Method for realizing pulse activation and synaptic weight simultaneously by monolithic integration of DFB-SA

Abstract

The present invention discloses a method for simultaneously realizing pulse activation and synaptic weight in a monolithic integrated DFB-SA, including: under the condition of applying a reverse voltage bias to the saturable absorption region of the DFB-SA, by adjusting the external optical injection of the DFB-SA and performing time-varying adjustment on the gain current of the DFB-SA, to simultaneously realize non-linear pulse activation and synaptic linear weighting on the monolithic integrated DFB-SA; wherein, the gain current includes a static bias current and a time-varying current; the excitatory threshold of the DFB-SA decreases with the increase of the gain current, the refractory period of the DFB-SA decreases with the increase of the gain current, and the time-domain cumulative characteristic of the DFB-SA is controlled by the gain current; the time-varying current is used to form the synaptic weight. Based on the method of the present invention, non-linear pulse activation of the SNN can be directly realized in the optical domain, and the use of linear weighting devices in the SNN network can be reduced.

Description

Technical Field

[0001] The present invention belongs to the field of optical computing, and particularly relates to a method for realizing pulse activation and synaptic weight simultaneously by monolithic integration of DFB (Distributed Feedback)-SA (Saturable Absorber). Background Art

[0002] The rapid development of deep neural networks has made breakthroughs in application fields such as computer vision and natural language processing. These applications generate a large amount of data that needs to be processed, which requires high-speed, high-throughput, and low-latency advanced processors. However, Moore's Law is approaching saturation, and traditional digital computers face the von Neumann bottleneck due to the separation of memory and processor units. Neuromorphic computing, which simulates the structure and mechanism of the brain, has become a competitive computing paradigm in the post-Moore era. Although great progress has been made in electronic neuromorphic computing systems based on the well-known Complementary Metal Oxide Semiconductor (CMOS) technology, they are still limited in terms of processing speed and energy efficiency.

[0003] Photonic neuromorphic computing has inherent advantages such as ultra-high speed, large bandwidth, and massive parallelism. It is a promising alternative method and has become a research hotspot in the field of neural computing. In addition to advanced neuromorphic computing hardware, new neural network models have also attracted extensive attention. Spiking neural networks (SNNs) propose to use discrete spike signals to achieve low-power computing on neuromorphic hardware. Therefore, implementing SNNs in photonic neuromorphic hardware is expected to be widely applied in scenarios with high latency requirements and power constraints, such as autonomous driving and edge computing.

[0004] In biological neural networks, neurons are connected by plastic synapses. As Figure 1 shown, dendrites receive external stimuli or presynaptic spikes, the soma converts external stimuli into response spikes or performs nonlinear pulse activation, the neural synapse adjusts the connection strength, and then transmits the weighted spikes to the dendrites of neighboring neurons. In an SNN, spiking neurons are the basic functional units of the SNN. Spiking neurons are connected by plastic synapses. Spiking neurons achieve nonlinear pulse activation, while synapses achieve linear weighting.

[0005] For linear computing, in the reported photonic neural networks, the coherent synaptic network based on Mach-Zehnder interferometer (MZI) and the incoherent synaptic network based on Microring Resonators (MRR) are two mainstream methods for realizing linear computing. This is because MZI and MRR can be fabricated using a silicon photonics platform compatible with CMOS. However, MZI and MRR have serious loss problems and are not suitable for realizing nonlinear computing. The nonlinear computing of the reported photonic neural networks is still mainly realized electronically, relying on high-speed photodetectors and analog-to-digital converters to convert the optical linear computing results back to the digital domain.

[0006] For nonlinear computing, the leaky integrated-and-fire (LIF) model has become one of the most popular simplified spiking neuron models for constructing SNNs due to its simplicity. The optical implementation of LIF neurons has received a lot of attention, such as vertical-cavity surface-emitting lasers (VCSELs) based on polarization switches, VCSELs with saturable absorbers (VCSEL-SA), graphene excitatory lasers, hybrid-integrated phase-change materials (PCMs) and MRRs, microcavity lasers, integrated distributed feedback (DFB) lasers and photodetector (PD) neurons, passive microring resonators, and Fabry-Perot cavity lasers with integrated built-in saturable absorbers (FP-SA).

