Shift Register Neural Network Weight Storage

The patent applies dynamics by making the weight representation adaptable between binary and multi-value formats. The system dynamically switches between binary weights for inference and multi-value weights for learning, allowing the weight storage system to optimize for either precision or speed depending on the operational mode. This is achieved through a weight storage unit that can hold multi-value weights and a weight conversion unit that converts them to binary when needed.

The patent changes the parameter of weight precision from fixed to variable. By using multi-value weights during learning and converting to binary weights during inference, the system adjusts the precision parameter dynamically. This allows high precision (10-bit or higher) during learning for accurate gradient updates, then switches to binary precision during inference for fast processing.

The patent segments the weight storage and processing functions into distinct units: a weight storage unit for holding multi-value weights, a weight conversion unit for transforming weights between formats, and separate learning and inference processing paths. This segmentation allows each component to be optimized independently, reducing overall system complexity while maintaining high precision weight storage.

The patent introduces a weight conversion unit as an intermediary between the multi-value weight storage and the binary weight processing required for inference. This intermediary component translates between the two weight formats, allowing the system to maintain high-precision multi-value weights for learning while efficiently converting to binary weights for fast inference, thus bridging the gap between precision and speed requirements.

The patent implements periodic action by alternating between learning mode and inference mode with distinct weight representations. During learning periods, high-precision multi-value weights are used and updated. During inference periods, these weights are converted to binary and used for fast processing. This periodic switching between precision modes allows the system to achieve both high learning accuracy and fast inference speed.

The patent applies preliminary action by pre-converting multi-value weights to binary weights before inference operations. The weight conversion unit prepares binary weight representations in advance, so that when inference is needed, the system can immediately use these pre-converted binary weights without computational delay, thus maintaining both learning accuracy and inference speed.