Neural Network Data Transmission System End-to-End Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing data transmission and processing systems are inefficient due to individually optimized components that do not achieve optimal end-to-end performance, and end-to-end training of neural networks for such systems is impractical for longer block lengths.

Innovation Solution

A method and device that configure a data transmission and processing system using a neural network-based machine learning model trained via stochastic gradient descent, optimizing all components jointly for end-to-end performance, including multiple dense layers and normalization layers in transmitters and receivers, and allowing for distributed detection and classification systems.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If individually optimized functional blocks are used in transmitter and receiver, then each block can be designed and optimized independently, but the end-to-end performance is not optimal

Engineering Contradiction:
ImproveIndependent optimization of functional blocksVSAvoidEnd-to-end performance
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent merges multiple individually optimized functional blocks (source coding, channel coding, modulation, pulse shaping) into a unified end-to-end trainable system. By combining these separate components into a single trainable model, the system achieves optimal end-to-end performance while maintaining the functional separation benefits during the training phase.

Inventive Principle:
Principle #5Merging (Combining)

2Reliability

If end-to-end training of neural network is applied to transmission system, then better end-to-end performance is achieved, but the system does not scale to practical block lengths (e.g., 100 bits)

Engineering Contradiction:
ImproveEnd-to-end performanceVSAvoidTraining complexity for large block lengths
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the transmission system into multiple functional blocks (source coding, channel coding, modulation, pulse shaping) that can be trained independently or in groups, rather than training the entire system as a single monolithic neural network. This segmentation reduces the training complexity and enables scalability to practical block lengths while still achieving good end-to-end performance through coordinated optimization.

Inventive Principle:
Principle #1Segmentation

3Loss of information

If full raw data is transferred over communication channel, then complete information is available at receiver, but the burden on communication system increases

Engineering Contradiction:
ImproveData completenessVSAvoidCommunication system burden
Core Design Contradiction:
Loss of informationVSLoss of energy

Solution Approach 1:

The patent applies parameter changes by transforming the data representation through multiple processing stages (source coding, channel coding, modulation) that optimize the data format for transmission. This allows the system to transfer essential information efficiently by changing the parameters and structure of the data, reducing the burden on the communication system while maintaining data completeness.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3393083B1Method and device for configuring a data transmission and processing system
Publication Date: 2021.09.29 NOKIA TECHNOLOGIES OY
  • EP3393083B1 patent drawingFigure 1~7
  • EP3393083B1 patent drawingFigure 4(a)~4(b)
  • EP3393083B1 patent drawingFigure 5~6

AI summary

Embodiments relate to a method for configuring a data transmission and processing system (1), executed by a configuration device (5), wherein the system (1) comprises at least one transmitter (22), at least one receiver (31), and a communication channel (4) between the transmitter and the receiver, the method comprising: - training (S1) a machine learning model (6) of the data transmission and processing system in function of training data, wherein the machine learning model specify at least a transmitter model (7) including a transmitter neural network (12, 13), a channel model (8), and a receiver and application model (34) including a receiver and application neural network (15, 16), wherein the training data specify respective sets of input vector (s) and output vector (d), wherein an output vector is different from the corresponding input vector, - configuring (S2) at least one of the transmitter and the receiver in function of the trained machine learning model.