 411-419.
Report, "Optische Neurocomputer" 1989, by Janosch, pp. 38-44.
Electronics Letters, 19 Jul. 1990, vol. 26, No. 15, "Novel Learning Method for Analogue Neural Networks," pp. 1136-1137.
Electronics Letters, 19 Jul. 1990, vol. 26, No. 15, "16-Channel parallel Optical Interconnect Demonstration with InGaAs/InP MQW Modulator Array," pp. 1126-1127.
Applied Optics, 1 Nov. 1988, vol. 27, No. 1, "Rapid Communications, parallel optoelectronic realization of neural networks models using CID technology" by Agranat et al., pp. 4354-4355.
IEEE International Conference on Neural Networks, Jul. 24-27, 1988, "Bimodal Stochastic Optical Learning Machine" by Farhat et al., pp. 365-372.
1989 Optical Society of America, "Programmable quadratic associative memory using holographic lenslet arrays" by Jang et al., pp. 838-840.
Electronics Letters, 19 Jul. 1990, vol. 26, No. 15, "High Speed Opto-electronic Neural Network," pp. 1110-1112.
Applied Optics, 10 Jun. 1990, vol. 29, No. 17, "Adaptive-clustering optical neuralnet" by Casasent et al., pp. 2603-2615.
PCT Search Report of PCT/GM 90/01782.
British Telecommunications public limited company
Dorvil Richemond
MacDonald Allen R.
LandOfFree
Method and apparatus for training a neural network depending on does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and apparatus for training a neural network depending on , we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and apparatus for training a neural network depending on will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-294201