Study of Synaptic Behavior in Doped Transition Metal ... - IEEE Xplore

0 downloads 0 Views 2MB Size Report
Nov 20, 2013 - Saptarshi Mandal, Student Member, IEEE, Branden Long, and Rashmi Jha, Member, IEEE. Abstract—We report the synaptic characteristics of ...
IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 60, NO. 12, DECEMBER 2013

4219

Study of Synaptic Behavior in Doped Transition Metal Oxide-Based Reconfigurable Devices Saptarshi Mandal, Student Member, IEEE, Branden Long, and Rashmi Jha, Member, IEEE

Abstract— We report the synaptic characteristics of novel two-terminal reconfigurable devices fabricated using a doped transition metal oxide. These devices demonstrate short-term plasticity, frequency-dependent synaptic augmentation, long-term potentiation, and long-term depression, and have a potential to show spike timing-dependent plasticity that are macroscopically similar to a biological synapse. The underlying mechanism behind the observed synaptic characteristics was studied using charge transport characterization. Based on this study, a fundamental correlation between the governing device physics and the synaptic characteristic has been established. We believe that by carefully engineering the dopants, the synaptic transmission of these devices can be modulated, which will provide a viable route to replicate the functional diversity of a biological neural system on chip. Index Terms— Memristive device, plasticity, synapse, transition metal oxide.

I. I NTRODUCTION

T

HE ability to perform low-power extreme-scale robust neuromorphic computing, inspired by a biological brain, has eluded the scientific community for a long time. One of the challenges lies in achieving the stringent power versus area requirement for nanoelectronic devices and circuits to implement the biologically inspired computational paradigms on a silicon chip [1]–[3]. To address this need, recent research activities have reported synaptic characteristics in memristive devices based on phase change memory, resistive random access memory, and conductive bridge memory devices [4]–[6]. Theoretical and experimental studies have demonstrated potential for implementing spike-timing dependent plasticity (STDP) using memristive devices using their ability to achieve multiple nonvolatile resistive states, which makes these devices interesting for further studies [7], [8]. In this paper, we report how trap dynamics in Mn-doped HfO2 based memristive devices can be used to achieve synaptic characteristics. Unlike previous work on transition metal oxide

Manuscript received August 28, 2012; revised May 19, 2013, August 21, 2013, and September 25, 2013; accepted October 16, 2013. Date of publication November 8, 2013; date of current version November 20, 2013. This work was supported by the National Science Foundation under Grant 1125743. The review of this paper was arranged by Editor F. Ayazi. S. Mandal and B. Long are with the Department of Electrical Engineering and Computer Science, University of Toledo, Toledo, OH 43606-3390 USA (e-mail: [email protected]; [email protected]). R. Jha was with IBM Semiconductor Research and Development Center, East Fishkill, NY 12533 USA. She is now with the Department of Electrical Engineering and Computer Science, University of Toledo, Toledo, OH 436063390 USA (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TED.2013.2288327

(TMO)-based memristive devices that reported diffusion of oxygen vacancies to be responsible for the synaptic behavior [9], our studies indicate that short-term plasticity (STP) can be achieved by trapping/detrapping dynamics of traps while longterm potentiation (LTP) can be achieved by the creation of new traps under positive bias and long-term depression (LTD) can be achieved by the annihilation of traps under negative bias. We believe that these fundamental studies will advance the existing knowledge in this area, which will be important for engineering the future synaptic devices. II. E XPERIMENTAL M ETHOD A. Fabrication The deposition of the stack was done in an RF magnetron sputtering system on a 2-in p-type Si wafer. After standard cleaning of Si wafer in 1% HF dip, 3-nm layer of Ti was deposited initially as an adhesion promoter layer. A blanket layer of Ru (100 nm) was then deposited as the bottom electrode (BE). The BE was protected by a shadow mask during the deposition of the doped-TMO, which consisted of Mn-doped HfO2 in this experiment. The doped-TMO was deposited by reactive co-sputtering of Mn and Hf in O2 :Ar environment. The sputtering power of Hf and Mn was adjusted to incorporate 9% of Mn in Mn:HfO2 . The substrate temperature was maintained at 300 °C during the depositions. The Mn:HfO2 thickness was targeted to be 6 nm. Mn, being a low valence cation, was expected to introduce oxygen vacancy (Vo ) related trap states in HfO2 [10]–[12]. Thereafter, 120-nm TiO2-x was deposited in situ by reactive sputtering of Ti in O2 :Ar mixture at 300 °C temperature. TiO2-x is relatively thick in this paper; however, the thickness can be optimized based on the process conditions. The top electrodes (TEs) consisting of Ru was deposited in situ without breaking the vacuum to avoid any unintentional interfacial layer. The Ru TE was patterned using standard photolithography and etched using Ru wet-etchant to define the TE. B. Testing The electrical characterization was performed using a Keithley 4200 semiconductor characterization system and 4225 pulse modulation unit. The sample was loaded in a Lakeshore cryogenic probe-station, and measurement was done under vacuum pressure of 7.5e–05 torr. The temperature for cryogenic testing ranged from 250 to 400 K. The TE was biased while the BE was either grounded or connected to the drain of an nMOS in a one-transistor-one-resistor

