Mimicking Biological Synaptic Functionality with an Indium Phosphide

Jan 12, 2018 - ... as self-driving vehicles or Internet of Things applications. In biology, the synapse serves as an active memory unit in the neural ...
0 downloads 0 Views 2MB Size
Subscriber access provided by Gothenburg University Library

Article

Mimicking Biological Synaptic Functionality with an Indium Phosphide Synaptic Device on Silicon for Scalable Neuromorphic Computing Debarghya Sarkar, Jun Tao, Wei Wang, Qingfeng Lin, Matthew Yeung, Chenhao Ren, and Rehan Kapadia ACS Nano, Just Accepted Manuscript • DOI: 10.1021/acsnano.7b08272 • Publication Date (Web): 12 Jan 2018 Downloaded from http://pubs.acs.org on January 12, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

ACS Nano is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

Mimicking Biological Synaptic Functionality with an Indium Phosphide Synaptic Device on Silicon for Scalable Neuromorphic Computing Debarghya Sarkar*, Jun Tao*, Wei Wang, Qingfeng Lin, Matthew Yeung, Chenhao Ren, Rehan Kapadia1 Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, California 90089, United States *Equal Contribution from Authors 1

[email protected]

Keywords: synaptic device, neuromorphic computing, metaplasticity, spike timing dependent plasticity, templated liquid phase growth, indium phosphide, Si back-end processing.

Abstract: Neuromorphic or “brain-like” computation is a leading candidate for efficient, faulttolerant processing of large-scale data, as well as real-time sensing and transduction of complex multivariate systems and networks such as self-driving vehicles or Internet of Things applications. In biology, the synapse serves as an active memory unit in the neural system, and is the component responsible for learning and memory. Electronically emulating this element via a compact, scalable technology which can be integrated in a 3-D architecture is critical for future implementations of neuromorphic processors. However, present day 3-D transistor implementations of synapses are typically based on low-mobility semiconductor channels or

1 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

technologies that are not scalable. Here, we demonstrate a crystalline InP based artificial synapse for spiking neural networks that exhibits elasticity, short-term plasticity, long-term plasticity, metaplasticity, and spike timing dependent plasticity, emulating the critical behaviors exhibited by biological synapses. Critically, we show that this crystalline InP device can be directly integrated via back end processing on a Si wafer using a SiO2 buffer without the need for a crystalline seed, enabling neuromorphic devices that can be implemented in a scalable and 3-D architecture. Specifically, the device is a crystalline InP-channel field effect transistor (FET) that interacts with neuron spikes by modification of the population of filled traps in the MOS structure itself. Unlike other transistor based implementations, we show that it is possible to mimic these biological functions without the use of external factors (e.g. surface adsorption of gas molecules), and without the need for the high electric fields necessary for traditional flash based implementations. Finally, when exposed to neuronal spikes with a waveform similar to that observed in the brain, these devices exhibit the ability to learn without the need for any external potentiating/depressing circuits, mimicking the biological process of Hebbian learning. The Arithmetic/Logic Unit (ALU) in any modern day Central Processing Unit (CPU) is the fundamental building block responsible for accurate computation and logic operations. While the requirement of accuracy and precision in computation is necessary for many applications, a dramatic increase in the need to extract general relations from large sets of high-dimensional data has driven the use of machine learning and multi-dimensional optimization algorithms in fields such as computer vision, voice recognition, natural language processing, and autonomous systems. To obtain useful results from these approaches, it is typically necessary to train the algorithms on large clusters with large amounts of data, requiring hours, or even days for commercially relevant applications. However, in this respect, the human brain supersedes any

2 ACS Paragon Plus Environment

Page 2 of 30

Page 3 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

supercomputer, being six to nine orders of magnitude more power efficient than today’s digital logic computers for these types of applications. A distinctive feature of the human brain is the closely intertwined nature of the memory and logic units which enables this efficient computing, unlike today’s computers where the CPU and memory units though themselves being sufficiently fast, are largely restricted by data transport among them, commonly referred to as the von Neumann architecture bottleneck. To this end, researchers have started developing computing architectures that mimic the functioning of the brain, referred to as neuromorphic computing. While the exact functioning of the brain is an area of active neuroscientific research, to a good approximation, it may be said that the neurons act as the logic units operating in a threedimensional multi-connected network supported by their end connections called synapses, acting as the memory units.1-4 Neural networks have been long studied and implemented in software,5 and some prototypes of hardware neuromorphic systems have already been built using existing CMOS technology.6-12 Neuronal spiking13-15 and synaptic behavior16-20 have been emulated using CMOS based circuits containing 6 to 12 transistors depending on the specific functionality and robustness of the design. But an omnipresent usage would require hardware implementation with reduced footprint and energy consumption, and increased parallelism and interconnections, by several orders of magnitude.4 Architectures based on single device to act as individual units of the neuromorphic computing block are therefore heavily studied.1-3 While there have been fewer reports of neuronal behavior from single device,21-22 emulation of synaptic activity from single device has been widely demonstrated, pioneered by Diorio et al. in 1996 based on hot-electron injection and electron tunneling in a floating-gate Si MOS transistor.23 Memristor-based synaptic devices from phase-change materials,24 ferroelectricity,25-26 ferromagnetism,22 and nanoionics27-