[0007] However, the separate design and fabrication of photonic spiking neurons and photonic synapses have hindered the further scalability of photonic SNNs. In the structure composed of photonic synapses and electronic spiking neurons, the frequent optoelectronic (OE) and electro-optic (EO) conversions, as well as analog-to-digital (AD) and digital-to-analog (DA) conversions, have hindered the optical implementation of multi-layer or deep neural networks. Summary of the Invention

[0008] To solve the above problems existing in the prior art, the present invention provides a method for simultaneously realizing pulsed activation and synaptic weights with a monolithic integrated DFB-SA.

[0009] The technical problems to be solved by the present invention are realized through the following technical solutions:

[0010] A method for simultaneously realizing pulsed activation and synaptic weights with a monolithic integrated DFB-SA, comprising:

[0011] Under the condition of applying a reverse voltage bias to the saturable absorption region of the DFB-SA, by adjusting the external optical injection of the DFB-SA and performing time-varying adjustment on the gain current of the DFB-SA, to simultaneously realize nonlinear pulsed activation and synaptic linear weighting on the monolithic integrated DFB-SA;

[0012] Among them, the gain current includes a static bias current and a time-varying current; the excitability threshold of the DFB-SA decreases with the increase of the gain current, the refractory period of the DFB-SA decreases with the increase of the gain current, and the time-domain cumulative characteristic of the DFB-SA is controlled by the gain current; the time-varying current is used to form synaptic weights.

[0013] Preferably, a bias device is used to combine the static bias current and the time-varying current to form the gain current.

[0014] Preferably, the wavelength of the external optical injection is determined according to the spectrum of the free-running DFB-SA.

[0015] Preferably, the external optical injection is an external stimulus pulse sequence with a pulse repetition frequency of 0.5 GHz.

[0016] Preferably, under the condition of applying a reverse voltage bias to the saturation absorption region of the DFB-SA, by adjusting the external optical injection of the DFB-SA and performing time-varying adjustment on the gain current of the DFB-SA, nonlinear pulse activation and synaptic linear weighting are simultaneously achieved on a monolithic integrated DFB-SA, including:

[0017] Using a DC power supply to generate the static bias current and the reverse voltage bias;

[0018] Using a tunable laser to generate an optical carrier;

[0019] Using a first arbitrary waveform generator to generate an electrical stimulation signal;

[0020] Inputting the optical carrier into a Mach-Zehnder modulator, and using the Mach-Zehnder modulator to perform electro-optic conversion on the electrical stimulation signal to form an optical signal;

[0021] Using a circulator to inject the optical signal into the DFB-SA;

[0022] Using a second arbitrary waveform generator to generate a time-varying current;

[0023] Using a bias device to combine the static bias current and the time-varying current to form the gain current of the DFB-SA;

[0024] By adjusting the first arbitrary waveform generator, the external optical injection of the DFB-SA is adjusted, by adjusting the DC power supply, the static bias current is adjusted, and by adjusting the second arbitrary waveform generator, the time-varying current is adjusted, so as to simultaneously achieve nonlinear pulse activation and synaptic linear weighting on a monolithic integrated DFB-SA.

[0025] Preferably, a first polarization controller is used between the tunable laser and the Mach-Zehnder modulator to adjust the polarization state of the optical path between the tunable laser and the Mach-Zehnder modulator; a second polarization controller is used between the Mach-Zehnder modulator and the circulator to adjust the optical path between the Mach-Zehnder modulator and the circulator.

[0026] Preferably, the method for simultaneously realizing pulse activation and synaptic weight on a monolithic integrated DFB-SA provided by the present invention is applied to the construction of an optical pulse neural network device; the optical pulse neural network device is constructed based on multiple photon neuro-synaptic units, and the photon neuro-synaptic unit is realized based on the DFB-SA by using any one of the methods for simultaneously realizing pulse activation and synaptic weight on a monolithic integrated DFB-SA.

[0027] Preferably, in the optical pulse neural network device, both the synaptic weight and the excitatory threshold realized by the DFB-SA are adjustable parameters.

