Memristors as synaptic emulators hold promise for neuromorphic computing
The seemingly simple two-terminal device presented in a Nature Materials journal in a paper titled “Memristors with diffusive dynamics as synaptic emulators for neuromorphic computing” is not only capable of emulating spike-timing-dependent plasticity (STDP) but also paired-pulse facilitation (PPF) followed by paired-pulse depression (PPD) and when combined with a non-volatile element (a drift-type memristor), spike-rate-dependent plasticity (SRDP) could be obtained. All these phenomenon are observed in biological synapses, initiating both short- and long-term plasticity of the synapses and forming the basis of memory and learning.
As well as enabling a substantial reduction in footprint, complexity and energy consumption compared to three-terminal CMOS synaptic circuits, the two-terminal device doesn’t require complex circuitry to simulate synaptic behaviour.
The diffusive memristors described in the paper consist of two platinum or gold inert electrodes sandwiching a switching layer of a dielectric film with embedded silver nanoclusters (SiOxNy:Ag, HfOx:Ag or MgOx:Ag). Devices were first built with a footprint of 10×10µm, the researchers then demonstrated similar switching behaviours for nano-device only 100nm by 100nm.
The resistance ratio between the conducting and insulating states was five orders of magnitude in SiOxNy:Ag and over ten orders in HfOx:Ag devices, the highest ever reported for a threshold switching device, the researchers wrote. They also reported very sharp turn-on slopes, around 10mV per decade in MgOx:Ag and SiOxNy:Ag and around 1mV per decade in HfOx:Ag.
The researchers demonstrated the key role of atomic silver for the bio-realistic synaptic behaviour of these diffusive memristors. Both In situ high-resolution transmission electron microscopy and nanoparticle dynamics simulations demonstrated that Ag atoms disperse under electrical bias and regroup spontaneously under zero bias due to interfacial energy minimization within the dielectric, closely resembling the synaptic influx and extrusion of Ca2+ ions in the pre- and postsynaptic compartments of biological synapses.
Here, the threshold switching is unipolar in nature (no need for a reverse bias), hence significantly different from non-volatile drift-type memristors.
Under an applied pulse, the devices exhibited threshold switching to a low resistance state after an incubation period related to the growth and clustering of silver nanoparticles to eventually form conduction channels. Following channel formation, the current jumped abruptly by several orders of magnitude, and then slowly increased further under bias as the channel thickened.
As the voltage pulse ended, the device relaxed back to its original high-resistance state over a characteristic time (which decreased as the ambient temperature increased). This characteristic time (tens of milliseconds) was on the same order as the response of bio-synapses, the researchers noted. Both the incubation period and the characteristic time were also functions of the voltage pulse parameters, operation history, silver concentration, host lattice, device geometry, humidity and other factors, all of which could be tuned to achieve the desired dynamics for neuromorphic systems.
Interestingly, although a single short pulse could not excite enough Ag particles to form a complete conducting path between the two terminals, applying subsequent pulses at intervals shorter than the diffusion relaxation time characteristic time (before the particles are reabsorbed), pushed more particles into the gap between terminals resulting in a gradual increase in device conductance, similar to the paired-pulse facilitation (PPF) phenomenon in bio-synapses (until a conducting bridge is formed).
Another interesting effect is that as the electric field pumps more particles towards one of the device terminals, the number of particles at the other terminal decreases. More consecutive pulses at high frequency (5,000Hz) end up widening the distribution gap at the depleted terminal, eventually altering the device’s conductance under excessive stimulation. This is another typical synaptic behaviour known as paired-pulse depression (PPD) following initial PPF.
The paper continues emulating the long-term plasticity of synapses by combining a diffusive memristor in series with a Pt/TaOx/Ta/Pt drift memristor as a non-volatile element. Connected to pulsed voltage sources similar to a synapse between pre- and postsynaptic neurons, the combined element benefited from the intrinsic timing mechanism offered by the diffusive memristor, rendering SRDP and STDP behaviours without any complex pulse engineering or spike overlapping.
IBM imitates neurons in cognitive computing advance
Lawrence Livermore, IBM partner on brain-like supercomputer
IBM neurocomputer detailed
Military startup launches neural processor for machine learning
BrainChip CEO gives details of neural network architecture