3 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

40

have been demonstrated. Time dependent threshold voltage shift arising from charging and

discharging of traps in the semiconductor,41 at the insulator-semiconductor interface,42-43 or ion migration in the gate dielectric44 in MOSFETs has more recently been utilized for synaptic emulation. However, due to the fundamental limitations on the growth of high-quality crystalline material on amorphous materials, current 3-D approaches often use low-mobility solution deposited materials,45-46 nanocrystalline channels,24, 41 or physical transfer processes42-43 that are not scalable. Here, we demonstrate a scalable, back-end compatible, artificial synapse with a crystalline indium phosphide (InP) channel grown directly on Si/SiO2 wafer with templated liquid phase growth.47-50 We show these synaptic devices emulate a range of behaviors such as elasticity, short-term and long-term plasticity, metaplasticity, and spike timing dependent plasticity. Results and Discussion A synapse51 is a 20-40 nm junction between two neurons (schematically shown in Figure 1 (a)) which aids in the communication of signals between neurons. These connections are not hard-coded but evolve over time, and are plastic, depending on the activity of the connecting neurons. Based on the nature of pre- and post-synaptic activity, the synaptic weight may remain unchanged (elasticity), increase (potentiation), or decrease (depression). The change in the synaptic connectivity (also called synaptic plasticity) is regarded as a key ingredient in learning and memory. Synaptic plasticity leads to future data processing along the same pathway to be accordingly affected: a potentiated synapse would enable better transduction of information from pre-synaptic action potentials, and a depressed synapse would lead to a constrained transduction of information. The difference in post synaptic current (PSC) determines the synaptic weight

4 ACS Paragon Plus Environment

Page 4 of 30

Page 5 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

change that occurred owing to the coupling neuronal activity. The synaptic weight change is also time dependent, with a gradual decay over time. Based on the lifetime of the plasticity, two categories of plasticity are defined: short-term plasticity and long-term plasticity. While the exact times scales defining each is not precisely defined, short-term plasticity often refers to behavior on the milliseconds to seconds time frame, while long-term plasticity refers to behavior on the time frame of seconds to years. The device architecture used in this work to emulate the synaptic behavior, is schematically shown in Figure 1 (b). The device is a top-gated 100 nm thick InP nanowire transistor fabricated on Si/SiO2 wafer with 60 nm Al2O3 as the gate dielectric. The Al2O3-InP stack may be regarded as the synapse connecting the cell membrane of the pre-synaptic neuron (gate electrode) to the cell membrane of the post-synaptic neuron (drain electrode). Pre-synaptic action potentials are applied as pulses to the gate electrode, which leads to a post-synaptic current (PSC) in the drain electrode based on the strength of the synapse. The synaptic strength is the conductance of the semiconductor channel. Application of a voltage pulse on the gate leads to charging or discharging of traps in the oxide or at the oxide/semiconductor interface.43 This in turn changes the electrostatics of the channel and leads to a shift in the threshold voltage, so that the drain current (PSC) at the same constant gate voltage before and after the gate voltage pulse, may be different. The magnitude of the hysteresis depends on the pulse amplitude and width. The amplitude determines the electric field within the oxide, and the population of charge carriers in the semiconductor, with higher amplitudes accessing higher energy traps. The pulse width determines the length of time under which the oxide is under tension, changing the number of traps that are charged or discharged. The synaptic strength evolves as a function of the nature of the pulse and the strength of the synapse before the application of the pulse.

5 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Here we demonstrate the implementation of these artificial synapses using a scalable platform, which enables arrays of crystalline InP devices to be grown on Si/SiO2 wafers with near unity yield. To carry this out, we utilize templated liquid phase (TLP) growth. Briefly, an array of 100 nm thick In nanowires and capping SiO2 layer is defined on the Si/SiO2 substrate using photolithography, evaporation, and lift-off processes, schematically shown in Figure 2(ab). A thin 4-6 nm layer of molybdenum oxide (MoOx) is deposited below the indium to aid in wetting of In. At the growth temperature, the indium nanowires turn to liquid but maintain their geometry due to the capping layer. In the presence of phosphine and hydrogen, InP precipitates out of the In liquid, leading to an array of InP nanowires directly grown on an Si/SiO2 substrate, as shown in Figure 2(c). After the nanowires are grown, we fabricate device arrays, as shown schematically in Figure 2(d). Using lithography, we carry out source drain metallization with GeAu-Ni contacts, followed by a 380 oC rapid thermal annealing (RTA). 60 nm of Al2O3 is deposited using atomic layer deposition (ALD) as the gate dielectric, and Ni is used as the gate metal. Details of the TLP growth of InP FET channels and device fabrication are discussed in the Methods section. The output characteristics of a representative 2 × 4 array of devices are shown in Figure S1(a-h), with an average extracted effective mobility of ~200 cm2/V-s at 3 V gate voltage overdrive with peak effective mobilities of ~500 cm2/V-s. This is more than 10x higher than most CNT network devices, which exhibit peak mobilities of ~5-20 cm2/V-s.52-53 The top row of devices all have channel length of L=10 µm, while the bottom row all have channel lengths of L=25 µm. Critically, we demonstrate the ability to deterministically place and grow devices on an amorphous substrate with high device yield. The effect of amplitude of gate pulse in changing the electrostatics of the channel and thus the behavior of the artificial synapse, is shown in Figure 3. We interpret the amplitude of

