阅读说明：本技术 一种基于三端突触器件的联想学习神经网络阵列 (Association learning neural network array based on three-terminal synapse device ) 是由 田博博 闫梦阁 朱秋香 段纯刚 于 2021-08-12 设计创作，主要内容包括：本发明公开了一种基于三端突触器件的联想学习神经网络阵列,其特点是利用四个铁电突触与两个IF神经元构建的阵列式神经网络系统,输入具有时间差的脉冲序列来训练该网络,IF神经元通过调节网络系统输入的脉冲信号,改变铁电突触栅极的脉冲信号频率,从而调控三端SRDP突触的权重,实现联想学习,并通过单次迭代实现联想识别。本发明与现有技术相比具有对任意的特征信息进行联想学习,并且通过单次迭代实现联想识别,电路结构简单,较好的模拟了生物体内联想学习的行为,有效解决了忆阻器件的重复性和均匀性问题,促进类脑神经网络的进一步发展。(The invention discloses an association learning neural network array based on a three-terminal synapse device, which is characterized in that an array neural network system constructed by four ferroelectric synapses and two IF neurons is utilized to input a pulse sequence with time difference to train the network, the IF neurons change the pulse signal frequency of a ferroelectric synapse grid through regulating a pulse signal input by the network system, so that the weight of the three-terminal SRDP synapses is regulated and controlled, association learning is realized, and association identification is realized through single iteration. Compared with the prior art, the method has the advantages that the association learning is carried out on any characteristic information, the association recognition is realized through single iteration, the circuit structure is simple, the association learning behavior in a living body is well simulated, the problems of repeatability and uniformity of a memristor are effectively solved, and the further development of the brain-like neural network is promoted.)

1. An association learning neural network array based on a three-terminal synapse device is characterized in that an array neural network system constructed by four three-terminal SRDP synapses and two IF neurons is adopted, a pulse sequence training network with time difference is input to perform association learning based on the three-terminal SRDP synapse network array, and association recognition is realized through single iteration, and drains of the two three-terminal SRDP synapses in the same row of the three-terminal SRDP synapse network array are connected in parallel to be an input end of the network system; the source electrodes of two three-terminal SRDP synapses in the same column of the three-terminal SRDP synapse network array are connected in parallel to be the input end of the IF neuron; the output ends of the two IF neurons are connected in parallel and then are connected with the grid of four three-end SRDP synapses in the array, and the grid is the output end of the network system; the network system inputs a pulse sequence training network with a time difference; the IF neuron changes the frequency of the pulse signal of the three-terminal SRDP synaptic grid by adjusting the pulse signal input by the network system, so that the weight of the three-terminal SRDP synapse is regulated, association learning is realized, and association identification is realized through single iteration.

2. The array of three-terminal synaptic device-based associative learning neural network according to claim 1, wherein two inputs of said network system are an analog bone stimulation signal input and a ring stimulation signal input, respectively.

3. The associative learning neural network array based on the three-terminal synapse device of claim 1, wherein the three-terminal synapse device is a ferroelectric synapse of a channel material layer, a drain electrode, a source electrode, a ferroelectric functional layer and a gate electrode sequentially fabricated on a substrate, the drain electrode and the source electrode are disposed at two sides of the channel material layer between the substrate and the ferroelectric functional layer; the channel material layer is a transition metal chalcogenide layer; the ferroelectric functional layer is an organic ferroelectric polymer.

4. The three-terminal synapse device-based associative learning neural network array of claim 1 or claim 3, wherein the weight of the three-terminal SRDP synapses is achieved by adjusting a ferroelectric local field generated by a ferroelectric polymer through a pulse signal applied by a gate, thereby regulating the conductance of the three-terminal SRDP synapses.

5. The three-terminal synapse device-based associative learning neural network array according to claim 1 or claim 3, wherein the gate of the three-terminal SRDP synapse is weighted by a high frequency pulse signal and has no change in conductance when a low frequency pulse signal is input.

