WO2002025395A2 - Optical computing - Google Patents

Optical computing Download PDF

Info

Publication number
WO2002025395A2
WO2002025395A2 PCT/US2001/029343 US0129343W WO0225395A2 WO 2002025395 A2 WO2002025395 A2 WO 2002025395A2 US 0129343 W US0129343 W US 0129343W WO 0225395 A2 WO0225395 A2 WO 0225395A2
Authority
WO
WIPO (PCT)
Prior art keywords
switching elements
array
optical switching
optical
inputs
Prior art date
Application number
PCT/US2001/029343
Other languages
French (fr)
Other versions
WO2002025395A3 (en
Inventor
Edward A. Rietman
Original Assignee
Starlab Nv/Sa
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Starlab Nv/Sa filed Critical Starlab Nv/Sa
Priority to AU2001292821A priority Critical patent/AU2001292821A1/en
Publication of WO2002025395A2 publication Critical patent/WO2002025395A2/en
Publication of WO2002025395A3 publication Critical patent/WO2002025395A3/en

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means

Definitions

  • This invention relates to optical computing.
  • Optical switching is used in routing Internet and telecommunication traffic.
  • the optical switch accepts signals from one array of optical fibers and routes the signals to another array of optical fibers.
  • the switching operation should be done quickly, accurately, and with little cross-talk.
  • Optical switching can be achieved by various types of spatial light modulators (SLMs).
  • SLMs spatial light modulators
  • a spatial light modulator is an optical device that modulates the intensity of a beam of light passing through or reflected from it. Exactly as the name implies, a spatial light modulator modulates the light in a spatial pattern (i.e., a specific pattern in 2-dimensional space).
  • a lens projects an image from a one-dimensional array of optical fibers onto the film.
  • a second lens projects the light collected from this film onto another one-dimensional array of optical fibers. Because the pattern of light and dark pixels on the film has the ability to modulate the light in a specific spatial arrangement, the end result is that the optical signal is transferred from one fiber to the second optical fiber.
  • Each of the pixels on the film can be considered a light modulating or optical switching element and the film can be considered an SLM array of the pixels.
  • the film together with the lenses, the optical fibers, and the controller can be considered an optical switch.
  • SLM arrays An overview of three types of SLM arrays and their use in an all-optical telecommunication network is presented in Fairley, P., "The Microphotonics Revolution” MIT Technology Review, 38-44, July- August, 2000.
  • One type of SLM array is based on microscopic bubbles (magneto-optic effect) with controllable orientation. The bubbles act as ratable mirrors that change deflection of light beams.
  • Another approach uses a spatial light modulating liquid crystal display.
  • the third described approach involves micro-mirrors in the form of individually addressable reflective films that can be controlled by electrostatic charges in the 10- Volt range.
  • SLM arrays are typically of low resolution and require arrays of emitters and detectors. For these reasons, Kornfeld, C. D.; Frye, R. C; Wong, C. C. and Rietman, E. A., "An Optically Programmed Neural Network,” IEEE Second International Conference on Neural Networks, Vol. 2, p. 357-364, 1988, used a pre-computed photographic film mask for interconnection.
  • Their optical neural network computer 10 shown in Figure 1, included a photoconductive array 12 for detection of light and a computer controlled CRT 14 for projecting light to the arrays.
  • Adaptive connections between operational amplifier- based artificial neurons 16 are made by modulating the intensity of beams of light on the photoconductive array.
  • the array can be updated at video frequency (about 60 Hz).
  • each tiny mirror can be positioned (rotated) by electrical signals to send an incident light beam to another mirror.
  • the light beam is reflected off a larger mirror and back onto a micro-mirror.
  • the second micro-mirror is positioned by electrical signals to send the optical signal (i.e. the light beam) to a second optical fiber.
  • the signal intensity transmitted from one micro-mirror to a second micro-mirror is optimized accurately.
  • the invention features apparatus that includes an optical switch having an array of controllable optical switching elements, circuitry adapted to control each of the optical switching elements in the array in accordance with an adaptive computation to be performed on inputs to produce outputs, the control of individual switching elements not being optimized, and optical devices adapted to emit and collect light that carries the inputs and outputs to and from the switching elements.
  • the optical switching elements may include mirrors (e.g., micro-mirrors) the orientations of which are controlled (along two orthogonal dimensions) by the circuitry.
  • the orientations may be controlled to represent respective elements of a matrix that is part of the computation.
  • the computation may include a vector-matrix multiplication in which the inputs and outputs represent vectors, and the matrix may be represented by control of the optical switching elements in the array.
  • the adaptive computation may include a pattern recognizing neural network in which the neural network connections represented by the array are pre-computed or are learned.
  • the network may include a Hopfield network or a Boltzmann machine.
  • the circuitry may control the switching elements over a continuous analog range, or in a digital mode.
  • optical switch array in particular the micro-n irror array, is faster than typical SLM interconnection systems and has better resolution.
  • Figure 1 is a diagram of an optical computer.
  • Figure 2 is a diagram of a micro-mirror array and other elements.
  • Figure 3 is an example of patterns for pattern recognition.
  • Figure 4 is a diagram of an optical computer.
  • Figure 5 is of a micro-lens array and optical splitters.
  • Figure 6 is a graph of temperature curves.
  • FIG 7 is a diagram of portions of a Boltzmann machine.
  • micro-mirror arrays 20 one of which is shown in figure 2 can be used as the SLM interconnection components for an adaptive computing system.