6 ACS Paragon Plus Environment

Page 6 of 30

Page 7 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

gate pulses to be proportional to the probability of exocytosis for neurotransmitter release.43 Thus, a higher amplitude pulse signifies higher probability of release of neurotransmitters than that of a lower amplitude pulse. Physically, a higher gate voltage pulse causes higher energy traps to be accessed. Figures 3 (a) and (b) show the transfer characteristics of the synaptic device at 3 V source-drain voltage for different peak gate voltages (5 V and 0.7 V respectively), starting from 0 V. It is found that a peak gate voltage of +5 V increases the threshold voltage, lowering Ids (Figure 3 (a)), while a peak gate voltage of -5 V decreases the threshold voltage, increasing Ids. From this, we show that positive gate voltage pulses lead to depression, a reduction in the PSC, while negative voltage pulses lead to potentiation. However, as shown in Figure 3 (b), it is seen that a peak voltage of +0.7 V has nearly no effect on the PSC, leading to elastic behavior of the synapses, where a pulse is passed however no synaptic weight change occurs. To show the dynamics of the synapse, representative post-synaptic current (PSC) curves are plotted in Figure 3 (c). The current for  < 0 is the PSC before arrival of the spikes, for  > 0 is the PSC after arrival of the spikes, and the dotted line indicates the time of spike arrival. Here, we use a rectangular pulse train of 40 pulses of width 500 µs, and 50% duty cycle with 0 V baseline. The three behaviors were obtained by the application of pulses with varying amplitudes. A pulse amplitude of 0.1 V results in elasticity, -5 V gives potentiation and +5 V gives depression. We also plot the synaptic weight change for different peak pre-synaptic potentials in Figure 3 (d). We define the short-term plasticity (STP) as the difference between the average PSC over the first 1 s after pulse, and that before pulse, normalized to the PSC before pulse. The long-term plasticity (LTP) is defined as the difference between the normalized average PSC between 10 s and 40 s after pulse, and that before pulse. The magnitude of plasticity increases with increasing voltage amplitude, and can be modeled with exponential functions of the form,

7 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

∆ = ∗

 

where  is the peak gate voltage, and  is the activation voltage for the traps. The physical implication of a non-zero plasticity is that a future signal would be transduced with the modified weight of the synapse. This leads to selectivity of pathways for transduction of a signal to converge to an optimal solution in a neural network. It may be noted that the synaptic weight change at time t is actually a function of the synaptic weight at t and the neuronal activity at that time.54 Mathematically, it may be represented as ∆ =  ,  , where  is the synaptic weight and  is the neuronal activity at time . Thus, for the same neuronal activity, the synaptic weight change would depend on the initial condition of the synapse. In other words, for the same gate voltage pulse(s), the change in the drain current (arising from the change in the electrostatics of the channel) would depend on the initial electrostatic condition of the channel and the threshold voltage of the device.37, 39 This phenomenon is referred to as metaplasticity and is one of the most important characteristics of biological synapses. It allows a highly non-linear optimization of the synaptic weight based on the sequence of signals it transduces, to in turn give rise to an optimal selectivity of the signals being transduced. Metaplasticity of our synaptic device was emulated by applying the same gate voltage pulse but preceded by different “priming” gate voltage pulses to emulate prior brain activity. The gate voltage pulse trains consisted of 10 pulses of 5 ms width and 6 ms period. As a demonstration, we chose +2.5 V pulse as a depressing priming pulse, and -2.5 V pulse as a potentiating priming pulse. The PSC plots for two representative potentiating and depressing cases are depicted in Figures 4 (a) and (b). The same -4.5 V pulse train gives a significantly 8 ACS Paragon Plus Environment

Page 8 of 30

Page 9 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

higher relative change of PSC when it follows depressing priming, than when it succeeds a potentiating priming. Similarly, a depressing priming allows a less depression for a +4.5 V pulse train than a potentiating priming. Figure 4 (c) shows the short-term synaptic weight change extracted from the PSCs for gate voltages ranging between -5 V to +5 V, for three different initial conditions of no priming, depressing priming, and potentiating priming. It is observed that a reduction in potentiation weight change occurs for an initially potentiated synapse, while an increased potentiation occurs for an initially depressed synapse. Similarly, an initially depressed synapse leads to having less depression than an initially potentiated synapse. As may be realized, the feedback mechanism embedded in metaplasticity allows the synapse to optimally tune its weight such that it regressively stays within bounds instead of diverging towards either extreme. For the same initial state of the synapse, the synaptic weight change is also dependent on the number of neurotransmitters being transmitted simultaneously.43 The number of neurotransmitters is in turn, dependent on the number of action potentials that have reached the pre-synaptic terminal in quick succession (each action potential typically releasing same number of neurotransmitters). To emulate the pulse number dependent plasticity of synapses, we excite the synaptic device with +5 V peak (for depression) or -5 V peak (for potentiation) and 0 V baseline pulse trains of 5 ms width and 5.5 ms period, consisting of varying number of pulses, the results of which are shown in Figure 5. Some representative potentiation and depression curves (for pulse trains consisting of 1, 20, and 100 pulses) are plotted in Figure 5 (a). A general trend of increased potentiation or depression is observed as we increase the number of incident action potentials. The PSC in each case was fitted with a double exponential function29 given by:  =  + 0.5 ∗ #$ −   ∗  &/( + &/) 

