Neuromorphic architectures offer great promise for achieving computation capacities beyond conventional Von Neumann machines. The essential elements for achieving this vision are highly scalable synaptic mimics that do not undermine biological fidelity. Here we demonstrate that single solid-state TiO2 memristors can exhibit non-associative plasticity phenomena observed in biological synapses, supported by their metastable memory state transition properties. We show that, contrary to conventional uses of solid-state memory, the existence of rate-limiting volatility is a key feature for capturing short-term synaptic dynamics. We also show how the temporal dynamics of our prototypes can be exploited to implement spatio-temporal computation, demonstrating the memristors full potential for building biophysically realistic neural processing systems.
Connections among neurons are more than simple cables that transmit signals: they are known to exhibit temporal, revertible dynamics in short time scales. These features are termed short-term plasticity (STP) and are well described both experimentally1,2,3,4,5 and in the context of specific models6,7,8,9,10. The role of STP in neuronal computation is thought to be related to temporal processing, see for instance11 or the work by Carvalho and Buonomano12, where STP was shown to enhance the discrimination ability of a single neuron, i.e. a tempotron13, when presented with forward and reverse patterns. Synapses with short-term plasticity are also optimal estimators of presynaptic membrane potentials14, and correlate to specific brain connectivity configurations15,16, that are hypothesized to emerge via learning processes17,18,19. Most interestingly, the vast majority of artificial and neuromorphic brain-like systems focus on stable modifications of connections, known as long-term plasticity, which are assumed to be the basis of memory, or do not make use of the computational power that short-term plasticity may provide, but rather demonstrate a behavior akin to short term dynamics. Moreover, they largely ignore the fact that synapses are inherently unreliable and there is often a large variance in their response to a specific signal, also apparent in short-term dynamics.
Among the several candidates for fabricating brain-like, neuromorphic systems, memristors20 are particularly promising: their characteristic signature of hysteresis is typically noticed in systems and devices that possess certain inertia, manifesting memory, including neural systems. Notwithstanding the several efforts for implementing these mechanisms via Complementary Metal-Oxide-Semiconductor (CMOS) topologies and emerging nanoscale cells, that were able to capture short-term plasticity21,22,23,24, memristors have the potential to alleviate challenges imposed by CMOS implementations25. They can reduce energy consumption and size by exploiting their simple (two terminal) architecture and small footprint26, their capacity to store multiple bits of information per cell27 and the miniscule energy required to write distinct states28. Memristors have been shown to exhibit properties akin to long-term plasticity, such as Spike-timing Dependent Plasticity (STDP)29, along with STDP variations30,31, in compact and large scale cross-bar architectures32,33,34. The majority of such approaches relies on non-volatile memory-state transitions based upon phase-change31,35 mechanisms or the diffusion of ionic-species within an active core36,37,38.
The functional properties of memristive devices are however associated with irreversible rate-limiting electro/thermo-dynamic changes that quite often bring them in far from equilibrium conditions, rendering a rate-limiting volatility39. While the majority of researchers focus on demonstrating how non-volatile conductance changes in memristive devices can resemble the STDP induced changes in real synapses, only few studies exist that leverage metastable effects in memristive devices40 towards reproducing short-term synaptic dynamics41. For example, Chang et al.42 as well as Lim et al.43, linked their devices response to short-term memory, with the latter study correlating their results to the Bienestock-Cooper-Munro (BCM) rule; a key long-term plasticity rule. Similar characteristics were also explored in the context of inference44. More relevant to our work, Zhu et al.45 employed nanogranular-based SiO2 devices and demonstrated pair-pulsed facilitation, which is a form of short-term synaptic plasticity. Nonetheless, in most cases the equivalence between the physics of memristive devices and the physics governing the behavior of biological synapses has been shown only at an abstract qualitative level. Here we focus on demonstrating how single TiO2 memristors are capable of capturing short-term synaptic plasticity by adopting the same experimental protocols and models41,46 used to validate the response of real synapses. We also show for the first time how this particular short-term plasticity properties of our memristive devices can be exploited for implementing spatio-temporal computation, following biologically realistic neural computation paradigms12,13,47.