0018-9383 © 2013 IEEE

4220

IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 60, NO. 12, DECEMBER 2013

Fig. 1. Two-terminal Ru(BE)/Mn:HfO2 /TiO2-x /Ru(TE) reconfigurable device showing (a) device schematic with 6-nm Mn:HfO2 . 120-nm TiO2-x is shown in inset. The TE was biased while the BE was grounded in the measurements. Initial I –V characteristics indicate asymmetric behavior. (b) CV profile at 100 KHz. The BE was biased since it behaves as gate for the MOS structure. TiO2-x being n-type semiconductor, shows accumulation-type behavior under positive bias, while negative bias results in deep depletion.

configuration. 100 μm × 100 μm sized devices were used for cryogenic testing while 30 μm × 30 μm sized devices were used for all other characterization. III. R ESULTS AND D ISCUSSION A. Initial Characteristics The device configuration is shown in the inset of Fig. 1(a). The initial current–voltage (I –V ) characteristics of the device are shown in Fig. 1(a). The dc bias was applied on TE while BE was grounded in this measurement. Multiple devices were measured across the 2-in wafer to ensure device repeatability. The asymmetric nature of the I –V can be observed. The capacitance–voltage (CV) measurements were performed by applying bias on the BE. The CV data at 100 KHz is shown in Fig. 1(b). The CV analysis indicated MOS diode type behavior with TiO2-x as an n-type semiconductor, the Mn:HfO2 as a gate dielectric, and Ru BE as the gate electrode. A positive bias on the BE causes an accumulation of electrons in TiO2-x at the Mn:HfO2 /TiO2-x interface as evident from the saturating trend of CV in positive bias. The dissipation factor was well below unity suggesting low leakage in this voltage range. Voltage sweep toward negative bias leads to depletion and further negative sweep leads to the deep depletion in TiO2-x as evident from the nonsaturating CV in negative biases. Using the accumulation portion of the CV, Mn:HfO2 thickness was extracted as 6.24 nm using a dielectric constant of 20 as reported for HfO2 [14]. The asymmetric nature of initial I –V in Fig. 1(a) can be explained based on this analysis where current due to negative bias on TE is due to injection of accumulation electrons from TiO2-x toward BE through Mn:HfO2 . The current with positive bias on TE is due to the injection of electrons from BE toward TiO2-x . However, since TiO2-x is depleted in this case and barrier height (b ) of Ru on Mn:HfO2 is larger than b of TiO2-x on Mn:HfO2 , current in the positive bias is smaller than current in the negative bias. To determine the mechanism of conduction in these devices, I –V was measured at several temperatures ranging from 250 to 400 K, as shown in Fig. 2(a) A strong dependency of current on temperatures can be observed. Fig. 2(b) and (c) shows ln J versus 1/kT fittings indicating transport to be