9 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The PSC approaches the steady state value  from #$ at a rate given by the double exponentials: one of them is used to fit the very fast decay right after the pulse, while the other gives the relatively slower decay rate until the PSC virtually reaches the steady-state value. The PSC decay time constant (inverse of the slower decay rate), referred to in the graph as the synaptic relaxation time constant, is plotted in Figure 5 (b). It is also seen that the relaxation rate initially decreases and gradually saturates with increase in the number of action potentials stimulating the synapse. Figure 5 (c) and (d) respectively show the long-term and short-term synaptic weight change, where the weight change definitions have been as done earlier. As expected, a gradual increase in the plasticity of the synapse is observed with increasing number of action potentials, as more number of traps are populated, until gradually all or most of the traps that can be affected by the pulse amplitude, are influenced, so that the plasticity gradually saturates. The same reasoning may be applied for the trend observed for the relaxation rate as well. Considering that faster traps are populated earlier than slower traps, with increasing number of pulses we gradually excite higher number of slower traps, increasing the decay time constant. Once an action potential is generated in the post-synaptic neuron, it travels down the length of the axon. However, a fraction of this signal is reflected and impinges on the output terminal of the synapse, potentially changing the synaptic strength. Synaptic modification arising from the correlated activity of the pre-synaptic and post-synaptic neuron was first introduced by Hebb55 in 1949, and later refined by other researchers such as Bienenstock-Cooper-Munro56 in 1982. One of the most well-known versions of Hebbian (or BCM) learning is spike timing dependent plasticity (STDP), some of the first demonstrations of which were in classic neuroscience experiments such as by Markram et al.,57 Bi and Poo,58 and modeled by van

10 ACS Paragon Plus Environment

Page 10 of 30

Page 11 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

Rossum et al.59 STDP has been emulated by the neuromorphic device community over the years, mostly based on memristor technology. However, STDP in MOSFET based synapses has not clearly been demonstrated using solely time-correlated activity of the pre- and post-synaptic neuron. Instead, due to the nature of the pulse waveforms used and the behavior of MOSFETs, it is necessary to use differing pulses for depression and potentiation, which would require external circuitry to enable learning. While many different STDP behaviors exist, a commonly implemented STDP behavior is the potentiation of the synapse when the post-synaptic action potential succeeds the reflected pre-synaptic action potential, and depression when the reflected post-synaptic action potential precedes the pre-synaptic action potential. The highest magnitude of potentiation and depression occurs for the shortest time gap (∆t) between the pre- and post-synaptic action potentials, and the synapse tends to stay elastic for increasing time spacing between the pre- and post-synaptic action potentials. Such an asymmetric and time-correlated behavior may well be reasoned based on causality principles. Here we utilize engineering of hysteresis in a family of transfer characteristics to demonstrate STDP in our synaptic device. The pre-synaptic action potential is designed as a pulse that goes from 0 to +2.75 V to -5 V to 0 V with a total width of 10 ms. The reflected post-synaptic action potential is designed as a pulse with a baseline of 1.5 V, ramping up to 2.1 V in 100 ms, ramping down to 0.9 V in 20 ms, and again ramping up to 1.5 V in 100 ms. The peak values of the pre-synaptic potential were chosen such that it results in elasticity at the 1.5 V baseline post-synaptic potential (Figure S1 (b)), which is the boundary condition for high |∆t|. For post-synaptic potentials less than 1.5 V, the same pre-synaptic pulse gives depression, with increasing plasticity for lower potentials (representative hysteresis shown in Figure S1 (a)). For post-synaptic potentials greater than 1.5 V, the pre-synaptic pulse gives

11 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

potentiation, with increasing plasticity for higher potentials (representative hysteresis shown in Figure S1 (c)). This effect has been utilized to accordingly design the shape of the post-synaptic potential. The ramp times have been chosen to closely match that found in STDP of typical synapses. The upper panel of Figure 6 (a) shows the waveform simulating the reflected postsynaptic action potential, while the lower panel depicts the pre-synaptic pulse. The relative positioning of the two pulses can be directly interpreted as positive and negative ∆t as has been shown in Figure 6 (a). The long-term synaptic weight change for different ∆t was calculated, and has been plotted in Figure 6 (b). It is seen that as expected from the waveform design, we get maximum potentiation (depression) for the minimum positive (negative) ∆t, which gradually decays down to give elasticity for longer time intervals. It may also be noted that other different STDP behaviors such as those demonstrated by Kuzum et al.,24 may also be easily reproduced in our synaptic device by only changing the post-synaptic potentials, for the same pre-synaptic potential as used here. Conclusion In summary, we demonstrated the scalable fabrication of an array of nanowire FETs integrated on Si/SiO2 wafer by direct growth of InP nanowire channels and reported their synaptic functionality for spiking neural network based neuromorphic computation. Modeling the gate electrode as the pre-synaptic axon terminal, the drain electrode as the post-synaptic dendrite, and the gate oxide-semiconductor channel as the synapse, we have controlled the charging and discharging of interfacial traps in the MOS structure to modulate the channel conductance interpreted as the synaptic weight. Non-linear synaptic weight change dependent on pre- and post-synaptic neuronal activity, such as elasticity, short- and long-term plasticity, metaplasticity, spike number dependent plasticity, and spike timing dependent plasticity, have

