Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural...

35
Optical Neural Optical Neural Network Network By: By: Roghayeh Roghayeh Saeidi Saeidi

Transcript of Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural...

Page 1: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Optical Neural NetworkNetwork

By:By:RoghayehRoghayeh SaeidiSaeidi

Page 2: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 2

MotivationMotivation

-- Massive interconnectionsMassive interconnections-- Parallel processing operationParallel processing operation-- Inherent additive propertiesInherent additive properties

Fast optical computingFast optical computingPower efficient mannerPower efficient manner

ONN Features:ONN Features:

MotivationMotivation

Page 3: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 3

All Optic SystemsAll Optic Systems

A packet switched telecommunication network. A packet switched telecommunication network. -- In each network router data packets are routed according to theIn each network router data packets are routed according to their ir destination address.destination address.-- The routers should perform a pattern detection task on the headThe routers should perform a pattern detection task on the header er detect sequence and classify the packets by their destination addetect sequence and classify the packets by their destination address.dress.

MotivationMotivation

Page 4: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 4

OutlineOutline

IntroductionIntroductionOptical MediumOptical MediumONN ExamplesONN Examples

BP BP TraningTraning of an ONNof an ONNA Multilayer ONNA Multilayer ONN

ReferencesReferences

OutlineOutline

Page 5: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

IntroductionIntroduction

Optical Neural NetworkOptical Neural Network

Page 6: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 6

IntroductionIntroduction

In an Optical NN:In an Optical NN:Sources are modulated light beams Sources are modulated light beams Synaptic Multiplications are due to attenuation Synaptic Multiplications are due to attenuation of light passing through an optical medium: of light passing through an optical medium: Geometric or Holographic.Geometric or Holographic.Target neurons sum signals from many source Target neurons sum signals from many source neurons.neurons.

IntroductionIntroduction

Page 7: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 7

HistoryHistoryONN was suggested by ONN was suggested by FarhatFarhat ad ad PissaltisPissaltis in 1987.in 1987.Many of the early Many of the early ONNsONNs had electrohad electro--optical componentsoptical components

Computers were slowComputers were slowWith converters between optical beams and electrical currentWith converters between optical beams and electrical currentThe full theoretical advantages of fullyThe full theoretical advantages of fully--optical NN couldnoptical NN couldn’’t be t be displayed. displayed.

With the advent of allWith the advent of all--optical components it became possible to optical components it became possible to create complete allcreate complete all--optical ONN.optical ONN.

They are not as accurate as their electronic equivalents, They are not as accurate as their electronic equivalents, That it is not easy to integrate a large number of optical unitsThat it is not easy to integrate a large number of optical unitson a chipon a chip

A common misconception: Optical A common misconception: Optical NNsNNs work much faster than work much faster than Electronic Electronic NNsNNs. . The future of ONN is yet uncertain.The future of ONN is yet uncertain.

IntroductionIntroductionHistoryHistory

Page 8: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Elements of Optical Elements of ONNONN

Optical Neural NetworkOptical Neural Network

Page 9: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 9

Optical Elements of ONNOptical Elements of ONN

Laser (1960)Laser (1960)LensesLensesLiquid Crystals (1888)Liquid Crystals (1888)

Optical Elements:Optical Elements:

Page 10: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 10

LaserLaser

A A laserlaser is a device that emits light through a is a device that emits light through a process called stimulated emission. process called stimulated emission. The term "laser" is an acronym for The term "laser" is an acronym for Light Light Amplification by Stimulated Emission of RadiationAmplification by Stimulated Emission of Radiation..Laser light is usually spatially coherent, which Laser light is usually spatially coherent, which means that the light either is emitted in a narrow, means that the light either is emitted in a narrow, low divergence beam, or can be converted into low divergence beam, or can be converted into one with the help of optical components such as one with the help of optical components such as lenses.lenses.

Optical Elements:Optical Elements:LaserLaser

Page 11: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 11

Liquid CrystalLiquid Crystal

-- The study of liquid crystals began in 1888 when an Austrian botThe study of liquid crystals began in 1888 when an Austrian botanist anist observed that a material had two distinct melting point.observed that a material had two distinct melting point. --Liquid crystals are a phase of matter whose order is intermediatLiquid crystals are a phase of matter whose order is intermediate e between that of a liquid and that of a crystalbetween that of a liquid and that of a crystal. . --The molecules are typically rodThe molecules are typically rod--shaped organic moieties about 25 shaped organic moieties about 25 Angstroms in length and their ordering is a function of temperatAngstroms in length and their ordering is a function of temperatureure. .