Switching dynamics of TiO2 memristors
Hysteresis is typically noticed in elements that possess certain inertia; manifesting memory48. Verily, considerably larger systems are also known to have similar non-linear signatures49. Particularly in nanoscale memristors this inertia is ascribed to Joule heating50, the electrochemical migration of oxygen vacancies51,52,53, the lowering of Schottky barrier heights at interfacial states54, the phase-change55,56 and the formation/rupture of conductive filaments57,58 in a device’s active core. Excellent reviews that cover distinct switching mechanisms exist27,59,60, which overall can be classified in three categories61: (1) displacement of ionic species, (2) phase-change and (3) the formation of conductive filaments. Recently we demonstrated that substantial resistive switching is only viable through the formation and annihilation of continuous conductive percolation channels62 that extend across the whole active region of a device, no matter what the underlying physical mechanism is. Innately, in the case of TiO2 memristors the underlying functional mechanism is a manifestation of all three categories: ionic-species (O−2 vacancies53 and/or Tin+ interstitials63) are translocated within the active core that cause the formation of locally reduced Magnelli phases57, which in turn extend along the TiO2 core as current percolating branches that can be annihilated via Joule heating60.
Our memristive device qualitatively represents a synapse (inset I of Fig. 1a), with its conductance corresponding to the notion of a synaptic efficacy modulated via the arrival of a spike, i.e. a pulse applied pre-synaptically to the device’s top electrode (TE), shown in inset II of Fig. 1a. The post-synaptic current entering the artificial neuron, from the device’s bottom electrode (BE), is proportional to the memristive conductance. Figure 1a depicts a microphotograph of one of our fabricated crossbar type TiO2-based memristors (fabrication details are given in Methods). The device comprises two Pt electrodes (TE and BE) that are separated by a stoichiometric TiO2 active core (cross-section is shown in inset II of Fig. 1a). Following an electroforming step (depicted in Figure S1), the devices’ electrical characteristics were first investigated via positive/negative ±2 V voltage sweeps, resulting into a bipolar mode of switching: positive sweeps cause low- (LRS) to high-resistive state (HRS) transitions, while negative ones cause HRS to LRS transitions. The corresponding current-voltage (I-V) characteristics, with the classical pinched-hysteresis memristor signature is shown in the supplementary material Figure S2a.
In order to induce a non-volatile (stable) resistive transition, a critical energy barrier Ei has to be exceeded that will allow toggling between long-term thermodynamically stable (non-volatile) states. It is interesting to note that the required activation energy Ei for toggling between stable states depends upon the previous state of the device, causing it to act as a non-linear accumulator. This is demonstrated in Figure S2c by employing subsequent identical voltage pulses that result into a non-uniform modulation of the effective resistance of our prototypes. Particularly, the notation used in Fig. 1b denotes that a non-uniform lowering39 (increasing) of the barrier occurs as the applied electric field elicits a HRS to LRS (LRS to HRS) non-volatile transition. In the case however the energy provided to the system is less than the corresponding activation energy Ei, a transient (volatile) response can be temporally induced, with the initial equilibrium state being eventually restored, as depicted in Fig. 1c. Such a response is equivalent to the revertible, use-dependent modifications observed in synaptic connections, known as “short-term” plasticity (STP). This state volatility occurs in memristors, due to metastable phase-transitions within the functional active core and precedes the induction of any long-term phase-change in the device’s bulk via forming (field-driven) or annihilating (thermally) stable conductive percolation channels. The kinetics of this process are concurrently governed by mass diffusion29 and nucleation64 processes that also form the basis of STP events in the atomic switch40. The accumulating nature of these events elicit activity-dependent state modulations that can cause unpredictable switching trends, which in combination with the exhibited volatility add a substantial probabilistic component in the devices’ switching dynamics65. Here, we employ this inherent rate limiting volatility of our TiO2-based memristive devices for emulating the temporal (short-term) behavior of real synapses. In what follows, we utilise our memristive synapses with an Integrate-and-Fire (IF) neuron to practically exploit short-term plasticity for spatiotemporal computation.
Non-associative, short-term synaptic plasticity in single TiO2 memristors
The transient conductance response of the same TiO2 memristor is first utilized to model STP changes in synapses, as shown on Fig. 2a,b. Figure 2c shows the voltage pulse pattern used to produce these responses (see supplementary material Figure S3 for a detailed description of the experimental setup used here that evolved into the one described in66). Each voltage pulse induces an increase in the conductance of the device, which then tends to slowly decay to its original state. Subsequent pulses have a similar effect that however depend on the previous resistive state of the device, the conductance peaks can be lower or higher in magnitude than the first peak, similar to short term plasticity mechanisms, as observed in biological synapses41.
