Resistive MAC Array Correction for Signed Neural Weights

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network calculations struggle with handling positive and negative weights using variable resistance elements, leading to complex apparatus design and size issues, as well as limitations in weight sign discrimination during learning.

Innovation Solution

A multiply and accumulate calculation device utilizing a variable resistor array and a reference array with detection units and correction calculation units to digitally or analogically process current outputs, allowing for the consideration of positive and negative weights through a single element, thereby simplifying and miniaturizing the apparatus.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate elements are used to store positive and negative weights, then weight sign discrimination is achieved, but device complexity and apparatus size increase

Engineering Contradiction:
Improveweight sign discriminationVSAvoidapparatus complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges the functions of positive and negative weight storage into a single variable resistance element by utilizing dual polarizations of the same material. This eliminates the need for separate element arrays for positive and negative weights, thereby reducing device complexity while maintaining weight sign discrimination capability through material polarization states.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The variable resistance element is designed to perform multiple functions: storing both positive and negative weights, maintaining stability during read operations, and enabling weight updates through polarization control. This multi-functionality is achieved by using the same element structure with dual polarization states, reducing the need for specialized components.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Volume of moving object

If dual polarization material is used for weight storage, then apparatus size is reduced, but reading accuracy may be affected by material characteristics

Engineering Contradiction:
Improveapparatus sizeVSAvoidreading accuracy
Core Design Contradiction:
Volume of moving objectVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the reading operation detects the polarization state of the variable resistance element and uses this information to accurately determine the stored weight value. The system compensates for material characteristics by continuously monitoring the electrical characteristics and adjusting the read interpretation accordingly, ensuring reading accuracy despite material variations.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent utilizes changes in electrical parameters (resistance, conductance) of the dual polarization material to represent different weight values. By carefully selecting and controlling these electrical parameters during write and read operations, the system achieves accurate weight storage and retrieval while maintaining compact apparatus size.

Inventive Principle:
Principle #35Parameter changes

3Speed

If variable resistance elements are used for analog calculation, then calculation speed is improved, but handling negative conductance values becomes problematic

Engineering Contradiction:
Improvecalculation speedVSAvoidnegative weight handling
Core Design Contradiction:
SpeedVSAdaptability or versatility

Solution Approach 1:

The patent employs asymmetric dual polarization states in the variable resistance element to represent positive and negative weight values. The material exhibits different electrical characteristics for each polarization state, allowing the system to naturally represent negative values without requiring negative conductance. This asymmetric design enables analog calculation speed while successfully handling negative weight values through material property differentiation.

Inventive Principle:
Principle #4Asymmetry

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

Enables accurate calculation of positive and negative weights with reduced apparatus complexity and size, enhancing the efficiency and compactness of neural network operations.

Implementation Method 1

a variable resistance element whose resistance value is changeable by application of voltage

Methodology Applied
Scientific EffectVariable resistance: Electrical Resistance

Implementation Method 2

a detection unit that detects a value obtained by multiplying the input value by the weight, by detecting a current value flowing through the variable resistor column

Methodology Applied
Scientific EffectOhm's law: Ohm's Law

Data Source

PatentUS12547373B2Multiply and accumulate calculation device, neuromorphic device, and multiply and accumulate calculation method
Publication Date: 2026.02.10 TDK CORP
  • US12547373B2 patent drawing
  • US12547373B2 patent drawing
  • US12547373B2 patent drawing

AI summary

A multiply and accumulate calculation device including a variable resistor array unit having a plurality of variable resistance elements, a reference array unit having a reference resistance element having a fixed resistance value, a signal input unit that generates an input signal from input data, and inputs the input signal to the variable and reference resistance elements, a first detection unit that detects a current flowing through the variable resistor array unit, based on the input signal applied to the variable resistance elements, a second detection unit that detects a current flowing through the reference array unit, based on the input signal applied to the reference resistance element, and a correction calculation unit that performs a predetermined calculation on the output from the first detection unit, based on the output from the second.