  • the micro-mirrors (the optical switching elements) in the micro-mirror arrays can be updated at a high frequency, for example, 1.5 MHz (see Goossen, K. W., Walker, J. A., Frigo, N., Iannone, P., Arney, S. C, Macdonald, M., Ruel, R., Bishop, D. and Boie, B., "Integrated Mechanical Anti-Reflection Switch (MARS) Device for Fiber-to-the-home Applications," Page 169-178 in Proceedings on Sensors, April 1996, Anaheim, CA.)
  • the optical switch that includes the array 20 can be made with bundles of optical fibers 22 instead of free-space optical connections.
  • incoming and outgoing optical signals are brought into and out of the mirror array using the optical fibers and micro-lenses 28 to emit and collect the light to and from the mirrors 30.
  • a light beam 34 is emitted from one of the lenses 36 and is reflected at a corresponding micro-mirror 38. From there, the light beam travels to a larger mirror 40 which then transfers the beam to a second micro-mirror 42 in the array.
  • Mirror 42 reflects the beam to an outgoing optical fiber 44 via a corresponding micro-lens 46.
  • the overall result is the transfer of an optical signal from one optical fiber 48 to another 44.
  • Mirror alignment is not a trivial issue.
  • the mirrors are positioned to achieve the desired reflections by electrostatic signals applied to appropriate control wires 49 (Goossen, 1996).
  • Each of the micro-mirrors and the larger mirror has two control wires for x-axis rotation and two control wires for y-axis rotation. So to transfer any given optical signal from one fiber to another requires twelve control voltages to be optimized for maximum optical signal intensity. For critical applications in optical switching for telecommunication, this would require a high degree of precision in maximizing the signal intensity.
  • the need for optimization of each of the array elements can be circumvented in adaptive (e.g., neural network) computing applications because the network is inherently fault tolerant.
  • the lack of precision in positioning allows the mirror arrays to be operated as analog devices thus enabling construction of an analog neural network for optical signal processing.
  • the system can be used with analog electrical signals converted to optical for conventional signal processing and for analog electrical neural network computing.
  • optical switches we can exploit the inherent sloppiness of the mirrors because the optical signal from several mirrors is added and used in the computation. If one mirror is slightly out of position then another in the array will compensate for it. So we do not need to carefully tune each mirror. Instead we tune all of them at the same time to give the desired output.
  • vector-matrix multiplication For purposes of later discussion, we define vector-matrix multiplication as follows:
  • y ⁇ ⁇ ij ⁇ j This is implemented in our optical switch using n inputs (the x's) and n outputs (the y's).
  • the total number of mirrors used in the array would be n(nA-2), one for each of the elements in the matrix and one for each of the input and output vectors.
  • the elements in the resulting matrix (M) are thresholded by setting any value over one equal to one, and all the elements on the diagonal are set equal to zero, irrespective of their values.
  • a ⁇ l ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ l ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ l ⁇ 3 ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ 3 ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ 3 ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ 3 ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ 0 ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇ , ⁇ l ⁇
  • the resulting matrix (M) is the connection matrix (A) for the Hopfield neural network:
  • This matrix becomes the connection matrix for our optical switch array.
  • the output vector is the product of the input vector and the above matrix.
  • the input vector could be a corrupted version of the pattern to be recognized and the output will be the completed vector pattern.
  • the computer system includes an analog signal input 70 to a neural network 72 and a digital computer 74 that controls the neural network.
  • the neural network 72 includes an optical switch array 76 linked by optical fibers 78 to optical emitters and detectors 80.
  • the positions of the mirrors are controlled by electrostatic signals provided by mirror control electronics 82.
  • the electronics 82 are governed by signals from the computer 74 to provide the desired matrix computations using the optical switch array.
  • the analog signal output 84 is provided from the optical detectors 80.
  • Some implementation details are shown in Figure 5.
  • the fan-out of the input signals is equal to the number of output signals for the system.
  • the bias input is used to adjust the optical signals so that each signal has a "background" level of signal. This bias signal is added to an initial noise signal so the initial light intensity at the mirror is shifted according to the bias. Then the learning algorithm is used to modify this shifted bias signal.
  • the learning algorithm includes a search in a dynamical system space.
  • This search space is the space of connections, and can be done synchronously (i.e., all connections in the array at the same time) or asynchronously (one connection at a time).
  • the output from the individual neurons is computed using a statistical probability as follows: First the same vector-matrix operation
  • a Boltzman machine implementation is shown in figure 7. Though it could be used for pattern recognition, like the Hopfield network, it may also be used for regression mapping and non-square matrices may be involved. For example, in the case shown in figure 7, there are three inputs 90, 92, 94, one bias input 96, and three output nodes 98. Each input is fed through a 1 :3 splitter 100 to form three beams. The three beams are delivered respectively through three lenses 102 to three corresponding mirrors in the array. The mirrors are oriented to cause reflection onto the large mirror 102 and then onto the three output mirrors in appropriate combinations to generate the desired outputs 98.
  • n represents the inputs and m the outputs then this will require (nA-J)m+n mirrors in the array, one for each element in the input and output vectors and in the matrix, to complete the Boltzmann machine for a non-square matrix.
  • SLM arrays other than micro-mirror arrays can be used.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Neurology (AREA)
  • Mechanical Light Control Or Optical Switches (AREA)
  • Optical Communication System (AREA)

