Neural Network Receiver Configuration Switching

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks (NNs) used in communication receivers are not configurable for different tasks, requiring multiple NNs to be trained for various modulation schemes and channel conditions, leading to increased hardware resources and inefficiency.

Innovation Solution

Implementing a method that allows a single neural network to represent multiple receiver configurations through superposition of models using quantized context matrices, which reduces memory and computational requirements by using the same set of weights for different receiver configurations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If multiple dedicated neural networks are trained for different modulation schemes and channel conditions, then receiver performance across various settings is improved, but hardware resources (memory, FPGA surface) increase significantly

Engineering Contradiction:
Improvereceiver performanceVSAvoidhardware resources
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent merges multiple dedicated neural networks into a single unified neural network that can handle multiple modulation schemes and channel conditions. This is achieved by training one neural network with diverse training data covering different modulation schemes (BPSK, QPSK, 16QAM, 64QAM) and channel conditions, allowing the network to generalize across tasks without requiring separate hardware implementations for each configuration.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified neural network is designed to be universal, performing multiple functions by adapting to different modulation schemes and channel conditions through its training data. The network structure remains constant while its learned parameters enable it to function as different specialized receivers depending on the input signal characteristics, eliminating the need for multiple dedicated networks.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If multiple dedicated neural networks are implemented for different system configurations, then task-specific performance is improved, but device complexity increases

Engineering Contradiction:
Improvetask-specific performanceVSAvoidnumber of neural networks
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent combines multiple task-specific neural networks into a single multi-functional network. Instead of implementing separate networks for BPSK, QPSK, 16QAM, and 64QAM demodulation, the system uses one unified network trained on all these modulation schemes, reducing the number of network instances from multiple to one while maintaining task-specific performance through its generalized learning capability.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified neural network dynamically adapts its behavior based on the input signal characteristics. By analyzing the received signal properties, the network automatically adjusts its processing to match the appropriate modulation scheme and channel conditions, providing dynamic task switching without requiring multiple static network implementations.

Inventive Principle:
Principle #15Dynamics

3Quantity of substance

If a single neural network is used for all receiver configurations, then hardware resources are reduced, but the network must be trained to generalize across different tasks which increases training complexity

Engineering Contradiction:
Improvehardware resourcesVSAvoidtraining complexity
Core Design Contradiction:
Quantity of substanceVSEase of manufacture

Solution Approach 1:

The patent applies preliminary action by preparing comprehensive training data in advance that covers all desired modulation schemes and channel conditions. The training dataset is constructed to include diverse scenarios (different SNR levels, fading channels, various modulations) so that when the single neural network is trained on this pre-prepared data, it automatically learns to generalize across all tasks without requiring complex runtime decision-making or multiple specialized networks.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4026056B1Signal demapping
Publication Date: 2024.07.31 NOKIA TECHNOLOGIES OY
  • EP4026056B1 patent drawingFigure 1a~1b
  • EP4026056B1 patent drawingFigure 2~3
  • EP4026056B1 patent drawingFigure 4

AI summary

Aspects and embodiments provide an apparatus and method which can be used to assist with selection of one of a plurality of available receiver configurations to demap a received signal at a receiver in a wireless communication network. Apparatus described may comprise trainable logic which may take the form of a neural network. The trainable logic is configured to quantize a plurality of receiver configurations. The trainable logic comprises: determination logic configured to determine a plurality of available receiver configurations and associate each receiver configuration with a context matrix; codebook logic configured to select a quantisation codebook to be used by the trainable logic for the context matrices; and learning logic configured to learn from a training dataset comprising a plurality of received signal samples relevant to reconstruction of a transmitted message; the learning logic being configured to generate from the training dataset: a set of superposed parameters and context matrices corresponding to the plurality of available receiver configurations and a set of contextual parameters for each context, the contextual parameters being generatable from the superposed parameters by combination with the context matrix associated with each receiver configuration quantize the context matrices according to the quantisation codebook; and repeat the generation of superposed parameters, context matrices and quantization of context matrices until a stop criterion is met. Arrangements may recognise that by providing context matrices for each receiver configuration and aligning those context matrices with a simple codebook, it becomes more computationally efficient to perform a context (receiver configuration) switch.