A neural synaptic device with optical fiber structure
By combining phase change materials in optical fiber structure, the regulation of synaptic weights and STDP learning rules are realized, and the signal interference and energy consumption problems of electrical synaptic devices are solved, providing an efficient and low-power all-photon synaptic solution. , supports simulation of biological synapses.
Patent Information
- Application Number
- CN202111021706.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-01
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-09-01
AI Technical Summary
Due to signal interference and high energy consumption, existing electrical synaptic devices are difficult to effectively simulate the functions of biological synapses. The complexity of the electrical system hinders their wide application, and a synaptic solution that can overcome these defects is needed.
Adopt fiber-structured synaptic devices based on phase change materials, adjust synaptic weights through pulse width to realize STDP learning rules, use the combination of optical fiber and phase change materials, use the fiber annular device and delay line to achieve the superposition and delay of pulses, and use the The optical properties of the phase change material regulate synaptic weights and prevent oxidation through the Au cover.
It realizes synaptic function with simple structure, low power consumption, fast computing, and feedback regulation weights. It overcomes the defects of electrical synapses and provides an efficient all-photon synaptic scheme, supports STDP learning rules, and regulates synaptic weights. The number of times reached 1012 times.
Smart Images

Figure CN113723602B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of optical fiber intelligent devices, and in particular to a neural synapse device based on an optical fiber structure of a phase change material. Background Art
[0002] Neuromorphic computing is the computing power support for the third generation of artificial intelligence (AI). It imitates the information processing mode of the human brain, integrates storage and computing, and is expected to break the "von Neumann" constraint architecture. It is a new type of computing with autonomous learning capabilities and real-time processing of unstructured information. The core technology of neuromorphic computing is the implementation of "spiking neural network (SNN)". SNN is similar to biological neural network, and is composed of various types of neurons and synapses. In the human brain, there are about 10 neurons. 11 There are about 10 synapses. 15 Synapses are the key part of information transmission between neurons. They dominate the structure of neural networks and are responsible for a large amount of parallelism, structural plasticity, and robustness of the brain. Therefore, synaptic devices that mimic the function and plasticity of biological synapses will become an important part of brain-inspired computing systems.
[0003] Phase change materials (PCM) have the advantages of fast write / erase speed, large difference in electrical and optical properties between amorphous and crystalline states, good reversible cycle, and good thermal stability. They are technically mature and the best choice for realizing synaptic functions. Therefore, many synaptic schemes with PCM have been proposed. For example, in electrical engineering, synaptic devices with resistance changes induced by electric pulses, although they can perform synaptic functions, are hindered by the inherent signal interference and high energy consumption of electrical systems. In order to avoid defects caused by electricity, Harish Bhaskaran et al. proposed a fully integrated all-optical synapse scheme on a silicon nitride / silicon dioxide platform (Zengguang Cheng, Carlos Ríos, Wolfram HP Pernice, et al. On-chip photonic synapse [J]. Science. Advances, 2017, 3: e1700160.), which overcomes the shortcomings of electrical synapses, but its manufacturing process is complicated and has high requirements for equipment, making it difficult to be widely used. Optical fiber has good light coupling, simple structure, low attenuation, is similar to neurons in structure and function, and can overcome electrical defects. We believe that the combination of optical fiber and PCM can better realize the synaptic function of neurons.
[0004] At present, optical fiber has been widely used in the field of optical fiber communication and optical fiber sensing due to its high speed, large bandwidth, good scalability, high transmission quality, low cost, strong anti-interference ability, etc. The neural synapse scheme based on optical fiber structure of phase change material proposed in this patent has the advantages of simple structure, low power consumption, fast operation, feedback adjustment weight, and can realize STDP learning rule. Summary of the invention
[0005] The purpose of the present invention is to propose a neural synaptic device based on a phase change material optical fiber structure, which adjusts the weight of the synapse by inputting a single pulse with a fixed power and different pulse widths to realize the STDP learning rule of the synapse.