12 ACS Paragon Plus Environment

Page 12 of 30

Page 13 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

been emulated by engineering the hysteresis of the channel conductance. Future work may involve realization of a single device spiking neuron and integration of the synapse and neuron arrays to form an on-chip compact neuromorphic circuit. Further innovation in the synaptic device architecture may also be done to expand the synaptic behavioral space such as to include selectivity and responsivity of synapses to different pre- and post-synaptic activity. Methods Templated liquid phase growth The Si/SiO2 substrate was thoroughly cleaned with acetone, isopropyl alcohol and deionized water, and blow-dried with dry nitrogen. FET channel templates were photolithographically patterned on the substrate. 4-6 nm MoOx followed by 100 nm Indium (99.99995%) was thermally evaporated using alumina covered tungsten boats, on the patterned samples connected to a cryo-cooled (liquid nitrogen) stage. 100 nm SiO2 was electron-beam evaporated using SiO2 pellets (99.99%), as the capping layer. FET templates of MoOx/In/SiO2 were then obtained using standard lift-off procedure. TLP growth of InP was done in a single-zone hot-wall tube furnace at furnace temperatures in the range 550-5750C. 99.9999% PH3, diluted ex-situ with 99.999% H2 to achieve the desired PH3 partial pressure, was used as the P precursor. Device fabrication The capping SiO2 layer on TLP grown InP FET channels was etched in 1:10 hydrofluoric acid (HF, 48-50%) dissolved in water. Source-drain pads were then patterned using photolithography. 6 nm Ge, 20 nm Au, and 80 nm Ni were electron-beam evaporated, followed by lift-off. A 380 oC rapid thermal annealing (RTA) was done to reduce contact resistance by Ge alloying with InP at the source-drain contact regions. 60 nm of Al2O3 was then deposited using

13 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

atomic layer deposition (ALD) as the gate dielectric. After that, gate electrodes were patterned using photolithography, followed by 80 nm Ni electron-beam evaporation and lift-off. Sourcedrain contact vias were then opened by photolithography and etching the Al2O3 on the contact pads. Device characterization The synaptic devices were excited with pulses generated from a Tektronix AWG 2021, and a time dependent measurement was done with an Agilent B1500A semiconductor device analyzer. Supporting Information The Supporting Information is available free of charge on the ACS Publications website at DOI:. Output characteristics of an array of transistors are shown in Figure S1; hysteresis engineering of transfer characteristics showing depression, elasticity, and potentiation for different Vds is shown in Figure S2. Acknowledgements R.K. acknowledges funding from the National Science Foundation (Award #1610604). D.S. thanks the support by the USC Annenberg Graduate Fellowship. J.T. thanks the support by the USC Provost Graduate Fellowship. The authors gratefully acknowledge the use of the ALD facility in Dr. Stephen Cronin’s lab.

Figure Captions Figure 1. (a) Schematic of a biological synapse showing communication of signal between neurons by release of neurotransmitters from the pre-synaptic neuron causing diffusion of Na+ ions into the post-synaptic neuron. (b) Schematic of the indium phosphide channel field effect

14 ACS Paragon Plus Environment

Page 14 of 30

Page 15 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

transistor on silicon as the synaptic device, where a voltage spike in the pre-synaptic neuron terminal (gate) results in a change in the current in the post-synaptic terminal (drain) based on the synaptic dynamics (oxide-semiconductor charge traps).

Figure 2. (a-d) Schematic of templated liquid phase growth of crystalline InP nanowire arrays on Si/SiO2 and subsequent fabrication of nanowire FETs.

Figure 3. Spike Amplitude Dependent Plasticity. (a) Hysteretic transfer characteristics showing depression for +5 V and potentiation for -5 V. (b) Hysteretic transfer characteristics showing elasticity for +0.7 V and potentiation for -0.7 V. (c) Transient post-synaptic current before and after application of pre-synaptic pulse leading to elasticity (0.1 V), potentiation (-5V), and depression (+5V). (d) Short-term and long-term synaptic weight change for different values of pre-synaptic voltage pulse. (All curves reported with Vds = 3 V)

Figure 4. Synaptic metaplasticity. (a) Increased potentiation of synapse succeeding a depressing priming compared to potentiating priming. (b) Increased depression of synapse succeeding a potentiating priming compared to depressing priming. (c) Short-term synaptic weight change for different values of pre-synaptic voltage pulse without priming, and with depressing and potentiating priming. (All curves reported with Vds = 3 V)

Figure 5. Spike Number Dependent Plasticity. (a) Transient post-synaptic current before and after application of pre-synaptic pulse for varying pulse number, nAP=1, 20, 100. (b) Variation of short-term synaptic relaxation time constant with different number of action potentials. (c) Long-

15 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

term, and (d) short-term synaptic weight change for different number of potentiating and depressing action potentials. (All curves reported with Vds = 3 V)

Figure 6. Spike Timing Dependent Plasticity. (a) Waveforms representing back-reflected postsynaptic action potential (upper row) and pre-synaptic action potential (lower row) for presynaptic action potential preceding (left column) and succeeding (right column) the reflected post-synaptic action potential. (b) Long-term synaptic weight change for different values of time offset between the pre- and post-synaptic action potentials.

References: 1.

Kuzum, D.; Yu, S.; Wong, H. S. P. Synaptic Electronics: Materials, Devices and

Applications. Nanotechnol. 2013, 24, 382001. 2.

Jeong, D. S.; Kim, I.; Ziegler, M.; Kohlstedt, H. Towards Artificial Neurons and

Synapses: A Materials Point of View. RSC Adv. 2013, 3, 3169-3183. 3.

Advances in Neuromorphic Hardware Exploiting Emerging Nanoscale Devices; Suri, M.

Ed., Springer, 2017. 4.

Neuromorphic Computing: from Materials to Systems Architecture. Report of a

Roundtable Convened to Consider Neuromorphic Computing Basic Research Needs; Schuller, I. K.; Stevens, R., Eds.; U.S. Department of Energy, 2015. 5.

