Neural Network Channel Encoding for Polar Codeword Complexity

The patent divides the large codeword space corresponding to high information bit rates into multiple smaller sub-codeword spaces. The neural network processes these segments independently, reducing the traversal complexity from exponential in the total number of information bits to exponential in smaller segments. This allows high data transmission capacity to be achieved without proportionally increasing the traversal complexity.

The patent introduces a new dimension of organization by structuring the codebook into multiple sub-codebooks or partitions. This dimensional reorganization allows the neural network to navigate the codeword space more efficiently by operating in lower-dimensional subspaces rather than directly in the full high-dimensional space, thereby reducing complexity while maintaining capacity.

The patent segments the codeword space coverage task into multiple smaller sub-tasks, each handled by a smaller neural network component or learning process. Instead of training one large neural network to cover the entire codeword space, the system uses multiple smaller networks or a hierarchical structure that processes segments independently. This reduces the learning complexity from training a single large network to training multiple smaller networks, while still achieving comprehensive coverage through aggregation.

The patent applies partial action by having the neural network learn and process only specific sub-codeword spaces or partitions at a time, rather than attempting to learn the entire codeword space in one go. This partial coverage approach reduces the immediate learning burden while maintaining the ability to achieve complete coverage through multiple passes or hierarchical processing of different partitions.