Abstract

Each of the optical switching elements in an array of controllable optical switching elements is controlled in accordance with an adaptive computation to be performed on inputs to produce outputs. Light that carries the inputs and outputs is emitted to and collected from the optical switching elements.

Description

OPTICAL COMPUTING
Background
This invention relates to optical computing.
Optical switching is used in routing Internet and telecommunication traffic. For this purpose, the optical switch accepts signals from one array of optical fibers and routes the signals to another array of optical fibers. Ideally the switching operation should be done quickly, accurately, and with little cross-talk.
Optical switching can be achieved by various types of spatial light modulators (SLMs). A spatial light modulator is an optical device that modulates the intensity of a beam of light passing through or reflected from it. Exactly as the name implies, a spatial light modulator modulates the light in a spatial pattern (i.e., a specific pattern in 2-dimensional space).
For example, one could achieve optical switching using a piece of photographic film (a very simple spatial light modulator) that has been exposed with different intensities in different regions (pixels). To use the piece of film for optical switching, a lens projects an image from a one-dimensional array of optical fibers onto the film. A second lens projects the light collected from this film onto another one-dimensional array of optical fibers. Because the pattern of light and dark pixels on the film has the ability to modulate the light in a specific spatial arrangement, the end result is that the optical signal is transferred from one fiber to the second optical fiber. Each of the pixels on the film can be considered a light modulating or optical switching element and the film can be considered an SLM array of the pixels. The film together with the lenses, the optical fibers, and the controller can be considered an optical switch. An overview of three types of SLM arrays and their use in an all-optical telecommunication network is presented in Fairley, P., "The Microphotonics Revolution" MIT Technology Review, 38-44, July- August, 2000. One type of SLM array is based on microscopic bubbles (magneto-optic effect) with controllable orientation. The bubbles act as ratable mirrors that change deflection of light beams. Another approach uses a spatial light modulating liquid crystal display. The third described approach involves micro-mirrors in the form of individually addressable reflective films that can be controlled by electrostatic charges in the 10- Volt range.
Using certain kinds of spatial light modulating (SLM) arrays for adaptive computing has been described by Collins, W. C; Stilwell, Jr., P. D.; and Athale, R. A, "Optical Matrix-Matrix Multiplier Based on Outer Product Decomposition," US Patent 4,569,033; by Budil, M. "Optical Vector Multiplier for Neural Networks," US Patent 5,784,309; and by Horn, P. and Arimoto, A., "Optical Neural Network System," US Patent 5,394,257. SLM arrays are typically of low resolution and require arrays of emitters and detectors. For these reasons, Kornfeld, C. D.; Frye, R. C; Wong, C. C. and Rietman, E. A., "An Optically Programmed Neural Network," IEEE Second International Conference on Neural Networks, Vol. 2, p. 357-364, 1988, used a pre-computed photographic film mask for interconnection.
Later, Rietman, E. A.; Frye, R. C; Wong, C. C. and Kornfeld, C. D., "Amorphous Silicon Photoconductive Arrays for Artificial Neural Networks," Appl Opt. 28, 3474-3478, 1989, and Rietman, E. A; Frye, R. C. and Wong, C. C, "Signal Prediction by an Optically Controlled Neural Network," Appl. Opt, 30, 950-957, 1991, reported using a computer controlled projection cathode ray tube (CRT) and an array of amorphous silicon photoconductors. Their optical neural network computer 10, shown in Figure 1, included a photoconductive array 12 for detection of light and a computer controlled CRT 14 for projecting light to the arrays. Adaptive connections between operational amplifier- based artificial neurons 16 are made by modulating the intensity of beams of light on the photoconductive array. The array can be updated at video frequency (about 60 Hz).
