A floating-gate MOS learning array with locally computed weight updates

Abstract

We have demonstrated on-chip learning in an array of floating-gate MOS synapse transistors. The array comprises one synapse transistor at each node, and normalization circuitry at the row boundaries. The array computes the inner product of a column input vector and a stored weight matrix. The weights are stored as floating-gate charge; they are nonvolatile, but can increase when we apply a row-learn signal. The input and learn signals are digital pulses; column input pulses that are coincident with row-learn pulses cause weight increases at selected synapses. The normalization circuitry forces row synapses to compete for floating-gate charge, bounding the weight values. The array simultaneously exhibits fast computation and slow adaptation: The inner product computes in 10 μs, whereas the weight normalization takes minutes to hours.