DLP®-Driven, Optical Neural Network Results and Future Design Emmett Redd Professor Missouri State University Neural Network Applications • • • • • • • • Stock Prediction: Currency, Bonds, S&P 500, Natural Gas Business: Direct mail, Credit Scoring, Appraisal, Summoning Juries Medical: Breast Cancer, Heart Attack Diagnosis, ER Test Ordering Sports: Horse and Dog Racing Science: Solar Flares, Protein Sequencing, Mosquito ID, Weather Manufacturing: Welding Quality, Plastics or Concrete Testing Pattern Recognition: Speech, Article Class., Chem. Drawings No Optical Applications—We are starting with Boolean • most from www.calsci.com/Applications.html Optical Computing & Neural Networks • Optical Parallel Processing Gives Speed – Lenslet’s Enlight 256—8000 Giga Multiply and Accumulate per second – Order 1011 connections per second possible with holographic attenuators • Neural Networks – – – – – – Parallel versus Serial Learn versus Program Solutions beyond Programming Deal with Ambiguous Inputs Solve Non-Linear problems Thinking versus Constrained Results Optical Neural Networks • Sources are modulated light beams (pulse or amplitude) • Synaptic Multiplications are due to attenuation of light passing through an optical medium (30 fs) • Geometric or Holographic • Target neurons sum signals from many source neurons. • Squashing by operational-amps or nonlinear optics Standard Neural Net Learning • We use a Training or Learning algorithm to adjust the weights, usually in an iterative manner. x1 … xN y1 S … … S yM Learning Algorithm Target Output (T) Other Info FWL-NN is equivalent to a standard Neural Network + Learning Algorithm FWL-NN S S S S + Learning Algorithm Optical Recurrent Neural Network Signal Source (Layer Input) Synaptic Medium (35mm Slide) Postsynaptic Optics Target Neuron Summation Presynaptic Optics Squashing Functions Micromirror Array Recurrent Connections Layer Output A Single Layer of an Optical Recurrent Neural Network. Only four synapses are shown. Actual networks will have a large number of synapses. A multi-layer network has several consecutive layers. Definitions • Fixed-Weight Learning Neural Network (FWL-NN) – A recurrent network that learns without changing synaptic weights • Potency – A weight signal • Tranapse – A Potency modulated synapse • Planapse – Supplies Potency error signal • Zenapse – Non-Participatory synapse • Recurron – A recurrent neuron • Recurral Network – A network of Recurrons Optical Fixed-Weight Learning Synapse Tranapse x(t) Σ y(t-1) W(t-1) T(t-1) x(t-1) Planapse Page Representation of a Recurron U(t-1) Bias B(t) A(t) Tranapse Tranapse Σ O(t-1) Tranapse Planapse Optical Neural Network Constraints • • • • • • Finite Range Unipolar Signals [0,+1] Finite Range Bipolar Attenuation[-1,+1] Excitatory/Inhibitory handled separately Limited Resolution Signal Limited Resolution Synaptic Weights Alignment and Calibration Issues Optical System DMD or DLP® Design Details and Networks • • • • Digital Micromirror Device 35 mm slide Synaptic Media CCD Camera Synaptic Weights - Positionally Encoded - Digital Attenuation • Allows flexibility for evaluation. Recurrent AND Unsigned Multiply FWL Recurron DMD/DLP®—A Versatile Tool • Alignment and Distortion Correction – Align DMD/DLP® to CCD——PEGS – Align Synaptic Media to CCD——HOLES – Calculate DMD/DLP® to Synaptic Media Alignment——Putting PEGS in HOLES – Correct projected DMD/DLP® Images • Nonlinearities Stretch, Squash, and Rotate x1 y 1 x12 D x1 y1 y12 1 yP xP 2 xP y P 2 yP 1 xP 1 None x1y0 x0y1 Linear x2y0 x1y1 x0y2 x3y0 x2y1 x1y2 x0y3 etc. Quadratic Cubic etc. Where We Are and Where We Want to Be DMD Dots Image (pegs) Slide Dots Image (holes) 100 100 200 200 300 300 400 400 500 500 600 600 700 700 800 800 900 900 1000 200 400 600 C 800 1000 1200 1000 200 400 600 C' 800 1000 1200 CCD Image of Known DLP® Positions DMD Dots Image (pegs) Automatically Finds Points via Projecting 42 Individual Pegs 100 200 300 400 500 600 C=H●D 700 H = C ● DP 800 900 1000 200 400 600 800 1000 1200 CCD Image of Holes in Slide Film Slide Dots Image (holes) Manually Click on Interference to Zoom In 100 200 300 400 500 600 700 800 900 1000 200 400 600 800 1000 1200 Mark the Center ZOOM OF DIFFRACTION PATTERN #1. Click on center 10 20 30 40 50 60 70 80 10 20 30 40 50 60 70 80 90 Plans for Automatic Alignment • Methods and Details about automatically finding the holes. Might add a couple of more slides, more matrix manipulation, and even an Excel spreadsheet demonstration. Eighty-four Clicks Later Slide Dots Image (holes) C' = M ● C 100 M = C' ● CP 200 300 400 500 600 700 800 900 1000 200 400 600 800 1000 1200 DLP® Projected Regions of Interest D' = H-1●M-1●H●D 100 200 and 300 C' = H ● D' 400 500 600 700 100 200 300 400 500 600 700 800 900 1000 Nonlinearities 1 Measured light signals vs. Weights for FWL Recurron 0.9 0.8 0.7 0.6 Opaque slides aren’t, ~6% leakage. 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Neural Networks and Results • Recurrent AND • Unsigned Multiply • Fixed Weight Learning Recurron Recurrent AND IN IN-1 1-cycle delay IN IN-1 Recurrent AND Neural Network Source Neurons (buffers) Bias (1) -14/16 logsig(16*Σ) IN 10/16 Σ 1 0 10/16 IN IN-1 Terminal Neurons -1/2 Σ 2/2 IN-1 linsig(2*Σ) 1 0 Synaptic Weight Slide Weights 0.5 10/16 -1/2 -14/16 10/16 2/2 Recurrent AND Demo • MATLAB Pulse Image (Regions of Interest) 100 200 300 400 500 600 700 800 900 1000 200 400 600 800 1000 1200 Output Swings Larger than Input 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 10 20 30 40 50 60 70 80 90 100 Synaptic Weight Slide Unsigned Multiply Results 1 About 4 bits 0.9 0.8 0.7 0.6 Blue-expected 0.5 Black-obtained 0.4 Red-squared error 0.3 0.2 0.1 0 0 5 10 15 20 25 Page Representation of a Recurron U(t-1) Bias B(t) A(t) Tranapse Tranapse Σ O(t-1) Tranapse Planapse FWL Recurron Synaptic Weight Slide Optical Fixed-Weight Learning Synapse Tranapse x(t) Σ y(t-1) W(t-1) T(t-1) x(t-1) Planapse Future: Integrated Photonics • Photonic (analog) • i. Concept • α. Neuron • β. Weights • γ. Synapses Photonics Spectra and Luxtera Continued • • • • • • • • • • ii. Needs α. Laser β. Amplifier (detectors and control) γ. Splitters δ. Waveguides on Two Layers ε. Attenuators ζ. Combiners η. Constructive Interference θ. Destructive Interference ι. Phase Photonics Spectra and Luxtera DLP®-Driven, Optical Neural Network Results and Future Design Emmett Redd & A. Steven Younger Missouri State University EmmettRedd@MissouriState.edu Source Pulse 100 200 300 400 500 600 700 100 200 300 400 500 600 700 800 900 1000