In a micro-mirror array (such as the ones available from Lucent Technologies), each tiny mirror can be positioned (rotated) by electrical signals to send an incident light beam to another mirror. When used for switching applications, like telephony, the light beam is reflected off a larger mirror and back onto a micro-mirror. The second micro-mirror is positioned by electrical signals to send the optical signal (i.e. the light beam) to a second optical fiber. For switching applications, the signal intensity transmitted from one micro-mirror to a second micro-mirror is optimized accurately.
Summary
In general, in one aspect, the invention features apparatus that includes an optical switch having an array of controllable optical switching elements, circuitry adapted to control each of the optical switching elements in the array in accordance with an adaptive computation to be performed on inputs to produce outputs, the control of individual switching elements not being optimized, and optical devices adapted to emit and collect light that carries the inputs and outputs to and from the switching elements.
Implementations of the invention may include one or more of the following features. The optical switching elements may include mirrors (e.g., micro-mirrors) the orientations of which are controlled (along two orthogonal dimensions) by the circuitry. The orientations may be controlled to represent respective elements of a matrix that is part of the computation. The computation may include a vector-matrix multiplication in which the inputs and outputs represent vectors, and the matrix may be represented by control of the optical switching elements in the array. The adaptive computation may include a pattern recognizing neural network in which the neural network connections represented by the array are pre-computed or are learned. The network may include a Hopfield network or a Boltzmann machine. The circuitry may control the switching elements over a continuous analog range, or in a digital mode.
Among the advantages of the invention are one or more of the following. The inherent analog nature of the optical signals is exploited. (Of course, this analog computation application does not preclude the application in digital optical computing using the vector-matrix multiplication application.) The use of the optical switch array, in particular the micro-n irror array, is faster than typical SLM interconnection systems and has better resolution.
Other advantages and features will become apparent from the following description and from the claims.
Description
Figure 1 is a diagram of an optical computer.
Figure 2 is a diagram of a micro-mirror array and other elements.
Figure 3 is an example of patterns for pattern recognition.
Figure 4 is a diagram of an optical computer.
Figure 5 is of a micro-lens array and optical splitters.
Figure 6 is a graph of temperature curves.
Figure 7 is a diagram of portions of a Boltzmann machine. As shown in figure 2, micro-mirror arrays 20 (one of which is shown in figure 2) can be used as the SLM interconnection components for an adaptive computing system. The micro-mirrors (the optical switching elements) in the micro-mirror arrays can be updated at a high frequency, for example, 1.5 MHz (see Goossen, K. W., Walker, J. A., Frigo, N., Iannone, P., Arney, S. C, Macdonald, M., Ruel, R., Bishop, D. and Boie, B., "Integrated Mechanical Anti-Reflection Switch (MARS) Device for Fiber-to-the-home Applications," Page 169-178 in Proceedings on Sensors, April 1996, Anaheim, CA.)
The optical switch that includes the array 20 can be made with bundles of optical fibers 22 instead of free-space optical connections. As shown in figure 2, incoming and outgoing optical signals are brought into and out of the mirror array using the optical fibers and micro-lenses 28 to emit and collect the light to and from the mirrors 30. For example, a light beam 34 is emitted from one of the lenses 36 and is reflected at a corresponding micro-mirror 38. From there, the light beam travels to a larger mirror 40 which then transfers the beam to a second micro-mirror 42 in the array. Mirror 42 reflects the beam to an outgoing optical fiber 44 via a corresponding micro-lens 46. The overall result is the transfer of an optical signal from one optical fiber 48 to another 44.
