Shared-Weight Transmitter Training Across Real Communication Channels

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing end-to-end communication systems face challenges in training due to the inability to backpropagate gradients through actual channels, leading to suboptimal performance from channel model mismatches.

Innovation Solution

A method where multiple transmitters share trainable weights, and a central receiver updates these weights based on a loss function using pseudo-randomly generated channel symbols and perturbations, enabling training without a channel model by broadcasting optimized weights to all transmitters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If end-to-end training is performed with trainable parameters at transmitter and receiver, then communication system performance is improved, but the inability to backpropagate gradients through actual channels causes channel model mismatches leading to suboptimal training

Engineering Contradiction:
Improvecommunication system performanceVSAvoidtraining accuracy
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The patent introduces a receiver as an intermediary that centralizes the training computation. The receiver collects signals from multiple transmitters, computes the loss function, and generates updated weights that are then broadcast back to all transmitters. This mediator approach allows gradient computation to occur at the receiver side where channel information is available, eliminating the need for transmitters to perform complex gradient calculations through actual channels.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent combines the training functions of multiple transmitters into a single receiver-based computation process. Instead of each transmitter independently attempting to backpropagate through its own channel, all transmitters share the same trainable weights that are updated collectively by the receiver based on aggregated loss from all channels. This merging approach simplifies the training architecture and resolves the backpropagation issue.

Inventive Principle:
Principle #5Merging (Combining)

2Ease of operation

If each transmitter performs independent training with its own channel model, then training can proceed without centralized coordination, but channel model mismatches and lack of knowledge about actual channel conditions lead to suboptimal weight updates

Engineering Contradiction:
Improvedistributed training capabilityVSAvoidweight update accuracy
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent implements a feedback mechanism where the receiver computes the actual loss based on received signals and known channel symbols, then broadcasts updated weights back to all transmitters. This closed-loop feedback ensures that weight updates are based on actual channel performance rather than theoretical models, improving the accuracy of training while maintaining a relatively simple distributed operation mode.

Inventive Principle:
Principle #23Feedback

3Manufacturing precision

If centralized weight update computation is performed at the receiver using knowledge of channel symbols and perturbations, then training accuracy is improved, but computational overhead at the receiver increases

Engineering Contradiction:
Improveweight update accuracyVSAvoidcomputational overhead
Core Design Contradiction:
Manufacturing precisionVSPower

Solution Approach 1:

The patent segments the training computation by having the receiver perform only the essential loss computation and weight update generation, while the actual signal processing and feature extraction remain distributed across multiple transmitters. This segmentation allows centralized computation of only the critical weight updates without duplicating the full computational burden at the receiver, thus improving accuracy while controlling computational overhead.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4173243B1Training in communication systems
Publication Date: 2025.11.26 NOKIA TECHNOLOGIES OY
  • EP4173243B1 patent drawingFigure 1~3
  • EP4173243B1 patent drawingFigure 4
  • EP4173243B1 patent drawingFigure 5~6

AI summary

An apparatus, method and computer program is described comprising: receiving, at a receiver (46) of a transmissions system (40), transmitted signals from each of a plurality of transmitters (42a, 42b, 42c), wherein each transmitter communicates with the receiver over one of a plurality of channels (44a, 44b, 44c) of the transmission system, wherein each transmitter includes a transmitter algorithm having at least some trainable weights, wherein each transmitter algorithm has the same trainable weights and wherein each of the transmitted signals is based on a perturbed channel symbol generated at the respective transmitter, wherein the channel symbols and perturbations are known to the receiver; updating said weights of said transmitter algorithm, at the receiver, based on a loss function; providing said updated weights to each transmitter of the transmission system; and repeating the receiving and updating until a first condition is reached.