JP2020067984A - Information processing system and information processing method - Google Patents

Abstract

To provide an information processing system capable of efficiently realizing time series data processing by machine learning using a physical system and an information processing method.SOLUTION: An information processing system 1S includes: a signal conversion part 2 that retains at least a part of previously input signals as an internal state and converts the input signals into a non-linear state using the internal state; a signal input part 1 that inputs the input signals into the signal conversion part; a signal output part 3 that reads an output signal which is a signal converted by the signal conversion part 2; and a calculation unit 101 that calculates a linear combination of the read output signals and the optimum combination strength. The signal conversion part 2 includes an optical amplifier 107 that nonlinearly converts an input signal as an optical signal into an output signal as an optical signal.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing system and an information processing method for realizing time series data processing by machine learning using a physical system.

  In the brain of a living body, many neurons form a network by synaptic connections, and it is considered that these are responsible for information processing. A system that imitates such a structure on a computer is called a neural network.

  Models of neural networks are roughly classified into two types. A model in which the signal flow is unidirectional and has a loop-free structure is called a feedforward neural network. On the other hand, a model having self-feedback inside and a structure in which signals propagate in both directions is called a recurrent neural network.

  Neural networks perform optimization by changing the connection strength of each neuron, and acquire the ability to solve various tasks. Optimization methods are roughly classified into two types. The method of learning by using the correct answer data given in advance is called supervised learning, and the method of extracting the data structure without using the correct answer data is called unsupervised learning.

  Reservoir computing is one of the forms for realizing supervised learning by a recurrent neural network (see Non-Patent Document 1). Reservoir computing is performed by the following procedure. The input layer, the intermediate layer, and the output layer form a neural network. Based on the output results obtained by inputting the teacher data, the bond strengths of the input layer to the intermediate layer, the intermediate layer to the intermediate layer, and the intermediate layer to the output layer are optimized. To do. In the case of the recurrent neural network, the learning algorithm becomes complicated and a huge amount of calculation is required to optimize the connection strength. On the other hand, in reservoir computing, the bond strength from the input layer to the intermediate layer and the bond strength from the intermediate layer to the intermediate layer are fixed at fixed values, respectively, and the bond strength from the intermediate layer to the output layer is fixed. Learn only for. In that case, the learning algorithm is simplified and the amount of calculation required for optimization can be reduced. In a situation where learning is performed in this manner, the intermediate layer is called a reservoir.

  Here, in order to efficiently perform time-series data processing by reservoir computing, the reservoir needs to satisfy three characteristics (see Non-Patent Document 2).

  The first feature is "non-linearity". It is shown that the input signal can be transformed nonlinearly and transformed into a higher dimensional state space. Such non-linearity can be realized by increasing the number of neurons contained in the reservoir or recursive coupling.

  The second feature is the "short-term memory effect". It shows that the state of the reservoir is affected by the latest past input signal, but at the same time is not affected by the old past input signal. This property is necessary when performing time-series data processing in which the latest past information becomes more important, such as voice recognition.

  The third characteristic is "reproducibility". It shows that the same output signal is always obtained for the same input signal. If reproducibility is not obtained, the binding strength cannot be optimized by learning.

  In order to realize a reservoir satisfying the above three characteristics, there are a method using a computer simulation and a method using a physical system. In order to perform complicated time series data processing, higher-order nonlinearity is required. When computer simulation is used, a huge amount of calculation is required to obtain high-order nonlinearity. Therefore, attempts have been made to perform reservoir computing using a physical system.

  Here, there is a technique using a loop structure using an electric circuit and an optical fiber, which utilizes the nonlinearity and reproducibility of the light intensity modulator. For example, the technique described in Non-Patent Document 3 has high reproducibility because it is composed of optical components that show a simple response. Further, there is disclosed a technique of using a feedback loop by a semiconductor laser and return light, which utilizes the nonlinearity and reproducibility of the semiconductor laser. For example, the technique described in Non-Patent Document 4 generates chaos when returning light is input to the semiconductor laser, and thus high nonlinearity is obtained.

