Analog Floating-Gate Synapses for General-Purpose VLSI Neural Computation

VLSI neural circuits can play an important role in real-time image and signal processing. To make the neural computing hardware more powerful, compact and electrically programmable synapses are sorely needed. A floating-gate-based synapse structure fabricated by a standard double-polysilicon CMOS process is presented. Simulation and experimental results on conductance programmability and charge retention demonstrate the capability of this programmable synapse circuit. With this new programmable synapse circuit, a neural chip of 100 K synapses complexity can be constructed using one megabit static-RAM fabrication technologies.