CN111291881B

CN111291881B - Mode identification device and method based on cascade VCSELs photonic nerves

Info

Publication number: CN111291881B
Application number: CN202010148127.3A
Authority: CN
Inventors: 邓涛; 高子叶; 唐曦; 林晓东; 吴正茂; 夏光琼; 田涛; 倪敏
Original assignee: Southwest University
Current assignee: Southwest University
Priority date: 2020-03-05
Filing date: 2020-03-05
Publication date: 2022-10-18
Anticipated expiration: 2040-03-05
Also published as: CN111291881A

Abstract

The invention provides a pattern recognition device and method based on cascade VCSELs photonic nerves, wherein the method comprises the following steps: step 1: outputting the waveform of the pulse sequence; and 2, step: modulating continuous waves generated by the tunable laser by using the waveform obtained in the step 1 to obtain an optical pulse sequence; and step 3: dividing the optical pulse sequence signal into three paths, and injecting the three paths of signals into three cascaded VCSELs photon neurons of a neural network respectively; and 4, step 4: and collecting signals output by each neuron, performing processing operation by using a computer, and adjusting and controlling pulse signals input to the neurons according to an operation result so as to realize pattern recognition of a pulse sequence. The invention can realize the photon neural network of pulse sequence identification through three cascade VCSELs photon neurons.

Description

Mode identification device and method based on cascade VCSELs photonic nerves

Technical Field

The invention relates to the technical field of information processing, in particular to a pattern recognition device and method based on cascade VCSELs photonic nerves.

Background

There are approximately 1000 million neurons in the human brain, and each neuron is connected to at least 10000 synapses. The computing power of brain reaches 10 ²⁰ MAC/s (MAC stands for multiply-add operation, which is similar to the floating point computing FLOP of a computer, but is more suitable for digital signal processors), and its power consumption is only 2 × 10 ^-10 nJ/MAC, but the energy consumption of the fastest ultra-computer in the world reaches 1nJ/FLOP at present, and thus, the human brain has very strong computing power. Therefore, exploring the activity mechanism of the cranial nerves and reasonably utilizing the complex information processing capability thereof is a hot issue which is always concerned by people. The artificial neural network that has been developed along with this has gradually become a research hotspot in the field of artificial intelligence. Such as: hua is Shang Teng series AI chip, IBM TrueNorth, stanford university Neurogrid, manchester university Neuromorphic chip, etc. These milestone-like efforts have demonstrated the great potential of artificial neural network technology and have greatly driven the development of future information processing devices. However, these artificial neural network models focus mainly on electrical implementations such as CMOS analog circuits, neurosynaptic digital chips, and modern very large scale integrated circuits (VLSI). Although these artificial neural techniques can effectively process partial activity behaviors of biological nerves, the artificial neural system based on the electrical approach is limited by bandwidth, distance, power consumption, and the like in practical application.

In recent years, a photon neural model based on a semiconductor laser is concerned because the photon neural model can generate ultrafast impulse response of 8 orders of magnitude higher than biological nerves, can break through the limitation of the bandwidth of an electronic device and greatly reduces the energy consumption of a system. In particular, VCSELs have several unique advantages, such as low cost, low power consumption, easy integration into two-dimensional/three-dimensional arrays, high coupling efficiency with optical fibers, and compatibility with existing optical fiber systems, and thus, the VCSELs-based photonic neural model has received much attention in recent years. The Hurtado team at Sicklede university, english, barland team at Niss university, france, prucnal team at Princeton, and domestic Beijing university of transportation, western Ann university of electronic technology, southwestern university, etc. also performed a lot of work on spiking dynamics based on the photon nerves of VCSELs. However, currently, related research mainly focuses on spiking dynamics behavior of single or two VCSELs photonic nerves, and research for realizing pattern recognition based on VCSELs photonic neural networking is still in a starting stage. Obviously, the research on the pulse dynamics characteristic based on VCSELs photonic nerves and the application of the pulse dynamics characteristic to the fields of pattern recognition and the like have great practical application value.

