Optical synapse

Abstract

An optical synapse comprising a memristive device for non-volatile storage of a synapse weight dependent on a device resistance, and an optical modulator for volatile modulation of optical transmission in a waveguide. The memristive device and the optical modulator are connected in a control circuit operable to provide a programming signal to the memristive device in a write mode to program the synapse weight, and to provide an electrical signal dependent on the synapse weight to the optical modulator in a read mode, whereby the optical transmission is controlled in a volatile manner according to the programmed synapse weight.

Description

Technical Field

[0001] This invention generally relates to optical synapses, and more specifically to optical synapses for modulating light transmission in neuromorphic networks. Background Technology

[0002] Neuromorphic networks perform computational tasks in a manner inspired by the biological architecture of the nervous system. In the human brain, information is processed by a complex network of neurons interconnected by synapses. Biological neurons receive input signals from other neurons and generate output signals ("action signals" or "spikes") when the neuronal state (or "membrane potential") crosses a threshold level. These spikes are transmitted to other neurons via synapses, which alter their connection strength ("plasticity" or "synaptic weights") due to neuronal activity. Neuromorphic networks mimic biological function through a series of network layers comprising nodes corresponding to neurons, interconnected by synapses that relay weighted signals between connected nodes in adjacent layers according to stored synaptic weights.

[0003] Neuromorphic networks can be employed in numerous applications in science and technology, including computer vision, speech recognition, audio / video analytics, medical diagnostics, genetic analysis, and drug design. A network is configured for a given application by programming synaptic weights. These weights can be programmed via an iterative training process, during which the network is exposed to a training dataset specific to the application and the weights are progressively updated as the network “learns” from the training data. The trained network with fixed weights, defined by the training operations, can then be applied for inference based on new (unseen) data for the application in question.

[0004] In the electronic implementation of neuromorphic networks, information is encoded in electrical signals, and synapses modulate voltage / current levels based on stored synaptic weights. In the optical implementation, information is encoded in the optical power and / or phase of photonic signals transmitted through a waveguide. Synaptic function is implemented by optical synapses, which modulate optical transmission in the waveguide based on stored synaptic weights. Synaptic weights can be programmed via programming signals provided to the synapses in a programmed (or “write”) operating mode (e.g., during network training). The programmed weights can then be applied in a “read” operating mode to modulate transmission in the network, such as during inference.

[0005] There are numerous ways to realize synapses in the electronic domain, including circuits utilizing the memristor properties of nanodevices based on resistive memory cells. These devices utilize various physical mechanisms, such as the resistive properties of phase-change memory (PCM) materials and filamentary, electrochemical, or ferroelectric switching in oxide materials, to non-volatilely store synaptic weights that depend on the device's programmable resistance. Dense cross-bar arrays of such memristor devices provide massively parallel, highly area- and energy-efficient neural networks that can be efficiently implemented in integrated circuits. However, proposals for realizing synapses in the optical domain are scarce. Two examples of optical synapses are described in U.S. Patent Application Publication No. 2018 / 0267386A1, which discloses synapses with "direct" optical weights based on PCM materials (i.e., weights stored non-volatilely in the optical domain).

[0006] There is a great need for improved implementation methods of optical synapses. Summary of the Invention

[0007] One aspect of this disclosure provides an optical synapse comprising: a memristor device for non-volatilely storing synaptic weights dependent on device resistance; and an optical modulator for volatile modulation of optical transmission in a waveguide. The memristor device and the optical modulator are connected in a control circuit operable to provide a programming signal to the memristor device in a write mode to program the synaptic weights, and to provide an electrical signal dependent on the synaptic weights to the optical modulator in a read mode, thereby volatilely controlling optical transmission according to the programmed synaptic weights.

[0008] The optical synapse embodying this disclosure utilizes indirect optical weighting. Synaptic weights are stored separately in the electronic domain by programming the non-volatile resistance of the memristor device. Modulation is performed volatilely in the optical domain based on the weight-related electrical signal provided to the optical modulator. This coupling of the non-volatile memristor device and the volatile optical modulator provides an extremely efficient photonic synapse, in which the optical properties of the synapse can be tuned independently of the electrical characteristics of the non-volatile memory. The optical modulator can be engineered for optimized optical response, and the memristor device can be engineered for optimized weight storage characteristics, thereby enhancing synaptic performance. The synapse embodying this disclosure can be readily fabricated as an integrated photonic structure. The improved synaptic efficiency provides a high-performance, energy-efficient integrated structure for optical neural networks. Furthermore, the use of two coupled systems for weighted storage and optical modulation allows for dynamic adjustment of the independent systems to achieve various synaptic plasticity effects. These and other advantages will be explained in more detail below.