governed by trap-assisted mechanisms such as trap-assisted tunneling (TAT) in both the positive and negative biases [13]. The intercept and slope of these fittings are provided in insets. Fig. 2(d) shows the band-diagram with negative bias. To extract the concentration of traps and trap energy state, the experimental I –V curves in negative bias at different temperatures were fitted to a generalized TAT model described in [15] using numerical simulations. The fitting parameters resulting in the best fitting [Fig. 2(e)] are shown in Table. I. This model indicated a trap density of 5 × 1021 cm−3 at an energy level of 1.3 eV from the conduction band of HfO2 . When the device was swept to a positive bias and back, an I –V hysteresis was observed, as shown in Fig. 3(a). The device conductance increased with each sweep and finally saturated. Interestingly, a region of overlap was always observed between any two consecutive I –V hysteresis loops. This indicates a transient behavior of the devices wherein decay of conductance occurs when the bias is removed. Hysteresis in I –V during negative sweeps was not observed. However, application of negative pulse decreased the conductance of the device, as shown in Fig. 3(b). The inset shows hysteresis in positive direction when the device was driven to saturation by sweeps 1–6. Thereafter, the conductance of the device in saturation was reduced using a −3 V pulse of given pulsewidth. After application of negative pulse, the device was swept again to 4 V and as indicated by sweep 7, the device conductance is reduced. Each time, the device was driven to saturation and a different pulsewidth was used to reduce the conductance. The decrease in conductance as a function of pulsewidth is shown in Fig. 3(b). After observing these initial characteristics where conductance of device could be reconfigured under positive and negative bias, the remainder of this paper focuses on understanding the origin of this behavior and if these devices can be used as biologically inspired synaptic elements in neural circuits. B. Model of Synaptic Behavior The conduction through a biological synapse involves a remarkably complex process. A synaptic transmission begins when an action potential arrives at the presynaptic terminal. This causes the transmitter molecule to enter the cleft and bind to the receptor on the post synaptic neuron. Thus, ion channels open, which modifies the postsynaptic conductance [16]–[18]. The synaptic conductance gs is given by gs = g0 P

(1)

where g0 is the open channel conductance, and P, the open channel probability, can be given by (2) P = Prel Ps

(2)

where Prel is the probability of transmitter release and Ps is the probability that the postsynaptic channel opens. The conductance of the synapse rises during the time when channels open and decays when the process of transmission is complete and channels close. Several mathematical models have been proposed to model the conductance of a biological synapse [19], [20]. A generic model describing the synapse

MANDAL et al.: STUDY OF SYNAPTIC BEHAVIOR

4221

Fig. 2. Conduction mechanism. (a) I –V characteristics at different temperatures. 100 × 100 μm2 devices were used for this testing. Current is dependent on temperature. (b) and (c) ln (J) versus 1/kT fitting for different positive and negative bias points, respectively. The trend is similar to a TAT mechanism at low electric fields. (d) Band diagram showing TAT under negative bias to TE. It is assumed that electrons travel along the TiO2-x conduction band till the HfO2 /TiO2-x interface. Here, electrons tunnel into one of the Mn-induced trap states in the HfO2 with probability P1. The electron is emitted from this trap to the BE with probability Ptrapezoid as given for a trapezoidal barrier. (e) J versus E-field for negative bias. The generalized TAT model was fitted to experimental data using numerical simulations. The best fits were obtained for the parameters shown in the inset and Table I. From here, the trap level and trap density in Mn:HfO2 were extracted.

TABLE I F ITTING PARAMETERS FOR TAT IN N EGATIVE B IAS

conductance consists of a sum of two exponentials—one for the rising phase with time constant τrise and one for the decaying phase with time constant τdecay , given by  −(t−t0 )  −(t−t0 ) τdecay τrise gs (t) = g0 f e −e (3) where f is the normalization factor to ensure that the amplitude equals g0 and t0 is the time when a spike arrives at the presynaptic terminal. The probability of release of the neurotransmitter molecule and the magnitude of the resulting conductance at the postsynaptic terminal depends on the history of synaptic activity. STP consists of a number of phenomena that affect the probability that a presynaptic action potential opens the postsynaptic channels. In facilitating synapses, it is well known that the low-frequency spikes are transmitted at a lower probability than high frequency bursts [21]–[23]. On the other hand, LTP involves structural changes in synapses, which are extremely persistent and usually contribute to the learning process. One such learning algorithm is based on STDP, where

Fig. 3. I –V characteristic of the sample indicating (a) hysteresis and increase in conductance with subsequent positive sweeps that finally saturates. Sweep rate of 1.6 V/s was used. (b) Percentage decrease in the conductance with negative pulse after the increase in the conductance of the devices saturated with positive sweep voltage. Inset: observed I –V characteristic. Sweeps 1–6 drive the device to saturation. Sweep 7 is the sweep after a 600 ms −3 V pulse was applied.