Mirror alignment is not a trivial issue. The mirrors are positioned to achieve the desired reflections by electrostatic signals applied to appropriate control wires 49 (Goossen, 1996). Each of the micro-mirrors and the larger mirror has two control wires for x-axis rotation and two control wires for y-axis rotation. So to transfer any given optical signal from one fiber to another requires twelve control voltages to be optimized for maximum optical signal intensity. For critical applications in optical switching for telecommunication, this would require a high degree of precision in maximizing the signal intensity. However, the need for optimization of each of the array elements can be circumvented in adaptive (e.g., neural network) computing applications because the network is inherently fault tolerant. Also, the lack of precision in positioning allows the mirror arrays to be operated as analog devices thus enabling construction of an analog neural network for optical signal processing. Or the system can be used with analog electrical signals converted to optical for conventional signal processing and for analog electrical neural network computing. To use these optical switches as network and adaptive computation engines we can exploit the inherent sloppiness of the mirrors because the optical signal from several mirrors is added and used in the computation. If one mirror is slightly out of position then another in the array will compensate for it. So we do not need to carefully tune each mirror. Instead we tune all of them at the same time to give the desired output.
Vector-matrix multiplication definition
For purposes of later discussion, we define vector-matrix multiplication as follows:
If we are given a vector x of ^-dimensions, we can find_y, the product of this vector and an n X n matrix, A, from the following relation:
y = Ax
The resulting vector j has ^-dimensions. In term-by-term form this equation is
Figure imgf000008_0001
yι = ∑ ijχj This is implemented in our optical switch using n inputs (the x's) and n outputs (the y's). The total number of mirrors used in the array would be n(nA-2), one for each of the elements in the matrix and one for each of the input and output vectors.
Hopfield networks example
As a first example of a neural network built from an optical switch array consider the following: We want a pattern recognition system that will recognize the three specific 16-bit stored patterns 60, 62, 64 shown in figure 3. In building a Hopfield neural network for this purpose, we need preliminarily to compute the product of each of the three pattern vectors (v) and its transpose (v7) and the sum of the resulting products.
M = ∑ W patterns
The elements in the resulting matrix (M) are thresholded by setting any value over one equal to one, and all the elements on the diagonal are set equal to zero, irrespective of their values.
As a specific example of these operations, we will compute the matrix for the example shown in figure 3. Using the programming language Mathematica ® (Wolfram Research, Champaign, Illinois) we get the following (where a, b, and c are the initial pattern vectors):
a={{l},{0},{0},{0},{l},{0},{0},{0},{l}3{0},{0},{0},{l},{l},{l},{l}}
b={{l},{l},{l},{l},{0},{0},{0},{l},{0}5{0},{0},{l},{0},{0},{0},{l}}
C={{0},{0},{0},{0},{1},{1},{1},{1},{1},{1},{1},{1},{0},{0},{0},{0}}
ml=a.Transpose[a]
m2=b.Transpose[b] m3=c.Transpose[c]
s=ml+m2+m3
{{2,1,1,1,1,0,0, 1,1,0,0, 1, 1, 1,1,2}, {1, 1,1,1,0,0,0,1,0,0,0,1,0,0,0, 1},
{1, i, i, 1 0, 1,0,0,0,1,0,0,0, 1}, {1, i, i, 0; 0, 1,0,0,0,1,0,0,0, 1},
{1,0,0, 2. 1, 1,2, 1, 1, 1, 1, 1,1, 1}, {0, o, 0, 1 1, 1, 1, 1, 1, 1, 0, 0, 0, 0}, {0, o, 0, 1 1, 1, 1, 1, 1, 1, 0, 0, 0, 0},
{i, i, i, 1 1,2,1,1,1,2,0,0,0,1},
{1,0,0, 2. 1, 1,2,1, 1,1,1,1,1, 1}, {0, o, 0, 1 1, 1, 1, 1, 1, 1, 0, 0, 0, 0}, {0, o, 0, 1 1, 1, 1, 1, 1, 1, 0, 0, 0, 0},
{i, i, i, 1 1,2,1,1,1,2,0,0,0, 1},
{1,0,0, 1 0,0,1,0,0,0,1,1,1,1}, {1,0,0, 1 0,0,1,0,0,0, 1, 1,1, 1}, {1,0,0, 1 0,0, 1,0,0,0, 1, 1, 1, 1}, {2, 1, 1, 1,1 0,1,1,0,0, 1, 1,1,1,2}}
Now we threshold the values above one and set the diagonal elements to zero. The resulting matrix (M) is the connection matrix (A) for the Hopfield neural network:
{{0,1,1,1,1,0,0,1,1,0,0,1, 1,1,1, 1}, {1,0,1,1,0,0,0, 1,0,0,0,1,0,0,0, 1}, {1, 1,0,1,0,0,0, 1,0,0,0,1,0,0,0, 1},
{1, 1,1,0,0,0,0, 1,0,0,0,1,0,0,0, 1}, {1,0,0,0,0,1,1, 1,1,1,1,1, 1, 1,1, 1}, {0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0}, {1, i, i, i, i, 1,0,1,1,1,1,0,0,0,1},
{1,0,0,0,1, 1,1,0,1,1,1,1,1,1,1}, {0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0},
{1, i, i, i, i, 1,1,1,1,1,0,0,0,0,1},
{1,0,0,0,1,00,0,1,0,0,0,0, 1,1,1}, {1,0,0,0,1,00,0,1,0,0,0, 1,0,1,1}, {1,0,0,0,1,00,0,1,0,0,0, 1, 1,0,1},
{1,1, 1,1,1,00, 1, 1, 0, 0, 1, 1, 1, 1, 0}}
This matrix becomes the connection matrix for our optical switch array. In this case, with «=16 (for 16 inputs and 16 outputs) we would need a total of n(n+2) mirrors (1088), and the computation is done exactly like the vector-matrix product example. The output vector is the product of the input vector and the above matrix.