[0028] Preferably, the response speed of the photon neuro-synaptic unit is at the sub-nanosecond level.

[0029] The method for simultaneously realizing pulse activation and synaptic weight on a monolithic integrated DFB-SA provided by the present invention, under the condition of applying a reverse voltage bias to the saturated absorption region of the DFB-SA, realizes non-linear pulse activation and synaptic linear weighting on the monolithic integrated DFB-SA by adjusting the external optical injection of the DFB-SA and performing time-varying adjustment on the gain current of the DFB-SA. Among them, in terms of non-linear calculation, the excitatory threshold, refractory period, and time-domain cumulative characteristics of the photon pulse neuron can be realized by adjusting the gain current of the monolithic integrated DFB-SA; in terms of linear calculation, the synaptic weight can be formed by adjusting the time-varying current in the gain current.

[0030] The following will further describe the present invention in detail with reference to the accompanying drawings. Description of the Drawings

[0031] Figure 1 is a schematic diagram of a biological neural network;

[0032] Figure 2 is a micrograph of a monolithic integrated DFB-SA;

[0033] Figure 3 is a schematic structural diagram of a monolithic integrated DFB-SA;

[0034] Figure 4 is a schematic diagram of the external stimulus processing process of a monolithic integrated DFB-SA;

[0035] Figure 5It is the spectrum of the monolithic integrated DFB-SA when operating freely;

[0036] Figure 6 It is a schematic structural diagram of an experimental platform used in the embodiments of the present invention;

[0037] Figure 7 The pulse neuron characteristics of the monolithic integrated DFB-SA at different gain currents are shown, which are: (a) excitatory threshold, (b) refractory period, (c) time-domain cumulative characteristics;

[0038] Figure 8 The three-part experimental data of the continuously adjustable weighting of the response spike pulses of the monolithic integrated DFB-SA at different static gain currents are shown, which are: (a) the weighted output of a single response spike pulse at different gain currents, (b) the weighted response spike pulse trains at three representative gain currents, (c) the response spike pulse amplitude as a function of the gain current.

[0039] Figure 9 The three-part experimental data of simultaneously implementing pulse activation and synaptic weight in the monolithic integrated DFB-SA are shown, which are: (a) the response spike pulse output under time-varying external optical stimuli, (b) the response spike pulse output under discrete time-varying gain currents, (c) the response spike pulse output under continuous time-varying gain currents. Detailed implementation manners

[0040] The present invention will be further described in detail below in conjunction with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0041] In order to directly implement the non-linear pulse activation of the SNN in the optical domain and reduce the use of linear weighting devices in the SNN network, the embodiments of the present invention provide a method for the monolithic integrated DFB-SA to simultaneously implement pulse activation and synaptic weight. This method is implemented based on a monolithic integrated DFB-SA chip, which allows the simultaneous implementation of linear weighting and non-linear pulse activation, and is called a photonic pulse neuron-synaptic chip. Its structure is as Figure 2 and Figure 3 shown: Anti-reflection (AR) and high-reflection (HR) coatings are respectively coated on the two surfaces of the laser to improve the laser emission power, and an SA region is designed near the HR side. Figure 2 It is a micrograph of the DFB-SA chip.

[0042] Based on the above DFB-SA chip, the method for the monolithic integrated DFB-SA provided by the embodiments of the present invention to simultaneously implement pulse activation and synaptic weight includes:

[0043] Under the condition of reverse voltage biasing in the saturation absorption region of the DFB-SA, by adjusting the external optical injection of the DFB-SA and making time-varying adjustments to the gain current of the DFB-SA, nonlinear pulse activation and synaptic linear weighting are simultaneously achieved on a monolithic integrated DFB-SA, that is, pulse coding and linear weighting are simultaneously achieved on a single DFB-SA chip.

[0044] Among them, the gain current includes a static bias current I G0 and a time-varying current I M ; the excitatory threshold of the DFB-SA decreases with the increase of the gain current I G =I G0 +I M , the refractory period of the DFB-SA decreases with the increase of the gain current I G , the time-domain cumulative characteristic of the DFB-SA is controlled by the gain current I G ; the time-varying current I M is used to form synaptic weights.