[0009] The control circuitry of the preferred embodiment includes a decoupling circuit for electrically decoupling the optical modulator from the memristor device in write mode. This decoupling protects the optical modulator from the programming signal applied to the memristor device. The memristor device is also advantageously connected in a voltage divider such that the electrical signal supplied to the optical modulator in read mode depends on the output signal of the voltage divider. This allows the voltage separation between the memristor device and the optical modulator to be tuned to an optimized operating range for the modulator in read mode.

[0010] Optical modulators can be absorptive or refractive. Particularly effective embodiments below use refractive optical modulators, some utilizing electro-optic effects and others thermo-optic effects. Specifically, the optical modulator may include an electro-optic modulator, and the electrical signal provided in readout mode may include a drive voltage for the electro-optic modulator. In a preferred embodiment, the electro-optic modulator utilizes a ferroelectric material to implement an optical phase shifter. The optical phase shifter includes a ferroelectric material optically coupled to a waveguide, disposed between a pair of electrodes for applying an electrical signal in readout mode. Here, the aforementioned decoupling circuitry is included in the control circuitry to suppress ferroelectric domain switching in the ferroelectric material during write mode.

[0011] In other embodiments, the optical modulator includes a thermo-optic modulator, and the electrical signal applied in the readout mode includes a drive current for a heater of the thermo-optic modulator. The thermo-optic modulator can be implemented using an optical phase shifter, wherein the heater includes a metal layer in thermal contact with the waveguide, the metal layer being disposed between a pair of contacts for applying the electrical signal.

[0012] Electro-optic modulators and thermo-optic modulators can be effectively integrated with memristor devices in a monolithic integrated circuit to provide integrated optical synapses embodying this disclosure.

[0013] The synapse embodying this disclosure can be adapted to achieve various synaptic plasticity effects by controlling one or both of the volatile and non-volatile subsystems. In particular, because the optical modulator is volatile, this subsystem can be controlled to simulate the short-term plasticity (STP) characteristics of biological synapses. Here, the control circuitry may include synaptic control logic responsive to a synaptic control signal. The synaptic control logic is operable to modify the electrical signal supplied to the optical modulator in readout mode according to the synaptic control signal to achieve the desired short-term plasticity effect. The non-volatile subsystem can be controlled independently to achieve various long-term plasticity (LTP) effects. In particular, the synaptic control logic is also operable to modify the programming signal supplied to the optical modulator in write mode to achieve the desired long-term plasticity (LTP) effect. The simulation of the synaptic control signal and STP / LTP characteristics is explained in more detail below.

[0014] Another aspect of this disclosure provides a neuromorphic network including a plurality of nodes, wherein pairs of nodes are interconnected via respective waveguides for transmitting optical signals between each pair of nodes, wherein each waveguide includes optical synapses as described above.

[0015] Embodiments of this disclosure will now be described in more detail by way of illustrative and non-limiting examples with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a schematic circuit diagram illustrating the optical synapse of this disclosure;

[0017] Figure 2 It shows Figure 1 An exemplary implementation of a memristor device in a circuit;

[0018] Figure 3 It shows Figure 1 An exemplary implementation of an optical modulator in a circuit;

[0019] Figure 4A and 4B It shows Figure 3 Structure and operation of an embodiment of an optical phase shifter in a modulator;

[0020] Figure 5 It shows the use of optical synapses and Figure 4A A more detailed circuit diagram of the phase shifter;

[0021] Figure 6A and 6B The structure and operation of another embodiment of an optical phase shifter are shown;

[0022] Figure 7 The structure of the integrated optical synapse embodying this disclosure is shown;

[0023] Figure 8A and 8B Another integrated synaptic structure embodying this disclosure is shown;

[0024] Figure 9A and 9B The illustration shows a plasmonic structure embodying a plasmonic modulator in a synapse according to the present disclosure.

