Communication Autoencoder Training With Quantized Fixed-Point Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing communication systems face challenges in practical hardware implementation due to high memory requirements and computational complexity of neural networks, which are often not viable with available hardware resources, particularly in communication systems lacking graphic processing units (GPUs).

Innovation Solution

Implementing neural networks with compressed representations using quantization and fixed-point arithmetic, and employing evolutionary computation to train autoencoders for end-to-end communication systems, reducing the complexity and memory requirements while maintaining performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If neural networks are implemented with high precision floating point arithmetic for accurate communication system training, then measurement precision is improved, but device complexity and memory requirements increase significantly

Engineering Contradiction:
ImproveprecisionVSAvoidcomplexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the numerical precision parameter from high-precision floating point arithmetic to low-precision fixed-point arithmetic. This parameter change allows the neural network to operate with reduced computational complexity and memory requirements while maintaining sufficient accuracy for communication system training, directly resolving the contradiction between precision and device complexity

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent employs quantized fixed-point representations that can be implemented using simpler, more resource-efficient hardware compared to floating-point units. This approach uses computationally lighter operations that consume less memory and processing power, effectively replacing complex floating-point arithmetic with simpler fixed-point operations

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

2Measurement precision

If neural networks are implemented with high precision floating point arithmetic, then measurement precision is improved, but the use of energy increases due to GPU requirements

Engineering Contradiction:
ImproveprecisionVSAvoidenergy
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent changes the numerical precision parameter from high-precision floating point arithmetic to low-precision fixed-point arithmetic. This parameter change allows the neural network to operate with reduced computational complexity and memory requirements while maintaining sufficient accuracy for communication system training, directly resolving the contradiction between precision and device complexity

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent employs quantized fixed-point representations that can be implemented using simpler, more resource-efficient hardware compared to floating-point units. This approach uses computationally lighter operations that consume less memory and processing power, effectively replacing complex floating-point operations

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

3Device complexity

If quantization is applied to reduce precision and computational complexity, then device complexity is reduced, but measurement precision deteriorates

Engineering Contradiction:
ImprovecomplexityVSAvoidprecision
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent changes the numerical precision parameter from high-precision floating point arithmetic to low-precision fixed-point arithmetic. This parameter change allows the neural network to operate with reduced computational complexity and memory requirements while maintaining sufficient accuracy for communication system training, directly resolving the contradiction between precision and device complexity

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3912094B1Training in communication systems
Publication Date: 2026.03.25 NOKIA TECHNOLOGIES OY
  • EP3912094B1 patent drawingFigure 1~2
  • EP3912094B1 patent drawingFigure 3~4
  • EP3912094B1 patent drawingFigure 5~6

AI summary

An apparatus, method and computer program is described comprising:evaluating some or all of a current population of algorithms according to a metric, each algorithm of the population implementing a transmission system;selecting a subset of the algorithms of the current population based on the metric; generating an updated population of algorithms from said subset; and repeating the evaluating, selecting and generating, based on the updated population, until a first condition is reached.