International Publication No. 2017/131081 JP-A-8-18139

H. Jaeger and H. Haas, Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science 304, 78 (2004). L. Appeltant et al., Information processing using a single dynamical node as complex system. Nature Communications 2, 468 (2011). Y. Paquot et al., Optoelectronic Reservoir Computing. Scientific Reports 2, 287 (2012). J. Nakayama et al., Laser dynamical reservoir computing with consistency: an approach of a chaos mask signal, Optics Express 24, 8679 (2016).

  However, in general, it is difficult to achieve stable control of a physical system that exhibits a complicated response, and thus it is difficult to achieve both high nonlinearity and reproducibility.

  An object of the present invention is to solve the above problems and provide an information processing system and an information processing method that efficiently realize time series data processing by machine learning using a physical system.

  In order to achieve the above object, the information processing system according to the present invention holds at least a part of a signal input in the past as an internal state, and a signal conversion unit that nonlinearly converts the input signal in the internal state, A signal input unit that inputs the input signal to the signal conversion unit, a signal output unit that reads an output signal that is a signal converted by the signal conversion unit, a linear combination of the read output signals, and an optimum coupling strength. And a calculation unit for calculating, wherein the signal conversion unit includes an optical amplifier that nonlinearly converts the input signal that is an optical signal into the output signal that is an optical signal.

  According to the present invention, it is possible to provide an information processing system that efficiently realizes time series data processing by machine learning.

  Problems, configurations, and effects other than those described above will be clarified by the following description of modes for carrying out the invention.

It is a figure which shows the structural example of the information processing system which concerns on Example 1 of this invention. It is a figure which shows the structural example of a light intensity modulator. It is a figure which shows the structural example of an optical amplifier. 5 is a flowchart illustrating an operation example of a calculation unit of the information processing system according to the first embodiment of the present invention. It is a figure which shows the structural example of the information processing system which concerns on Example 2 of this invention. It is a figure which shows the simulation result by the information processing system which concerns on Example 2 of this invention. It is a figure which shows the structural example of the information processing system which concerns on Example 3 of this invention. It is a figure which shows the structural example of the computer which implement | achieves the computer of the information processing system which concerns on each Example of this invention.

  Embodiments of the present invention will be described below with reference to the drawings. The following description and drawings are examples for explaining the present invention, and are appropriately omitted and simplified for clarity of the description. The present invention can be implemented in various other modes. Unless otherwise specified, each component may be singular or plural.

  When there are a plurality of constituent elements having the same or similar functions, the same reference numerals may be given with different subscripts. However, when it is not necessary to distinguish between these plural components, the subscripts may be omitted in the description.

  Further, in the following description, a process performed by executing a program may be described, but the program is executed by a processor (for example, a CPU or GPU) so that a predetermined process can be performed as appropriate for a storage resource ( For example, since the processing is performed using a memory) and / or an interface device (for example, a communication port), the main body of processing may be the processor. Similarly, the main body of processing executed by executing the program may be a controller having a processor, a device, a system, a computer, or a node. The main body of the processing performed by executing the program may be the arithmetic unit, and may include a dedicated circuit (for example, FPGA or ASIC) for performing the specific processing.

  Further, the program may be installed from a program source to a device such as a computer. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is the program distribution server, the program distribution server may include a processor and a storage resource for storing the distribution target program, and the processor of the program distribution server may distribute the distribution target program to another computer. Further, in the following description, two or more programs may be realized as one program, or one program may be realized as two or more programs.

  The information processing system of the present invention includes a signal conversion unit. The signal conversion unit is mounted by a physical system including an optical amplifier and functions as a reservoir. A signal input unit that inputs a signal to the signal conversion unit, a signal output unit that reads an output signal that is a signal converted by the signal conversion unit, a linear combination of the read output signals, and an optimal connection strength of the neural network are calculated. By including the calculation unit, it operates as a reservoir computer. In the reservoir computing, the part that realizes the input layer is the signal input part, the part that realizes the intermediate layer or the reservoir is the signal conversion part, the part that realizes the output layer is the signal output part, and the part that learns the coupling strength is the calculation part. Become.