[0025] Figure 10 It is a schematic representation of synaptic connections in a neural network;

[0026] Figure 11A and 11B It shows Figure 10 The timing-dependent plasticity effect of spikes in the system; and

[0027] Figure 12The realization of tunable synaptic dynamics in optical synapses of neuromorphic networks is shown. Detailed Implementation

[0028] Figure 1 The schematic diagram illustrates the basic structure of the optical synapse embodying this disclosure. Synapse 1 includes a memristor device 2, represented as a variable resistor R1, and an optical modulator, schematically indicated by 3. The memristor device 2 is adapted to non-volatilely store a synaptic weight w, which depends on the device's resistance R1. Specifically, the synaptic weight w corresponds to the conductance of device 2 and therefore varies inversely with the resistance R1. The optical modulator 3 is adapted to volatilely modulate optical transmission in a waveguide 4. The storage device 2 and the optical modulator 3 are connected in a control circuit that includes a drive circuit 5 connected between device 2 and modulator 3, and a power supply circuit for providing signals to device 2 during operation. Specifically, in a programming or "write" operation mode, a programming signal (here, a pulse voltage signal V) can be used. W A voltage is applied to the memristor device to program the resistor R1, and thus the synaptic weight w. In "read" operation mode, the read voltage V... R An output signal, dependent on resistance R1 (and therefore on synaptic weight w), is applied to memristor device 2 and then to drive circuit 5. Drive circuit 5 then provides an electrical drive signal to optical modulator 3 based on the weight-dependent output from device 2, which in turn modulates the transmission in waveguide 4. Thus, optical transmission in the waveguide is controlled in a variable manner according to the programmed synaptic weight w.

[0029] Figure 1 The synapse provides a non-volatile optical synapse, where the synaptic weights w are stored separately in the electronic domain, while optical correlation modulation is performed volatilely in the optical domain. The drive signal provided to modulator 3 can be a voltage signal V. drive or current signal I drive The non-volatile optical state W of a synapse can be expressed as W = f(V drive ) or W = g(I drive ), where: f and g are functions of the optical device; V drive and I drive It is a function of the non-volatile electrical state of memristor device 2, and the electrical and optical subsystems of the synapse can be engineered independently for optimized characteristics. The optical contrast in modulator 3 can be optimized independently of the non-volatile electrical subsystem, and device 2 can be adapted to optimized weighted-memory characteristics independently of the volatile optical subsystem.

[0030] The programmable resistor of memristor device 2 can utilize various physical mechanisms known in the art, including filamentary switching, electrochemical switching, ferroelectric switching, and resistive properties of PCM materials. Device 2 may include one or more resistive memory elements or cells, such as PCM cells and resistive RAM (RRAM) cells, including oxide / metal oxide RRAM cells, conductive bridge RRAM cells, and carbon RRAM cells. The basic operating principle of these cells relies on a reversible, non-volatile change in the resistive properties of one or more material layers disposed between two electrodes. The total resistance can be changed by applying a programming pulse via the electrodes, and thus the conductivity of the cell can be altered. As an example, Figure 2 The formation of conductive filaments in a filamentary RRAM cell is illustrated. Cell 10 includes a high-resistivity oxide layer 11, such as HfO2, between a pair of electrodes 12. Applying programming pulses through the electrodes 12 causes conductive filaments 13 to gradually form in the high-resistivity layer 11; the left-hand diagram shows the high-resistivity cell state. With successive programming pulses, the length of the filaments increases, thereby reducing the effective resistance of the oxide layer between the electrodes. The right-hand diagram shows the low-resistance state, where the filaments have bridged the oxide layer. The cell's resistance can be "read" by applying a low read voltage across the electrodes, and the resulting current through the cell depends on the programmed resistance.

[0031] Many physical mechanisms can be employed in resistive memory elements, including the migration of metal ions, oxygen ions, metal precipitates, or graphene clusters through various material layers, as well as Joule heating of PCM materials. One or more cells can be arranged in a variety of known circuit configurations to provide desired programmable resistive characteristics. Therefore, a memristor device 2 can typically include one or more resistive memory cells of any desired type, and the device properties can be readily tuned to give desired weighted memory characteristics.

