Information Processing with Pulsed Neural Networks Ulrich Ramacher
- Slides: 51
Information Processing with Pulsed Neural Networks Ulrich Ramacher Corp. Research, Infineon Technologies AG, Munich ulrich. ramacher@infineon. com Prof. Ramacher, CPR ST 10 -Sep-20 Page 1
Dilemma (1) Prof. Ramacher, CPR ST 10 -Sep-20 Page 2 lack of robustness
Dilemma (2) Prof. Ramacher, CPR ST 10 -Sep-20 Page 3 lack of architecture information
Dilemma (3) • Non-invasive long-term recording of neural tissue. • High-density sensor with 128 Sensors in 1 1 mm². • Extended CMOS-process with biocompatible high-k surface dielectric. • Self-calibration circuitry and pre-amplification on chip. • Applications in neurobiology and drug discovery Sensor [m. V] Neuron [m. V] Infineon Neurochip 20µm Time [ms] Prof. Ramacher, CPR ST 10 -Sep-20 Page 4 Gap between signal processing by few cells and information processing by large arrays
Program 1. Use pulsed IAF neurons and adaptive synapses (inhibitory, excitatory) 2. Use experimental evidence 3. Start from simple, complete vision systems 4. Find basic network structures for information process 5. Find quantitative ansatz for information processing 6. Demonstrate usefulness for cell phones Prof. Ramacher, CPR ST 10 -Sep-20 Page 5
Neuron Model or : Rule 1: Reaching the threshold, a neuron resets membrane poten and starts sending a pulse of 1 ms Rule 2: A neuron either receives or sends deterministic dynamical system Prof. Ramacher, CPR ST 10 -Sep-20 Page 6
Experiment 1 Prof. Ramacher, CPR ST 10 -Sep-20 Page 7
Frequency Experiment 1 Prof. Ramacher, CPR ST 10 -Sep-20 Page 8
Observations • „Firing patterns don‘t come to an end“ • information processing is not a function of time Does frequency of firing patterns characterize t Inet = - pn x ln pn Is entropy a reproducible measure of informatio Prof. Ramacher, CPR ST 10 -Sep-20 Page 9
Experiment 2: Prof. Ramacher, CPR ST 10 -Sep-20 Page 10
Experiment 2 Prof. Ramacher, CPR ST 10 -Sep-20 Page 11
Findings • distribution function of fp‘s is independent of initial cond Inet = - pn x ln pn is a reproducibly measurable quantity Prof. Ramacher, CPR ST 10 -Sep-20 Page 12
Entropy as a function of mean synaptic weight Each color corresponds to a different realization of wij Network size: 40 neurons Prof. Ramacher, CPR ST 10 -Sep-20 Page 13
Network size: 40 neurons ISI Prof. Ramacher, CPR ST 10 -Sep-20 Page 14 Fire Rate
Synapse Model (proposed by U. Ramacher, April 99) Adaptation rule : local, causal, simple Prof. Ramacher, CPR ST 10 -Sep-20 Page 15
µ negative • 20 x 24 pixels , each connected to a neuron by a constant synapse • adaptive synapses connecting neighboured neurons Prof. Ramacher, CPR ST 10 -Sep-20 Page 16 synapses = coupled system of damped oscillators
µ positive 20 x 24 pixel, 10% noise synapses = coupled system of damped Prof. Ramacher, CPR STexponentially rising and falling „elements“ 10 -Sep-20 Page 17
Spot Detector = Illumination Encoder Prof. Ramacher, CPR ST 10 -Sep-20 Page 18
Pixels: 64 x 64 Prof. Ramacher, CPR ST 10 -Sep-20 Page 19
Projection onto the receptive field of a simple cell in the visual cortex LGN primary visual cortex (V 1) light retina Dr. Arne Heittmann Prof. Ramacher, CPR ST 10 -Sep-20 Page 20 stimulus optic nerve
convergence Dr. Arne Heittmann Prof. Ramacher, CPR ST 10 -Sep-20 Page 21
Modeling the Experiment Bright Stimulus Dark Stimulus y 0 Dy x 0 Dx Stimulus Stimuli projected onto RF of a Simple Cell Prof. Ramacher, CPR ST 10 -Sep-20 Page 22 P: measured pulse rate Gi: Gabor function H: Spot Intensity B: Background Intensity
Pulse difference detector (1) i 1>i 2 i 1<i 2 Prof. Ramacher, CPR ST 10 -Sep-20 Page 23
Pulse difference detector (2) Dynamics of the Synapse W 41: Prof. Ramacher, CPR ST 10 -Sep-20 Page 24
Pulse difference detector (3) Number of pulses Characteristic Neuron 2 Neuron 1 Neuron 4 i 2 Prof. Ramacher, CPR ST 10 -Sep-20 Page 25 i 1 = 0. 5 Q =1 WK 0 = 0. 08 td = 1 ms T = 0. 5 s