[0045] Figure 4 Shows the pulse action principle in the DFB-SA: the gain region of the DFB-SA is driven by a current source, while the SA region is reverse-driven by a voltage source. The bias current in the gain region is denoted as the gain current I G , and the reverse bias voltage in the SA region is denoted as V SA . The gain current I G uses the time-varying current I G (t) to perform a linear weighting function, that is, the time-varying weight is loaded in the gain current I G . After the external optical wave E inj (t) is injected into the DFB-SA, the interaction between photons and electrons in the gain region and the SA region of the DFB-SA further simulates a LIF-type pulsed neuron, and the response output of the DFB-SA is denoted by E o (t).

[0046] For LIF neurons, time-domain integration, excitatory threshold, and refractory period are important information processing mechanisms for photon pulse neurons. In the process of implementing the present invention, the inventor first tested three information processing mechanisms of photon pulse neurons to achieve the plasticity of the neuron excitation threshold of the DFB-SA chip.

[0047] Specifically, first, the voltage-current characteristics of the prepared DFB-SA chip were tested: under the condition of a room temperature of 25 °C, when V SA =0V, the threshold current of the DFB-SA is about I G =86 mA, when V SAWhen = -0.4V, the threshold current of the DFB-SA is I G = 94 mA. The spectrum of the DFB-SA in the free-running state is as Figure 5 shown, and the side mode suppression ratio is about 50.9 dB. The wavelength of the externally injected light is determined according to the spectrum of the free-running DFB-SA. Specifically, the wavelength corresponding to the main peak in the spectrum is selected as the wavelength of the externally injected light.

[0048] Then, the neuronal excitatory threshold plasticity of the DFB-SA chip was tested. Before the test, an experimental platform as shown in Figure 6 was built to simulate and implement the method provided in the embodiments of the present invention. The inset in the upper right corner is a chip diagram after wire bonding of the DFB-SA.

[0049] Based on Figure 6 the experimental platform shown, in an optional implementation manner, the method for simultaneously implementing pulse activation and synaptic weight on the monolithic integrated DFB-SA provided in the embodiments of the present invention can be achieved through the following steps:

[0050] (a) Use a DC power supply (DC) to generate a static bias current and a reverse voltage bias;

[0051] (b) Use a tunable laser (TL) to generate an optical carrier;

[0052] (c) Use a first arbitrary waveform generator to generate an electrical stimulation signal;

[0053] (d) Input the optical carrier into a Mach-Zehnder modulator (MZM), and use the Mach-Zehnder modulator to perform electro-optic conversion on the electrical stimulation signal to form an optical signal;

[0054] (e) Use a circulator (CIRC) to inject the optical signal into the DFB-SA;

[0055] (f) Use a second arbitrary waveform generator to generate a time-varying current; the second arbitrary waveform generator can be the same arbitrary waveform generator (AWG) as the first arbitrary waveform generator.

[0056] (g) Use a bias tee to combine the static bias current with the time-varying current to form the gain current of the DFB-SA to drive the gain region of the DFB-SA;

[0057] (h) Adjust the external optical injection of the DFB-SA by adjusting the first arbitrary waveform generator, adjust the static bias current by adjusting the DC power supply, and adjust the time-varying current by adjusting the second arbitrary waveform generator to simultaneously implement non-linear pulse activation and synaptic linear weighting on the monolithic integrated DFB-SA.

[0058] The output of the DFB-SA is coupled into an optical spectrum analyzer (OSA) using an optical coupler (OC) for analysis. Meanwhile, a photodetector (PD) is used for photoelectric conversion, and then an oscilloscope (OSC) is used for recording.

[0059] In addition, a first polarization controller (PC1) can be further used between the tunable laser and the Mach-Zehnder modulator to adjust the polarization state of the optical path between the tunable laser and the Mach-Zehnder modulator; and a second polarization controller (PC2) is used between the Mach-Zehnder modulator and the circulator to adjust the optical path between the Mach-Zehnder modulator and the circulator.