[0032] The optical modulator 3 can be an absorption or refractive modulator, thereby modulating the light transmission in the waveguide 4 by the volatile change of the waveguide's absorption coefficient or refractive index (i.e., the real part of the complex refractive index). A preferred embodiment utilizes a refractive optical modulator. These can be efficiently implemented using an optical phase shifter, which is driven by an electrically driven signal V. drive or I drive The phase of light in a portion of a waveguide can be altered by controlling changes in the refractive index. For example, the resulting phase modulation can be converted into amplitude modulation using an interferometer. Figure 3An embodiment of such an optical modulator 20 based on a Mach-Zehnder interferometer (MZI) is shown. The waveguide 21 is divided into two arms 22a and 22b, and an optical phase shifter, schematically indicated as 23, is provided in one of the arms 22b. The phase shifter 23 includes a pair of electrodes 24 that respond to a drive signal V via a mechanism explained below. drive or I drive This involves changing the effective refractive index of at least one material in waveguide arm 22b. The change in refractive index affects the propagation coefficient n of the optical mode in the waveguide. b The change. For lossless balanced MZI (with waveguide arms 22a and 22b of length l, and absorption coefficient α = 0), the effective propagation coefficient n of arms 22a and 22b is... a n b The difference affects the intensity of light transmitted through waveguide 21 according to the following formula:

[0033] I out =I in [1+cos(2πl(n a -n b ) / λ)] / 2

[0034] Where I in and I out These are the light input and output intensities, respectively, and λ is the wavelength of the transmitted light.

[0035] The refractive index modulation in phase shifter 23 can utilize electro-optic or thermo-optic effects. Figure 4A The structure of phase shifter 23 in an electro-optic modulator embodying the synapse of this disclosure is shown. The figure illustrates... Figure 3 The cross-sectional view of the AA line shows that the phase shifter comprises a ferroelectric material, here a barium titanate (BaTiO3) (“BTO”) layer 25, which is optically coupled to an arm 22b of a silicon waveguide formed on the BTO layer 25. This structure realizes a hybrid BTO / Si waveguide in arm 22b of the modulator 20. The BTO layer 25 is disposed between electrodes 24, which apply a driving voltage V to the hybrid waveguide structure. drive The hybrid waveguide is embedded in an oxide material (here, SiO2) formed on a silicon substrate 26.

[0036] For ferroelectric materials such as BTO, the refractive index n changes according to the following formula when an electric field E is applied:

[0037] n(E) = n - rn 3 E / 2-ξn 3 E 2 / 2

[0038] Among them, item rn 3E / 2 represents the electro-optic (Pockels) effect with the Pockels coefficient r, and the term ξn 3 E 2 / 2 represents the Kerr effect with a Kerr constant ξ. The Pockels effect is dominant in such materials, and is further influenced by the presence of a static bias voltage V. drive The change in refractive index under the given condition can be expressed as:

[0039]

[0040] in, This is caused by the Pockels effect. For example, for a static field and a waveguide orientation at 45° relative to the principal crystal axis of BTO, Figure 4B It is shown as V in a 2mm long non-destructive MZI 20 drive The output intensity variation of the function, which uses an electrode spacing of 4.75 μm. Figure 4A Phase shifter.

[0041] Figure 5 A more detailed circuit configuration of a preferred embodiment of the optical synapse is shown. The synapse 30 includes a memristor device 31 (also represented by a variable resistor R1) and an optical modulator employing the phase shifter 23 described above. The memristor device 31 is connected to the resistor R2 in a voltage divider configuration. The output V of the voltage divider... O =V R R2 / (R1+R2) is provided to the subsequent drive circuit for use in read mode in response to DC bias V. R Generate modulator drive signal V drive The control circuitry of this embodiment also includes a decoupling circuit to electrically decouple the optical modulator from the memristor device 31 in write mode. The decoupling circuit includes a capacitor C and an operational amplifier voltage follower 32. The capacitor C is connected in series with the memristor device 31 to provide voltage pulse train V applied in write mode for programming the memristor device 31. W A short circuit is provided. Operational amplifier 32 (gain A1 = 1) provides a voltage buffer and additional decoupling for the modulator drive circuit. The second operational amplifier 33 provides voltage amplification with a gain A2 = 1 + R4 / R3. Here, resistor R5 represents the impedance load presented by optical phase shifter 23.

