Long-term CMOS static storage cell performing AD/DA conversion for analogue neural network implementations

The authors describe a simple n-bit CMOS static storage cell intended to store and restore synaptic weights in neural network implementations. The principle is based on the successive approximation of the synaptic weight current with binary-weighted current sources, and the result is directly stored into standard static RAM cells. The cell restores a quantised value of the weight current to a common I/O access node by means of a mismatch-insensitive bidirectional current mirror