Image Sensor Analog Signal Feeding Cross-Point Synapse Array

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network systems face challenges with significant signal degradation, increased power consumption, and processing time due to the conversion of analog image sensor inputs into digital signals before processing in cross-point array-based ANNs, which affects accuracy and efficiency in machine learning applications.

Innovation Solution

The integration of an image sensor device with an analog neural network synapse matrix that feeds normalized analog signals directly into the cross-point array, reducing signal degradation and eliminating the need for digital-to-analog and analog-to-digital conversions, thereby enhancing the accuracy and efficiency of image recognition and processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If analog signals are converted to digital signals before processing in cross-point array-based ANNs, then the system can be operated with digital electronics, but signal degradation occurs and processing time increases

Engineering Contradiction:
Improvesignal qualityVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent extracts the analog-to-digital conversion step from the signal processing chain by feeding normalized analog signals directly into the cross-point array-based ANN, eliminating the conversion bottleneck that causes both signal degradation and processing delays

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent introduces a normalization circuit as an intermediary component that conditions analog signals before they enter the cross-point array, ensuring signal quality is maintained without requiring digital conversion, thus preserving both signal integrity and processing speed

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of operation

If analog signals are converted to digital signals, then digital processing can be performed, but power consumption increases

Engineering Contradiction:
Improveprocessing capabilityVSAvoidpower consumption
Core Design Contradiction:
Ease of operationVSUse of energy by moving object

Solution Approach 1:

The patent removes the power-intensive analog-to-digital conversion stage from the system architecture, allowing analog signals to be processed directly by the cross-point array, thereby significantly reducing overall power consumption while maintaining full processing capability

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the digital electronic processing chain with an analog processing approach using normalized analog signals and cross-point array multiplication, substituting high-power digital operations with lower-power analog computations

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Productivity

If analog signals undergo digital conversion, then signal processing can be performed, but signal degradation occurs

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidsignal accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies preliminary normalization to analog signals before they enter the cross-point array, preparing the signals in advance to maintain their integrity throughout processing, thereby preserving signal accuracy while enabling efficient analog computation

Inventive Principle:
Principle #10Preliminary action

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach reduces signal degradation, lowers power consumption, and accelerates processing times, improving the overall performance and accuracy of neural network systems for image recognition and machine learning tasks.

Implementation Method 1

The pixel unit circuit generates a voltage output based on an input light

Methodology Applied
Scientific EffectPhotoelectric Effect: Photoelectric Effect

Implementation Method 2

The pixel unit circuit generates a voltage output based on an input light, the voltage output being sent to the set of cross-point devices for conversion to corresponding current signals

Methodology Applied
Scientific EffectElectrical Conduction: Conduction (electrical)

Data Source

PatentUS10708522B2Image sensor with analog sample and hold circuit control for analog neural networks
Publication Date: 2020.07.07 INTERNATIONAL BUSINESS MACHINE CORPORATION
  • US10708522B2 patent drawing
  • US10708522B2 patent drawing
  • US10708522B2 patent drawing

AI summary

According to one or more embodiments of the present invention, an image processing system includes a cross-point synapse array that includes multiple row wires, multiple column wires, and multiple cross-point devices, a cross-point device at each intersection of the row wires and the column wires. The image processing system further includes an image sensor array that includes multiple pixel unit circuits, each pixel unit circuit is connected to a corresponding row wire of the cross-point synapse array, wherein the pixel unit circuit generates a voltage output based on an input light. The image processing system further includes a pixel unit controller that adjusts an exposure time of the pixel unit circuits based on voltage outputs from the pixel unit circuits respectively.