ArC Instruments’ ArC ONE has evolved from a long line of instruments used for accelerating research in memristive devices, circuits and systems. An important part of that mandate involves research into neuromemristive systems, i.e. electronic circuits exploiting the unique properties of memristive devices to mimic the general computational principles of the brain. The special architecture of these systems means they come with their own characteristic set of testing and operation requirements and ArC ONE is equipped to meet them.

A case use exemplifying ArC technology can be found in a recent demonstration of unsupervised learning using an artificial neural network employing memristive devices as artificial synapses [1]. The preparations for the experiment required the assessment of each and every memristive device for suitability in its use as an artificial synapse. This task required:

a) the execution of basic functionality tests (is this device switching to a sufficiently high standard? – Figure 1) and

b) an assessment of device performance under the specific input waveforms used to drive plasticity in the artificial neural network (Is this device capable of implementing the learning rule we designed for it? – Figure 2).

Simultaneously, the final experiment required the implementation of a complicated, but stereotyped routine capable of running the artificial neural network and applying the correct pre- and post-synaptic waveforms across the terminals of the artificial synapses (Figure 3).

Using the capabilities of the highly modularised and easy-to-use Python-based interface a bespoke module was designed that could carry out all required tasks from basic device characterisation to full network operation with the specification of a few parameters and at the click of a button. Moreover, the standardised format specification of the resulting test logfiles, created by the system with machine precision, allowed the development of highly automated plotting and analysis routines. This generated results as shown in figure 1(b). Thus, using the staple ArC technologies of software flexibility and array-level interfacing of memristive devices allowed an experiment that would have otherwise taken weeks to complete to be carried out within a few days.

The success of this example study has inspired us at ArC to look towards the next step: using our experience with neuromemristive systems to develop a much more general use and flexible software platform for quick and easy implementation of artificial neural networks using memristors as artificial synapses. Simultaneously, we strive to keep the software architecture as open and modular as possible in order to allow customers to develop their own tools when nothing but the most well-suited of tailor-made solutions will do the job. At ArC instruments we believe that neuromemristive systems hold great promise for the future development of electronics and we shall continue our efforts to accelerate their development.

[1] A. Serb, J. Bill, A. Khiat, R. Berdan, R. Legenstein, and T. Prodromakis, “Unsupervised learning in probabilistic neural networks with multi-state metal-oxide memristive synapses,” Nat. Commun., vol. 7, p. 12611, Sep. 2016.

[2] A. Serb, A. Khiat, and T. Prodromakis, “A biasing parameter optimiser for RRAM technologies,” IEEE Transactions on Electron Devices. 29-Jun-2015.