Neural Network Decoder for Low-Latency Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing error correction coding techniques, such as LDPC, Reed-Solomon, BCH, and Polar coding, increase processing complexity and resource usage, making them undesirable for applications requiring ultra-low power consumption and latency, such as IoT and tactile internet.
Innovation Solution
Multi-layer neural networks with nonlinear mapping and distributed processing capabilities are used to decode encoded data, reducing errors introduced by noise and minimizing resource usage by transferring complexity to an offline training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional error correction coding techniques (LDPC, Reed-Solomon, BCH, Polar coding) are used, then error correction capability is improved, but processing complexity and resource usage increase
Solution Approach 1:
The patent replaces traditional mechanical/mathematical decoding algorithms (LDPC, Reed-Solomon, BCH, Polar coding) with a neural network-based system. The neural network learns error correction patterns during training and performs decoding through neural activation functions, substituting complex algorithmic processing with trained neural model inference, thereby reducing processing complexity while maintaining error correction capability
Solution Approach 2:
The patent performs extensive training of the neural network offline before deployment. During this preliminary action, the neural network learns optimal error correction strategies for various encoding schemes. This pre-learning transfers complexity from the runtime decoding process to the offline training phase, enabling fast and simple inference during actual operation
2Reliability
If traditional error correction coding techniques are used, then error correction capability is improved, but power consumption and latency increase
Solution Approach 1:
The neural network decoder replaces computationally intensive traditional decoding algorithms with a trained neural model that performs decoding through simpler neural operations. This substitution dramatically reduces the computational resources required during runtime, leading to lower power consumption and reduced latency, especially critical for mobile and IoT applications
Solution Approach 2:
By performing the complex learning process offline during training, the patent transfers computational burden from runtime operation to preliminary preparation. The trained neural network model can then be deployed with minimal runtime computation, enabling ultra-low power consumption and low latency during actual decoding operations
3Reliability
If traditional error correction coding techniques are used, then error correction capability is improved, but resource usage increases
Solution Approach 1:
The patent substitutes traditional decoding algorithms that require extensive computational resources with a neural network-based system. The neural network, once trained, requires significantly fewer resources for decoding operations, reducing CPU/GPU utilization, memory access, and overall system resource consumption while maintaining or improving error correction performance
Data Source
AI summary
Examples described herein utilize multi-layer neural networks to decode encoded data (e.g., data encoded using one or more encoding techniques). The multi-layer neural networks include an encoder configured to encode input data using encoded bits in accordance with an encoding technique and to provide encoded input data, and a memory configured to receive the encoded input data from the encoder and configured to store the encoded input data. The multi-layer neural networks further include combiners configured to receive the encoded input data from the memory and further configured to combine the encoded input data among a set of predetermined weights. The combiners are further configured to provide encoded data with reduced noise, the noise introduced by the memory.