  "Reservoir computing" is one of the methods of supervised machine learning by a recursive neural network, and is suitable for processing time-series-dependent information such as speech. It is suitable for operation. In the "reservoir computing", an intermediate layer which is a recursive neural network having a fixed weighting among the input layer, the intermediate layer and the output layer forming the neural network is called "reservoir". In "reservoir computing", the weight of the linear combination from the input layer to the reservoir is fixed, the weight of the linear combination in the reservoir is also fixed, and only the weight of the linear combination from the reservoir to the output layer is updated by learning. .

  Each embodiment and each modification described below can be combined in part or in whole within the range of matching without departing from the scope of the technical idea of the present invention.

(Configuration of Information Processing System of Example 1)
FIG. 1 is a diagram showing a configuration example of an information processing system according to the first embodiment of the present invention. In the information processing system 1S according to the first embodiment, the signal input unit 1 includes a D / A converter 102, a light source 103, and a light intensity modulator 104. The respective parts are connected by an electric wire 105 shown by a dotted arrow or an optical fiber 106 shown by a solid arrow. The calculator (calculator) 101 includes a D / A converter 102 and an A / D converter 109, and can output and measure an electric signal. The electric signal output from the D / A converter 102 based on the audio data is input to the light intensity modulator 104. Light from the light source 103 is input to the light intensity modulator 104 and is subjected to intensity modulation based on an electric signal. It is assumed that the electrical signal based on the audio data is output from the D / A converter 102 and input to the light intensity modulator 104, but data other than the audio data may be used.

(Structure of intensity modulator)
FIG. 2 is a diagram showing a configuration example of the light intensity modulator. An electro-optic crystal 201 is enclosed in the light intensity modulator 104. When the input optical signal 202 is input, the output optical signal 203 is output through the optical waveguide interferometer formed in the electro-optic crystal 201. When the input electric signal 204 is applied to the electro-optic crystal 201, the refractive index of the optical waveguide changes, and the optical signals interfere with each other to cancel each other, and the intensity of the output optical signal 203 changes. With the above operation, the output optical signal 203 shows a non-linear response to the input electric signal 204. The intensity-modulated light output from the light intensity modulator 104 becomes an optical signal output from the signal input unit.

  Returning to the explanation of FIG. The signal converter 2 is composed of an optical amplifier 107. The optical signal subjected to the intensity modulation output from the signal input unit is input to the optical amplifier 107.

(Structure of optical amplifier)
FIG. 3 is a diagram showing a configuration example of the optical amplifier. The optical amplifier 107 is an EDFA (Erbium Doped optical Fiber Amplifier), in which an impurity-doped optical fiber 301, two wavelength division multiplexing optical couplers 302 and 303, and a pumping light source 304 are enclosed.

  Excitation light 305 is output from the excitation light source 304. The input optical signal 306 is multiplexed with the pumping light 305 by the wavelength division multiplexing optical coupler 302 and input to the impurity-doped optical fiber 301. The excitation light 305 excites the energy of the impurity ions in the impurity-doped optical fiber 301, and the input optical signal 306 causes stimulated emission of the energy of the impurity ions. The remaining pumping light 307 is demultiplexed by the wavelength division multiplexing optical coupler 303, and the input optical signal 306 amplified by stimulated emission is output as the output optical signal 308.

  With the above operation, the output optical signal 308 exhibits a non-linear response to the input optical signal 306. Since the excited state transits to the ground state in a certain time, it has a short-term memory effect. If the intensity of the input light and the internal energy distribution are the same, the light of the same intensity is output, so reproducibility is obtained.

  As described above, all the three features that the reservoir should have are achieved. The optical signal output from the optical amplifier 107 that has undergone the intensity conversion becomes an optical signal output from the signal conversion unit.

  The optical amplifier 107 is not limited to the EDFA, and may be a semiconductor optical amplifier that can expect a memory effect with a time width different from that of the EDFA.

  Returning to the explanation of FIG. The signal output unit 3 includes a photodetector 108 and an A / D converter 109. The optical signal output from the signal conversion unit 2 is input to the photodetector 108 and converted into an electrical signal. The electric signal is input to the A / D converter 109 included in the computer 101, and the measured value is recorded in the computer 101.

