Non-linear reconfigurable threshold logic gates based on nanostructured metallic films

Compared to Boolean gates, threshold logic gates (TLGs) can substantially simplify the circuit design for artificial neural networks. Architectures based on classical CMOS components or metal-oxide memristors to implement TLG circuits have been reported. Recently we proposed a TLG design, called Receptron, based on nonlinear weights, thus widening the spectrum of Boolean computable functions by a single device, while relying on random search protocols for training. Here we present a theoretical and an experimental characterization of the Receptron model to determine the connection between the structure of the weights and function computability, identifying sub-linearity as an enabling feature. Exploiting these results, we fabricated an improved version of a Receptron device, enhancing its sub-linearity and random search efficiency thanks to an ad hoc circuit. This Receptron-based device can be considered for the integration with other conventional logic components for higher levels of computational complexity.