In particular, the form of short-term plasticity emulated with this device is “short-term facilitation”. Figure 2a,b show two different cases of facilitation: in Fig. 2a we reproduce the classical form of short-term facilitation (here denoted as STP-F), where each input pulse has the effect of increasing the conductance, including its peak response. At the arrival of the first pulse the memristive conductance increases from the base line (blue line) allowing current to pass through to the neuron, causing an increase on the membrane potential of the connected neuron, shown by the black solid line. On the second spike, the conductance transiently increases even more and so at the third spike, leading to a behavior akin to short-term potentiation, where the synaptic efficacy transiently changes in short time scales. The observed phenomenon is clearly non-linear and cannot be explained by the linear summation of the input signals. We have deliberately chosen a fast membrane potential time constant so that no residuals are remaining from the previous spikes, demonstrating that each spike indeed contributes to the membrane voltage by a different amount. Modeling the measured conductance peaks by a reduced version of the Tsodyks-Markram model67 (see supplementary section), verified quantitatively the equivalence with biological synapses and also revealed that the time constants involved in the process are close to typical biological values. At a subset of our experiments, particularly when the initial conductance of the material is a relatively higher state, the same protocol will lead to a transient conductance at the second and third spike clearly lower than the initial one. We call this phenomenon saturation (STP-S), and we provide a hypothesis of how this deviation from the typically observed response of the memristors occurs. We argue that with every pulse, the potential alignment of filaments comprising of reduced TiO2 leads to a higher conductive state, following the model presented in Fig. 1b. In the case where the provided energy does not exceed the instantaneous Ei that would cause the device to undergo LTP or LTD (in the case of a significant energetic overshoot), such a transition is revertible and after a transient the conductance returns to its initial state. However, if the provided energy will be (is about to be) dissipated via an existing (partially formed) filament, at the same time annihilating it, we will most likely observe saturation in the response, as shown in Fig. 2b. We argue that the exhibited STP-S response stems from the mobility saturation of the available ionic resources (mainly O−2 vacancies) in the vicinity of a partially reduced TiO2−x volume, i.e. a partially formed filament, particularly when the device has been previously stimulated. Clearly, a finite number of mobile resources exist within the volume of interest that can play a role in reducing TiO2 from insulating towards (semi-) metallic phases, i.e. towards forming a conductive filament. And as single devices could in principle host multiple filaments46 within their functional cores, alike short-term plasticity phenomena can be triggered across a wide conductance spectrum (see Figure S8 in supplementary material).
In a concurrent experiment, the same device was subjected to a train of 3 consecutive voltage pulses of −4 V, 10 μs wide and inter-pulse interval tint = 400 ms. This sequence was repeated 600 times with a recovery interval between sequences trec = 10 s, to allow for the device’s state being restored. The pre-stimuli initial conductance was found to vary within 2.85 and 3.1 μS. This range was divided into 17 equal conductance bins with STP-F and STP-S events discriminated (as in Fig. 2a,b) and plotted with respect to the device’s initial conductance (Figure S9). It is interesting to note that during the experiment the initial conductance range increased to values above 2.95 μS, possibly due to the partial formation of a new stable filament. This effect yielded a new equilibrium conductance at which the device could settle. Nonetheless, for both stable-state conditions, STP-S events are more likely to manifest at higher conductance levels than STP-F events, as shown by the corresponding probabilities of STP-F (Fig. 2d) and STP-S (Fig. 2e) occurrences. This illustrates the strong probabilistic switching nature that under the classical ReRAM context will contribute substantially to the devices unreliability.
In relation to the occurrence of the STP events presented in Fig. 2d,e, we have recorded all events of a single device when repeatedly excited with a stimulating scheme comprising three voltage pulses of 4 V, 10 μs wide, tint = 200 ms and trec = 20 s, as illustrated in Fig. 2f. Facilitating (depressing-like) events have been colored mapped with blue (red), following a simple qualitative rule: if the initial conductance of the device is smaller (larger) than the immediate post-stimuli conductance then this event is considered as STP-F (STP-S). Initially, a steadily increase in the memristor’s conductance is observed. When however a critically high conductance is reached, depressing-like events are activated to restore the low conductance level; this trend occurs consistently when observed over a long period of time.
