Mixed-Signal Neuron Calibration for Low-Power Accuracy

The patent applies preliminary action by performing offset calibration before the neuron operates in its functional mode. A calibration circuit generates calibration currents that pre-compensate for offset errors, so when the neuron processes actual signals, the offset has already been corrected. This allows the neuron to maintain high accuracy while operating in power-efficient analog mode throughout computation.

The patent introduces an intermediary calibration circuit that mediates between the analog neuron core and the digital control system. This calibration circuit generates corrective calibration currents that are injected into the neuron, acting as an intermediary signal that compensates for offset errors without requiring the entire neuron to operate in high-power digital mode.

The patent segments the neuron operation into two distinct phases: a brief calibration phase that uses digital-to-analog converters for high accuracy, and a sustained computation phase that uses pure analog circuitry for low power consumption. This segmentation allows the system to get accuracy when needed and power efficiency during extended operation.

The patent implements periodic action by performing offset calibration only once or occasionally, rather than continuously. After the initial calibration establishes the correct operating point, the neuron can operate in analog mode without requiring continuous digital intervention, achieving power efficiency during the periodic intervals between calibrations.