Sparse-Readout Reservoir Computing for Optical Communication Power Reduction

The patent extracts only the necessary connections from the full readout layer by identifying and retaining significant weights while eliminating redundant ones. This is achieved through threshold-based pruning where connections below a certain weight threshold are removed, creating a sparse readout layer that maintains essential functionality while reducing power consumption.

The patent applies local quality by making different parts of the readout layer have different properties - some connections are retained while others are pruned based on their individual weight magnitudes. This creates a heterogeneous structure where important connections remain dense while less important areas become sparse, optimizing the trade-off between performance and power consumption locally.

The patent performs preliminary action by pre-calculating the sparse structure configuration using elastic linear regression before deployment. The significant weights are identified and the sparse pattern is determined in advance through offline training, which simplifies the runtime configuration and reduces the complexity of setting up the sparse readout layer in the actual communication device.

The patent introduces an intermediary elastic linear regression model that acts as a bridge between the full readout layer and the sparse readout layer. This intermediary model is trained to learn the mapping relationships and generate the sparse configuration, thereby reducing the direct complexity of configuring the sparse structure in the communication device itself.

The patent segments the training process into two distinct phases: an offline training phase where the elastic linear regression model is trained to generate sparse configurations, and an online deployment phase where the pre-trained model is used for inference. This segmentation allows complex training computations to be performed once offline, reducing the ongoing training complexity during actual device operation.

The patent performs the computationally intensive parameter generation and model training as a preliminary action before deployment. The elastic linear regression model is trained offline to learn the optimal sparse configurations, so that during actual communication device operation, only lightweight inference is required, reducing real-time computational power requirements.