Architecture of feature detector (proposed by A. Heittmann, 2003) 1 1 1 3 --- 3 1 2 3 4 3‘ 3‘ 3‘ 3 4 4 2 4‘ 4‘ 4‘ 4 2 --- 2 3‘ 4‘ 5‘ 5 6 µ>0 µ<0 Prof. Ramacher, CPR ST 10 -Sep-20 Page 26 constant
Shaping the response-profile of the detector section of the retina Prof. Ramacher, CPR ST 10 -Sep-20 Page 27 H: Spot Intensity B: Background Intensity
Results of a detector implementation Gabor-Wavelet Prof. Ramacher, CPR ST 10 -Sep-20 Page 28 measured profile - 256 Gradient detectors - size of receptive field: 17 x 17 Pixel - T=750 ms @ 1 ms Pulse-duration
ideal Simulated Filter responses, T=750 ms Real Imaginary 90° 0° 0° Prof. Ramacher, CPR ST 10 -Sep-20 Page 29
The Head-Detector Restriction - single scale (keep eye-distance fixed) Prof. Ramacher, CPR ST 10 -Sep-20 Page 30
Check for Robustness in Eye-Brow Zone Reference Image Filter Response, horizontal direction region of interest 20 x 20 Pixel Prof. Ramacher, CPR ST 10 -Sep-20 Page 31 new eye-brow image
A simple memory (1) , 1 zone Learning Phase Detector Layer K K Prof. Ramacher, CPR ST 10 -Sep-20 Page 32 Input Layer Memory Layer K 1 -1 connection between input and detector layer 1 -1 connection between input and memory layer full connection between detector and memory layer
Experiment 1: learned image in input and memory Activity (number of events), 2 ms window Pulse-patterns of input-layer Prof. Ramacher, CPR ST 10 -Sep-20 Page 33
A simple memory (2), 1 Zone Recognition Phase Detector Layer K K Prof. Ramacher, CPR ST 10 -Sep-20 Page 34 Gabor Layer Memory Layer K 1 -1 connection between input and detector layer full connection between detector and memory layer
Experiment 3: recognition of non-learned eye-brow Activity (number of events), 2 ms window Pulse-patterns of input-layer Prof. Ramacher, CPR ST 10 -Sep-20 Page 35
Experiment 2: Isolated neurons Activity (number of events), 2 ms window Pulse-patterns of input-layer Prof. Ramacher, CPR ST 10 -Sep-20 Page 36
Zone-Architecture I Detector Layer 1 Horizontal Orientations Detector Layer 2 Memory Layer Vertical Orientations : Gabor-Kernel : Input Image Prof. Ramacher, CPR ST 10 -Sep-20 Page 37
Zone-Architecture II Zone 5 Zone 6 Zone 1 Zone 7 Zone 2 Zone 3 Zone 4 Zuordnung: Zonen zu Bildregionen : Zone für Gabor-Wavlet mit horizontaler Orientierung Prof. Ramacher, CPR ST 10 -Sep-20 Page 38 : Zone für Gabor-Wavlet mit vertikaler Orientierung
Reference-Image and Test-Images Reference-Image: Test-Images: Prof. Ramacher, CPR ST 10 -Sep-20 Page 39 Face 0001 Face 0014
Face 0001 Normalized accumulated activity Activity-Diagram Memory Zone 1 Detector Memory Zone 2 Detector Memory Detector Zone 3 Memory Detector Zone 4 Memory Detector Time [ms] Prof. Ramacher, CPR ST 10 -Sep-20 Page 40 Zone 5
Face 0011 Normalized accumulated activity Activity-Diagram Memory Zone 1 Detector Memory Zone 2 Detector Memory Detector Zone 3 Memory Detector Zone 4 Memory Detector Time [ms] Prof. Ramacher, CPR ST 10 -Sep-20 Page 41 Zone 5
Face 0014 Normalized accumulated activity Activity-Diagram Memory Zone 1 Detector Memory Zone 2 Detector Memory Detector Zone 3 Memory Detector Zone 4 Memory Detector Time [ms] Prof. Ramacher, CPR ST 10 -Sep-20 Page 42 Zone 5
Binding of zones by synchrony Binding Layer … 1 3 2 4 6 5 7 Zuordnung: Zonen zu Neuronen des Spotdetektors zur Bindung Prof. Ramacher, CPR ST 10 -Sep-20 Page 43 Memory Layer
Results ~ 40 ms Spot_0001 ~ 40 ms Spot_0011 Prof. Ramacher, CPR ST 10 -Sep-20 Page 44 ~ 40 ms Spot_0014
Column Architecture Image plane Detector Feature 1 Memory Feature 1 Zone 1 Detector Feature 2 Memory Feature 2 robust recognition Zone 2 . . . Detector Feature n Memory Feature n Binding Object 1 Associative Memory Binding Object 2 Binding Object n Prof. Ramacher, CPR ST 10 -Sep-20 Page 45 Zone n ∙ ∙ ∙ ∙
