Neural Network Signal Classification With Bit-Matrix Load Reduction

The patent segments the classification task into two stages: first, a reduced Neural Network classifies signals into a smaller number of groups; second, a lookup table completes the classification by mapping group identifiers to final signal patterns. This segmentation reduces the Neural Network's output nodes from 256 (for 8-bit signals) to a smaller number of groups, thereby reducing processing load while maintaining accuracy through the combined two-stage approach.

The patent introduces a lookup table as an intermediary component between the reduced Neural Network and the final classification output. The Neural Network outputs a group identifier that serves as a key to the lookup table, which then provides the complete signal pattern classification. This intermediary allows the system to achieve full classification capability without requiring the Neural Network to directly output all 256 possible signal patterns.

The classification capability is segmented between the Neural Network and the lookup table. The Neural Network handles the complex feature extraction and grouping, while the lookup table handles the final mapping to complete signal patterns. This segmentation allows the Neural Network to have fewer output nodes (reducing processing load) while the system as a whole maintains full classification capability through the lookup table's complementary function.

The patent changes the parameter of the Neural Network's output dimension from the full signal pattern space (256 nodes for 8-bit signals) to a reduced group space (fewer nodes representing groups of signals). This parameter change is compensated by the lookup table, which maps the reduced-dimensional output back to the full signal pattern space, thereby maintaining classification capability while reducing the Neural Network's processing burden.