Neural Network Data Compression Profiles for Memory-Bound Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network processing methods face challenges in efficiently compressing data, leading to high memory requirements and performance limitations, especially in memory-bound operations, due to the complexity of hardware resources in devices like embedded systems.

Innovation Solution

A method involving a compression pipeline with lossy and lossless techniques, including quantization, pruning, and Huffman coding, is implemented to reduce memory traffic by splitting neural network data into lanes and applying optimal compression techniques based on statistical characteristics, thereby determining an optimal configuration for compression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If data is compressed using traditional methods, then memory requirements are reduced, but compression rate and performance are insufficient

Engineering Contradiction:
Improvememory requirementsVSAvoidcompression rate
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The patent divides neural network data into multiple lanes based on bit representations, where each lane can be independently compressed using optimal compression techniques. This segmentation allows parallel processing and significantly improves compression rate while reducing memory requirements through efficient data organization

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically determines optimal compression techniques and configurations based on statistical characteristics of the input data. By adaptively selecting compression methods for different data patterns, the system achieves higher compression rates without compromising memory efficiency

Inventive Principle:
Principle #15Dynamics

2Productivity

If hardware resources are increased to improve processing capability, then neural network performance is improved, but device complexity and cost increase

Engineering Contradiction:
Improveneural network performanceVSAvoidhardware resources
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent replaces complex hardware resources with software-based compression algorithms and statistical analysis. By using lossy and lossless compression techniques along with Huffman coding, the system achieves high neural network performance on resource-constrained devices without increasing hardware complexity

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system changes data representation parameters by quantizing neural network data and applying different compression techniques based on statistical characteristics. This allows high-performance neural network processing using standard hardware resources without requiring specialized high-performance components

Inventive Principle:
Principle #35Parameter changes

3Speed

If memory access is reduced to improve performance, then processing speed is improved, but data transmission efficiency decreases

Engineering Contradiction:
Improveprocessing speedVSAvoidmemory access burdens
Core Design Contradiction:
SpeedVSLoss of energy

Solution Approach 1:

The patent applies compression techniques to neural network data before it is stored or transmitted, reducing the amount of data that needs to be accessed from memory. This preliminary compression action significantly reduces memory access burdens and energy consumption while maintaining processing speed through efficient decompression

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3817236B1Neural network data processing method and apparatus
Publication Date: 2024.10.16 SAMSUNG ELECTRONICS CO LTD
  • EP3817236B1 patent drawingFigure 1
  • EP3817236B1 patent drawingFigure 2
  • EP3817236B1 patent drawingFigure 3

AI summary

A method and apparatus for processing data of a neural network. The method includes: obtaining one or more bit representations of data used for processing a neural network; generating a plurality of candidate profiles based on the bit representations; determining a final profile by comparing compression performances for each of the candidate profiles; and determining an optimal configuration for compressing data of the neural network based on the determined final profile.