The vision: a 3 D-Vision-Cube The Vision-Cube • 3 D-stacking-architecture • Low-power • Real time capabilities • Integration of sensors and information processing • Distributes layers of information processing to layers of the stack • Solves problem of connectivity Prof. Ramacher, CPR ST 10 -Sep-20 Page 46 CMOS-sensor, sensor array Analogue-pulse conversion Feature-Detection Gabor-Wavelets, different orientations … Object-recognition Object-detection
Design of a testchip for the „Synchrony detector“, Base-Chip for the 3 D-Stack Infineon 130 nm CMOS-Technology pixel memory (1 Pixel) local AER-subcircuit testcircuits for 3 Dinterconnects Test circuit Integrate-and. Fire-Neuron 3 D-vias for supply and digital control signals adaptive synapses layout of 1 neuron rows of 3 D-vias for layer-to-layer signal distribution array of 128 x 128 neurons, 64 k synapses control circuit AER-encodercircuit Prof. Ramacher, CPR ST 10 -Sep-20 Page 47 Power consumption: 3 m. W @1. 5 V (analog) , 40 -250 m. W @1. 5 V (digital) Size: 7. 6 mm x 7. 8 mm
Gabor-Feature-Detector chip (layout) including a pulse router for the 3 D-integration Infineon 130 nm CMOS-Technology Array of 128 x 128 (64 k) processing elements for pulse processing and routing Processing element: - photo-Detector - pulse-processing (gradient detection) - dynamic routing of pulses 3 D- wiring channel for vertical signal distribution SRAM for storing routing information Digital macro: - routing circuit - configuration - AER-circuit (event based) Prof. Ramacher, CPR ST 3 D- wiring 10 -Sep-20 Page 48 for power supply and control signals
3 D-Stacking Chip 3 SOLID Connection Chip 2 15 µm 7 µm Chip 1 14 µm Layer 7 12 µm Layer 6 Layer 5 12µm Prof. Ramacher, CPR ST 10 -Sep-20 Page 49 Bottom-Chip
Real 3 D !! Layer 7 Layer 6 Layer 5 Layer 4 Layer 3 Layer 2 Bottom Substrate, Layer 1 Prof. Ramacher, CPR ST 10 -Sep-20 Page 50 13µm 7µm
Conclusion • IAF neurons, adaptive synapses, only • network built for -- Gabor wavelet based feature cascade -- memory -- comparison of memory and detector plane • synchrony of neurons indicative for -- robust recognition of memory feature at detector plane (elastic matching) -- binding of features as object • built in 130 nm CMOS -- synchrony detector -- Gabor wavelet detector -- 3 D stack of 7 silicon chips Prof. Ramacher, CPR ST 10 -Sep-20 Page 51
- Efficient processing of deep neural networks pdf
- O,hf
- Pulsemaster pulsed electric field
- Pulsed coherent radar
- Visualizing and understanding convolutional networks
- Vc bound
- Freed et al 2001 ib psychology
- Audio super resolution using neural networks
- Convolutional neural networks for visual recognition
- Style transfer
- Mippers
- Stride in cnn
- Neural networks and learning machines 3rd edition
- Pixelrnn
- Neural network matlab toolbox
- Neural networks for rf and microwave design
- 11-747 neural networks for nlp
- Neural networks simon haykin
- Sparse convolutional neural networks
- On the computational efficiency of training neural networks
- Tlu neural network
- Neural networks and fuzzy logic
- Netinsights
- Convolutional neural networks
- Few shot learning with graph neural networks
- Deep forest: towards an alternative to deep neural networks
- Convolutional neural networks
- Neuraltools neural networks
- Lstm andrew ng
- Predicting nba games using neural networks
- Neural networks and learning machines
- The wake-sleep algorithm for unsupervised neural networks
- Audio super resolution
- Convolutional neural network alternatives
- Difference between datagram and virtual circuit operation
- Backbone networks in computer networks
- Neural processing
- Information theory and neural coding
- Bottom-up processing example
- Bottom up processing vs top down processing
- Bottom-up processing examples
- Neighborhood processing
- Secondary processing of wheat
- Point processing
- Histogram processing in digital image processing
- Parallel processing vs concurrent processing
- A generalization of unsharp masking is
- Point processing in image processing
- Morphological
- Top down vs bottom up processing
- Batch processing and interactive processing
- Ulrich von jungingen