Memristive Physically Evolving Networks Enabling the Emulation of Heterosynaptic Plasticity

Yuchao Yang, Bing Chen, Wei Lü

Advanced Materials · 2015 · 157 citations · 30 references

Concepts

Abstract

A nanoscale, solid-state physically evolving network is experimentally demonstrated, based on the self-organization of Ag nanoclusters under an electric field. The adaptive nature of the network is determined by the collective inputs from multiple terminals and allows the emulation of heterosynaptic plasticity, an important learning rule in biological systems. These effects are universally observed in devices based on different switching materials. Physically evolving networks (PENs), a form of “B-type unorganized machine” proposed by Turing in 1948,1 are connectionist networks based on adaptive reconfigurations of the physical network morphology and can serve as fundamental building blocks for parallel and efficient computation.1, 2 So far only preliminary examples of PEN have been demonstrated using millimeter-scale stainless steel balls in castor oil,2, 3 exhibiting spontaneous physical reorganization into ordered dendrite networks (so-called arbortron networks) upon application of an electric field. However, potential application of PENs requires the development of scalable solid-state systems. Here we report for the first time nanoscale, solid-state PENs based on memristive effects that enable the self-organization of Ag nanoclusters. In response to external electrical stimulations, the nanocluster networks adaptively evolve and simultaneously form high conductance regions (filaments) bridging different pairs of terminals depending on the stimulation activity. The physically modified conduction paths can be used to regulate information flow and form the basis for neuromorphic computing systems. As an example, we demonstrate heterosynaptic plasticity4, 5 in a network with multiple terminals, where the activity at the modulatory terminal strongly affects how the connectivity is formed between the pre- and postsynaptic terminals. Tremendous interest has been gained recently in the area of bio-inspired neuromorphic computing using emerging devices, where two-terminal elements such as memristors6-8 can be used to efficiently emulate synaptic functions.9-18 To date, encouraging progress has been made and a number of critical synaptic learning rules have been successfully demonstrated using memristive devices.12-18 Nevertheless, it should be pointed out that previous studies were limited to the emulation of input specific, homosynaptic plasticity effects, where the same pair of terminals are used to supply the enabling activities and sense the subsequent synaptic changes. On the contrary, in a PEN, the network configuration is instead determined by the collective activity from all inputs, allowing the system to exhibit much more complex behaviors. In particular, we show that such a multiterminal system can be used to emulate another fundamental form of synaptic plasticity widely observed in biological systems, that is, heterosynaptic plasticity where the synaptic plasticity between a pair of pre- and postneurons is induced by the activity of another modulatory interneuron.4, 5 Hetero­synaptic plasticity has been widely observed in central as well as peripheral synapses, in both invertebrates and vertebrates, and is believed to be an enabling phenomenon for important biological functions such as sensory perception, associative learning, synaptic runaway dynamics prevention, and long-term memory.4, 5 The ability to implement this important pheno­menon in nanoscale solid state devices will likely significatly advance the development of neuromorphic hardware systems. The solid-state PEN in our study is based on the self-organization of Ag nanoclusters in a dielectric (e.g., SiO2) under an electric field. The clusters were prepared by annealing a 1–2 nm thin Ag film on the dielectric (see the Experimental Section for details), and cross-sectional scanning electron microscopy (SEM) imaging shows that the clusters were partly incorporated into the dielectric layer after annealing (Figure S1, Supporting Information), in agreement with the transmission electron microscopy (TEM) observation of Ag incorporation and migration in dielectric in a similar structure.19, 20 Our previous in situ TEM studies have revealed that individual metal nanoclusters in dielectrics behave as bipolar electrodes upon application of an electric field, and effective cluster displacements along the field direction can be achieved by the movement of individual atoms through a sequence of ionization, ion migration, and reduction processes.21 These processes fundamentally underpin the various filament growth dynamics in cation-based memristive devices.20, 21 In particular, the evolution of the nano­clusters is determined