Figure imgf000011_0001
In this case, the input vector could be a corrupted version of the pattern to be recognized and the output will be the completed vector pattern.
As shown in figure 4, the computer system includes an analog signal input 70 to a neural network 72 and a digital computer 74 that controls the neural network. The neural network 72 includes an optical switch array 76 linked by optical fibers 78 to optical emitters and detectors 80.
The positions of the mirrors are controlled by electrostatic signals provided by mirror control electronics 82. The electronics 82 are governed by signals from the computer 74 to provide the desired matrix computations using the optical switch array. The analog signal output 84 is provided from the optical detectors 80. Some implementation details are shown in Figure 5. Here we see for a 3 X 3 network 50 how it is necessary to split (or fan-out) the input signals 52 into n=3 input fibers 54. This is because of the nature of vector matrix multiplication. The fan-out of the input signals is equal to the number of output signals for the system. The bias input is used to adjust the optical signals so that each signal has a "background" level of signal. This bias signal is added to an initial noise signal so the initial light intensity at the mirror is shifted according to the bias. Then the learning algorithm is used to modify this shifted bias signal.
Boltzmann machines examples
Unlike the Hopfield network, the connections in the Boltzmann machine (cf. Aarts E., and Korst, J., Simulated Annealing and Boltzmann Machines, John Wiley, New York, 1989; and Rietman, E. A., Exploring Parallel Processing, Windcrest Books, McGraw-Hill, 1990) are not pre- computed, but rather are "learned." One advantage of learning is that we can now exploit the inherent inefficiencies of the micro-mirrors. For example, the mirror arrays require fine-tuning to optimize the maximum signal. This implies that the signal is actually a scalar number within some range of the maximum. The ability to achieve a scalar number within a range enables the use of scalar connection values, rather than merely a binary choice between maximum or nothing as typical optical switching systems use (that is, we can achieve analog computation versus digital computation).
The learning algorithm includes a search in a dynamical system space. This search space is the space of connections, and can be done synchronously (i.e., all connections in the array at the same time) or asynchronously (one connection at a time). In the case of our mirror array, we change the voltages on the mirrors to reduce the difference between the output and some target value representing the pattern to be recognized.
The output from the individual neurons is computed using a statistical probability as follows: First the same vector-matrix operation
Figure imgf000013_0001
is conducted, except that the output at each neuron is updated according to the following statistical probability:
y,T))
Figure imgf000013_0002
where yt is essentially the net input to the ith neuron. The parameter T, for historical reasons, is called temperature. During the recall process the temperature is decreased in a process similar to freezing slowly. Figure 6 shows a family of probability curves for different temperatures.
A Boltzman machine implementation is shown in figure 7. Though it could be used for pattern recognition, like the Hopfield network, it may also be used for regression mapping and non-square matrices may be involved. For example, in the case shown in figure 7, there are three inputs 90, 92, 94, one bias input 96, and three output nodes 98. Each input is fed through a 1 :3 splitter 100 to form three beams. The three beams are delivered respectively through three lenses 102 to three corresponding mirrors in the array. The mirrors are oriented to cause reflection onto the large mirror 102 and then onto the three output mirrors in appropriate combinations to generate the desired outputs 98. If n represents the inputs and m the outputs then this will require (nA-J)m+n mirrors in the array, one for each element in the input and output vectors and in the matrix, to complete the Boltzmann machine for a non-square matrix.
Other implementations
Other implementations are within the scope of the following claims. SLM arrays other than micro-mirror arrays can be used..

