Self-contained error-compensated N-valued memory for neural applications

One of the main challenges faced in VLSI implementations of neural networks and other massively parallel analog signal processing systems is the long-term local storage of analog information. The author presents a self-contained multilevel memory which is capable of retaining an analog value indefinitely and has a built-in error control mechanism to give the system robustness against noise. The implementation requires small silicon area and the cells are modular to permit direct pile up