by the local ion supply and the local electric field. As a result, complex and adaptive evolution of the physical nanocluster configurations is expected if the supply is limited and the field can be applied from multiple terminals, as shown in Figure 1a. This configuration allows us to implement the concept of PEN in nanoscale solid-state devices based on such memristive effects. Figure 1b–c shows a couple of examples highlighting the basic operations of the multiterminal PEN system. To control the evolution of the nanoclusters, different voltage bias schemes are applied to the terminals. Over 60 devices have been tested and consistent results have been reliably obtained. In the case of Figure 1b, “V” and “0” were applied on the electrodes at diagonal positions, while the other two electrodes were biased at potential “1/2V.” Time-dependent current measurements (I–t) were performed for both the “V” terminal (noted as IV) and the “1/2V” terminals (noted as I1/2V and includes the total current from the two “1/2V” terminals) to monitor the evolution of the PEN. In this way, the configuration change in the complex network is reflected in the electric outputs. Figure 1i shows a significant increase of IV (from ≈1 pA to ≈10 nA) at ≈120 s, suggesting an electrical connection was formed between the two diagonal electrodes, as schematically illustrated in the inset of Figure 1i, while no changes were observed in I1/2V. Indeed, from the postswitching SEM observation (Figure 1b) one can see that the original randomly distributed Ag clusters (Figure 1a) have self-organized into a curved filament connecting the “V” and “0” terminals. The self-organization of the Ag clusters was further verified by atomic force microscopy (AFM) analysis (Figure 1d). During the cluster evolutions the Ag clusters do not move as a whole but rather act as bipolar electrodes and successively go through oxidation → migration → reduction processes,21 resulting in the growth of some clusters and the loss of others, thus driving the self-assembly process (Figure 1b,c). This process was further verified through 3D AFM imaging (Figure 1f) and Kelvin probe force microscopy (KPFM) analysis (Figure 1g), which confirmed that electrical connection was indeed established between the electrodes bridged by the filament. As has been observed in conventional cation-based memristive devices, the filament may be composed of discrete Ag clusters (Figure 1b,d) at weak programming conditions,20, 21 leading to a resistance higher than that of a single-point contact (12.9 kΩ). The memristive effect of the devices was further verified by the pinched hysteresis loops22 from two-terminal current–voltage (I–V) measurements in these multiterminal devices (Figure S2, Supporting Information). Notably, different from previously studied memristive devices, an infinite ion source, such as the active electrode used in electrochemical metallization cells,23 is not present in the PENs studied here. Instead, the filament is formed by the redistribution and self-organization of existing Ag ions/atoms with a finite supply. The spontaneous organization of Ag clusters is determined by the collective inputs from the terminals, as well as device geometric parameters such as the electrode spacing (L), and the number (n) and area (A) of the clusters. Figure 1h plots the dependence of the switching time (the time it takes to form a conduction path) on the ratio of V/L and the area covered by the clusters, from analysis of 43 devices. Overall the switching time decreases as V/L or A increases, as can be explained below. As V increases or L decreases, the effective electric field sensed by the clusters in the network will be increased; while as n or A increases, the available ion source is increased while the average gap size that needs to be bridged is decreased, thus affecting how the connections are made. In addition, it should be noted that the cluster dynamics and electrochemical processes can be influenced by ambient conditions especially moisture levels,24 which was verified by the significantly shorter time or less pulse numbers to achieve resistive switching in air (with moisture) compared with that in vacuum (with no or limited moisture), as shown in Figures S3 and S4 (Supporting Information). The nanocluster-based PEN system can thus produce diverse behaviors during network evolution based on the parameters discussed above. Primarily, one expects different input activity patterns fed into the multiterminal network should lead to different connectivity patterns. Indeed, Figure 1c shows results from another input configuration where “V” and “0” were applied on the top two neighboring terminals, while the lower two terminals were assigned to “1/2V.” Indeed, a different conduction path connecting the top two terminals (corresponding to the “V”–“0” terminal