Editorial: Focus on disordered, self-assembled neuromorphic systems

This NCE Focus Issue is motivated by the intriguingly neuromorphic properties of many-body systems self-assembled from nanoscale elementary components. The rationale behind this is that biological neural networks, including in particular their nanoscale synapses, are formed by bottom-up self-assembly, rather than top-down design. Self-assembled nanosystems inherit a disordered network structure and the nonlinear interactions between the networked elements can give rise to emergent properties, as espoused by the legendary Nobel laureate Phillip W Anderson in his famous article 'More is Different' (1972 Science 177 393).

This collection of articles covers self-assembled nano-electronic network systems exhibiting nonlinear electrical properties that are neuromorphic in one or more ways. For example, a common feature is resistive switching, driven by nanoscale physical processes that mimic synapses. Another common feature is recurrent network structure, which allows feedback and enables temporal signal processing. Importantly, it is the physical processes in these neuromorphic systems that enable computational tasks to be performed, by harnessing the collective nonlinear dynamics usually, but not exclusively, via physical reservoir computing (RC).

In the first article [1], Tanaka et al review studies to date investigating networks of carbon nanotubes (CNTs) complexed with chemically dynamic molecules. CNTs complexed with molecules exhibit resistive switching and dense networks of single-walled CNTs complexed with polyoxometalate (SWCNT-POM), in particular, also exhibit spontaneous generation of spikes. Several benchmark learning tasks (i.e. NARMA-10, waveform generation, memory capacity, Boolean logic gates) have been demonstrated using physical RC based on CNT network devices, and RC-based object classification has also been demonstrated in a robotic application. In a separate research article [2], Akai-Kasaya et al demonstrate how physical RC is implemented using the diverse nonlinear response signals from a dense SWCNT-POM network to perform waveform reconstruction and nonlinear autoregression, as well as the memory capacity task.

Montano et al present a grid-graph model of very dense Ag nanowire networks (NWNs) that can account for their observed emergent dynamics, including short-term memory and synaptic plasticity [3]. An analytic potentiation–depression rate balance equation is used to model memristive edge junctions on a regular grid graph model. In related work on Ag NWNs [4], the simulation study by Loeffler et al implements two RC learning tasks with different timescales (waveform transformation and memory capacity) to demonstrate how modularity in a relatively sparse network structure exploits intrinsic long and short-term memory (due to memristive junctions and recurrent loops, respectively) and enables multitasking.

Mirigliano et al present a study using Au thin films structured with nanoclusters that exhibit resistive switching behaviour [5]. Using the dense network of Au nanojunctions as a physical substrate, they demonstrate a 'receptron' device (a generalisation of the perceptron) for binary classification of Boolean functions and classification of nonlinearly separable functions. Note that this is an example of in materio computation and does not invoke RC. Similarly, Ruiz-Euler et al present a study using boron-doped silicon in a dopant network processing unit (DNPU) as a hardware neural network emulator [6]. The dopant (B) forms a network in the host (Si), allowing electrons to hop between two terminals. Using multiple DNPUs as high-capacity nodes, binary classification is demonstrated with a feed-forward DNPU network and simulated MNIST classification with 10 DNPU nodes.

We hope this collection provides new insights into the intrinsic neuromorphic properties of physical substrates based on self-assembled nano-networks and how these properties can be harnessed for neuromorphic computing.