Voltage Amplitude Control for RPU Learning Rate

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current methods for training Deep Neural Networks (DNNs) are computationally intensive and face challenges in scaling due to limitations in device specifications of non-volatile memory technologies, which hinder efficient voltage control of learning rates for Resistive Processing Units (RPUs) during DNN training.

Innovation Solution

The method involves controlling the learning rate of RPUs by varying the amplitude of input voltage signals applied to each node in a 2D array of resistive processing units, allowing for efficient voltage pulse height control to tune the learning rate without increasing training time or compromising accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If device specifications of non-volatile memory technologies are used for RPU implementation, then memory storage capability is improved, but learning rate control precision deteriorates

Engineering Contradiction:

Improvememory storage capabilityVSAvoidlearning rate control precision

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent changes the control parameter from pulse width (time domain) to voltage amplitude (electrical domain). By varying the amplitude of voltage pulses applied to the RPU devices, the learning rate is precisely controlled without being constrained by device timing specifications. This parameter transformation resolves the contradiction by decoupling learning rate control from the inherent timing limitations of non-volatile memory devices.

Inventive Principle:

Principle #35Parameter changes

2Ease of operation

If pulse width modulation is used to control learning rate, then learning rate adjustment is achieved, but training time increases

Engineering Contradiction:

Improvelearning rate adjustmentVSAvoidtraining time

Core Design Contradiction:

Ease of operationVSLoss of time

Solution Approach 1:

The patent replaces the mechanical/time-based control mechanism (pulse width modulation) with an electrical field-based mechanism (voltage amplitude modulation). Instead of varying the duration of voltage pulses to control learning rate, the invention varies the amplitude of fixed-duration pulses. This substitution eliminates the direct coupling between learning rate adjustment and training time, as voltage amplitude changes do not extend the pulse duration or update cycle.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Productivity

If voltage amplitude variation is used to control learning rate, then training speed is improved, but device reliability may deteriorate

Engineering Contradiction:

Improvetraining speedVSAvoiddevice reliability

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent introduces dynamic voltage amplitude control where the amplitude of voltage pulses is varied during training to achieve different learning rates. This dynamic control allows the system to optimize training speed by using higher amplitudes when needed while maintaining device reliability through controlled operation within safe voltage ranges. The system adapts voltage levels dynamically rather than using fixed high amplitudes, balancing speed and reliability.

Inventive Principle:

Principle #15Dynamics

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 enables faster and more accurate DNN training by maintaining constant update cycle duration while achieving a wide range of learning rates, independent of array size, thus overcoming the limitations of previous methods.

Implementation Method 1

a conductance state of the array of resistive processing units is varied according to the amplitude received at each of the resistive processing units of the array of resistive processing units

Methodology Applied

Scientific EffectResistive switching: Electrical Resistance

Data Source

PatentUS11263521B2Voltage control of learning rate for RPU devices for deep neural network training

Publication Date: 2022.03.01 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

A device, system, product and method of controlling resistive processing units (RPUs), includes applying an input voltage signal to each node of an array of resistive processing units, and controlling a learning rate of the array of resistive processing units by varying an amplitude of the input voltage signal to the array of resistive processing units. A conductance state of the array of resistive processing units is varied according to the amplitude received at each of the resistive processing units of the array of resistive processing units. The controlling of the amplitude of input voltage signal is according to a processor of a control device.