[0006] The neural synapse scheme of the optical fiber structure of the present invention is implemented as follows:
[0007] A neural synapse scheme with an optical fiber structure comprises: a pulse light source (1), a phase modulator (2), a 50:50 coupler (3), a circulator (4), a synapse structure (5), a 90:10 beam splitter (6), a photodetector (7), a delay line (8), and an optical amplifier (9). The pulse light source (1) is used to generate pulse light that meets the requirements, such as a certain wavelength, power, pulse width and frequency; the phase modulator (2) is used to eliminate the interference between the presynaptic pulse and the postsynaptic pulse; the 50:50 coupler (3) is used to superimpose the presynaptic pulse with the processed postsynaptic pulse; the circulator (4) is used to transmit the pulse from a1 to a2 to the synaptic structure (5), and then transmit the reflected pulse from a2 to a3 to the postsynapse; the 90:10 beam splitter (6) is used to divide the postsynaptic pulse into two beams of 90% and 10%, 90% of which flows into the delay line (8) and 10% flows into the photodetector (7); the photodetector (7) is used to detect the shape of the postsynaptic pulse and determine the pulse width at this time; the delay line (8) is a tunable delay line, which is used to delay the postsynaptic pulse to different degrees, and then pass through the optical amplifier (9) for power amplification, and amplify to the same power as the presynaptic pulse.
[0008] Furthermore, the synaptic structure (5) is composed of an optical fiber (5-1), a phase change material (5-2), and a covering layer (5-3). The phase change material (5-2) directly covers the end face of the optical fiber (5-1), and the covering layer (5-3) directly covers the phase change material (5-2). The optical fiber (5-1) can be a single-mode optical fiber or a multi-mode optical fiber. The phase change material (5-2) can be a sulfur compound composed of any two or more of the three elements of Ge, Sb, and Te, and its optical properties can be adjusted by the incident light pulse. The covering layer (5-3) can be an Au film to prevent the phase change material from oxidizing and improve the reflectivity of the synaptic structure. The phase change material (5-2) and the covering layer (5-3) can be plated on the end face of the optical fiber (5-1) by radio frequency magnetron sputtering and direct current magnetron sputtering, respectively. The optical properties of the phase change material (5-2) vary according to the degree of crystallization. The higher the degree of crystallization, the higher the reflectivity of the communication band light. The synaptic weight is adjusted by the crystallization degree of the phase change material (5-2), and the crystallization degree is adjusted by inputting pulsed light of fixed power and different pulse widths. The optical fiber synaptic structure (5) uses an optical fiber circulator for light injection and light collection, wherein the received light pulse is the light pulse reflected by the phase change material (5-2) and the covering layer (5-3). The synaptic structure (5) can use different optical signals to adjust and read the synaptic weight, and the number of times the synaptic weight can be adjusted is up to 10 12 .
[0009] Furthermore, for the presynaptic and postsynaptic pulses passing through the 50:50 coupler (3), only the overlapping portion of the pulse widths of the two pulses can exceed the threshold for adjusting the weight of the synaptic structure (5), and the non-overlapping portion will not exceed the threshold and will not cause a change in the weight.
[0010] The STDP learning rule depends on the order and interval of external signal pulses arriving at the synapse. When the time sequence of pulses arriving at the synapse changes, the synaptic weight can change from enhancement (inhibition) to inhibition (enhancement). This phenomenon is called asymmetric STDP in the brain, and on the contrary, it is called symmetric STDP; when the time interval between two trains of pulses arriving at the synapse changes, the synaptic weight changes exponentially; STDP is considered to be the most important learning mechanism in the Hebbian theory, and plays an important role in information encoding, learning and memory. Its learning formula is:
[0011] ;
[0012] In the formula, ΔW represents the change in synaptic weight, Δt represents the relative time difference between the arrival of the presynaptic pulse signal and the postsynaptic pulse signal, τ and + / - represent the time constant and the exponential factor, and a is a constant.
[0013] In the present invention, different delay times Δt of the delay line (8) can be adjusted to achieve pulse widths with different overlapping degrees, so that the synapses can obtain different weights, thereby realizing the STDP learning rule.
[0014] The present invention utilizes a combination of phase change materials and optical fibers to construct a fully integrated all-photonic synaptic device. Phase change materials have the advantages of fast write / erase speed, good reversible cycle, good thermal stability, and large differences in electronic and optical properties between amorphous and crystalline states. They are technically mature and the best choice for realizing synaptic functions. The phase change material responds differently to different pulses. Under the stimulation of different pulse widths, the crystallinity of the phase change material changes, affecting its reflectivity, absorptivity, refractive index and other properties, thereby maintaining long-term stability. Under pulse width modulation (PWM), multi-level reflection states of the synaptic device are realized, and pulse timing dependent plasticity (STDP) is realized using an adjustable delay circuit.