Demuth, H. B.; Beale, M. H.; De Jess, O.; Hagan, M. T. In Neural Network Design;

Martin Hagan, 2014. 6.

Mead, C. Neuromorphic Electronic Systems. Proc. IEEE 1990, 78, 1629-1636.

16 ACS Paragon Plus Environment

Page 16 of 30

Page 17 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

7.

Mead, C.; Ismail, M. In Analog VLSI Implementation of Neural Systems. Springer

Science & Business Media, 2012. 8.

Merolla, P. A.; Arthur, J. V.; Alvarez-Icaza, R.; Cassidy, A. S.; Sawada, J.; Akopyan, F.;

Jackson, B. L.; Imam, N.; Guo, C.; Nakamura, Y.; Brezzo, B.; Vo, I.; Esser, S. K.; Appuswamy, R.; Taba, B.; Amir, A.; Flickner, M. D.; Risk, W. P.; Manohar, R.; Modha, D. S. A Million Spiking-Neuron Integrated Circuit with a Scalable Communication Network and Interface. Sci. 2014, 345, 668-673. 9.

Hsu, J. IBM's New Brain, IEEE Spectrum 2014, 51, 17-19.

10.

Furber, S. B.; Galluppi, F.; Temple, S.; Plana, L. A. The Spinnaker Project. Proc. IEEE

2014, 102, 652-665. 11.

Benjamin, B. V.; Gao, P.; McQuinn, E.; Choudhary, S.; Chandrasekaran, A. R.; Bussat,

J.-M.; Alvarez-Icaza, R.; Arthur, J. V.; Merolla, P. A.; Boahen, K. Neurogrid: A Mixed-AnalogDigital Multichip System for Large-Scale Neural Simulations. Proc. IEEE 2014, 102, 699-716. 12.

Furber, S., Large-Scale Neuromorphic Computing Systems. J. Neural Eng. 2016, 13,

051001. 13.

Farquhar, E.; Hasler, P., A Bio-Physically Inspired Silicon Neuron. IEEE Trans. Circuits

Syst. 2005, 52, 477-488. 14.

Rangan, V.; Ghosh, A.; Aparin, V.; Cauwenberghs, G. A Subthreshold aVLSI

Implementation of the Izhikevich Simple Neuron Model. In 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC).; IEEE, 2010; pp 4164-4167.

17 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

15.

Indiveri, G.; Linares-Barranco, B.; Hamilton, T. J.; Van Schaik, A.; Etienne-Cummings,

R.; Delbruck, T.; Liu, S.-C.; Dudek, P.; Häfliger, P.; Renaud, S., Neuromorphic Silicon Neuron Circuits. Front. Neurosci. 2011, 5, 73. 16.

Chicca, E.; Indiveri, G.; Douglas, R. An Adaptive Silicon Synapse. In Proceedings of the

2003 International Symposium on Circuits and Systems (ISCAS).; IEEE, 2003; pp I-81-84. 17.

Bartolozzi, C.; Indiveri, G. Synaptic Dynamics in Analog VLSI. Neural Comput. 2007,

19, 2581-2603. 18.

Friesz, A. K.; Parker, A. C.; Zhou, C.; Ryu, K.; Sanders, J. M.; Wong, H. S. P.; Deng, J.;

A Biomimetic Carbon Nanotube Synapse Circuit. In Biomedical Engineering Society (BMES) Annual Fall Meeting. 2007, 2, 29. 19.

Arthur, J. V.; Boahen, K. Learning in Silicon: Timing is Everything. In Advances in

Neural Information Processing Systems 18; Weiss, Y.; Schölkopf, B.; Platt, J.C., Eds.; MIT Press, 2006; pp 75-82. 20.

Azghadi, M. R.; Iannella, N.; Al-Sarawi, S. F.; Indiveri, G.; Abbott, D., Spike-Based

Synaptic Plasticity in Silicon: Design, Implementation, Application, and Challenges. Proc. IEEE 2014, 102, 717-737. 21.

Pickett, M. D.; Medeiros-Ribeiro, G.; Williams, R. S. A Scalable Neuristor Built with

Mott Memristors. Nat. Mater. 2013, 12, 114-117. 22.

Krzysteczko, P.; Münchenberger, J.; Schäfers, M.; Reiss, G.; Thomas, A. The Memristive

Magnetic Tunnel Junction as a Nanoscopic Synapse Neuron System. Adv. Mater. 2012, 24, 762-766. 23.

Diorio, C.; Hasler, P.; Minch, A.; Mead, C. A. A Single-Transistor Silicon Synapse.

IEEE Trans. Electron Devices 1996, 43, 1972-1980.

18 ACS Paragon Plus Environment

Page 18 of 30

Page 19 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

24.

Kuzum, D.; Jeyasingh, R. G. D.; Lee, B.; Wong, H. S. P. Nanoelectronic Programmable

Synapses Based on Phase Change Materials for Brain-Inspired Computing. Nano Lett. 2011, 12, 2179-2186. 25.

Chanthbouala, A.; Garcia, V.; Cherifi, R. O.; Bouzehouane, K.; Fusil, S.; Moya, X.;

Xavier, S.; Yamada, H.; Deranlot, C.; Mathur, N. D. A Ferroelectric Memristor. Nat. Mater. 2012, 11, 860-864. 26.