We further studied the effect of the amplitude and rate of the stimulating scheme in controlling the short-term dynamics of our prototypes. Figure 3a shows the transient conductance change for a two-pulses (spikes) input for inter-spike intervals ranging from 20 ms up to 200 ms. This change is recoverable after a period of time as measurements after 1–120 s show (see also supplementary material Figure S10). A clear correlation is found between the conductance decay and the interpulse timing, with the decay time constant being smaller for lower pulsing rates, as also illustrated in Fig. 3a. This adheres with the notion that when repeated training of an event occurs within a short period of time it becomes more difficult to forget this event. Details about the fitting and parameter extraction methods are found in the supplementary material. Figure 3b depicts the contribution of each pulse stimulus to the device’s conductance as a function of the stimulus amplitude; large amplitudes contribute a higher conductance modulation. The volatile behavior of our prototypes can be reproduced by an equivalent SPICE circuit model that we have presented previously68, as demonstrated for example in Fig. 3c.
Exploiting metastable switching dynamics for processing spatio-temporal spike patterns
Both the short- and long-term plasticity mechanisms that characterize the ReRAM memristors described in this work have temporal properties and dynamics that are well within the range of biological cortical synapses. In addition to being an extremely useful property for directly emulating the properties of real synapses (e.g., for bio-hybrid systems and basic neuroscience research), this feature is very appealing for neuromorphic electronic systems. Neuromorphic systems comprise large arrays of neural processing elements in which memory and computation are co-localized, and in which time represents itself: the synapse and neuron circuits in these architectures process input spikes as they arrive, and produce output responses in real-time. Consequently, in order to interact with the environment and process real-world sensory signals efficiently, these systems must use computing elements that have biologically plausible time constants. By combining the advantages of ReRAM memristors with the properties of subthreshold analog neuromorphic VLSI circuits, it is therefore possible to build extremely compact and low-power neural processing systems that can interact with the environment in real-time69. In69 the authors demonstrated a hybrid CMOS/memristor circuit that can use the conductance changes of ReRAM memristors to produce post-synaptic currents that have dynamics and properties very similar to the ones measured from real synapses (e.g. in terms of current amplitudes, time constants, etc.) When interfaced to neuromorphic silicon neuron circuits70 one can implement efficient neural processing systems.
For example, one of the basic requirements of neural processing systems is their ability to recognize different patterns encoded in the temporal sequences of spikes produced by multiple neurons, e.g. at the sensory periphery. It is essential therefore to be able to distinguish different sets of spatio-temporal spike patterns quickly and efficiently. Here we show, perhaps, the first example of how memristors can be used to achieve spatiotemporal computation by performing an experiment with a static resistor and a memristor connected to a circuit that implements an exponential IF neuron model71 (see also supplementary material, Figures S4–7). In Fig. 4a we show the neural network diagram of the circuit designed to discriminate between two spatio-temporal patterns: the first pattern is represented by the sequence of events AB, when the spike-train labeled A appears before the one labeled B; the second pattern is represented by the sequence BA when the spike trains are sent in reversed order. The spike trains consist of three −4 V, 10 μs wide pulses with inter-spike interval tint = 250 ms. For the pattern AB, the spike train is applied first to the static (RS) and then to the dynamic (M) synapse, as illustrated in Fig. 4b. The same train of pulses, but in reversed order, is applied in the case of the pattern BA as in Fig. 4f. This sequence detector was designed to produce a spike at the membrane of the IF neuron when the pattern BA occurs, based on the short-term facilitating response of the memristor, and no spike when pattern AB occurs. The experiment was conducted by first applying five AB patterns, followed by five BA patterns and then repeating this full sequence eight times.
The success rate of the discrimination task is 67.5% with 15% false positive, as illustrated in Fig. 4h. It was found that the variability in success is a cumulative effect of the response of the memristive synapse and the inherent noise of the system coupled with a low membrane voltage threshold. The performance of this network is compatible with the fact that biological neurons are inherently unreliable. Achieving reliability with unreliable circuits via redundancy strategies is a well-established concept in nature72,73 that can be exploited also in these types of applications. A control experiment was performed in a similar fashion where the memristor was replaced by a static resistor in parallel with a capacitor (Figure S11), showing that in the absence of short-term facilitation (memristive synapse) we are unable to reliably discriminate between the events AB and BA (see Fig. 4h).
