The exponential, Moore's Law, progress of electronics may be continued beyond the 10-nm frontier if the currently dominant CMOS technology is replaced by hybrid CMOL circuits combining a silicon MOSFET stack and a few layers of parallel nanowires connected by self-assembled molecular electronic devices. Such hybrids promise unparalleled performance for advanced information processing, but require special architectures to compensate for specific features of the molecular devices, including low voltage gain and possible high fraction of faulty components. Neuromorphic networks with their defect tolerance seem the most natural way to address these problems. Such circuits may be trained to perform advanced information processing including (at least) effective pattern recognition and classification. We are developing a family of distributed crossbar network (CrossNet) architectures that permit the combination of high connectivity neuromorphic circuits with high component density. Preliminary estimates show that this approach may eventually allow us to place a cortex-scale circuit with about 10(10) neurons and about 10(14) synapses on an approximately 10 x 10 cm(2) silicon wafer. Such systems may provide an average cell-to-cell latency of about 20 nsec and, thus, perform information processing and system training (possibly including self-evolution after initial training) at a speed that is approximately six orders of magnitude higher than in its biological prototype and at acceptable power dissipation.

SUMMARYThis paper reviews recent important results in the development of neuromorphic network architectures ('CrossNets') for future hybrid semiconductor=nanodevice-integrated circuits. In particular, we have shown that despite the hardware-imposed limitations, a simple weight import procedure allows the CrossNets using simple two-terminal nanodevices to perform functions (such as image recognition and pattern classiÿcation) that had been earlier demonstrated in neural networks with continuous, deterministic synaptic weights. Moreover, CrossNets can also be trained to work as classiÿers by the faster error-backpropagation method, despite the absence of a layered structure typical for the usual neural networks. Finally, one more method, 'global reinforcement', may be suitable for training CrossNets to perform not only the pattern classiÿcation, but also more intellectual tasks. A demonstration of such training would open a way towards artiÿcial cerebral-cortex-scale networks capable of advanced information processing (and possibly self-development) at a speed several orders of magnitude higher than that of their biological prototypes.

SUMMARYExtremely dense neuromorphic networks may be based on hybrid 2D arrays of nanoscale components, including molecular latching switches working as adaptive synapses, nanowires as axons and dendrites, and nano-CMOS circuits serving as neural cell bodies. Possible architectures include 'free-growing' networks that may form topologies very close to those of cerebral cortex, and several species of distributed crossbar-type networks, 'CrossNets' (including notably 'InBar' and 'RandBar'), with better density and speed scaling. Numerical modelling show that the speciÿc signal sign asymmetry used in CrossNets allows self-excitation of recurrent networks with long-range cell interaction, without a symmetry-breaking global latchup. Our next goal is to develop methods of globally supervised teaching of extremely large networks with no external access to individual synapses. Such development would open a way towards cerebral-cortex-scale networks (with ∼ 10 10 neural cells and ∼ 10 14 synapses) capable of advanced information processing and self-evolution at a speed several orders of magnitude higher than their biological prototypes.

No abstract