Single-electron latching switches as nanoscale synapses

Nanoscale latching switches based on controllable single-electron transfer and trapping may serve as a synaptic basis for extremely dense and fast self-evolving BiWAS (binary weight, analog signal) neural networks. We have designed and simulated two devices of this type, a "propagating" switch and a "branching" switch, as well as multi-entry switching nodes based on their combination. We have also carried out a preliminary study of two architectures of neural networks based on 2D arrays of the switching nodes: a "free-growing" network in which the shape of axonic and dendritic trees may be very complex, and a "randomized distributed crossbar" network in which axons and dendrites are implemented as straight wire segments. The latter network scales much better, but the former one may be more adequate for input parts of very large scale networks.