Zuo, F.; Panda, P.; Kotiuga, M.; Li, J.; Kang, M.; Mazzoli, C.; Zhou, H.; Barbour, A.;

Wilkins, S.; Narayanan, B. Habituation Based Synaptic Plasticity and Organismic Learning in a Quantum Perovskite. Nat. Commun. 2017, 8, 240. 27.

Chang, T.; Jo, S.-H.; Lu, W. Short-Term Memory to Long-Term Memory Transition in a

Nanoscale Memristor. ACS Nano 2011, 5, 7669-7676. 28.

Ohno, T.; Hasegawa, T.; Tsuruoka, T.; Terabe, K.; Gimzewski, J. K.; Aono, M. Short-

Term Plasticity and Long-Term Potentiation Mimicked in Single Inorganic Synapses. Nat. Mater. 2011, 10, 591-595. 29.

Wang, Z. Q.; Xu, H. Y.; Li, X. H.; Yu, H.; Liu, Y. C.; Zhu, X. J. Synaptic Learning and

Memory Functions Achieved Using Oxygen Ion Migration/Diffusion in an Amorphous InGaZnO Memristor. Adv. Funct.l Mater. 2012, 22, 2759-2765. 30.

Mayr, C.; Staerke, P.; Partzsch, J.; Cederstroem, L.; Schüffny, R.; Shuai, Y.; Du, N.;

Schmidt, H. Waveform Driven Plasticity in BiFeO3 Memristive Devices: Model and Implementation. In Advances in Neural Information Processing Systems 25; Pereira, F.; Burges, C. J. C.; Bottou, L.; Weinberger, K. Q. Eds.; Curran Associates, Inc., 2012; pp 1700-1708.

19 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

31.

Subramaniam, A.; Cantley, K. D.; Bersuker, G.; Gilmer, D. C.; Vogel, E. M. Spike-

Timing-Dependent Plasticity Using Biologically Realistic Action Potentials and LowTemperature Materials. IEEE Trans. Nanotechnol. 2013, 12, 450-459. 32.

Williamson, A.; Schumann, L.; Hiller, L.; Klefenz, F.; Hoerselmann, I.; Husar, P.;

Schober, A. Synaptic Behavior and STDP of Asymmetric Nanoscale Memristors in Biohybrid Systems. Nanoscale 2013, 5, 7297-7303. 33.

Ambrogio, S.; Balatti, S.; Nardi, F.; Facchinetti, S.; Ielmini, D. Spike-Timing Dependent

Plasticity in a Transistor-Selected Resistive Switching Memory. Nanotechnol. 2013, 24, 384012. 34.

Shi, J.; Ha, S. D.; Zhou, Y.; Schoofs, F.; Ramanathan, S. A Correlated Nickelate Synaptic

Transistor. Nat. Commun. 2013, 4, 2676. 35.

Du, C.; Ma, W.; Chang, T.; Sheridan, P.; Lu, W. D. Biorealistic Implementation of

Synaptic Functions with Oxide Memristors Through Internal Ionic Dynamics. Adv. Funct. Mater. 2015, 25, 4290-4299. 36.

Kim, S.; Du, C.; Sheridan, P.; Ma, W.; Choi, S.; Lu, W. D. Experimental Demonstration

of a Second-Order Memristor and its Ability to Biorealistically Implement Synaptic Plasticity. Nano Lett. 2015, 15, 2203-2211. 37.

Tan, Z. H.; Yang, R.; Terabe, K.; Yin, X. B.; Zhang, X. D.; Guo, X. Synaptic

Metaplasticity Realized in Oxide Memristive Devices. Adv. Mater. 2016, 28, 377-384. 38.

Wang, Z.; Joshi, S.; Savel’ev, S. E.; Jiang, H.; Midya, R.; Lin, P.; Hu, M.; Ge, N.;

Strachan, J. P.; Li, Z. Memristors with Diffusive Dynamics As Synaptic Emulators for Neuromorphic Computing. Nat. Mater. 2017, 16, 101-108. 39.

Zhu, X.; Du, C.; Jeong, Y.; Lu, W. D. Emulation of Synaptic Metaplasticity in

Memristors. Nanoscale 2017, 9, 45-51.

20 ACS Paragon Plus Environment

Page 20 of 30

Page 21 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

40.

La Barbera, S.; Vuillaume, D.; Alibart, F. Filamentary Switching: Synaptic Plasticity

Through Device Volatility. ACS Nano 2015, 9, 941-949. 41.

Alibart, F.; Pleutin, S.; Guérin, D.; Novembre, C.; Lenfant, S.; Lmimouni, K.; Gamrat,

C.; Vuillaume, D. An Organic Nanoparticle Transistor Behaving As a Biological Spiking Synapse. Adv. Funct. Mater. 2010, 20 (2), 330-337. 42.

Tian, H.; Guo, Q.; Xie, Y.; Zhao, H.; Li, C.; Cha, J. J.; Xia, F.; Wang, H. Anisotropic

Black Phosphorus Synaptic Device for Neuromorphic Applications. Adv. Mater. 2016, 28, 49914997. 43.

Arnold, A. J.; Razavieh, A.; Nasr, J. R.; Schulman, D. S.; Eichfeld, C. M.; Das, S.

Mimicking Neurotransmitter Release in Chemical Synapses via Hysteresis Engineering in MoS2 Transistors. ACS Nano 2017, 11, 3110-3118. 44.

Xu, W.; Min, S.-Y.; Hwang, H.; Lee, T.-W. Organic Core-Sheath Nanowire Artificial

Synapses with Femtojoule Energy Consumption. Sci. Adv. 2016, 2, e1501326. 45.