the relative timing between pre and postsynaptic spike alters the conductance of the synapse for long periods [24]–[26]. To observe these characteristics, the devices were excited with constant voltage pulses of 4.5 V for 400, 500, and 700 ms. Fig. 4(a) shows a transient increase in the current to the steady state value during application of these pulses. This increase in current can be attributed to the detrapping of electrons from existing traps and generation of new donor traps (V2+ o ), possibly at TiO2-x /Mn:HfO2 interface and bulk Mn:HfO2 , under positive voltage stress. The decay in current (I ), with time (t), after the removal of the voltage pulse, is shown in Fig. 4(b). The current was measured at 0.5 V read voltage. Clearly, log(I ) shows a linear dependency on log(t) for all three conditions. In accordance with the TAT mechanism,

4222

IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 60, NO. 12, DECEMBER 2013

C. STP and LTP

Fig. 4. Transient behavior of device. (a) Increase in the conductance of the device in response to pulses of amplitude 4.5 V and 400-, 500-, and 700-ms pulsewidths, and (b) decay in conductance after the device was excited with the pulses in (a). The decay in conductance was much slower when devices were excited with longer pulses.

this observation can be explained using the tunneling front model [27]–[29]. After the removal of dc voltage stress, the electrons from TiO2-x can tunnel and get trapped by V2+ o near the TiO2-x /Mn:HfO2 interface. The trapping of electrons in V2+ o leads to the discharging current, which can be analyzed using tunneling front model. In the tunneling front model, the current, due to the motion of the tunneling front, is given by (4) as q N (x (t)) A −1 t (4) I (t) = 2β where N(x(t)) is the concentration of traps distributed inside the TMO, x(t) is the distance of the tunneling front from an electrode, q is the electronic charge, β is the tunneling constant, and A is the area of the device. Using this equation, it is evident that log(I ) versus log(t) will follow a linear relation, and that there will be a slope of −1 if the traps (N(x(t))) are uniformly distributed in the TMO. Though our data shows a linear relation between log(I ) versus log(t) with a negative slope, the slope is not −1, which indicates a nonuniform distribution of traps generated during the voltagepulse stressing. Therefore, the equation was modified to reflect this change q N (x (t)) A −1−γ t (5) I (t) = 2β where γ is the correction factor. From the curve-fitting in Fig. 4(b), γ was extracted to be, −0.48 for 400-ms pulse, −0.85 for 500-ms pulse, and −0.97 for 700-ms pulse. Different values of γ indicate that voltage pulses with longer pulsewidths generate higher trap densities in the dielectric, which can lead to a nonvolatile change in the conductance of the device. It is interesting to note that, similar power-law decay has been reported for short-term memory in biological systems [30], [31]. According to these studies, the strength of the memory (m), over time (t), is given by m = at −b

(6)

where a and b are constants. Synaptic strength has a direct correlation with the memory of the organism, and changes in synaptic conductance reflect macroscopically on its memory [32].

STP of a synapse depends on several factors—the most widely observed is the augmentation of synaptic strength, which is dependent on the frequency of presynaptic excitations [33]–[35]. The amount of transmitter released by action potentials increases with the pulse stimulations at higher frequency than at low frequency. This is usually attributed to the influx of presynaptic Ca2+ ions during the tetanus conditioning. The tetanic stimulation consists of facilitation, augmentation, and potentiation. A higher frequency of stimuli results in an increased concentration of residual Ca2+ ions, which ultimately leads to potentiation [36]. To observe similar characteristic in the proposed devices, voltage pulses with fixed amplitude of 4.5 V and frequencies of 5, 7, 10, and 15 Hz were applied. The frequency was increased while keeping the pulsewidth constant. The time–period was decreased to replicate the high-frequency (HF) bursts of action potential at the presynaptic terminal. Fig. 5(a)–(d) shows the measured current (I ) versus time (t) for ten pulses of frequency 5, 7, 10, and 15 Hz, respectively. At low frequencies of excitation, the conductance change is not appreciable; however, HF excitations show a significant change in conductance. It is interesting to note that HF excitations lead to a higher change in conductance than a constant voltage stress for the same effective time. This indicates that more traps are generated with HF excitation, possibly due to frequent charging and discharging of traps, compared with a constant voltage stress. Fig. 5(e) shows the decay in the current with time after ten cycles of pulsing with the given frequency. This data was collected after exciting the virgin devices with voltage pulses of 4.5 V amplitude and 30-ms pulsewidth at a given frequency for ten cycles and then measuring the current for 500-s postexcitation. A new device was used for each condition to avoid any residual effect. An initial decay of the conductance after excitation can be observed due to discharging current (explained earlier); however, an overall increase in the conductance as a function of frequency can be observed, which does not decay completely to the initial state over a time-frame of at least 500 s studied in this paper. Therefore, unlike a biological synapse where STP lasts for a few milliseconds, the transient device state decays more slowly and does not return to the initial conductance state completely, as in Fig. 5(e). However, the similarity of behavior in terms of frequency-dependent augmentation is evident. Fig. 5(f) shows the current (measured at 0.5 V read voltage after 500 s) plotted against the pulse frequency. D. Implementation of STDP The learning ability of a biological brain is related to the change in the plasticity of synapses. The Hebbian synaptic learning rule has been widely demonstrated by STDP in a broad spectrum of ganglia [37]–[40]. In STDP, the plasticity of a synapse is modulated based in the order of pre- and postsynaptic firing within a critical window of tens of milliseconds. When the presynaptic neuron fires the signal before the postsynaptic neuron, the amplitude of the neural signal is additive and the synaptic conductance is increased, leading to LTP. Similarly, when postsynaptic neuron fires the signal before the

