Sparse Neural Network Data Rearrangement for Fewer Invalid Operations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network processing technologies face inefficiencies due to the performance of unnecessary operations on data with invalid values, leading to increased computational complexity and resource utilization.

Innovation Solution

The method involves identifying sparsity in input data based on valid information, rearranging the data to minimize the presence of invalid values, and applying specific rules to generate output data through convolution operations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If neural network processing performs operations on all input data including invalid values, then complete data processing is achieved, but computational complexity and resource utilization increase

Engineering Contradiction:
Improveprocessing completenessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent extracts and removes invalid values (zero elements) from the input data matrix before processing. By separating valid data from invalid data and eliminating unnecessary computations on zero elements, the system reduces computational complexity while maintaining processing completeness on valid data only.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs preliminary rearrangement of the input data matrix to group valid values together and position them optimally before the main processing operation. This preliminary action reorganizes data to minimize subsequent computational operations, reducing the number of arithmetic operations required while ensuring all valid data is processed.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If neural network processing performs operations on all input data including invalid values, then complete data processing is achieved, but resource utilization increases

Engineering Contradiction:
Improveprocessing completenessVSAvoidresource utilization
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts invalid values (zero elements) from the computation process entirely. By removing these elements before processing, the system eliminates wasted computational resources and energy consumption on operations that would produce zero results, while still ensuring complete processing of all valid data.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs preliminary data reorganization to optimize the processing sequence and minimize resource usage. By rearranging valid values to contiguous positions and eliminating zero elements beforehand, the system reduces the total number of computational operations required, thereby lowering energy consumption and resource utilization.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If data is rearranged based on sparsity distribution, then unnecessary operations are minimized, but additional processing steps are required

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidprocessing steps
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent performs a preliminary rearrangement of the input data matrix to group valid values together and eliminate zero elements before the main processing operation. This single preliminary reorganization step reduces the number of arithmetic operations required in subsequent processing, improving overall efficiency without requiring multiple complex processing stages.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3789892B1Method and apparatus for processing data
Publication Date: 2026.02.18 SAMSUNG ELECTRONICS CO LTD
  • EP3789892B1 patent drawingFigure 1
  • EP3789892B1 patent drawingFigure 2
  • EP3789892B1 patent drawingFigure 3

AI summary

A method of processing data includes identifying a sparsity of input data, based on valid information included in the input data, rearranging the input data, based on a form of the sparsity, and generating output data by processing the rearranged input data.