6. The associative learning neural network array based on three-terminal synapse devices of claim 1, wherein impulse response patterns of said IF neurons are: the external input signal raises the potential inside the IF neuron, which then fires a pulse when the potential exceeds a threshold voltage, and then returns to the initial state.

7. The neural network array for association learning based on three-terminal synapse devices of claim 1, wherein the association learning is realized by software simulation, comprising the following steps:

s1: inputting a pulse sequence at the input end of the simulated bone stimulation signal, and testing the conductance changes of four three-terminal SRDP synapses and the pulse release of two IF neurons;

s2: with time difference delta between two inputs_tTesting the conductance changes of the four three-terminal SRDP synapses and the pulse issuance of the two IF neurons;

s3: the same pulse sequence as in step S1 is input to the simulated bone stimulation signal input, and the conductance changes of the four three-terminal SRDP synapses and the pulse firing of the two IF neurons are again tested.

8. The three-terminal synapse device-based associative learning neural network array according to claim 7, wherein the impulses of the IF neurons represent salivary secretion from dogs or ear erection.

9. The array of associative learning neural networks based on three-terminal synapse devices of claim 2, wherein the simulated bone stimulation signal input and the simulated ring stimulation signal input are provided with a series of voltage pulses representing food stimulation signals and ring stimulation signals in a Barlow dog experiment.

Technical Field

The invention relates to the technical field of a pulse neural network, in particular to an association learning neural network array based on a three-terminal synapse device.

Background

Associative learning is one of the important ways to realize learning behavior in a living body, and is embodied in that, due to the connection of intrinsic features, when an event comes into consciousness, another event is associated, or complete information is recalled through pieces of information, which has important significance in pattern recognition, image and voice processing. And the basis for realizing associative learning is synaptic plasticity in organisms. In recent years, memristors, as bionic synapses, are the most potential electronic devices for realizing brain-like computation at present, and have successfully simulated various basic functions of the neural synapses, such as long-range retention/inhibition (LTP/D), pulse time-dependent plasticity (STDP), pulse frequency-dependent plasticity (SRDP), and the like. A neural network array built by using the memristor successfully simulates some basic learning behaviors in a living body, such as a Barlow dog conditioned reflex experiment and the like.

In the neural network array for simulating associative learning in the prior art, the fully-connected neural network is completed through multiple iterations, so that the information processing speed is greatly limited.

Disclosure of Invention

The invention aims to provide an association learning neural network array based on a three-terminal synapse device aiming at the defects of the prior art, the association learning neural network array formed by SRDP synapses and neurons is adopted, synapse weights are adjusted to proper weights by changing the pulse signal frequency of an input end and the excitation and training of IF neuron pulse signals, the association learning of the SRDP synapses is realized on any characteristic information, the association recognition is realized through single iteration, the problems of repeatability and uniformity of a memristor device are effectively solved, the method is simple and convenient, the behavior of the association learning in a living body is simulated by using a relatively simple circuit structure, the further development of a brain-like neural network is promoted, and the association learning neural network array has important significance and wide application prospect for information processing.

The purpose of the invention is realized as follows: an associative learning neural network array based on a three-terminal synapse device is characterized in that an array type neural network system constructed by four three-terminal SRDP synapses (ferroelectric synapses) and two Integrated-fire (IF) neurons is utilized, a pulse sequence with time difference is input to train the network, and therefore associative learning based on the three-terminal SRDP synaptic network array is achieved, and associative identification is achieved through single iteration.

The drains of two ferroelectric synapses in the same row of the three-terminal SRDP synapse network array are connected in parallel to be respectively connected with a bone stimulation signal input end and a bell stimulation signal input end; the source ends of two ferroelectric synapses in the same column are used as the input ends of the IF neuron, the output ends of the neurons are connected to the grid ends of the four ferroelectric synapses, and the weight of the ferroelectric synapses is changed by changing the pulse signal frequency of the two input ends and the pulse emission of the IF neuron, namely the spiking neural network is trained. The frequency of the input signal is determined by SRDP performance parameters of the ferroelectric synapse, and the synapse weight is regulated to a proper weight through excitation and training of pulse signals with different frequencies, so that association learning based on three-terminal SRDP is realized.