The occurrence of false negatives, which in turn render the spiking probability of the neuron to be less than 100% is due to the probabilistic response of our DUT being facilitating (Fig. 2d) or saturating (Fig. 2e), with Fig. 5c illustrating such a case. As the contribution to the membrane potential decreases with each subsequent input pulse, the membrane potential does not reach the pre-set threshold, prohibiting the neuron from spiking, as depicted in Fig. 5b. It is however interesting to see that the opposite also holds; we have measured false positive cases where the neuron erroneously spikes when presented with an AB event. This is represented in Fig. 5d–f and is again due to the synapse exhibiting an STP-S response rather than STP-F. Our proof-of-concept example of memristor-based sequence detectors, as presented in Fig. 4, is only a simple scenario for highlighting the potential of this application. This concept is clearly amenable both to up-scaling, where multiple memristors and/or more complicated spatio-temporal patterns are employed, as well as the use of different neuronal models or even circuit parameters (e.g. membrane potential thresholds); opening remarkable opportunities for advancing the complexity of this system. Figure S12 demonstrates the same concept with aid of two volatile memristive synapses, illustrating a coincidence detector circuit.
In this work we presented detailed and quantitative parallels between memristive devices and biophysically realistic models of synaptic dynamics. In particular we showed how meta-stable memory transitions that are typically seen as non-ideal effects that contribute to the large variability observed in emerging ReRAM, are in fact a salient feature for the establishment of truly biomimetic synapses that can faithfully reproduce short-term synaptic dynamics. In addition to reproducing plasticity mechanisms at a phenomenological level, these devices can be biased to exhibit both stochastic properties and biologically plausible temporal dynamics. These features can be exploited for developing non-von Neuman computing architectures in which memory and computation are co-localized, and where massively parallel circuits can process signals in real-time. Their ability to implement biologically realistic time constants would enable the construction of neural computing systems that can efficiently process real-world biologically relevant sensory signals and interact with the environment. Encompassing all these features in a single compact low-power device offers enormous potential for the development of real-time neuromorphic computing systems. Our approach paves the way towards performing a Turing-test for neuromorphic engineering, where emerging synaptic mimics are tested as in electrophysiology experiments and fitted by biophysically realistic models, while experts are invited to comment on whether the exhibited response stems from a real or an artificial system.
All device prototypes exploited in this work were fabricated by the following process flow. 200 nm of SiO2 was thermally grown on top of 4-inch Si wafer, with 5 nm Ti and 30 nm Pt layers deposited via electron-gun evaporation to serve as the bottom electrodes (Ti is used as an adhesion layer). An RF magnetron sputtering system was used to deposit the active TiO2 core from a stoichiometric target, with 30 sccm Ar flow at a chamber pressure of P = 10−5 mbar. Finally, all top Pt electrodes were deposited by electron-gun evaporation. A lift-off process was employed for patterning purposes prior each metal deposition. Good lift-off was accomplished via using two photoresist layers, LOR10 and AZ 5214E respectively, and conventional contact optical photolithography methods were used to define all layers. All finalized wafers were then diced, to attain 5 × 5 mm2 memristor chips, which were wire-bonded in standard packages for measurements. Preliminary characterization of all samples took place on wafer by employing a Wentworth semi-automatic prober and a Keithley SCS-4200 semiconductor characterization suite.
The cross-section of our memristor prototypes appearing on the inset of Fig. 1a is a 256 × 256 pixel EDX map of a pristine (as-fabricated) device. This map was taken at 50 μs dwell time, 1.2 nA beam current and 8 mins acquisition time on a FEI Titan G2 ChemiSTEM 80–200 microscope.
How to cite this article: Berdan, R. et al. Emulating short-term synaptic dynamics with memristive devices. Sci. Rep. 6, 18639; doi: 10.1038/srep18639 (2016).
We acknowledge the financial support of the eFutures XD EFXD12003–4, the CHIST-ERA ERA-Net and EPSRC EP/J00801X/1, EP/K017829/1 and FP7-RAMP.