MANDAL et al.: STUDY OF SYNAPTIC BEHAVIOR

4223

Fig. 5. Frequency-dependent augmentation of synaptic plasticity. (a) I versus t plot with frequency 5 Hz for ten cycles. (b) I versus t plot with frequency 7 Hz for ten cycles. (c) I versus t plot with frequency 10 Hz for ten cycles. (d) I versus t plot with frequency 15 Hz for ten cycles. (e) I versus t plot to demonstrate the decay in conductance after the excitations. This plot was generated after exciting the devices with the specified frequency pulses for ten cycles and then measuring the current at 0.5 V read voltage for 500 s. (f) I versus frequency of excitation plot. This plot indicates current at the end of 500 s of part (e) plotted against the frequency of excitation. A sharp transition in conductance can be observed at 15 Hz and higher frequencies.

Fig. 6. STDP like behavior. (a) To simulate presynaptic action potential before postsynaptic depolarization (positive t), a train of 60 positive pulses is applied leading to LTP. Sixty negative pulses are applied to show LTD. The width of the pulse is given as a function of t in (7). To simulate presynaptic action potential leading to LTP (LTD), positive (negative) pulse can be applied to TE, while postsynaptic depolarization can be simulated by negative (positive) pulse on BE. An equivalent pulse with a width defined by τ was applied in the test. (b) Conductance increase or decrease is larger with increasing pulsewidth. From (7), it is evident that as pulsewidth is increased, the spike timing decreases. To check the nonvolatile nature of LTP, the same devices were measured a year later using a read bias of 1 V and it is evident that the device tends to retain some conductance, despite the transient decay. Hence, the devices have the potential to show STDP type learning when connected to a circuit that can generate pulses defined by (7).

activation of excitatory postsynaptic pulse, the overall synaptic strength is reduced leading to LTD. Since the plasticity depends on the state and time of both the pre- and post-synapse, it is almost impossible to synthesize an electrical device that is able to independently emulate this synaptic characteristic. Hence, several testing procedures have been developed and modified to emulate such behavior [4], [41], [42]. Spike-timing-based learning rules have been shown in vitro by stimulating a synapse with pairs of presynaptic action potential and postsynaptic depolarization current pulses. For LTP, the depolarizing current is applied at the postsynaptic membrane after the presynaptic stimuli, while for LTD the depolarizing pulse precedes the presynaptic action

potential. To demonstrate the potential for achieving STDP using the proposed devices, we have assumed that, for LTP, the presynaptic potential corresponds to a positive voltage pulse at the TE and the postsynaptic depolarization is represented by a negative pulse at the BE. For LTD, the polarities of the voltages are reversed. The overlap region between the pre- and post-spiking is responsible for the greatest change in conductance of the devices. The input to the device is defined as a pulse that has a width (τ ) equal to the overlap region, shown in Fig. 6(a). The τ is then expressed as a function of the relative spike timing (t) as τ = ω − |t|

(7)