pair) was formed, accompanied by an abrupt increase in IV (Figure 1c,e,j). Compared to the previous case, the conduction path direction is changed from nearly diagonal to nearly horizontal, in response to different inputs. Beyond this first-order control from input patterns, the network evolution is also affected by local geometric and materials parameters, as the physiochemical processes strongly depend on the “local” field and physiochemical properties. For example, Figure 1k–m shows examples of different PEN connectivity patterns observed from ≈40 devices under the same input scheme. Despite the nominally smaller voltage drop between “1/2V”–“0” terminals compared with “V”–“0,” filaments are sometimes observed to form between one of the “1/2V”–“0” terminal pairs instead (Figure 1k; also see SEM image in Figure S5, Supporting Information). This can be explained from the stochastic nature of the physiochemical processes involved25 and the inhomogeneity of the field and cluster distribution. Similarly, filaments could also be formed between the “V”–“1/2V” terminal pairs (inset of Figure 1l). In this case, simultaneous increases in IV and I1/2V were observed with opposite signs (Figure 1l), as the current across the filament would flow out of the “V” terminal and into the “1/2V” terminal. A more rare case is the simultaneous formation of multiple filaments in the PEN, e.g., filaments formation across both “V”–“0” and “1/2V”–“0” terminal pairs, leading to increases in both IV and I1/2V with the same sign of current change, as shown in Figure 1m. Finally, Figure 1n shows results obtained from another input scheme where one of the electrodes was assigned to “V” and all three others were assigned to “0.” As expected, a filament was formed between one of the neighboring “V”–“0” terminal pairs (see SEM image in Figure S6, Supporting Information) instead of between the diagonal pair due to the lower electric field in the diagonal direction. The dynamic, adaptive evolution of the PENs based on well-understood memristive effects allows them to implement more complex functions compared to simple two-terminal devices. To date, previous studies on memristive synapses have been focused on mimicking homosynaptic plasticity Figure 2a), while heterosynaptic plasticity requires another modulatory terminal to regulate synaptic facilitation/depression between the pre- and postsynaptic neurons (Figure 2b). In particular, when the homosynaptic stimulus is paired with the modulatory stimulus, a significantly enhanced, nonadditive synaptic facilitation will be obtained, which is believed to be the underlying mechanism for classical conditioning in biological systems.4, 26, 27 The hetero­synaptic cellular structure can be readily emulated in a PEN system with three terminals, as shown in Figure 2c. Here the Ag nanoclusters are dispersed randomly between the presy­naptic and postsynaptic terminals (Pt electrodes), while the modulatory terminal is separated from the nanoclusters by a gap of >100 nm to prevent direct filament connection with that terminal. To more realistically mimic the synaptic behavior, voltage pulses representing spikes, instead of constant bias, were used in subsequent studies. Detailed input and output signals are schematically sketched in Figure 2e. Here presynaptic (Vpre) and modulatory (Vmod) voltage pulses were applied as input stimuli to the device, while the change in the synaptic weight was monitored by the postsynaptic current output (Ipost). Due to the extremely low operation current of the device (≈nA) and the large parasitic capacitance in the off-chip measurement setup, relatively wide voltage pulses (≈1 s) were used to ensure accurate reading of the output currents. Much faster operations can be achieved in the future through integrated current sensing circuitry that can significantly reduce the parasitic capacitance. Figure 2f shows Ipost as a function of applied pulse number for the cases of homosynaptic (where only the presynaptic pulses were applied with Vpre = 10 V, Vmod = 0) and heterosynaptic (where the presynaptic pulses were paired with modulatory pulses with Vpre = 10 V, Vmod = 2, 4, 6, 8 V) inputs. One can see that even though synaptic facilitation can be achieved in all cases, corresponding to the self-organized filament formation between the presynaptic and postsynaptic terminals (Figure 2d, more examples of filament/dendrite formations can be found in Figure S7, Supporting Information), the number of pulses it took to reach the same facilitation level (e.g., weight change by a ratio of 102) was significantly reduced when the presy­naptic input was paired with the modulatory input. Furthermore, this effect becomes more pronounced as the paired modulatory input becomes stronger (e.g., increased Vmod amplitude; Figure 2f). This behavior can be seen more