The associative learning is realized through software simulation, and specifically comprises the following steps:

s1: a pulse sequence with fixed frequency is input at an input end, and the conductance changes of four ferroelectric synapses and the pulse sending of two neurons are tested.

S2: with time difference delta between two inputs_tThe conductance changes of four ferroelectric synapses and the pulsing of two neurons are tried.

S3: the same pulse sequence as in step S1 is input at the input, and the conductance changes of the four ferroelectric synapses and the pulse firing of the two neurons are tested again.

The three-terminal synapse device is a channel material layer, a drain electrode, a source electrode, a ferroelectric functional layer and a ferroelectric synapse of a grid electrode which are sequentially prepared on a substrate, wherein the drain/source electrodes are arranged on two sides of the channel material layer between the substrate and the ferroelectric functional layer; the channel material layer is a transition metal chalcogenide layer; the ferroelectric functional layer is an organic ferroelectric polymer.

The ferroelectric synapse utilizes the conductance of the device to characterize the weight of the ferroelectric synapse, namely, when a high-frequency pulse signal is input to a grid of the synapse of the three-terminal SRDP, the weight of the synapse is increased, and when a low-frequency pulse signal is input, the conductance of the synapse is not changed.

The impulse response mode of the IF neuron is as follows: an external input signal raises the potential inside the neuron, and when the potential exceeds a threshold voltage, the neuron will fire a pulse and then return to the initial state.

The connection mode of the array type neural network is as follows: the drain terminals of the ferroelectric synapses in the same row are connected together and respectively used as two input terminals of the network system, the source terminals of the ferroelectric synapses in the same column are connected together and used as the input terminals of the IF neuron, and the output terminals of the IF neuron are connected to the gate terminals of the four ferroelectric synapses.

And the output ends of the neurons are respectively used as two output ends of the network system.

Compared with the prior art, the invention has the following advantages and obvious technical effects:

1) the SRDP function of synapse is realized based on the ferroelectric effect, and the problems of repeatability and uniformity of a memristor are solved due to the accurate control of an external electric field on the ferroelectric domain turnover.

2) The association learning neural network array based on the three-terminal synapse device simulates the association learning behavior in organisms by using a relatively simple circuit structure, and promotes the further development of the brain-like neural network.

3) On the basis of an association learning neural network array of a three-terminal synapse device, the method is expanded to a larger-scale three-terminal neural network, association learning is carried out on any characteristic information, and association recognition is realized through single iteration.

Drawings

FIG. 1 is a schematic structural view of the present invention;

FIG. 2 is a schematic diagram of a synapse device structure;

FIG. 3 is a diagram showing input/output results and conductance changes before neural network training;

FIG. 4 is a diagram illustrating input/output results and conductance variations of a neural network training process;

FIG. 5 is a diagram illustrating input and output results after neural network training.

Detailed Description

The invention will be described and illustrated in further detail with reference to specific embodiments:

example 1

Referring to FIG. 1, the present invention consists of a three terminal SRDP synapse (ferroelectric synapse) G₁₁、G₁₂、G₂₁And G₂₂And IF neuron N₃₁And N₃₂The constructed array type neural network system inputs a pulse sequence training network with time difference to perform association learning based on a three-terminal SRDP synaptic network array, and association recognition is realized through single iteration. Two ferroelectric synapses G in the same row of the three-terminal SRDP synapse network array₁₁、G₁₂（G₂₁、G₂₂) The drain electrodes of the grid-connected grid; two ferroelectric synapses G in the same column of the three-terminal SRDP synapse network array₁₁、G₂₁（G₁₂、G₂₂) Is connected in parallel as an IF neuron N₃₁And N₃₂An input terminal of (1); the IF neuron N₃₁And N₃₂After being connected in parallel with the ferroelectric synapse G in the array₁₁、G₁₂、G₂₁And G₂₂And are the output terminals 1 and 2 of the network system; the network system inputs a pulse sequence training network with a time difference; the IF neuron N₃₁And N₃₂Changing ferroelectric synapse G by modulating a pulse signal input by a network system₁₁、G₁₂,、G₂₁And G and₂₂the frequency of the pulse signal of the grid electrode is adjusted and controlled, thereby the ferroelectric synapse G₁₁、G₁₂,、G₂₁And G and₂₂the association learning is realized, and the association identification is realized through single iteration.

