SUPERCONDUCTING NEURAL COMPUTER

We fabricated neural-based superconducting integrated-circuits by usin g Nb/AlOx/Nb Josephson junctions, and demonstrated circuit operation o f 2-bit neural-based A/D converter as an example of optimization probl em. We used, in the present superconducting neural circuits, fluxon pu lses as the neural impulses and a Josephson junction as a threshold el ement to achieve an advanced operation. The integrative time delays du e to the inductance of superconducting circuits correspond to those du e to capacitance in real neurons. The values of resistor by which Jose phson transmission lines are connected represent fixed synaptic streng thes. The preliminary experimental result suggests that variable criti cal currents of d.c.SQUID may provide synapses with variable strength. In our implementation scheme, complex circuit design can easily be re alized by connecting Josephson transmission lines. In order to design complex circuits like a Hopfield neural network, we propose a new meth od which can be used to analyze the properties of local minima in the energy function. A novel design technique to eliminate these local min ima in the networks has been developed. It is useful for hardware stud ies on artificial neural networks using superconductor and semiconduct or.