[0015] Beneficial Effects
[0016] The present invention has the following beneficial effects compared with the prior art:
[0017] 1. The present invention proposes a method of combining phase change materials with optical fibers to construct an optical fiber neural synapse structure, endow the optical fiber with "intelligence", and realize the STDP learning rule of the synapse, opening up a new direction for simulating biological synapses;
[0018] 2. Compared with electrical synapses, the all-optical fiber synapses based on phase change materials of the present invention have the advantages of simple structure, high efficiency and low power consumption. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] 1. Figure 1 A schematic diagram of a synapse scheme for an optical fiber structure;
[0020] 2. Figure 2 A schematic diagram of the synaptic structure.
[0021] 3. Figure 3 A schematic diagram of pulse width modulation (PWM) in the present invention;
[0022] 4. Figure 4 It is an exponential relationship diagram of the synaptic weight of the present invention as the time delay (Δt) between the presynaptic pulse and the postsynaptic pulse changes. DETAILED DESCRIPTION
[0023] The technical method of the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. The described specific embodiments are only used to explain the present invention and are not intended to limit the present invention.
[0024] like Figure 1 , 2As shown, in the example, the wavelength of the pulse light is selected from the communication band 1550nm, the power is selected from 95mW, and the pulse width is selected from 70ns; the optical fiber (5-1) of the synaptic structure is selected from single-mode optical fiber, the phase change material (5-2) is selected from Ge2Sb2Te5 (GST), and the cover layer (5-3) is selected from Au film. The amorphous state of GST has low reflectivity, and the crystalline state has high reflectivity. When GST is heated to about 420K by external light / electric pulses, it will change into an FCC structure and present a crystalline state. When a higher energy pulse heats the material to above the melting temperature (about 890K) and quickly cools it, it can return to the amorphous state. Therefore, pulses with different energies and different pulse widths can be used to switch GST to obtain different crystalline states of GST; different crystalline states of GST have different reflectivities and corresponding synaptic weights.
[0025] The pulse output by the pulse light source (1) is defined as In1, the postsynaptic pulse is defined as In2, and the presynaptic pulse output through the coupler (3) is defined as Out. The corresponding pulse powers are defined as P In1 , P In2 and P out The threshold switching power of GST is defined as P th The pulse output by the pulse light source (1) is connected to the phase modulator (2) to eliminate the interference between In1 and In2, and 50% of P In1 and 50% P In2 The presynaptic pulse is coupled together and passes through the circulator (4) a1 to a2. Then, the reflected pulse, i.e., the postsynaptic pulse, passes through the circulator (4) a2 to a3 and is then split into two beams by the beam splitter (6). 10% of the pulses are transmitted to the photodetector (7) to observe the pulse shape, and 90% of the pulses are transmitted through the optical fiber delay line (8) to delay the postsynaptic pulse by nanoseconds. The pulses are then amplified by the erbium-doped fiber optical amplifier (9) to make the power of the postsynaptic pulse consistent with the input pulse power, so that P In2 =P In1 <P th , P Out =P In1 +P In2 When the delay time of In1 and In2 is less than the pulse width, the overlapped part of the two pulses is P out , and P out >P th , the presynaptic pulse has an effect on the synapse, causing its weight to change. When the delay time of In1 and In2 is greater than the pulse width and less than the pulse interval, the two pulses do not overlap and the synaptic weight does not change. Figure 3 As shown, by adjusting the delay line (8), the τ of Out is changed from 0 to τ max (pulse width) and the corresponding pulse power is higher than P th , update the synaptic weights. Figure 4 The given delay time is exponentially related to the synaptic weight, satisfying the learning rule of STDP.