[0042] As an example, the memristor device 31 can be implemented using RRAM memory cells with a resistance range of 10kΩ to 100kΩ. Programming the device resistance R1 within this range can be achieved via a pulse train V in write mode. W This is achieved using a pulse generator that generates different numbers of programming pulses (e.g., 10ns pulses with amplitudes ranging from 2V to 4V). The voltage V is read.R It is set to a sufficiently low level, such as 0.2V, to avoid altering the programming state of device 31 in read mode. Using this configuration, V in read mode... O = 0.07V to 0.17V, and V drive =A1.A2.V O =1.82V to 4.42V, where R3 = 1kΩ and R4 = 25kΩ.

[0043] The voltage divider in the synapse control circuit allows the memristor device output to be tuned to the desired operating range of the optical phase shifter for readout operations. The decoupling circuitry in synapse 30 suppresses the transmission of programming pulses to the optical phase shifter 23, protecting the modulator from the potentially adverse effects of these pulses. In particular, a potential problem with using ferroelectric Pockels material in phase shifter 23 is the high-voltage pulse V... W Polarization reversal can occur in the ferroelectric domains of these materials, resulting in a non-volatile change in refractive index. To suppress this ferroelectric domain switching, the applied bias voltage must remain positive or negative, and the programming pulse must be decoupled from the phase shifter electrodes. Figure 5 The control circuit thus provides an implementation scheme for a volatile optical modulator using ferroelectric sPockels material.

[0044] Figure 6A An alternative phase shifter for the thermo-optic modulator in an embodiment of this disclosure is shown. The figure shows a schematic cross-section of the MZI structure corresponding to line AA in the inset view. The phase shifter 40 includes a metal layer disposed between electrodes 42 as a heater 41, which is driven by a drive current I applied via contacts. drive Heating is achieved through Joule heating. The heater 41 can be formed of, for example, tungsten, aluminum, copper, titanium, or carbon. The heater 41 is in thermal contact with the waveguide 43 (here, silicon), where the application of heat causes a change in refractive index through a thermo-optical effect. This structure is embedded in an oxide material (here, SiO2) formed on a silicon substrate 44.

[0045] Apply drive current I drive The change in waveguide temperature T ΔT caused by heater 41 leads to a corresponding change in the refractive index n of silicon. The change in waveguide temperature ΔT varies with the change in heater temperature ΔT. H Increase, where ΔT H ∝(I driv e) 2 R H , where R H This is the resistance of the heater (which can be temperature-dependent). The resulting refractive index of the silicon waveguide can be approximated as:

[0046]

[0047] yes This is the thermo-optic coefficient. As shown in the figure, for a silicon waveguide at room temperature with a wavelength λ = 1.55 μm, Figure 6B This demonstrates the non-destructive use of MZI 20 (with) for a length of 2mm. Figure 6A The change in output intensity of the phase shifter as a function of waveguide temperature.

[0048] use Figure 6A The optical synapse of the phase shifter can be used Figure 5 This is achieved through a control circuit, where operational amplifier 33 is selected to drive the current signal I within a range of several mA. drive In this embodiment, the decoupling circuit prevents the phase shifter structure from heating in response to programming pulses in write mode.

[0049] Memristor devices and optical modulators can be easily integrated with synaptic control circuits into a single integrated circuit, such as... Figure 7 This is shown schematically. It illustrates an integrated optical synapse 50, where a metal interconnect (e.g., interconnect 51) bridges the electrical and photonic circuitry of the synapse. A memristor device (here, an RRAM memory cell 52) is formed in one of a pair of electrical contacts 53 of the control circuitry 54 in this embodiment. The optical modulator here corresponds to... Figure 4A The structure shown.

[0050] Figure 8A An alternative structure for an integrated synapse is shown, in which a memristor device 60 is formed directly on a metal stack of one of the electrodes 61 of an electro-optic phase shifter, and a silicon waveguide 62. In this embodiment, the memristor can be fabricated during back-end processing of the integrated circuit structure. Figure 8B Another configuration is shown, in which the memristor device 70 is deposited directly on the BTO layer of the phase shifter. Similar structures, in particular... Figure 7 and 8A The structure can be combined with Figure 6A It is used together with a thermo-optical phase shifter.

