Fixed-Point Neural Network Quantization for Real-Time Inference

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural networks face challenges in processing large amounts of data in real-time due to the complexity of floating-point operations, which hinders efficient analysis and information extraction, especially in resource-constrained devices like mobile devices.

Innovation Solution

A method is introduced to generate a fixed-point quantized neural network by analyzing the statistical distribution of floating-point parameter values, determining fractional lengths for biases and weights, and bit-shifting to optimize operations, thereby reducing the number of operations required for processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If floating-point operations are used in neural networks, then calculation precision is maintained, but processing speed and computational efficiency deteriorate

Engineering Contradiction:
Improvecalculation precisionVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent applies parameter changes by converting the data representation format from floating-point to fixed-point. This involves changing the numerical precision parameters and quantization levels, transforming the neural network operations to use fixed-point arithmetic which is computationally more efficient while maintaining adequate precision for the application

Inventive Principle:
Principle #35Parameter changes

2Reliability

If floating-point neural networks are used, then accuracy is maintained, but computational burden and resource consumption increase

Engineering Contradiction:
ImproveaccuracyVSAvoidcomputational burden
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent changes the numerical representation parameters from floating-point to fixed-point format, which reduces the computational complexity and resource consumption. The fixed-point quantization parameters are carefully selected to maintain the necessary accuracy while significantly reducing the computational burden on resource-constrained devices

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent employs fixed-point arithmetic which can be implemented using simpler, less resource-intensive hardware operations compared to floating-point arithmetic. This approach uses computationally cheaper operations that are sufficient for the application requirements, reducing overall resource consumption

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Ease of operation

If floating-point operations are performed, then processing capability is sufficient, but device complexity and operational difficulty increase

Engineering Contradiction:
Improveoperational simplicityVSAvoidhardware requirements
Core Design Contradiction:
Ease of operationVSDevice complexity

Solution Approach 1:

The patent transforms the operational parameters from floating-point to fixed-point representation, which simplifies the hardware requirements and makes the system easier to operate. Fixed-point operations can be performed using simpler digital logic circuits that are more accessible and easier to implement in resource-constrained environments

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3438890B1Method and apparatus for generating fixed-point quantized neural network
Publication Date: 2024.01.10 SAMSUNG ELECTRONICS CO LTD
  • EP3438890B1 patent drawingFigure 1
  • EP3438890B1 patent drawingFigure 2
  • EP3438890B1 patent drawingFigure 3

AI summary

A method of generating a fixed-point quantized neural network includes analyzing a statistical distribution for each channel of floating-point parameter values of feature maps and a kernel for each channel from data of a pre-trained floating-point neural network, determining a fixed-point expression of each of the parameters for each channel statistically covering a distribution range of the floating-point parameter values based on the statistical distribution for each channel, determining fractional lengths of a bias and a weight for each channel among the parameters of the fixed-point expression for each channel based on a result of performing a convolution operation, and generating a fixed-point quantized neural network in which the bias and the weight for each channel have the determined fractional lengths.