4224

IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 60, NO. 12, DECEMBER 2013

where ω is a specific timing parameter defined to ensure that LTP and LTD transitions occur in the window of tens of milliseconds range. In our case, ω was chosen as 50 ms. A train of 60 pulses at 10 Hz were used for both potentiation and depression. τ was calculated by (7) for different t. The pulse amplitude was chosen positive for potentiation and negative for depression. Fig. 6(b) shows the relative change in conductance levels as a function of t. To obtain this response, devices in the initial conductance level were selected and pulses of amplitude 4.5 V were applied for potentiation. Pulse train of −2 V was applied in the case of depression to the devices that were already at the highest conductance states. As explained in the earlier sections, the application of positive voltage stress on TE creates more Vo leading to LTP while application of negative voltage stress leads to the annihilation of Vo leading to LTD. The current was measured using 0.5 V dc read voltage after 20 min of the excitation to allow for the device to stabilize after any transient decay. Since the change in device conductance increases with τ , the maximum increase or decrease is observed for larger pulsewidths and hence smaller t, as defined by (7). To verify that the devices were indeed in LTP, the conductance of the same devices were measured a year later. For LTP, a read voltage of 1 V was used for the year later measurement since 0.5 V read did not show a significant change in the conductance from the initial conductance of the devices. We believe this is because, although new traps are generated during the STDP learning process, the traps get filled during the decay process over time. Hence, 1 V read bias was necessary to activate the traps to take part in conduction. Therefore, the devices in LTP are nonvolatile in this regard. For LTD, the read voltage was kept at 0.5 V since the device with maximum LTD was at a conductance level that could be distinguished using 0.5 V read. This observation clearly indicates ability for an analog reconfiguration of conductance in these devices and provides opportunity to achieve STDP by integrating these devices in neural circuits where interface circuitry can sense t and output an appropriate τ to be applied across these devices. IV. C ONCLUSION In this paper, we demonstrated that doped-TMO can closely mimic the characteristics of a biological synapse. We also demonstrated STP as a function of frequency of excitation. In addition, we demonstrated LTP and LTD using these reconfigurable devices. The devices have a potential to show STDP type learning behavior as shown by a specialized test. The manifestation of these synaptic characteristics were explained and modeled based on the experimentally observed device physics. Our studies open new routes to achieve biologically inspired synaptic devices on chip and replicate the functional diversity of biological synapses by modulating the dopants. However, these devices currently operate at higher voltages than desirable for biologically inspired artificial synapses and for compatibility with the CMOS devices. Therefore, future work in this area lies in optimizing the device and material parameters to match the low-power specifications of a biological synapse, studying the upper limits