Kim, S.; Yoon, J.; Kim, H.-D.; Choi, S.-J. Carbon Nanotube Synaptic Transistor Network

for Pattern Recognition. ACS Appl. Mater. Interfaces 2015, 7, 25479-25486. 46.

Kim, S.; Choi, B.; Lim, M.; Yoon, J.; Lee, J.; Kim, H.-D.; Choi, S.-J. Pattern Recognition

Using Carbon Nanotube Synaptic Transistors with an Adjustable Weight Update Protocol. ACS Nano 2017, 11, 2814-2822. 47.

Kapadia, R.; Yu, Z.; Wang, H.-H. H.; Zheng, M.; Battaglia, C.; Hettick, M.; Kiriya, D.;

Takei, K.; Lobaccaro, P.; Beeman, J. W. A Direct Thin-Film Path Towards Low-Cost LargeArea III-V Photovoltaics. Sci. Rep. 2013, 3, 2275.

21 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

48.

Kapadia, R.; Yu, Z.; Hettick, M.; Xu, J.; Zheng, M. S.; Chen, C.-Y.; Balan, A. D.;

Chrzan, D. C.; Javey, A. Deterministic Nucleation of InP on Metal Foils with the Thin-Film Vapor–Liquid–Solid Growth Mode. Chem. Mater. 2014, 26, 1340-1344. 49.

Chen, K.; Kapadia, R.; Harker, A.; Desai, S.; Kang, J. S.; Chuang, S.; Tosun, M.; Sutter-

Fella, C. M.; Tsang, M.; Zeng, Y. Direct Growth of Single-Crystalline III–V Semiconductors on Amorphous Substrates. Nat. Commun. 2016, 7, 10502. 50.

Lin, Q.; Sarkar, D.; Lin, Y.; Yeung, M.; Blankemeier, L.; Hazra, J.; Wang, W.; Niu, S.;

Ravichandran, J.; Fan, Z.; Kapadia, R. Scalable Indium Phosphide Thin-Film Nanophotonics Platform for Photovoltaic and Photoelectrochemical Devices. ACS Nano 2017, 11, 5113-5119. 51.

Purves, D. E.; Augustine, G. J.; Fitzpatrick, D. E.; Katz, L. C. Neuroscience; Sinauer

Associates, 1997. 52.

Cao, C.; Andrews, J. B.; Kumar, A.; Franklin, A. D. Improving Contact Interfaces in

Fully Printed Carbon Nanotube Thin-Film Transistors. ACS Nano 2016, 10, 5221-5229. 53.

Chen, K.; Gao, W.; Emaminejad, S.; Kiriya, D.; Ota, H.; Nyein, H. Y. Y.; Takei, K.;

Javey, A. Printed Carbon Nanotube Electronics and Sensor Systems. Adv. Mater. 2016, 28, 4397-4414. 54.

Striedter, G. F. Neurobiology: A Functional Approach; Oxford University Press, 2015.

55.

Hebb, D. O. The Organization of Behavior: A Neuropsychological Approach; John Wiley

& Sons, 1949. 56.

Bienenstock, E. L.; Cooper, L. N.; Munro, P. W. Theory for the Development of Neuron

Selectivity: Orientation Specificity and Binocular Interaction in Visual Cortex. J. Neurosci. 1982, 2, 32-48.

22 ACS Paragon Plus Environment

Page 22 of 30

Page 23 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

57.

Markram, H.; Lübke, J.; Frotscher, M.; Sakmann, B. Regulation of Synaptic Efficacy by

Coincidence of Postsynaptic APs and EPSPs. Sci. 1997, 275, 213-215. 58.

Bi, G.-q.; Poo, M.-m. Synaptic Modifications in Cultured Hippocampal Neurons:

Dependence on Spike Timing, Synaptic Strength, and Postsynaptic Cell Type. J. Neurosci. 1998, 18, 10464-10472. 59.

Van Rossum, M. C. W.; Bi, G. Q.; Turrigiano, G. G. Stable Hebbian Learning from

Spike Timing-Dependent Plasticity. J. Neurosci. 2000, 20, 8812-8821.

23 ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Paragon Plus Environment

Page 24 of 30

Page 25 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(a)

ACS Nano

From pre‐synaptic  neuron

(b)

From  pre‐synaptic neuron

Vesicle Gate Source

Synapse

InP

SiO2

Ion channel To post‐synaptic  neuron Neurotransmitter Na+ Ion

ACS Paragon Plus Environment

Si

Al2O3 Drain

To  post‐ synaptic neuron

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Paragon Plus Environment

Page 26 of 30

(a)

(c)

(b)

(d)

40

STP Data STP Model LTP Data LTP Model

20 0 -20 -40

Potentiation -60

-4

-2

Elasticity

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

Synaptic Weight Change (%)

Page 27 of 30

Depression

0

Vgs (V)

ACS Paragon Plus Environment

2

4

ACS Nano

(a)

(b)

(c)

60

Synaptic Weight Change (%)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 28 of 30

40 20 0 -20 -40 -60

No priming Depressing priming Potentiating priming

-4

-2

0

Vgs (V)

ACS Paragon Plus Environment

2

4

Page 29 of 30 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Nano

(a)

(c)

(b)

(d)

ACS Paragon Plus Environment

ACS Nano 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(a)

t > 0

t < 0

(b)

ACS Paragon Plus Environment

Page 30 of 30