clearly from the monotonously decreasing trend in the number of presynaptic stimulations needed for facilitation as a function of Vmod, as shown in Figure 2g, demonstrating strong modulatory effects on the synaptic plasticity. The purely modulatory role of the modulatory input was verified by the persistently low current level through the modulatory terminal (<pA) throughout the measurements (Figure 2h). These behaviors resemble heterosynaptic plasticity observed in biological systems, whose key attribute is the significantly enhanced, nonadditive synaptic facilitation in the case of paired stimuli.4, 26, 27 Indeed, in the case when only homo­synaptic stimuli were applied (Vpre = 10 V, Vmod = 0) it required 70 repeated stimulations to achieve the desired facilitation, a number that is significantly larger than the cases of paired stimulations (43, 20, 18, and 12 stimulations for Vmod = 2, 4, 6, and 8 V, respectively; Figure 2f,g). Additionally, synaptic facilitation was not observed after >300 repeated stimulations in the case where only modulatory stimulations were applied, even with a large Vmod value of 10 V (Vpre = Vmod = 10 Figure Supporting Information). The nonadditive of facilitation by the of homosynaptic and modulatory stimuli shown in Figure is consistent with the basic of hetero­synaptic plasticity in biological systems.4, 26, 27 In to the role of modulatory inputs, when modulatory stimuli were paired with presynaptic stimuli (Vpre = 10 V, Vmod = synaptic facilitation was even with the application of >300 pulses (Figure the role of modulatory inputs in synaptic In addition, we the synaptic weight changes are as from the measurements after the stimuli have been (Figure the nature of both the and The physical processes in PEN evolutions were further studied using The device is shown in Figure the can be found in the Experimental Figure and S2, Supporting Information). Figure shows the when and modulatory stimuli are paired with presynaptic stimulation experimentally observed the of facilitation when modulatory stimuli are paired with presynaptic can be The evolution of the Ag during the corresponding to by the in Figure is shown in Figure along with the corresponding electric field we that consistent with the nature of the Ag cluster evolution the cluster evolution at with the electric field (Figure In the homosynaptic case = the of the filament strong electric field (Figure As a result, Ag from clusters persistently move to and the leading to the growth of the filament and connection between the presynaptic and postsynaptic terminals. The of the modulatory terminal in this case only a weak electric field the terminal (Figure Notably, the application of modulatory stimuli 0) the electric field at the of the from to during the filament thus for the filament growth Additionally, the of the filament path is also the modulatory terminal due to the changed electric field (Figure the electric field direction is more clearly shown by the field in Figure Supporting Information). On the contrary, the application of modulatory stimuli 0) an electric field that is between the presynaptic and modulatory terminals with a weak electric field the postsynaptic terminal (Figure also see the electric field in Figure Supporting Information). As a result, the ion and filament growth between the presynaptic and modulatory terminals instead of the postsynaptic synaptic facilitation between the pre- and postsynaptic terminals (Figure Similarly, the of modulatory stimuli on synaptic can be were performed on PEN that have been with an existing connection between the pre- and postsynaptic terminals. the signals on the pre- and postsynaptic terminals were with the presynaptic terminal and a input applied to the postsynaptic and the sign of Vmod was also to achieve as schematically illustrated in Figure Figure shows synaptic current as a function of the number of applied stimulations when different modulatory stimuli were applied with inputs = 10 V, Vmod = Compared with the in the homosynaptic case = 10 V, Vmod = the application of modulatory stimuli paired with pulses clearly the process (Figure similar modulatory role of Vmod in the behavior was by as shown by the in Figure The filament process that to synaptic is shown in Figure The filament was found to the of the filament at the due to the electric field in that (Figure Notably, compared with the homosynaptic case (Figure the application of modulatory stimuli an electric field the modulatory terminal (Figure see electric field in Figure Supporting Information), which to Ag out of the filament and to the filament observed in Figure In if the sign of the modulatory stimuli synaptic to be device current in increased after an drop (Figure instead of Ag from the filament the modulatory stimuli Ag into the similar to the facilitation case, synaptic