Weight Matrix Circuit Linearizing Non-Linear Resistive Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Resistive memory devices with non-linear current-voltage characteristics in weight matrix circuits lead to inaccuracies and inefficiencies in neural network training and inference processes, affecting power consumption and calculation speed.
Innovation Solution
A weight matrix circuit with an input circuit connected to resistive memory devices, utilizing an operational amplifier and resistor to achieve a linear current-voltage characteristic, and a weight matrix input circuit with an operational amplifier and resistor to compensate for the non-linearity of resistive memory devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If resistive memory devices with non-linear current-voltage characteristic are used in weight matrix circuit, then device complexity is reduced and integration is improved, but calculation accuracy deteriorates due to non-linear characteristics
Solution Approach 1:
An input circuit is introduced as an intermediary component between the external signal source and the resistive memory devices. This input circuit includes a conversion element that transforms the input signal in a manner that compensates for the non-linear current-voltage characteristic of the resistive memories, enabling accurate vector-matrix multiplication despite the non-linear device properties
Solution Approach 2:
The input circuit modifies the parameters of the input signal (voltage or current) to account for the non-linear characteristics of the resistive memory devices. By pre-distorting the input signal parameters, the overall system achieves linear input-output relationship necessary for accurate neural network calculations
2Ease of manufacture
If non-linear current-voltage characteristic of resistive memories is present, then manufacturing ease is improved, but training and inference accuracy deteriorates
Solution Approach 1:
The input circuit serves as a mediator that decouples the manufacturing simplicity of non-linear resistive devices from the computational accuracy requirements. It translates the inherently non-linear device response into accurate linear computational results, allowing the use of easily manufacturable non-linear resistive memories while maintaining precision
3Use of energy by moving object
If resistive memories with non-linear characteristic are used directly, then power consumption is reduced, but calculation speed and accuracy deteriorate
Solution Approach 1:
The calculation function is segmented between the input circuit (signal conditioning) and the resistive memory array (parallel multiplication). This segmentation allows the low-power advantage of resistive memories to be preserved while the input circuit handles the non-linearity correction, enabling both energy efficiency and accurate rapid computation
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
The solution improves the accuracy and speed of neural network calculations while reducing power consumption by linearizing the current-voltage characteristic of the weight matrix, enhancing training and inference processes.
Implementation Method 1
an input circuit configured to be connected to each of the input lines; and an output circuit configured to be connected to each of the output lines. The input circuit is connected to the resistive memory devices such that the weight matrix circuit has a linear current-voltage characteristic
Implementation Method 2
n×m resistive memory devices each connected to the n input lines and the m output lines and each having a non-linear current-voltage characteristic
Data Source
AI summary
Provided are a weight matrix circuit and a weight matrix input circuit. The weight matrix circuit includes a memory array including n input lines, m output lines, and n×m resistive memory devices each connected to the n input lines and the m output lines and each having a non-linear current-voltage characteristic, an input circuit connected to each of the input lines, and an output circuit connected to each of the output lines. The input circuit is connected to the resistive memory devices such that the weight matrix circuit has a linear current-voltage characteristic.