Neural Decoding of Concatenated Codes for NAND Flash Reliability

The patent replaces traditional algebraic decoding algorithms with a neural network-based system that uses soft computing to decode error correction codes. The neural network processes log-likelihood ratios and syndrome information to directly output corrected codewords, eliminating the need for complex algebraic algorithms like Berlekamp-Massey or Euclidean algorithms, thereby reducing decoding complexity while maintaining or improving error correction capability under stress conditions

The patent changes the operational parameters of error correction by using soft information (log-likelihood ratios) instead of hard decisions, and by adjusting the neural network's internal parameters (weights and biases) through training on synthetic data that models various stress conditions. This allows the system to adapt to changing channel conditions and maintain high reliability without increasing decoding complexity

The patent performs preliminary training of the neural network offline using synthetic data that models various stress conditions and error patterns. This pre-computes the optimal decoding behavior, so that during actual programming and decoding operations, the neural network can quickly process received signals without requiring complex real-time computations, thereby maintaining high programming speed while achieving strong error correction capability

The patent uses synthetic training data that copies and models various stress conditions and error patterns that might occur in real NAND flash memory operations. By training on these copied scenarios, the neural network learns to handle real stress conditions efficiently, achieving strong error correction without the overhead of complex real-time processing that would slow down programming operations