[0060] Exemplarily, Figure 6 The models / parameters of the main modules can be seen in Table 1:

[0061] Table 1

[0062]

[0063]

[0064] Based on Figure 6 the platform shown, configure V SA = -0.4V. For a given external stimulus intensity and injection power, 5 different amplitude stimulus pulses are designed. By setting different gain currents, different non-linear pulse activations are obtained. As Figure 7 the data in part (a) shows, when I G = 88 mA, only the stimulus pulse with the highest stimulus intensity exceeds the excitatory threshold to successfully trigger the response spike pulse. As the gain current increases, the number of response spike pulses increases in a completely controllable manner (note that the synaptic weight corresponding to the external stimulus intensity here is fixed). Thus, it can be determined that different pulse responses are achieved by adjusting the excitatory threshold of the DFB-SA pulsed neuron, that is, the prepared DFB-SA has an inherent excitatory threshold plasticity, which is similar to the corresponding biological characteristics.

[0065] Similarly, fixing the external stimulus intensity, 10 pairs of external stimulus pulses are designed, and by continuously increasing the external stimulus pulse interval (ISI), the refractory period and the time-domain cumulative characteristics are explored. As Figure 7 the data in part (b) shows, for different pairs of external stimulus pulses, according to the different ISIs, the number of response spike pulses is 1 or 2. In addition, it can also be seen that when I G = 91 mA, the first 5 pairs of external stimulus pulses with relatively small ISIs will trigger a single response spike pulse, and the last 5 pairs of external stimulus pulses with larger ISIs will trigger two response spike pulses. As I GThe increase results in the reduction of the single response spike pulses to 4, 3, and 2 respectively, that is, with the increase of I G the refractory period of the DFB-SA gradually decreases with the increase of I.

[0066] To achieve the time-domain cumulative characteristic, the injection power externally injected into the photonic pulse neuron can be reduced to ensure that a single external stimulus pulse cannot trigger a response spike pulse alone. As Figure 7 shown by the data in part (c) of, when I G = 92.5 mA, only the first pair of external stimulus pulses with the smallest ISI can perform time-domain accumulation and exceed the excitatory threshold, thus causing response spike pulses. With the increase of the gain current, the number of response spike pulses also increases. It can be determined from this that by adjusting the gain current of the DFB-SA, the response spike pulses caused by the time-domain cumulative characteristic are controllable.

[0067] This verified inherent excitatory threshold plasticity of the neuron indicates that the excitatory threshold of the DFB-SA is adjustable, which is different from the traditional supervised training algorithm of SNN where only the weights are learnable and the response characteristics of the neuron itself remain unchanged. Therefore, a new type of hardware-aware training algorithm combining weight plasticity and excitatory threshold plasticity is expected to improve the performance of the photonic SNN.

[0068] Then, the synaptic plasticity of the DFB-SA chip, that is, the linear weighting characteristic, is tested. The bias current is statically changed during the experiment. A periodic pulse sequence is defined as the external stimulus, and the gain current is adjusted to ensure that the DFB-SA can respond as a LIF-like neuron. From Figure 8 the data in part (a) of, it can be seen that with the increase of the gain current (from 90 mA to 110 mA, with a step of 4 mA), the amplitude of the response spike pulse gradually increases. Figure 8 The data in part (b) of further shows the timing diagrams of the response spike pulse trains for three typical gain currents (I G = 90 mA, 100 mA, 110 mA). The results show that the larger the gain current, the larger the amplitude of the response spike pulse. In addition, for a given gain current, the weighted amplitude of the response spike pulse is almost constant in different periods, which indicates the stability of weighting. Figure 8 The data in part (c) of shows the function of the amplitude of the response spike pulse varying with the gain current (the bias current increases from 88 mA to 112 mA). It can be seen that the weighted amplitude almost linearly increases with the gain current. It can be determined from this that in the photonic SNN, the synaptic weight can be easily mapped to the gain current of the DFB-SA. By adjusting the bias current in the gain region of the DFB-SA, precise adjustment of the weighted amplitude of the optical response spike pulse can be achieved.