[0051] The aforementioned optical synapses can be fabricated using known materials processing techniques. These synapses can be fabricated as integrated photonic structures for neuromorphic networks, where the optical synapse array enables the vector-matrix computations required to propagate weighted signals across successive layers of the network. Control signals for programming and read-out mode operations can be generated by global signal generators within these structures. This integrated synapse array provides an extremely high-speed, low-power implementation of neuromorphic networks.

[0052] While exemplary components for efficiently integrating synaptic structures have been described above, various other components can be used for the synapses embodying this disclosure. For example, electro-optic phase shifters can employ other ferroelectric materials, such as lithium niobate or PZT (lead zirconate titanate), and can utilize the Pockels and / or Kerr effects for refractive index modulation. Electro-optic phase shifters can also utilize other mechanisms, such as PIn (p-type, intrinsic, n-type) diode structures, for refractive index variation. Optical modulators can also be implemented in other ways, such as using directional couplers and optical resonators such as ring resonators. Modulators may include additional material layers and can also modify other optical properties, such as optical absorption, as will be apparent to those skilled in the art. In all implementations, the characteristics of the non-volatile electrical subsystem and the volatile optical subsystem can be independently tuned for desired synaptic performance.

[0053] While the above embodiments use photonic modulators, plasma modulators can be used in other embodiments. This structure can operate at very low voltages and can be very small. Figure 9A and 9B Schematic cross-sections of exemplary BTO-based plasmonic structures are shown, which can be used for phase modulation due to the Pockels effect in the BTO layer. These figures illustrate the BTO-based plasmonic structures in cross-sections transverse to the waveguide direction. Both structures include BTO layers disposed for applying a voltage signal V. drive Between a pair of electrodes, Figure 9A The horizontal slot configuration is shown. Figure 9B The vertical slot configuration is shown. The plasmonic structure can be embedded in a photonic circuit and optically coupled to a photonic waveguide using a standard photonic-to-plasmonic coupler.

[0054] The desired synaptic plasticity effect can also be achieved by leveraging the independence of the electrical and optical subsystems, using a reference... Figure 10 , 11A And 11B explains these effects in more detail. Figure 10 This is a simplified diagram of synaptic connections between nodes in a neural network. Presynaptic node n i Through synapse ij Connected to postsynaptic node n j Presynaptic node n i Receive signals from other network nodes and generate signals to provide to the synapses. ij The action signal of the synapse. ij The weighted signal is relayed to the postsynaptic node n. j It then generates action signals based on weighted input signals from connected nodes in the network. (The input signals come from the postsynaptic node n.) jThe action signal is sent to other nodes in the network via synaptic connections. Presynaptic node n i and postsynaptic node n j They can generate action signals at arbitrary timings based on their respective input signals, and synapses s ij Synaptic efficacy, or weight, can be enhanced or diminished through network activity. This "plasticity" of synapses is crucial for memory and other brain functions. For example, it can be determined based on the activity of the presynaptic node n. i and / or postsynaptic node n j Action signals are used to modify synaptic weights, as shown in the figure from node n j to synapse ij The feedback connection is shown. Changes in synaptic weights can depend on the timing of these action signals, such as the relative timing of action signals at the presynaptic and postsynaptic nodes. Changes in synaptic weights can be instantaneous, known as the "short-term plasticity" (STP) effect, or long-term, known as the "long-term plasticity" (LTP) effect. Figure 11A and 11B Examples of the LTP and STP effects are shown respectively. Figure 11A Indicates how the change Δw in the synaptic weight w depends on the input from node n j and n i The relative timing of the action signal Δt = (t j -t i The magnitude of ΔW changes inversely to ΔT, and the sign of Δw depends on which action signal occurs first. This effect is known as spike timing-dependent plasticity (STDP). Figure 11B An example of the STP effect, known as synaptic fatigue, is shown, where the synaptic weight w decreases incrementally at times t1 and t2 upon arrival of the presynaptic action signal, gradually returning to its programmed long-term value over time after the last presynaptic signal. Therefore, synaptic efficacy dynamically decreases in response to more frequent presynaptic signals.