In addition, most of the current methods for realizing pulse sequence recognition are based on electrical methods, and few optical methods are used. Moreover, the research based on VCSELs photonic nerves mainly focuses on the research of basic functional characteristics of a single neuron at present, and the research on relevant practical application scenes of artificial intelligence is not many, which obviously does not meet the practical requirements.

Disclosure of Invention

The invention aims to solve the defects existing in the prior art, and provides a pattern recognition method and a pattern recognition device which skillfully cascade three VCSELs neurons and apply the advantages of the VCSELs photonic nerves to pulse sequences by utilizing the change of response thresholds of the neurons.

A pattern recognition apparatus based on cascade VCSELs photonic nerves, comprising: the optical pulse train comprises an arbitrary waveform generator capable of generating a pulse train, a Mach-Zehnder modulator, a tunable laser, an optical pulse train, a photoelectric detector PD, an oscilloscope OSC and a computer, wherein a signal output from the arbitrary waveform generator is modulated by the Mach-Zehnder modulator to obtain continuous waves generated by the tunable laser to obtain the optical pulse train;

the three cascade VCSELs photon neurons comprise a neuron 1, a neuron 2 and a neuron 3; the time delay between neuron 1 and neuron 2 is τ 12; the time delay between neuron 2 and neuron 3 is τ 23, and the time delay between neuron 1 and neuron 3 is τ 13; τ 13 = τ 12 + τ 23.

Further, as described above in the cascade VCSELs photonic nerve-based pattern recognition apparatus, the three-way signal includes one neuron, two neurons, and three neurons;

one path of the neuron is as follows: the signal output from the modulator MZM is divided into two paths after passing through the coupler FC1, wherein one path of signal is divided into two parts after passing through the coupler FC2, and one part of signal enters the VCSEL1 after passing through the polarization controller PC3, the adjustable attenuator VA3, the optical fiber delay line DL3 and the optical fiber circulator OC 1;

the two paths of the neuron are composed of two paths; one path is as follows: a part of signals divided by the coupler FC1 are divided into two parts again after passing through a coupler FC2, wherein one part of the signals enter the VCSEL2 after passing through a polarization controller PC2, an adjustable attenuator VA2, an optical fiber delay line DL2 and a coupler FC 4;

the other path is as follows: a signal output from the optical fiber circulator OC1 is divided into two parts by a coupler FC3, wherein one part enters the VCSEL2 after passing through a polarization controller PC4, an adjustable attenuator VA4, an optical fiber delay line DL4 and the coupler FC 4;

the neuron three paths consist of three paths; wherein the 1 st way is: the other path divided by the coupler FC1 enters the VCSEL3 after passing through a polarization controller PC1, an adjustable attenuator VA1, an optical fiber delay line DL1, a coupler FC7 and an optical fiber circulator OC 2;

the 2 nd way is: the signal output from the VCSEL2 enters the VCSEL3 after passing through a coupler FC5, a polarization controller PC5, an adjustable attenuator VA5, an optical fiber delay line DL5, a coupler FC6, a coupler FC7 and an optical fiber circulator OC 2;

the 3 rd path is: the other path of signal branched from the coupler FC3 is also input to the coupler FC6 after passing through the coupler FC8, the polarization controller PC6, the adjustable attenuator VA6 and the fiber delay line DL6, and the signal output from the coupler FC6 enters the VCSEL3 after passing through the coupler FC7 and the fiber circulator OC 2.

Further, as in the pattern recognition apparatus based on cascaded VCSELs photonic nerves described above, the signal of the coupler FC1 is divided into 3; 70% of the output signal passes through coupler FC2; the signals of the coupler FC2, the coupler FC3, the coupler FC4, the coupler FC6 and the coupler FC7 are divided into 5.