Claims

Claims
1. Apparatus comprising
an optical switch including
an array of controllable optical switching elements,
circuitry adapted to control each of the optical switching elements in the array in accordance with an adaptive computation to be performed on inputs to produce outputs, the control of individual switching elements not being optimized, and
optical devices adapted to emit and collect light that carries the inputs and outputs to and from the switching elements.
2. The apparatus of claim 1 in which the optical switching elements comprise mirrors.
3. The apparatus of claim 2 in which the orientations of the mirrors are controlled by the circuitry.
4. The apparatus of claim 3 in which the orientations are controlled in two orthogonal dimensions.
5. The apparatus of claim 1 in which the optical signals are controlled to represent respective elements of a matrix that is part of the computation.
6. The apparatus of claim 2 in which the optical switching elements comprise micro-mirrors.
7. The apparatus of claim 1 in which the computation comprises a vector-matrix multiplication.
8. The apparatus of claim 7 in which the inputs and outputs represent vectors, and the matrix is represented by control of the optical switching elements in the array.
10. The apparatus of claim 1 in which the adaptive computation comprises a pattern recognizing neural network.
11. The apparatus of claim 10 in which neural network connections represented by the array are pre-computed.
12. The apparatus of claim 10 in which the neural network connections represented by the array are learned.
13. The apparatus of claim 10 in which the network comprises a Hopfield network or a Boltzmann machine.
14. The apparatus of claim 1 in which the network in which the circuitry controls the switching elements over a continuous analog range.
15. A method comprising
controlling each of the optical switching elements in an array of controllable optical switching elements in accordance with an adaptive computation to be performed on inputs to produce outputs, and
emitting and collecting light that carries the inputs and outputs to and from the optical switching elements.
16. A method comprising
configuring an array of optical switching elements in accordance with an adaptive algorithm for recognizing patterns,
optically applying a version of one of the patterns to the array of optical switching elements, and optically receiving a resulting recognized one of the patterns from the array of optical switching elements.
17. A method comprising
configuring an array of optical switching elements in accordance with a vector matrix computation,
optically applying an input vector to the array of optical switching elements, and
optically receiving a resulting output vector from the array of optical switching elements.
PCT/US2001/029343 2000-09-21 2001-09-20 Optical computing WO2002025395A2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2001292821A AU2001292821A1 (en) 2000-09-21 2001-09-20 Optical computing