can be or depending on the activity of the modulatory stimuli (Figure allowing control of the synaptic it was found that can also be achieved by modulatory stimuli only in cases where a stimulus is applied to the pre- and postsynaptic pair with a input applied to the presynaptic terminal and the postsynaptic terminal Figure Due to the facilitation effect by the presynaptic the of Vmod has to be significantly increased in to the filament (Figure in with the scheme shown in Figure This behavior the critical role by the modulatory stimuli and the effects of presynaptic of the of heterosynaptic observed in Finally, we that PEN and the heterosynaptic effects can be obtained using different and (Figure and Figures Supporting Information). these two switching materials have from (the used in Figures 2, with a while a all devices similar hetero­synaptic that is, the synaptic facilitation was significantly by the application of paired presynaptic and modulatory The modulatory effect was also universally found to be more pronounced when the of the modulatory stimuli was The from the three can all be well with this as shown in Figure parameters can be found in S1, Supporting Information). The observation of hetero­synaptic in memristive PEN is of the same fundamental electrochemical processes at the atomic in such devices, and it for future in neuromorphic hardware by an to achieve the desired learning and In we experimentally demonstrated PENs in nanoscale, multiterminal memristive devices based on the self-organization of Ag nanoclusters under electric field. The physically evolving nature of the devices allows electrical connections to be adaptively made in such multiterminal devices and can enable more efficient neuromorphic computing As an example, heterosynaptic plasticity, an important learning rule found in biological systems, was demonstrated in a system where activity of the modulatory terminal strongly affects the synaptic facilitation/depression between the pre- and postsynaptic terminals. These effects were universally observed in devices based on different systems. Additionally, the memristive PENs exhibit electrical large ratio low current and them for hardware would to that the PEN devices studied are fundamentally different from other device such as the and or where the are from the terminal while the source and terminals are used for reading out the In our PEN devices the filaments are formed between the source and and postsynaptic terminals) instead and key to the device operation is the of the presynaptic inputs with the modulatory inputs. The devices in this were to demonstrate the of nanoscale, solid-state PENs and the effects of heterosynaptic plasticity based on the self-organization of metal clusters. studies will be needed for the development of more devices. For example, more (Figure S2, Supporting Information) and switching can be achieved through measurement that the parasitic and device by the device and the cluster to Figure and by using dielectric materials with higher ion The can be further by the device configuration into In of the two of the switching materials in this and are with while the electrodes and Ag clusters may to be by more for and into the we that and of the memristive PENs will lead to more complex behaviors and functions and can an efficient parallel computing from with nm or were by while were by electron the switching materials were prepared at with a of the of the switching electrodes were by electron using as the by electron and Ag with a of 1–2 nm in the device were and with the existing electrodes by a of electron electron and The was to annealing at in for 5 s, which the Ag thin into Ag nanoclusters. The electrical behaviors of the devices were tested using a system in a probe The SEM imaging of the devices was performed and after network evolutions by at while the of the filaments was by AFM The potential of the postswitching device was by parameters of the PEN devices the numbers and of the Ag nanoclusters were by after the original SEM to The was performed in using a In three physical processes were Ag from the clusters, in the and the reduction of Ag The electric field and current were based on a resistance network and the The evolution of metal atoms electric field, and current with time were and and parameters the can be found in Figures and and (Supporting Information). The heterosynaptic plasticity was with and the parameters are shown in (Supporting Information). The would to and for This was in by the through by the through and by the through The in this are of the and do not the or of the of or the for As a to our and this information by the materials are and may be for but are not or from information than should be to the The is not for the or of information by the than should be to the corresponding for the

References

30