Neural Network Variable Data Type Selection

Solution Overview

Problem

Neural networks face a trade-off between calculation time and recognition rate, where using floating-point numbers improves accuracy but increases memory and time, while fixed-point numbers reduce memory and time but lower accuracy, leading to a decrease in recognition rate.

Innovation Solution

An information processing device that calculates quantization errors and generates threshold values based on past and current recognition rates to dynamically determine the data type for variables in a neural network, switching between floating-point and fixed-point types to optimize recognition rate and calculation time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If floating-point numbers are used for neural network calculations, then recognition rate is improved, but calculation time and memory usage increase

Engineering Contradiction:

Improverecognition rateVSAvoidcalculation time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies dynamics by making the data type selection dynamic rather than static. The system automatically switches between floating-point and fixed-point numbers based on real-time quantization error thresholds and recognition rate requirements, allowing the neural network to adapt its precision level during different calculation phases or layers to optimize both speed and accuracy

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent implements local quality by applying different data types (floating-point vs. fixed-point) to different variables or layers within the neural network based on their specific quantization error characteristics. Not all variables are quantized uniformly; instead, each variable is evaluated individually to determine the appropriate data type, ensuring high precision where needed and speed where acceptable

Inventive Principle:

Principle #3Local quality

2Productivity

If fixed-point numbers are used for neural network calculations, then calculation time is reduced, but recognition rate decreases

Engineering Contradiction:

Improvecalculation speedVSAvoidrecognition rate

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies parameter changes by dynamically adjusting the precision parameter (data type) of neural network variables based on quantization error analysis. The system changes the parameter from fixed-point to floating-point representation when quantization error exceeds a threshold, thereby maintaining recognition rate while maximizing calculation speed where possible

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The system makes the data type selection dynamic rather than static, automatically switching between floating-point and fixed-point numbers based on real-time quantization error thresholds and recognition rate requirements, allowing the neural network to adapt its precision level during different calculation phases or layers

Inventive Principle:

Principle #15Dynamics

3Quantity of substance

If all variables are quantized to reduce memory usage, then memory amount is reduced, but recognition rate decreases

Engineering Contradiction:

Improvememory amountVSAvoidrecognition rate

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent implements local quality by applying different data types (floating-point vs. fixed-point) to different variables or layers within the neural network based on their specific quantization error characteristics. Not all variables are quantized uniformly; instead, each variable is evaluated individually to determine the appropriate data type, ensuring high precision where needed and speed where acceptable

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent applies parameter changes by dynamically adjusting the precision parameter (data type) of neural network variables based on quantization error analysis. The system changes the parameter from fixed-point to floating-point representation when quantization error exceeds a threshold, thereby maintaining recognition rate while maximizing calculation speed where possible

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS11809995B2Information processing device and method, and recording medium for determining a variable data type for a neural network

Publication Date: 2023.11.07 FUJITSU LTD

AI summary

An information processing device, includes a memory; and a processor coupled to the memory and configured to: calculate a quantization error when a variable to be used in a neural network is quantized, generate a threshold value based on reference information related to a first recognition rate obtained by past learning of the neural network and a second recognition rate that is obtained by calculation of the neural network, determine a variable of data type to be quantized among variables to be used for calculation of the neural network based on the calculated quantization error and the generated threshold value, and execute the calculation of the neural network by using the variable of data type.