[0069] Therefore, the weighting of the response to spike pulses can be achieved on a monolithic integrated DFB-SA. Next, the excitatory threshold plasticity and synaptic linear weighting of the DFB-SA chip under dynamic bias current are tested.

[0070] As Figure 9 shown by the data in part (a) of [], for comparison, a time-varying optical external stimulus is defined, and the output of the DFB-SA under different constant gain currents is considered. From Figure 9 the data in part (a) of [], it can be obtained that: even when the intensity of the external stimulus is different, at a given gain current, when the intensity of the external stimulus exceeds the excitatory threshold, the amplitude of the response spike pulse is almost the same. In addition, the excitatory threshold decreases with the increase of the gain current, and more optical external stimuli will trigger response spike pulses. However, for different gain currents, the amplitudes of the response spike pulses are different, and the larger the gain current, the larger the amplitude of the response spike pulse.

[0071] Then, considering that the intensity of the external optical stimulus is fixed, the gain current is adjusted, and the repetition frequency of the input external stimulus pulse train is 0.5 GHz. As Figure 9 shown by the data in part (b) of [], the gain current is adjusted in 5 discrete constant levels by an AWG, corresponding to 5 discrete weights. The time-varying current I M is combined with another static bias current I G0 through a bias, and the total gain current can be expressed as I G = I G0 + I M , where I G0 = 80 mA. From Figure 9 the data in part (b) of [], it can be obtained that the excitatory threshold in each case is different. When the injection power is 171 μW, only the highest I M will cause the output of the response spike pulse. When the injection power is 215 μW, the output of the DFB-SA is two weighted response spike pulses with distinguishable amplitudes. When the injection power is further increased, 3 or 4 clusters of weighted response spike pulses with different amplitudes can be obtained. In addition, by adjusting the gain current of the DFB-SA to a continuously time-varying gain current, as Figure 9 shown by the data in part (c) of [], the DFB-SA can simultaneously achieve excitatory threshold plasticity and linear weighting characteristics. Therefore, the embodiment of the present invention realizes high-speed (0.5 GHz) tunable dynamic weights based on a monolithic integrated DFB-SA.

[0072] In addition, the inventors have considered external stimuli with a repetition rate higher than 0.5 GHz and found that the linear weighting function degrades. The partial response spikes cannot be weighted properly, which is restricted by the refractory period of the DFB-SA. Therefore, external optical injection preferably uses an external stimulus pulse train with a pulse repetition frequency of 0.5 GHz. Of course, since the refractory period of the DFB-SA is mainly determined by the carrier lifetime, the limitation of the DFB-SA refractory period can be adjusted by shortening the carrier lifetime using a shorter cavity length.

[0073] In summary, the method for simultaneously realizing pulse activation and synaptic weight in the monolithic integrated DFB-SA provided by the embodiments of the present invention, under the condition of applying a reverse voltage bias to the saturated absorption region of the DFB-SA, realizes non-linear pulse activation and synaptic linear weighting on the monolithic integrated DFB-SA by adjusting the external optical injection of the DFB-SA and the time-varying adjustment of the gain current of the DFB-SA. Among them, in terms of non-linear calculation, the excitatory threshold, refractory period, and time-domain cumulative characteristics of the photon pulse neuron can all be realized by adjusting the gain current of the monolithic integrated DFB-SA; in terms of linear calculation, the synaptic weight can be formed by adjusting the time-varying current in the gain current.

[0074] The method for simultaneously realizing pulse activation and synaptic weight in the monolithic integrated DFB-SA provided by the embodiments of the present invention is mainly applied to the construction of an optical pulse neural network device; the optical pulse neural network device is constructed based on multiple photon neuro-synaptic units. The photon neuro-synaptic unit mentioned here is realized based on the DFB-SA using the method provided by the embodiments of the present invention. Compared with biological neurons with a time scale of milliseconds, the photon neuro-synaptic unit based on the DFB-SA in the embodiments of the present invention provides a faster operation speed in the sub-nanosecond range due to its shorter carrier lifetime. That is to say, the response speed of the photon neuro-synaptic unit in the embodiments of the present invention can reach the sub-nanosecond level.