of long-term retention times of different conductance states at room temperature and high temperatures, and reliability. ACKNOWLEDGMENT We would like to thank Dr. Scott C. Molitor in the Bioengineering department at the University of Toledo for useful discussions during this study. R EFERENCES [1] T. Shibata, H. Kosaka, H. Ishii, and T. Ohmi, “A neuron-MOS neural network using self-learning-compatible synapse circuits,” IEEE J. SolidState Circuits, vol. 30, no. 8, pp. 913–922, Aug. 1995. [2] C. Diorio, P. Hasler, B. A. Minch, and C. A. Mead, “A floating-gate MOS learning array with locally computed weight updates,” IEEE Trans. Electron Devices, vol. 44, no. 12, pp. 2281–2289, Dec. 1997. [3] J.-S. Seo, B. Brezzo, Y. Liu, B. D. Parker, S. K. Esser, R. K. Montoye, et al., “A 45 nm CMOS neuromorphic chip with a scalable architecture for learning in networks of spiking neurons,” in Proc. IEEE CICC, Sep. 2011, pp. 1–4. [4] S. H. Jo, T. Chang, I. Ebong, B. B. Bhadviya, P. Mazumdar, and W. Lu, “Nanoscale memristor device as synapse in neuromorphic systems,” Nano Lett., vol. 10, no. 4, pp. 1297–1301, Mar. 2010. [5] D. Kuzum, R. G. D. Jeyasingh, B. Lee, and H.-S. P. Wong, “Nanoelectronic programmable synapses based on phase change materials for brain-inspired computing,” Nano Lett., vol. 12, no. 5, pp. 2179–2186, Jun. 2011. [6] S. Yu, Y. Wu, R. Jeyasingh, D. Kuzum, and H. P. Wong, “An electronic synapse device based on metal oxide resistive switching memory for neuromorphic computation,” IEEE Trans. Electron Devices, vol. 58, no. 8, pp. 2729–2737, Aug. 2011. [7] B. Linares-Barranco and T. Serrano-Gotarredona. (2009) Memristance Can Explain Spike-Time-Dependent-Plasticity in Neural Synapses [Online]. Available: http://hdl.handle.net/10101/npre.2009.3010.1 [8] Y. V. Pershin and M. D. Ventra, “Experimental demonstration of associative memory with memristive neural networks,” J. Neural Netw., vol. 23, no. 7, pp. 881–886, Sep. 2010. [9] T. Chang, S.-H. Jo, and W. Lu, “Short-term memory to long-term memory transition in a nanoscale memristor,” ACS Nano, vol. 5, no. 9, pp. 7669–7676, Aug. 2011. [10] R. Meyer and R. Waser, “Observation of vacancy defect migration in the cation sublattice of complex oxides by 18O tracer experiments,” Phys. Rev. Lett., vol. 90, no. 10, pp. 105901-1–105901-4, Mar. 2003. [11] T. A. Lee and A. Navrotsky, “Enthalpy of formation of cubic yttriastabilized hafnia,” J. Mater. Res., vol. 19, no. 6, pp. 1855–1861, Mar. 2004. [12] I. V. Maznichenko, S. Ostanin, A. Ernst, and I. Mertig, “First-principles study of manganese-stabilized hafnia,” J. Magn. Magn. Mater., vol. 321, no. 7, pp. 913–916, Apr. 2009. [13] M. Houssa, M. Tuominen, M. Naili, V. Afanasev, A. Stesmans, S. Haukka, et al., “Trap-assisted tunneling in high permittivity gate dielectric stacks,” J. Appl. Phys., vol. 87, no. 12, pp. 8615–8620, Jun. 2000. [14] S. Yu, X. Guan, and H.-S. P. Wong, “Conduction mechanism of TiN/HfO x /Pt resistive switching memory: A trap-assisted-tunneling model,” Appl. Phys. Lett., vol. 99, no. 6, pp. 063507-1–063507-3, Aug. 2011. [15] D. M. Sathaiya and S. Karmalkar, “Thermionic trap-assisted tunneling model and its application to leakage current in nitrided oxides and AlGaN/GaN high electron mobility transistors,” J. Appl. Phys., vol. 99, no. 9, pp. 093701-1–093701-6, 2006. [16] R. S. Zucker, “Short-term synaptic plasticity,” Annu. Rev. Neurosci., vol. 12, pp. 13–31, Mar. 1989. [17] C. Rosemund, J. Clements, and G. L. Westbrook, “Nonuniform probability of glutamate release at a hippocampal synapse,” Science, vol. 262, no. 5134, pp. 754–757, Oct. 1993. [18] L. R. Squire, F. E. Bloom, and N. C. Spitzer, Fundamental Neuroscience. Amsterdam, The Netherlands: Elsevier, 2008. [19] O. Sacchi, M. L. Rossi, R. Canella, and R. Fesce, “Synaptic current at the rat ganglionic synapse and its interactions with the neuronal voltage-dependent currents,” J. Neurophysiol., vol. 79, no. 2, pp. 727–742, Feb. 1998.