Further, as described above, in the pattern recognition apparatus based on cascade VCSELs photonic nerves, the signals output from the neurons 1, 2, and 3 are converted into electrical signals by the PD1, PD2, and PD3, respectively;

the input signal of the PD1 comes from a part of signals branched by the coupler FC 8; the input signal of the PD2 comes from a part of signals branched by the coupler FC 5; the input signal of the PD3 comes from the signal output by the optical fiber circulator OC 2.

Further, as for the pattern recognition device based on cascade VCSELs photonic nerves, the signals of the couplers FC8 and FC5 are divided into 90; of which 10% of the signals are input to PD1 and PD2.

The invention also provides a pattern recognition method based on cascade VCSELs photonic nerves, which comprises the following steps:

step 1: outputting the waveform of the pulse sequence;

step 2: modulating continuous waves generated by the tunable laser by using the waveform obtained in the step 1 to obtain an optical pulse sequence;

and step 3: dividing the optical pulse sequence signal into three paths, and injecting the three paths of signals into three cascaded VCSELs photon neurons of a neural network respectively;

the three cascade VCSELs photon neurons of the neural network comprise a neuron 1, a neuron 2 and a neuron 3; the time delay between neuron 1 and neuron 2 is τ 12; the time delay between neuron 2 and neuron 3 is τ 23, and the time delay between neuron 1 and neuron 3 is τ 13; τ 13 = τ 12 + τ 23;

and 4, step 4: and collecting signals output by each neuron, performing processing operation by using a computer, and adjusting and controlling stimulation signals input to each neuron according to an operation result so as to realize pattern recognition of a pulse sequence.

Further, in the method for pattern recognition based on cascaded VCSELs photonic nerves as described above, the processing operation in step 4 includes:

training by using an unsupervised pulse time-dependent plasticity mechanism, and adjusting the weight and time delay of each external stimulation signal according to the comparison result of the computer processing result and the target signal in the training process, namely: inputting the pulse signal intensity of each neuron; and finally, the weight information determined after training is used for controlling corresponding parameters of the system, so that the mode identification of the pulse sequence is realized.

The invention forms a photon neural network capable of realizing the identification of any pulse sequence by three cascaded VCSELs photon neurons.

Specifically, after the signal output by the MZM modulator is divided into three paths, the three paths of signals can be ensured to be injected into three neurons without delay by respectively adjusting the three delay lines DL1, DL2 and DL 3. As the response thresholds of the three neurons gradually increase, the signal output by the neuron is injected into the neuron 2 after a delay τ 12, and when the stimulation signal output by the neuron 1 coincides with a certain signal injected into the neuron 2 from the modulator output, the neuron 2 generates a corresponding response; the response signals output from neurons 1 and 2 are then injected into neuron 3 after being delayed by τ 13 and τ 23, respectively, and it is guaranteed that a certain response signal output from neurons 1 and 2 can be simultaneously used to stimulate neuron 3 only when the delay time condition of τ 13 = τ 12 + τ 23 is satisfied, thereby enabling neuron 3 to respond to stimulation at a certain time.

Has the beneficial effects that:

1. and the identification of any pulse sequence can be realized.

2. VCSELs have some unique advantages of low cost, low energy consumption, easy integration into two-dimensional/three-dimensional arrays, high coupling efficiency with optical fibers, and the like;

3. all optical devices adopt 1550nm wave band, and can be compatible with the existing optical fiber system;

4. the VCSELs are used as photonic nerves, so that not only can the impulse response characteristic of biological nerves be simulated, but also the response speed of the VCSELs is 8 orders of magnitude faster than that of the biological nerves;

5. the core devices of the pattern recognition system adopted by the invention are all optical devices, so that the limitation of the bandwidth of an electronic device can be effectively broken through, and the working speed of the system is greatly improved;

6. the neuron weight adjustment can be completed by using a common adjustable attenuator;

7. the system has simple structure and is composed of three neurons and corresponding devices.

