Neural Network Activation Data Bit Width Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network processing methods are inefficient in reducing bit width for activation data, leading to high computational resource usage and processing time, especially in resource-constrained environments like mobile terminals.

Innovation Solution

A method that generates output maps of a neural network's current layer by performing convolution operations, determines a lightweight format based on activation data distribution, and shifts data to achieve low bit width, allowing for adaptive bit width reduction in both input and output maps.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional neural network processing is used without bit width reduction, then processing accuracy is maintained, but computational resource usage and processing time increase significantly

Engineering Contradiction:
Improveprocessing speedVSAvoidcomputational resource usage
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

The patent changes the bit width parameter of activation data from high precision (e.g., 32-bit floating point) to low precision (e.g., 8-bit integer) during neural network processing. This parameter transformation reduces computational resource usage and increases processing speed while maintaining acceptable accuracy through careful selection of quantization strategies and scaling factors.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses lightweight, low-precision data representations instead of heavy, high-precision data structures. By employing low-bit width activation data that requires less computational overhead and memory bandwidth, the system achieves faster processing with reduced energy consumption, effectively replacing expensive computational operations with cheaper alternatives.

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

2Productivity

If bit width reduction is applied to activation data, then computational efficiency improves, but additional operations are required for format conversion

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidprocessing operations
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent performs bit width reduction and format conversion operations in advance, before the main computational workload. By quantizing activation data to low bit width early in the processing pipeline and maintaining this format throughout subsequent layers, the system avoids repeated format conversions and reduces the overall complexity of processing operations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent maintains continuous use of low-bit width data format throughout the neural network processing pipeline. Once activation data is converted to low bit width, this format is preserved across multiple layers and operations, eliminating the need for repeated high-precision conversions and enabling sustained computational efficiency gains.

Inventive Principle:
Principle #20Continuity of useful action

3Quantity of substance

If low bit width format is used for activation data, then memory usage is reduced, but precision and accuracy may be compromised

Engineering Contradiction:
Improvememory usageVSAvoiddata precision
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent carefully manages the precision parameter by using low bit width formats (e.g., 8-bit integers) for activation data while employing appropriate scaling factors and quantization strategies. This approach significantly reduces memory usage compared to high-precision formats while maintaining sufficient accuracy for practical neural network applications through optimized parameter selection.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3528181B1Processing method of neural network and apparatus using the processing method
Publication Date: 2024.04.17 SAMSUNG ELECTRONICS CO LTD
  • EP3528181B1 patent drawingFigure 1
  • EP3528181B1 patent drawingFigure 2
  • EP3528181B1 patent drawingFigure 3

AI summary

A processing method using a neural network includes generating output maps of a current layer of the neural network by performing a convolution operation between input maps of the current layer and weight kernels of the current layer, determining a lightweight format for the output maps of the current layer based on a distribution of at least a portion of activation data being processed in the neural network, and lightening activation data corresponding to the output maps of the current layer to have a low bit width based on the determined lightweight format.