MANDAL et al.: STUDY OF SYNAPTIC BEHAVIOR

[20] B. Ermentrout, “Dynamical consequences of fast-rising, slow-decaying synapses in neuronal networks,” J. Neural Comput., vol. 15, no. 11, pp. 2483–2522, Nov. 2003. [21] J. E. Lisman, “Bursts as a unit of neural information: Making unreliable synapses reliable,” Trends Neurosci., vol. 20, pp. 38–43, Jun. 1997. [22] C. M. Hempel, K. H. Hartman, X. J. Wang, G. G. Turrigiano, and S. B. Nelson, “Multiple forms of short-term plasticity at excitatory synapses in rat medial prefrontal cortex,” J. Neurophysiol., vol. 83, pp. 3031–3041, May 2000. [23] E. S. Fortune and G. J. Rose, “Short-term synaptic plasticity contributes to the temporal filtering of electrosensory information,” J. Neurosci., vol. 20, pp. 7122–7130, Sep. 2000. [24] T. H. Brown, P. F. Chapman, E. W. Kairiss, and C. L. Keenan, “Longterm synaptic potentiation,” Science, vol. 242, pp. 724–728, Nov. 1988. [25] T. V. P. Bliss and G. L. Collingridge, “A synaptic model of memory: Long-term potentiation in the hippocampus,” Nature, vol. 361, pp. 31–39, Jan. 1993. [26] J. R. Whitlock, A. J. Heynen, M. G. Shuler, and M. F. Bear, “Learning induces long-term potentiation in the hippocampus,” Science, vol. 313, pp. 1093–1097, Aug. 2006. [27] D. J. Dumin and J. R. Maddux, “Correlation of stress-induced leakage current in thin oxides with trap generation inside the oxides,” IEEE Trans. Electron Devices, vol. 40, no. 5, pp. 986–993, May 1993. [28] R. Moazzami and C. Hu, “Stress-induced current in thin silicon dioxide films,” in Proc. IEEE IEDM, Dec. 1992, pp. 139–142. [29] H. Chen, J. Lan, J. Chen, and J. Y. Lee, “Time-dependent and traprelated current conduction mechanism in ferroelectric Pb(Zrx Ti1−x )O3 films,” Appl. Phys. Lett., vol. 69, pp. 1713–1715, Sep. 1996. [30] J. T. Wixted and S. K. Carpenter, “The Wickelgren power law and the Ebbinghaus savings function,” Psychol. Sci., vol. 18, pp. 133–134, Feb. 2007. [31] J. T. Wixted and E. B. Ebbesen, “On the form of forgetting,” Psychol. Sci., vol. 6, pp. 409–415, Nov. 1991. [32] S. J. Martin, P. D. Grimwood, and R. G. M. Morris, “Synaptic plasticity and memory: An evaluation of the hypothesis,” Annu. Rev. Neurosci., vol. 23, pp. 649–711, Mar. 2000. [33] C. F. Stevens and J. F. Wesseling, “Augmentation is a potentiation of the exocytotic process,” Neuron, vol. 22, pp. 139–146, Jan. 1999. [34] L. F. Abbott and W. G. Regehr, “Synaptic computation,” Nature, vol. 431, pp. 796–803, Oct. 2004. [35] B. Granseth and S. Lindström, “Augmentation of corticogeniculate EPSCs in principal cells of the dorsal lateral geniculate nucleus of the rat investigated in vitro,” J. Physiol., vol. 556, pp. 147–157, Apr. 2004. [36] H. Kamiya and R. S. Zucker, “Residual Ca2+ and short-term synaptic plasticity,” Nature, vol. 371, pp. 603–606, Oct. 1994. [37] K. S. Harris, J. C. Fiala, and L. Ostroff, “Structural changes at dendritic spine synapses during long-term potentiation,” Phil. Trans. R. Soc. London, vol. 358, pp. 745–748, Apr. 2003. [38] G. Bi and M. Poo, “Synaptic modifications in cultured hippocampal neurons: Dependence on spike timing, synaptic strength, and postsynaptic cell type,” J. Neurosci., vol. 18, pp. 10464–10472, Dec. 1998.

4225

[39] S. Song, K. D. Miller, and L. F. Abbott, “Competitive Hebbian learning through spike-timing-dependent synaptic plasticity,” Nature Neurosci., vol. 3, pp. 919–926, Sep. 2000. [40] L. F. Abbott and S. B. Nelson, “Synaptic plasticity: Taming the beast,” Nature Neurosci., vol. 3, pp. 1178–1183, Nov. 2000. [41] G. Indiveri, E. Chicca, and R. Douglas, “A VLSI array of low-power spiking neurons and bistable synapses with spike-timing dependent plasticity,” IEEE Trans. Neural Netw., vol. 17, no. 1, pp. 211–221, Jan. 2006. [42] G. S. Snider, “Spike-timing-dependent learning in memristive nanodevices,” in Proc. IEEE Int. Symp. Nanoscale Arch., Jun. 2008, pp. 85–92.

Saptarshi Mandal (S’13) has been pursuing the M.S. degree in electrical engineering from the University of Toledo, Toledo, OH, USA, since 2011. His current research interests include fabrication, characterization, and physics-based simulation of transition metal-oxide-based memristive devices for neuromorphic applications.

Branden Long received the M.S. and Ph.D. degrees in electrical engineering from the University of Toledo, Toledo, OH, USA, in 2009 and 2013, respectively. He is currently a Device Characterization Engineer with Intel, Hillsboro, OR, USA.

Rashmi Jha (S’05–M’11) received the Ph.D. degree in electrical engineering from North Carolina State University, Raleigh, NC, USA, in 2006. She has been an Assistant Professor with the University of Toledo, Toledo, OH, USA, since 2008.