Drawings

Fig. 1 is a schematic structural diagram of a pattern recognition device based on cascade VCSELs photonic nerves according to the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention are described clearly and completely below, and it is obvious that the described embodiments are some, not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without inventive step based on the embodiments of the present invention, are within the scope of protection of the present invention.

The invention forms a photonic neural network capable of realizing pulse sequence identification by three cascading VCSELs photonic neurons, the three VCSELs photonic neurons adopt a cascade structure, the pulse response thresholds of the three neurons are gradually increased, the time delay tau 12 between the neuron 1 and the neuron 2 is fixed, and the sum of the tau 12 and the tau 23 (the time delay between the neuron 2 and the neuron 3) is ensured to be equal to tau 13 (the time delay between the neuron 1 and the neuron 3), namely tau 13 = tau 12 + tau 23. As shown in fig. 1, the apparatus includes: an arbitrary waveform generator AWG capable of generating a target signal (pulse sequence), wherein the signal output from the AWG is modulated onto a continuous wave generated by a tunable laser through a Mach-Zehnder modulator, and the signal output from the modulator (the injection delay can be adjusted by adjusting delay lines DL1, DL2 and DL 3) is injected into three neurons of a neural network without delay. The signals output by each neuron are converted into electric signals by PD and then input into an oscilloscope OSC, the signals of the oscilloscope are collected and processed and operated by a computer, and the stimulation signals input into each neuron are regulated and controlled according to the operation result so as to realize the identification of the pulse sequence.

Specifically, the invention converts the electric pulse generated by the AWG into the optical pulse, and the core devices of the whole system are all optical devices, so that the band limitation of electronic devices can be broken through, the system speed is greatly improved, and the energy consumption of the system can be greatly reduced.

In addition, the three VCSELs photonic neurons include neuron 1, neuron 2, neuron 3; the time delay between neuron 1 and neuron 2 is τ 12; the time delay between neuron 2 and neuron 3 is τ 23, and the time delay between neuron 1 and neuron 3 is τ 13; τ 13 = τ 12 + τ 23.

The three-way signal comprises a neuron one way, a neuron two way and a neuron three way;

one way of the neuron is as follows: the signal output from the modulator MZM is divided into two paths after passing through the coupler FC1, wherein one path of signal is divided into two parts after passing through the coupler FC2, and one part of signal enters the VCSEL1 after passing through the polarization controller PC3, the adjustable attenuator VA3, the optical fiber delay line DL3 and the optical fiber circulator OC 1;

the two paths of the neuron consist of two paths; one path is as follows: one part of the signal divided by the coupler FC1 is divided into two parts after passing through a coupler FC2, wherein one part of the signal enters the VCSEL2 after passing through a polarization controller PC2, an adjustable attenuator VA2, an optical fiber delay line DL2 and a coupler FC 4;

the other path is as follows: the signal output from the optical fiber circulator OC1 is divided into two parts by a coupler FC3, wherein one part enters the VCSEL2 after passing through a polarization controller PC4, an adjustable attenuator VA4, an optical fiber delay line DL4 and the coupler FC 4;

the 2 nd route is: the signal output from the VCSEL2 enters the VCSEL3 after passing through a coupler FC5, a polarization controller PC5, an adjustable attenuator VA5, an optical fiber delay line DL5, a coupler FC6, a coupler FC7 and an optical fiber circulator OC 2;

the 3 rd path is: the other path of signal split from the coupler FC3 is also input to the coupler FC6 after passing through the coupler FC8, the polarization controller PC6, the adjustable attenuator VA6, and the fiber delay line DL6, and a signal output from the coupler FC6 enters the VCSEL3 after passing through the coupler FC7 and the fiber circulator OC 2.