[0055] Synaptic plasticity effects can be realized in networks employing synapses, which embodies the synaptic properties of this disclosure. Figure 12 The diagram is schematically shown. This illustrates an optical synapse 80 integrated in waveguide 81 for transmitting optical input signals from the presynaptic node n. i Transmission to the postsynaptic n of the neuromorphic network j Synapse 80 can generally be implemented as described above. However, the control circuitry for synapse 80 includes the synapse control logic indicated at 82 in the figure. First photodetector D pre Optical coupling is applied to waveguide 81 to detect signals from presynaptic node n. i The input signal. Second photodetector D post Optically coupled to waveguide 83, which transmits from postsynaptic node nj The optical output signal. Detector D pre and D post This can be implemented, for example, by a photodiode that relays an electrical signal to control logic 82 in response to an optical signal in waveguides 81 and 83. Control logic 82 responds to an optical signal from detector D. pre D post These synaptic control signals alter the operation of the control circuitry in synapse 80. Specifically, control logic 82 can be adapted to change the electrical drive signal supplied to the optical modulator in readout mode to achieve a desired short-term plasticity effect. drive or I drive Changes can be made, for example, by adjusting the read bias V. R Alternatively, this can be achieved by implementing resistors R2, R3, or R4 as variable resistors and dynamically adjusting their resistance values. This allows control of the volatile optical modulator to achieve the desired STP effect. This effect typically depends on the detector D. pre D post One or more synaptic control signals. For example, Figure 11B The STP effect can be determined from the detector D. pre This is achieved through synaptic control signals.

[0056] The synchronous control logic 82 can also be adapted to change the programming signal provided to the optical modulator in write mode based on the synaptic control signal from the detector. Here, the control logic 82 can change the programming signal to achieve the desired long-term plasticity effect. This can be achieved by controlling the pulse V provided to the memristor device in write mode. W The quantity, amplitude, and duration, or a combination thereof, of the programmed signal can be altered to change the long-term synaptic weight w. Similarly, this LTP effect typically depends on the signal from detector D. pre D post One or both of the synaptic control signals. For example, this can be achieved based on the relative timing of the synaptic control signals from two detectors. Figure 11A The LTP effect.

[0057] Synaptic control logic 82 can typically be implemented in hardware or software or a combination thereof, and a suitable implementation will be apparent to those skilled in the art. By dynamically adjusting the operation in this way, the non-volatile weighting and volatile optical modulation in synapse 80 can be independently tuned to achieve desired long-term and short-term synaptic dynamics.

[0058] Various embodiments of this disclosure have been described for illustrative purposes, but are not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope of the described embodiments. The terminology used herein has been chosen to best explain the principles of the embodiments, their practical application, or improvements to existing technologies in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An optical synaptic device, comprising: Memristor devices for non-volatile storage based on the synaptic weights of the device's resistance; as well as Optical modulators are used for volatile modulation of optical transmission in waveguides. as well as The memristor and the optical modulator are connected in a control circuit that is operable to provide programming signals to the memristor in a write mode to program synaptic weights, and to provide electrical signals dependent on the synaptic weights to the optical modulator in a read mode, thereby volatilely controlling optical transmission according to the programmed synaptic weights.

2. The optical synaptic device as claimed in claim 1, wherein, The control circuit includes a decoupling circuit for electrically decoupling the optical modulator from the memristor device in write mode.

3. The optical synaptic device as claimed in claim 1, wherein, The memristor device is connected in a voltage divider, and the electrical signal provided to the optical modulator in readout mode depends on the output signal of the voltage divider.

4. The optical synaptic device as claimed in claim 1, wherein, The memristor and optical modulator are integrated into a single integrated circuit.

5. The optical synaptic device as claimed in claim 1, wherein, The optical modulator includes an electro-optic modulator, and the electrical signal includes the driving voltage of the electro-optic modulator.

6. The optical synaptic device as claimed in claim 5, wherein, The electro-optic modulator includes an optical phase shifter comprising a ferroelectric material optically coupled to the waveguide and disposed between a pair of electrodes for applying the electrical signal, wherein the control circuitry includes a decoupling circuit for electrically decoupling the optical modulator from the memristor device in the write mode to suppress ferroelectric domain switching in the ferroelectric material.

7. The optical synaptic device as claimed in claim 6, wherein, The ferroelectric material includes one of barium titanate, lithium niobate, and lead zirconate titanate.