Referring to fig. 2, the three-terminal synapse device is a ferroelectric synapse prepared by a channel material layer 2, a drain and a source 4, a ferroelectric functional layer 5 and a gate 6 sequentially disposed on a substrate 1, wherein the drain 3 and the source 4 are disposed on two sides of the channel material layer 2 between the substrate 1 and the ferroelectric functional layer 5; the substrate 1 is SiO-containing₂An oxide layer heavily doped with p-Si;

the channel material layer 2 is an oxide layer upper transition MoS₂The two-dimensional semiconductor layer of (a); the drain electrode 3 and the source electrode 4 are Cr/Au drain/source electrodes formed on the two-dimensional semiconductor layer of the transition metal chalcogenide; the ferroelectric functional layer 5 is an organic ferroelectric polymer of polyvinylidene fluoride (PVDF) which is spin-coated on a metal drain source Cr/Au electrode; the grid electrode 6 is a transparent Al electrode formed on a polyvinylidene fluoride (PVDF) -based ferroelectric function layer.

Referring to fig. 1, the invention realizes associative learning and identification according to the following steps:

1) setting of initial conductance of synaptic devices

Setting the initial electrical conductivity state of each synapse to: g₁₁=1 nS, G₁₂=100 nS，G₂₁=100 nS， G₂₂=1 nS。

2) Pulse application prior to network training

Before training the neural network, a pulse sequence with a frequency of 500 Hz is applied to an input end 1 as a bone signal, no signal is input to an input end 2, and an IF neuron N is recorded₃₁And N₃₂Of the pulsed and ferroelectric synapse G₁₁、G₁₂、G₂₁And G₂₂Change in conductance of (c).

Referring to FIG. 3a, the pulse signals applied by two input terminals 1 and 2 of the neural network are shown, wherein the upper half is the pulse signal input by the bell input terminal, and the lower half is the pulse signal input by the bone input terminal (the input signal is zero, i.e. no signal input), which indicates that only the pulse signal is input at the bell input terminal alone.

Referring to FIG. 3b, IF neuron N is shown₃₁And N₃₂The pulse-firing condition, the upper half shows the IF neuron N₃₁Pulse firing conditions of (1), the lower half shows the IF neuron N₃₂Pulse delivery conditions of (1). Due to ferroelectric synapse G₁₁In a high conduction state, so that an input signal can pass through the ferroelectric synapse G₁₁Direct access to IF neuron N₃₁So that the IF neuron N₃₁There is a pulsing, i.e., when the bell signal is input, the neuron representing the raised ear responds. At the same time due to ferroelectric synapses G₁₂In a low conductive state, and thus inputThe signal cannot pass through the ferroelectric synapse G₁₂Reach IF neuron N₃₂Further IF neuron N₃₂No pulses are emitted. It is explained that only bell signal is input before training neural network, only IF neuron N representing ear is raised₃₁Responsive, to represent salivary IF neurons N₃₂No response.

Referring to FIG. 3c, a ferroelectric synapse G is shown₁₁、G₁₂、G₂₁And G₂₂The weight of (2) is changed. According to FIG. 3b and the neural network structure, the IF neuron N₃₁The emitted pulse is applied to the ferroelectric synapse G₁₁、G₁₂、G₂₁And G₂The frequency of the pulse signal of the grid is small and is not enough to cause obvious change of synaptic weights.