Optical Elements:Optical Elements:Liquid CrystalLiquid Crystal

Page 12: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 12

Liquid Crystals in ONN Liquid Crystals in ONN

SLM (Spatial Light Modulation)SLM (Spatial Light Modulation)OAOA--SLMSLMEAEA--SLMSLM

LCVT (Liquid Crystal Television)LCVT (Liquid Crystal Television)LCLV (Liquid Crystal Light Valve)LCLV (Liquid Crystal Light Valve)PCM (Phase Conjugate Mirror)PCM (Phase Conjugate Mirror)

Optical Elements:Optical Elements:Liquid CrystalLiquid Crystal

Page 13: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 13

Optically Addressed SLMOptically Addressed SLM

The “incoherent” light is detected (as intensity), by a photo-detector (as an electrical change distribution).This charge distribution affects the modulator, and so changes the Amplitude or Phase of the reflected coherent light.

Optical Elements:Optical Elements:OAOA--SLM1SLM1

Page 14: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 14

Practical Uses of Practical Uses of AOSLMsAOSLMs

Simplest applications is for real-time input to “4-f” optical processor

Optical Elements:Optical Elements:OAOA--SLM2SLM2

Page 15: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 15

Electrically Addressed SLMElectrically Addressed SLM

The hologram is generated on SLM by a The hologram is generated on SLM by a computer.computer.EASLM is similar to LCTV.EASLM is similar to LCTV.

Optical Elements:Optical Elements:EAEA--SLMSLMLCVTLCVT

PC

SLM2

Lens

InputSLM1

Page 16: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 16

LCLV: Optical LCLV: Optical ThresholdingThresholdingOptical Elements:Optical Elements:LCLV1LCLV1

Page 17: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 17

Sigmoid Curve fit for Sigmoid Curve fit for LCLVsLCLVsOptical Elements:Optical Elements:LCLV2LCLV2

Page 18: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 18

Phase Conjugate MirrorPhase Conjugate Mirror

A phase conjugate mirror is like a mirror, in that it reflects iA phase conjugate mirror is like a mirror, in that it reflects incident ncident light back towards where it came from, but it does so in a diffelight back towards where it came from, but it does so in a different rent way than a regular mirror. way than a regular mirror.

Optical Elements:Optical Elements:PCM1PCM1

Page 19: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 19

Phase Conjugate MirrorPhase Conjugate MirrorOptical Elements:Optical Elements:PCM2PCM2

Page 20: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

ExamplesExamples

Optical Neural NetworkOptical Neural Network

11-- BP Training of an ONNBP Training of an ONN22-- AA Multilayer ONN for Digit RecognitionMultilayer ONN for Digit Recognition

Page 21: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 21

BP Training of an ONN

Use optical BP in a feedUse optical BP in a feed--forward ONNforward ONNIs the first report of BP training in an optical system. Is the first report of BP training in an optical system. Uses a thermal nonlinear material as a neural processing Uses a thermal nonlinear material as a neural processing layer and a photorefractive crystal as a phase conjugate layer and a photorefractive crystal as a phase conjugate mirror to mirror to backpropagatebackpropagate the optical error. the optical error. The nonlinear material modulates the phase front of a The nonlinear material modulates the phase front of a forward propagating information beam by dynamically forward propagating information beam by dynamically altering the index of the refraction profile of the altering the index of the refraction profile of the material via a stronger weighting beam.material via a stronger weighting beam.

Example1Example1BP ONNBP ONN

Page 22: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 22

Basic Neural NetworkBasic Neural NetworkExample1Example1NetNet

Page 23: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 23

Two-Wave Mixing

In linear opticsIn linear opticsis a transient phenomenon that has no is a transient phenomenon that has no effect on anything else. effect on anything else.

In the transparent volume of a In the transparent volume of a nonlinear optical mediumnonlinear optical mediumthe interference pattern will cause a the interference pattern will cause a change in the refractive index of the change in the refractive index of the nonlinear medium in the shape of nonlinear medium in the shape of those same parallel planes.those same parallel planes. )(.)( 20 yInnyn +=

Example1Example1MixingMixing

Page 24: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 24

Network DescriptionNetwork Description

)(0 yE

InputSLM

)(yW

)(yO

NLM

)(.)( 20 yInnyn +=

Example1Example1DescriptionDescription

Page 25: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 25

Experimental Set upExperimental Set upExample1Example1Set UpSet Up

Page 26: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 26

BP Training ResultsBP Training Results

Mean Square Error (MSE)

Example1Example1ResultsResults

Page 27: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 27

A Multilayer ONN for Digit A Multilayer ONN for Digit RecognitionRecognition

Example2Example2MONNMONN

An optical An optical perceptronperceptron with a soft optical threshold is with a soft optical threshold is implemented, that is trained with an adapted BP implemented, that is trained with an adapted BP algorithm.algorithm.An optical An optical thresholdingthresholding perceptronperceptron is composed of two is composed of two parts:parts:

MatrixMatrix--vectorvector--Multiplier (MVM),Multiplier (MVM),A A thresholdingthresholding device.device.