8. The optical synaptic device as claimed in claim 6, wherein, The memristor and the optical modulator are integrated in a monolithic integrated circuit, and the memristor is formed on one of the electrodes.

9. The optical synaptic device of claim 6, wherein the memristor and the optical modulator are integrated in a monolithic integrated circuit, and the memristor is formed on a ferroelectric layer.

10. The optical synaptic device as claimed in claim 1, wherein, The optical modulator includes a thermo-optic modulator, and the electrical signal includes a drive current for a heater of the thermo-optic modulator.

11. The optical synaptic device of claim 10, wherein, The thermo-optic modulator includes an optical phase shifter, and the heater includes a metal layer that is in thermal contact with the waveguide and is disposed between a pair of contacts for applying the electrical signal.

12. The optical synaptic device as claimed in claim 11, wherein, The memristor and the optical modulator are integrated in a monolithic integrated circuit, and the memristor is formed on one of the contacts.

13. The optical synaptic device as claimed in claim 1, wherein, The optical modulator includes a Mach-Zehnder interferometer.

14. The optical synaptic device as claimed in claim 1, wherein, The optical modulator includes a plasmonic modulator.

15. The optical synaptic device as claimed in claim 1, wherein, The control circuit includes synaptic control logic that is responsive to a synaptic control signal. The synaptic control logic is operable to change the electrical signal provided to the optical modulator in the readout mode according to the synaptic control signal to achieve a desired short-term plasticity effect.

16. The optical synaptic device of claim 15, wherein: The waveguide is arranged to transmit an optical input signal from a presynaptic node to a postsynaptic node of a neuromorphic network, wherein the postsynaptic node transmits an optical output signal based on the optical input signal received therefrom; and The synaptic control signal depends on at least one selected from the group consisting of: the light input signal and the light output signal.

17. The optical synaptic device of claim 15, wherein, The synaptic control logic is further operable to modify the programming signal provided to the optical modulator in write mode according to the synaptic control signal to achieve the desired long-term plasticity effect.

18. The optical synaptic device of claim 17, wherein: The waveguide is arranged to transmit an optical input signal from a presynaptic node to a postsynaptic node of a neuromorphic network, wherein the postsynaptic node transmits an optical output signal based on the optical input signal received therefrom; and The synaptic control signal depends on at least one selected from the group consisting of: the light input signal and the light output signal.

19. A neuromorphic network system comprising multiple nodes, wherein, Node pairs are interconnected via corresponding waveguides for transmitting optical signals between each pair of nodes, wherein each waveguide includes an optical synapse comprising: Memristor devices for non-volatile storage based on synaptic weights of the device's resistance; and Optical modulators for volatile modulation of optical transmission in waveguides; and The memristor and the optical modulator are connected in a control circuit that is operable to provide a programming signal to the memristor in a write mode to program synaptic weights, and to provide an electrical signal to the optical modulator in a read mode that depends on the synaptic weights, thereby controlling optical transmission in a volatile manner according to the programmed synaptic weights.

20. The neuromorphic network system of claim 19, wherein, The control circuitry for each synapse includes a decoupling circuit for electrically decoupling the optical modulator of the synapse from the memristor device in the write mode.

21. The neuromorphic network system of claim 19, wherein, The memristor device of each synapse is connected in a voltage divider, and the electrical signal provided to the optical modulator of the synapse in readout mode depends on the output signal of the voltage divider.

22. The neuromorphic network system of claim 19, wherein, The optical modulator of each synapse includes an electro-optic modulator, and the electrical signal includes a driving voltage for the electro-optic modulator.

23. The neuromorphic network system of claim 19, wherein, The optical modulator of each synapse includes a thermo-optic modulator, and the electrical signal includes a drive current for a heater of the thermo-optic modulator.

24. The neuromorphic network system of claim 19, wherein, The control circuitry for each synapse includes synapse control logic responsive to a synapse control signal, operable to alter the electrical signal supplied to the optical modulator in readout mode according to the synapse control signal to achieve a desired short-term plasticity effect.

25. The neuromorphic network system of claim 24, wherein, The synaptic control logic is also operable to modify the programming signal of the optical modulator provided to the synapse in the write mode according to the synaptic control signal, in order to achieve the desired long-term plasticity effect.

Images