3) Pulse application during network training

When training the neural network, pulse sequences with the frequency of 500 Hz are applied to the input end 1 and the input end 2 respectively as the bone and the bell signals. In particular, the input pulses of the inputs 1 and 2 have a value δ_tTime difference of = 12 ms, recording IF neuron N₃₁And N₃₂Of the pulsed and ferroelectric synapse G₁₁、G₁₂、G₂₁And G₂₂Change in conductance of (c):

referring to FIG. 4a, the pulse signals applied by the input terminals 1 and 2 of the neural network are shown, wherein the upper half is the pulse signal input by the bell input terminal, and the lower half is the pulse signal input by the bone input terminal. As can be seen from FIG. 4a, input terminals 1 and 2 of the neural network are both inputted with pulse signals and have δ_tTime difference of = 12 ms.

Referring to FIG. 4b, IF neuron N is shown₃₁And N₃₂The pulse-firing condition, the upper half shows the IF neuron N₃₁Pulse firing conditions of (1), the lower half shows the IF neuron N₃₂Pulse delivery conditions of (1). At this time, the ferroelectric synapse G is input by the pulse signal at input terminals 1 and 2 of the neural network₁₁And G₂₂And is in a high conductive state. Thus, the input signal can pass through the ferroelectric synapses G, respectively₁₁And G₂₂Reach IF neuron N₃₁And N_32。Thus, IF neuron N₃₁And N₃₂There is a pulse issue.

Referring to FIG. 4c, a ferroelectric synapse G is shown₁₁、G₁₂、G₂₁And G₂The weight of (2) is changed. According to FIG. 4b and the neural network structure, the IF neuron N₃₁And N₃₂Is input in parallel to the ferroelectric synapse G₁₁、G_12、 G₂₁And G₂₂A gate electrode of (1). At this time, IF neuron N₃₁And N₃₂Superposition of outputs to the ferroelectric synapse G₁₁、G_12、 G₂₁And G₂₂The frequency of the pulse signal of the gate increases. Thus, the ferroelectric synapse G₁₁、G_12、 G₂₁And G₂₂The conductance of (c) is increased, i.e., the synaptic weight.

4) Pulse application after network training

After training the neural network, a pulse train with a frequency of 500 Hz is applied again as a bone signal at input 1, and no signal is input at input 2.

Referring to FIG. 5a, the pulse signals applied by the input terminals 1 and 2 of the neural network are shown, wherein the upper half is the pulse signal input by the bell input terminal, and the lower half is the pulse signal input by the bone input terminal (the input signal is zero, i.e. no signal input), which indicates that only the pulse signal is input at the bell input terminal alone.

Referring to FIG. 5b, IF neuron N is shown₃₁And N₃₂The pulse-firing condition, the upper half shows the IF neuron N₃₁Pulse firing conditions of (1), the lower half shows the IF neuron N₃₂Pulse delivery conditions of (1). At this time, only the bell signal is inputted, representing the salivary IF neuron N₃₂There is also pulse delivery. Due to the training process of FIG. 4, the ferroelectric synapse G₁₁、G_12、 G₂₁And G₂₂Both increases in conductance. At this time, the ferroelectric synapse G₁₂And G₂₁Also in a high conducting state. Therefore, the input bell signal can pass through the high-conductivity ferroelectric synapse G₁₂Reach IF neuron N₃₂In turn, cause IF neuron N₃₂The pulse of (2). That is, the neural network is trained to establish a connection between the bell signal and the bone signal, and when only the bell signal is input, the IF neuron N representing salivation₃₂There is also a response.

The invention realizes the function of associative learning, when the input end 1 and the input end 2 input pulse signals with time difference, the IF neuron N passes₃₁And N₃₂Is applied to the ferroelectric synapse G₁₁、G_12、 G₂₁And G₂₂The frequency of the pulse signal of the grid electrode is increased, namely the pulse interval time is shorter and is lower than the polarization relaxation time of the ferroelectric PVDF of the grid dielectric layer. Thus, the PVDF polarization state fails to return to its initial state, and continued application of the pulse causes ferroelectric synapse G₁₁、G_12、 G₂₁And G₂I.e. the weight, changes. After training, IF neuron N inputs only bone signals₃₁And N₃₂The pulse issuance of (a) indicates that the impulse neural network associates with the bell signal.

The invention is further described and not intended to be limited to the specific embodiments disclosed, but rather, the invention is to be accorded the full scope and equivalents thereof.