In particular, the fiber coupler FC is used to split the input light into two parts; the polarization controller PC is used for controlling the polarization state of input light; the adjustable attenuator VA is used for adjusting the intensity of input light; the optical fiber delay line DL is used for adjusting the delay time of input light; the optical fiber circulator OC is used to control unidirectional injection of incident light into and unidirectional output from VCSEL neurons. The signal output from the modulator MZM is divided into two paths after passing through a coupler FC1 of 3, wherein 30% of the signal enters the VCSEL3 after passing through a polarization controller PC1, an adjustable attenuator VA1, an optical fiber delay line DL1, a coupler FC7 and an optical fiber circulator OC2, the other 70% of the signal is divided into two parts after passing through a coupler FC2 of 5.

The signal trend of the device provided by the invention is as follows: firstly, the VCSEL1 nerve can respond spike signals under all external stimuli by adjusting a polarization controller PC3 and an adjustable attenuator VA 3; then, polarization states of two paths of signals injected into the VCSEL2 neurons are consistent through the polarization controllers PC4 and PC2, then the intensity of the two paths of injected signals is adjusted through adjusting VA2 and VA4, and meanwhile, time delay of the two paths of injected signals is adjusted through adjusting DL2 and DL4, so that the neuron VCSEL2 can respond to pulse signals when being stimulated by the two paths of signals at the same time; next, the polarization state of the three-way signal injected into the VCSEL3 neuron is then made uniform by the polarization controllers PC1, PC5, and PC6, and then the intensity of the three-way injected signal is adjusted by adjusting VA1, VA5, and VA6, while the time delay of the three-way injected signal is adjusted by DL1, DL5, and DL6 so that τ 13 = τ 12 + τ 23, thereby causing the neuron VCSEL3 to respond to a pulse signal when subjected to three-way stimulation simultaneously.

According to the invention, different response thresholds of each neuron are utilized, namely the response thresholds of the three neurons of the VCSEL1, the VCSEL2 and the VCSEL3 are gradually increased, so that the VCSEL1 can generate pulse response to all stimuli output by the MZM, the VCSEL2 can have pulse response only after being simultaneously subjected to two paths of stimulation signals, the VCSEL3 can generate pulse response only after being simultaneously subjected to 3 paths of stimulation signals, and the injection time of each path of stimulation signal can be adjusted through a delay line, so that specific mode identification can be realized on any pulse sequence by using the three cascaded VCSELs neurons.

step 1: outputting the waveform of the pulse sequence;

and 3, step 3: dividing the optical pulse sequence signal into three paths, and injecting the three paths of signals into three cascaded VCSELs photon neurons of a neural network respectively;

the three VCSELs photon neurons comprise neuron 1, neuron 2 and neuron 3; the time delay between neuron 1 and neuron 2 is τ 12; the time delay between neuron 2 and neuron 3 is τ 23, and the time delay between neuron 1 and neuron 3 is τ 13; τ 13 = τ 12 + τ 23;