[0026] The specific embodiments described above further describe the technical solutions and beneficial effects of the present invention in detail, and are only a part of the embodiments rather than all the embodiments. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A neural synaptic device with an optical fiber structure, comprising: A pulse light source (1), a phase modulator (2), a 50:50 coupler (3), a circulator (4), a synaptic structure (5), a 90:10 beam splitter (6), a photodetector (7), a delay line (8) and an optical amplifier (9); the pulse output by the pulse light source (1) is connected to the phase modulator (2), then passes through the 50:50 coupler (3), the circulator (4), and then passes through the 90:10 beam splitter (6) to be divided into two beams; wherein the function of the phase modulator (2) is to eliminate the interference between the presynaptic pulse and the postsynaptic pulse; the function of the 50:50 coupler (3) is to superimpose the presynaptic pulse with the processed postsynaptic pulse; the function of the circulator (4) is to transmit the pulse from a1 to a2 to the synaptic structure (5), and then transmit the reflected pulse from a2 to a 3 is transmitted to the postsynaptic site; the 90:10 beam splitter (6) is used to split the postsynaptic pulse into two beams of 90% and 10%, 90% of which flows into the delay line (8) and 10% of which flows into the photodetector (7); the photodetector (7) is used to detect the shape of the postsynaptic pulse and determine the pulse width at this time; the delay line (8) is a tunable delay line, which is used to delay the postsynaptic pulse to different degrees, and then pass through the optical amplifier (9) to amplify the power and amplify it to the same power as the presynaptic pulse; it is characterized in that: the synaptic structure (5) is composed of an optical fiber (5-1), a phase change material (5-2) and a covering layer (5-3), the phase change material (5-2) directly covers the end face of the optical fiber (5-1), and the covering layer (5-3) directly covers the phase change material (5-2); The optical fiber (5-1) is a single-mode optical fiber or a multi-mode optical fiber, the phase change material (5-2) is a chalcogenide compound, composed of any two or more of the three elements Ge, Sb, and Te, and its optical properties can be adjusted by an incident light pulse, the covering layer (5-3) is an Au thin film, which prevents the phase change material from being oxidized and improves the reflectivity of the synaptic structure, and the phase change material (5-2) and the covering layer (5-3) are respectively plated on the end face of the optical fiber (5-1) by radio frequency magnetron sputtering and direct current magnetron sputtering, wherein the optical properties of the phase change material (5-2) vary according to the degree of crystallization, and the higher the degree of crystallization, the higher the reflectivity to the communication band light, and the synaptic weight is adjusted by the degree of crystallization of the phase change material (5-2), and the degree of crystallization is adjusted by inputting pulse light of fixed power and different pulse widths; The device uses different optical signals to adjust and read synaptic weights, and the number of times the synaptic weight can be adjusted is up to 10 12 ; By adjusting the delay line (8) to different delay times to achieve pulse widths with different overlapping degrees, the synapses are given different weights, thus realizing the STDP learning rule of the synapses; The STDP learning rule depends on the order and interval of external signal pulses arriving at the synapse. When the time sequence of pulses arriving at the synapse changes, the synaptic weight changes from enhancement to inhibition or from inhibition to enhancement. This phenomenon is called asymmetric STDP in the brain. On the contrary, it is called symmetric STDP. When the time interval between two trains of pulses arriving at the synapse changes, the synaptic weight changes exponentially. This device realizes all-optical neural synapses with a simple structure and low attenuation, overcoming the defects of electrical synaptic signal interference and high energy consumption. In addition, optical fibers have natural similarities with neurons in both structure and function.
2. The neural synapse device according to claim 1, characterized in that: The synaptic structure (5) uses a circulator (4) for light injection and light collection, wherein the received light pulses are light pulses reflected by the phase change material (5-2) and the covering layer (5-3), and the circulator (4) is an optical fiber circulator.
3. According to any one of claims 1-2, the presynaptic and postsynaptic pulses passing through the 50:50 coupler (3) can exceed the threshold for adjusting the weight of the synaptic structure (5) only when the pulse widths of the two overlap, and the non-overlapping parts will not exceed the threshold and thus will not cause a change in the weight.
Citation Information
Patent Citations
Unit, device and method for simulating biological neuron and neuronal synapsis
CN103078054A
Electronic synapses for reinforcement learning
CN103282919A
Hardware friendly pulse neural network model based on STDP non-supervised learning algorithm
CN107092959A
SiNx-Based optical reading synaptic device structure and preparation method thereof
CN107634140A
Neuron-like all-optical memory device based on Ge2Sb2Te5
CN108470575A