[0075] Based on the above method provided by the embodiments of the present invention, it can be known that in the above optical pulse neural network device, both the synaptic weight and the excitatory threshold realized by the DFB-SA are adjustable parameters.

[0076] In addition, based on the photon neuro-synaptic unit in the embodiments of the present invention, the use of photon weighting components in the SNN network can be avoided / reduced, which is very desirable for further expanding the photon SNN hardware and is more likely to realize multi-layer or deep pulse neural networks.

[0077] It should be noted that the terms "first", "second", etc. are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure.

[0078] In the description of this specification, the description referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0079] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and implement other variations of the disclosed embodiments by viewing the accompanying drawings and the disclosure. In the description of the present invention, the term "including" does not exclude other components or steps, the term "one" or "a" does not exclude a plurality of cases, and the meaning of "a plurality" is two or more, unless otherwise specifically defined. In addition, certain measures are described in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0080] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for a monolithic integrated DFB-SA to simultaneously achieve pulse activation and synaptic weight, characterized in that Including: Under the condition of applying a reverse voltage bias to the saturation absorption region of the DFB-SA, by adjusting the external optical injection of the DFB-SA and making a time-varying adjustment to the gain current of the DFB-SA, non-linear pulse activation and synaptic linear weighting are simultaneously achieved on a monolithic integrated DFB-SA; Among them, the gain current includes a static bias current and a time-varying current; the excitatory threshold of the DFB-SA decreases with the increase of the gain current, the refractory period of the DFB-SA decreases with the increase of the gain current, and the time-domain cumulative characteristic of the DFB-SA is controlled by the gain current; the time-varying current is used to form synaptic weights; Among them, the method of simultaneously achieving non-linear pulse activation and synaptic linear weighting on a monolithic integrated DFB-SA by adjusting the external optical injection of the DFB-SA and making a time-varying adjustment to the gain current of the DFB-SA under the condition of applying a reverse voltage bias to the saturation absorption region of the DFB-SA includes: Using a DC power supply to generate the static bias current and the reverse voltage bias; Using a tunable laser to generate an optical carrier; Using a first arbitrary waveform generator to generate an electrical stimulation signal; Inputting the optical carrier into a Mach-Zehnder modulator, and using the Mach-Zehnder modulator to perform electro-optic conversion on the electrical stimulation signal to form an optical signal; Using a circulator to inject the optical signal into the DFB-SA; Using a second arbitrary waveform generator to generate a time-varying current; Using a bias unit to combine the static bias current and the time-varying current to form the gain current of the DFB-SA; By adjusting the first arbitrary waveform generator, the external optical injection of the DFB-SA is adjusted, by adjusting the DC power supply, the static bias current is adjusted, and by adjusting the second arbitrary waveform generator, the time-varying current is adjusted, so as to simultaneously achieve non-linear pulse activation and synaptic linear weighting on a monolithic integrated DFB-SA.

2. The method according to claim 1, characterized in that, Using a bias unit to combine the static bias current and the time-varying current to form the gain current.

3. The method according to claim 1, wherein The wavelength of the external optical injection is determined according to the spectrum of the free-running DFB-SA.

4. The method according to claim 1, wherein The external optical injection is an external stimulation pulse sequence with a pulse repetition frequency of 0.5 GHz.

5. The method according to claim 1, wherein Using a first polarization controller between the tunable laser and the Mach-Zehnder modulator to adjust the polarization state of the optical path between the tunable laser and the Mach-Zehnder modulator; Using a second polarization controller between the Mach-Zehnder modulator and the circulator to adjust the optical path between the Mach-Zehnder modulator and the circulator.

6. The method according to claim 1, wherein Applied to the construction of an optical pulse neural network device; the optical pulse neural network device is constructed based on multiple photonic neuron-synapse units, and the photonic neuron-synapse unit is realized by using the method according to any one of claims 1 to 5 based on the DFB-SA.

7. The method according to claim 6, characterized in that, In the optical pulse neural network device, both the synaptic weight and the excitatory threshold realized by the DFB-SA are adjustable parameters.

8. The method according to claim 6, characterized in that The response speed of the photonic neuron-synapse unit is at the sub-nanosecond level.

Images