Neural Network Decoding With Syndrome Checks for Lower BLER
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing approaches for training neural networks to decode linear block codes, such as Bose-Chaudhuri-Hocquenghem (BCH) or Polar codes, face challenges in reducing Block-Error-Rate (BLER) and require additional computations due to the presence of cycles and trapping sets in the code graph, with existing methods focusing on Bit-Error-Rate (BER) reduction without explicit methods for BLER reduction.
Innovation Solution
A novel loss metric is introduced that trains the neural network to minimize Block-Error-Rate (BLER) by performing a syndrome check at every even layer of the network, optimizing parameters only where the syndrome check is not met, and using stochastic gradient descent methods to minimize the loss function, which includes a cross-entropy loss function calculated based on intermediate output representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing approaches train neural networks using cross entropy loss functions to reduce Bit-Error-Rate (BER), then BER performance is improved, but Block-Error-Rate (BLER) reduction is not achieved and additional computations are required
Solution Approach 1:
The patent changes the loss function parameter from cross entropy (which optimizes BER) to a syndrome-check-based loss function (which optimizes BLER). This parameter change in the training objective enables the neural network to directly minimize block error rate while maintaining computational efficiency through early termination capabilities.
2Productivity
If the neural network is trained end-to-end for the entire unrolled graph, then comprehensive optimization is achieved, but additional computations and performance degradation occur
Solution Approach 1:
The patent segments the training process by introducing intermediate syndrome check points within the neural network architecture. Instead of treating the entire unrolled graph as a single end-to-end optimization problem, the network is divided into segments that can be independently evaluated against syndrome checks, enabling early termination and reducing overall computational complexity.
Solution Approach 2:
The patent applies preliminary action by performing syndrome checks at intermediate layers during training before the final output is produced. This allows the network to detect successful decoding early in the processing sequence and terminate further computations, preventing unnecessary computational overhead associated with complete end-to-end processing.
3Reliability
If syndrome checks are performed at every layer to ensure decoding accuracy, then BLER is reduced, but computational load increases
Solution Approach 1:
The patent applies partial action by performing syndrome checks only at specific intermediate layers rather than at every single layer. This selective approach provides sufficient decoding accuracy verification to reduce BLER while avoiding the excessive computational burden of checking at every possible layer, achieving an optimal balance between reliability and complexity.
Data Source
AI summary
Methods and apparatus for training a neural network to recover a codeword and for decoding a received signal using a neural network are disclosed. According to examples of the disclosed methods, a syndrome check is introduced at even layers of the neural network during the training, testing and online phases. During training, optimisation of trainable parameters of the neural network is ceased after optimisation at the layer at which the syndrome check is satisfied. Examples of the method for training a neural network may be implemented via a proposed loss function. During testing and online phases, propagation through the neural network is ceased at the layer at which the syndrome check is satisfied.