Page 28: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 28

Matrix Vector Multiplier (MVM)Matrix Vector Multiplier (MVM)Example2Example2MVMMVM

P: P: PolariserPolariser; ; GnGn: Gratings: Gratings

Page 29: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 29

Weight MappingsWeight Mappings

Behaviour of optical weights under different weight mappings:x-axis has calculated weights, and y-axis has corresponding optical weights or LCTV2 transmittances

Example2Example2Digit TestDigit Test

Page 30: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 30

Test inputsTest inputsExample2Example2Digit TestDigit Test

A set of handwritten digitsA set of handwritten digits

Page 31: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 31

Output ScreenOutput Screen

LCLV output images without and with write lightLCLV output images without and with write light

Example2Example2Digit TestDigit Test

Page 32: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 32

Digit RecognitionDigit Recognition

Recognition of a 0, 2, 3, 4, 5, 6, 7, and 9Recognition of a 0, 2, 3, 4, 5, 6, 7, and 9

Example1Example1Digit TestDigit Test

Page 33: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

ReferencesReferences

Optical Neural NetworkOptical Neural Network

Page 34: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 34

ReferencesReferences[1] J. E. Steck1, S. R. Skinner2, A. A. Cruz[1] J. E. Steck1, S. R. Skinner2, A. A. Cruz--Cabrera2, M. Yang, and E. C. Cabrera2, M. Yang, and E. C.

Behrman3, Behrman3, ““BackpropagationBackpropagation Training of an Optical Neural Network,Training of an Optical Neural Network,”” IEEE, IEEE, 2000.2000.

[2] I. [2] I. SaxenaSaxena, P. , P. MoerlandMoerland, E. , E. FieslerFiesler, A. R. , A. R. PourzandPourzand, and N. , and N. CollingsCollings, , ““An An Optical Optical ThresholdingThresholding PerceptronPerceptron with Soft Optical Threshold.with Soft Optical Threshold.”” 1997.1997.

[3] I. [3] I. SaxenaSaxena and P. and P. MoerlandMoerland, , ““An Adaptive Multilayer Optical Neural Network An Adaptive Multilayer Optical Neural Network Design,Design,”” 1996.1996.

[4] I. [4] I. SaxenaSaxena and P. and P. MoerlandMoerland, , ““Adaptive Multilayer Optical Neural Network Adaptive Multilayer Optical Neural Network with Optical with Optical ThresholdingThresholding,,”” 1994.1994.

[5] P. [5] P. MoerlandMoerland, E. , E. FieslerFiesler and I. and I. SaxenaSaxena, , ““Incorporating LCLV Incorporating LCLV Non_LinearitiesNon_Linearitiesin Optical Neural Network,in Optical Neural Network,””1995.1995.

[6] T. [6] T. YatagaiYatagai and M. and M. ItohItoh, , ““Development of an Optical Neural Network,Development of an Optical Neural Network,”” 1995.1995.[6] PCM: [6] PCM: http://cnshttp://cns--alumni.bu.edu/~slehar/Spatial/Spatial.htmalumni.bu.edu/~slehar/Spatial/Spatial.htm[7] Liquid Crystal: [7] Liquid Crystal: http://plc.cwru.edu/tutorial/enhanced/files/lc/Intro.htmhttp://plc.cwru.edu/tutorial/enhanced/files/lc/Intro.htm[8] [8] http://www.exo.net/~pauld/activities/liquidcrystal/liquidcrystalhttp://www.exo.net/~pauld/activities/liquidcrystal/liquidcrystal.html#how.html#how[9] Google Search for so many optic concepts in Optical Neural N[9] Google Search for so many optic concepts in Optical Neural Networks!etworks!

Page 35: Optical Neural Network - Sharif University of Technologyee.sharif.edu/~ann/Optical Neural Network.pdf · Optical Neural Network 3 All Optic Systems A packet switched telecommunication

Optical Neural Network 35