Specifically, since the response threshold of each neuron gradually increases, the delay times τ 13 and τ 23 can be adjusted by adjusting the delay lines DL5 and DL6, thereby ensuring τ 13 = τ 12 + τ 23. Therefore, the three neurons can respectively respond to partial pulse signals in the target pulse sequence. In the process of pattern recognition, firstly, signals output by each neuron are converted into electric signals through PD and then input into an Oscilloscope (OSC), then the signals are input into a computer through a data acquisition card for data processing, an unsupervised pulse time-dependent plasticity mechanism (namely STDP learning rule) is utilized for training, and in the training process, the weight of each external stimulation signal (realized by adjusting an adjustable attenuator), the time delay tau 13 between the neuron 1 and the neuron 3 and the time delay tau 23 between the neuron 2 and the neuron 3 are adjusted according to the comparison result of the processing result of the computer and a target signal. Finally, the weight information determined after training is used for controlling corresponding parameters of the system (such as the strength of a stimulation signal entering each neuron), so that the pattern recognition of the pulse sequence is realized. In addition, the error code information of the signal obtained after the processing of the computer and the target signal can be calculated by comparing the signal with the target signal, so that the target identification performance of the system can be effectively evaluated.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, and not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A pattern recognition device based on cascade VCSELs photonic nerves, characterized by comprising: the device comprises an arbitrary waveform generator capable of generating a pulse sequence, a Mach-Zehnder modulator, a tunable laser, an optical pulse sequence, an OSC (optical signal processor), a computer and a controller, wherein a signal output from the arbitrary waveform generator is modulated by the Mach-Zehnder modulator to obtain a continuous wave generated by the tunable laser to obtain the optical pulse sequence, the signal output from the Mach-Zehnder modulator is divided into three paths, the three paths of signals are respectively injected into three cascade VCSELs photon neurons without delay, the signal output from each neuron is converted into an electric signal by a photoelectric detector PD and then is input into the OSC, the signal of the OSC is collected and is processed and operated by the computer, and the pulse signal input into the neuron is adjusted and controlled according to the operation result to realize the mode identification of the pulse sequence;

the three cascade VCSELs photon neurons comprise a neuron 1, a neuron 2 and a neuron 3; the time delay between neuron 1 and neuron 2 is τ 12; the time delay between neuron 2 and neuron 3 is τ 23, and the time delay between neuron 1 and neuron 3 is τ 13; τ 13 = τ 12 + τ 23;

the two paths of the neuron are composed of two paths; one path is as follows: a part of signals divided by the coupler FC1 is divided into two parts again after passing through a coupler FC2, wherein one part of the signals enters the VCSEL2 after passing through a polarization controller PC2, an adjustable attenuator VA2, an optical fiber delay line DL2 and a coupler FC 4;

2. The cascade VCSELs photonic nerve-based pattern recognition device of claim 1, wherein the signal of coupler FC1 is divided into 3; 70% of the output signal passes through the coupler FC2; the signals of the couplers FC2, FC3, FC4, FC6 and FC7 are divided into 5.

3. The cascade VCSELs photonic nerve-based pattern recognition device of claim 1, wherein the signals output from the neuron 1, the neuron 2 and the neuron 3 are converted into electrical signals by PD1, PD2 and PD3, respectively;

4. The cascade VCSELs photonic nerve-based pattern recognition device of claim 3, wherein the signals of the couplers FC8 and FC5 are divided into 90; of which 10% of the signals are input to PD1 and PD2.

5. A pattern recognition method based on cascade VCSELs photonic nerves is characterized by comprising the following steps:

step 1: outputting the waveform of the pulse sequence;

the two paths of the neuron consist of two paths; one path is as follows: a part of signals divided by the coupler FC1 is divided into two parts again after passing through a coupler FC2, wherein one part of the signals enters the VCSEL2 after passing through a polarization controller PC2, an adjustable attenuator VA2, an optical fiber delay line DL2 and a coupler FC 4;

the 3 rd path is: the other path of signal split from the coupler FC3 is also input into the coupler FC6 after passing through the coupler FC8, the polarization controller PC6, the adjustable attenuator VA6 and the optical fiber delay line DL6, and a signal output from the coupler FC6 enters the VCSEL3 after passing through the coupler FC7 and the optical fiber circulator OC 2;

and 4, step 4: and collecting signals output by each neuron, performing processing operation by using a computer, and adjusting and controlling pulse signals input into the neurons according to an operation result so as to realize pattern recognition of a pulse sequence.

6. The method of claim 5, wherein the processing operation in step 4 comprises:

training by using an unsupervised pulse time-dependent plasticity mechanism, and adjusting the weight and injection delay of each external stimulation signal according to the comparison result of the computer processing result and the target signal in the training, namely: the stimulation intensity injected into each neuron; and finally, the weight and the time delay information determined after training are used for controlling corresponding parameters of the system, so that the mode identification of the pulse sequence is realized.