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US66671500A 2000-09-21 2000-09-21
US09/666,715 2000-09-21

Publications (2)

Publication Number Publication Date
WO2002025395A2 true WO2002025395A2 (en) 2002-03-28
WO2002025395A3 WO2002025395A3 (en) 2002-06-13

Family

ID=24675144

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2001/029343 WO2002025395A2 (en) 2000-09-21 2001-09-20 Optical computing

Country Status (2)

Country Link
AU (1) AU2001292821A1 (en)
WO (1) WO2002025395A2 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112180583A (en) * 2020-10-30 2021-01-05 中国工程物理研究院激光聚变研究中心 Self-adaptive optical system based on all-optical neural network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5448749A (en) * 1989-11-22 1995-09-05 Mitsubishi Denki Kabushiki Kaisha Data processing apparatus with optical vector matrix multiplier and peripheral circuits
GB2317487A (en) * 1996-09-21 1998-03-25 Rupert Charles David Young Hybrid digital/optical system for pattern recognition

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5448749A (en) * 1989-11-22 1995-09-05 Mitsubishi Denki Kabushiki Kaisha Data processing apparatus with optical vector matrix multiplier and peripheral circuits
GB2317487A (en) * 1996-09-21 1998-03-25 Rupert Charles David Young Hybrid digital/optical system for pattern recognition

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112180583A (en) * 2020-10-30 2021-01-05 中国工程物理研究院激光聚变研究中心 Self-adaptive optical system based on all-optical neural network

Also Published As

Publication number Publication date
WO2002025395A3 (en) 2002-06-13
AU2001292821A1 (en) 2002-04-02

Similar Documents

Publication Publication Date Title
JP3707804B2 (en) Optical information processing equipment
US4859012A (en) Optical interconnection networks
US5121231A (en) Incoherent/coherent multiplexed holographic recording for photonic interconnections and holographic optical elements
US8854423B2 (en) Multiplanar volumetric three-dimensional display apparatus
US5256869A (en) Free-space optical interconnection using deformable mirror device
US5056039A (en) Holographic interconnect system
US5132813A (en) Neural processor with holographic optical paths and nonlinear operating means
Collins et al. Deformable mirror device spatial light modulators and their applicability to optical neural networks
Saxena et al. Adaptive multilayer optical neural network with optical thresholding
GB2231989A (en) Neural networks
WO2002025395A2 (en) Optical computing
McAulay Spatial-light-modulator interconnected computers
CN116258624A (en) Optical tensor convolution computing system and method based on multi-imaging projection architecture
US6714568B2 (en) Active optical system for beam-steering a laser beam
Saxena et al. Optical implementations
Hermans et al. Towards trainable media: Using waves for neural network-style training
US5440671A (en) Neural net model and neural net optimizing method and system for realizing the same
US5412755A (en) Optical implementation of inner product neural associative memory
WO2000072104A1 (en) Optical processor architecture
Kirk et al. Design and fabrication of computer-generated holographic fan-out elements for a matrix/matrix interconnection scheme
Saxena et al. Adaptive Multilayer Optical Neural Network Design
Rietman et al. Signal prediction by an optically controlled neural network
KR102622313B1 (en) Opto-electronic module using spatial light modulator and optical computer including the same
Morris et al. Reconfigurable optical interconnects by a combined computer-generated hologram and spatial light modulator method
Pavlov An algebraic model for application of optical neuro-fuzzy systems to networking problems

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A2

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NO NZ PH PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG US UZ VN YU ZA ZW

AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG

AK Designated states

Kind code of ref document: A3

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NO NZ PH PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG US UZ VN YU ZA ZW

AL Designated countries for regional patents

Kind code of ref document: A3

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG

121 Ep: the epo has been informed by wipo that ep was designated in this application
REG Reference to national code

Ref country code: DE

Ref legal event code: 8642

122 Ep: pct application non-entry in european phase
NENP Non-entry into the national phase in:

Ref country code: JP