  The calculation unit is composed of the computer 101. The computer 101 includes an arithmetic processing unit such as a CPU (Central Processing Unit). FIG. 4 is a flowchart showing an operation example of the calculation unit of the information processing system according to the first embodiment of the present invention. First, the computer 101 inputs the teacher data to the light intensity modulator 104 through the D / A converter 102 (step S401). Next, in the computer 101, the signal input unit 1, the signal conversion unit 2, and the signal output unit 3 operate, and based on the measured values obtained through the A / D converter 109, the coupling strength set inside the computer 101. The output result is calculated from (step S402).

  Next, the computer 101 compares the output result acquired in step S402 with the answer of the teacher data to determine whether they match (step S403). If they do not match (step S403: YES), the computer 101 updates the coupling strength so that the error becomes small (step S404), and inputs the teacher data to the light intensity modulator 104 again. The coupling strength (linear weight) is optimized by repeating the loop of steps S401 → S402 → S403: NO → S404 → S401 until the output result and the answer of the teacher data match. Since the optimization is performed only for the last output layer, the optimization is completed in a shorter time than the machine learning method for optimizing all bond strengths.

  When the output result obtained after the optimization is completed and the answer of the teacher data match (step S403: YES), the computer 101 inputs unknown data (step S405). When the output result obtained after the optimization is completed and the answer of the teacher data match (step S403: YES), it means that the classification of the teacher data is completed. When the output result obtained after the optimization is completed and the answer of the teacher data match, the computer 101 also obtains the output result for the unknown data (step S406).

  As described above, according to the present embodiment, by utilizing the nonlinear dynamics of the optical amplifier, both high nonlinearity and reproducibility are achieved in time series data processing by machine learning using a physical system, and time series data processing is performed. Can be efficiently realized.

  In the first embodiment, the optical amplifier 107 alone is used as the signal conversion unit, but in the second embodiment, the optical amplifier is incorporated into the loop formed of the optical fiber to further improve the nonlinearity and the short-term memory effect of the signal conversion unit. You can control it freely.

  FIG. 5 is a diagram showing a configuration example of an information processing system according to the second embodiment of the present invention. In the information processing system 2S of the second embodiment, as compared with the information processing system 1S of the first embodiment, the optical intensity modulator 104, the optical amplifier 107, the optical coupler 501, the optical isolator 502, and the optical filter 503 are respectively provided in the optical fiber 106. It has a loop structure L2 configured by being connected by.

  A part of the optical signal branched from the electro-optical coupler 201 is input to the photodetector 108. The electrical signal output from the photodetector 108 is input to the A / D converter 109, and the computer 101 extracts the clock frequency. When the light phase of the loop structure L2 changes due to a change in temperature or the like, it appears as a change in the clock frequency. By controlling the clock frequency of the electric signal output from the D / A converter 102 based on the extracted clock frequency, the optical signal generated in the loop structure L2 is stabilized. After that, by inputting the teacher data and optimizing the coupling coefficient as in the first embodiment, the voice data can be classified.

(Simulation result of Example 2)
FIG. 6 is a diagram showing a simulation result by the information processing system according to the second embodiment of the present invention. An input electric signal 601 to the reservoir output from the D / A converter 102 and an output electric signal 602 from the reservoir input to the A / D converter 109 are shown. The input electrical signal 601 is a signal having a waveform with two amplitudes, a relatively large first amplitude and a relatively small second amplitude, whereas the output electrical signal 602 is compared to the input electrical signal 601. Resulting in a signal with a complicated waveform. This is because the output electric signal 602 is influenced by the past input electric signal 601, which indicates that the nonlinearity and the short-term memory effect are exhibited. It was also confirmed that when the input electric signal 601 has the same waveform, the output electric signal 602 also has the same waveform, and reproducibility is obtained.

  As described above, according to the present embodiment, the time series data processing can be efficiently realized by utilizing the nonlinear dynamics of the optical amplifier and the loop structure of the optical fiber.

  In the second embodiment, the optical signal is stabilized by the clock extraction and the control of the clock frequency by the computer 101. However, in the third embodiment, the reproduction mode synchronization is performed by the method described in Japanese Patent Laid-Open No. 2004-242, so that a further optical signal is obtained. Can be stabilized.

(Information Processing System of Example 2)
FIG. 7 is a diagram showing a configuration example of an information processing system according to the third embodiment of the present invention. The information processing system 3S of the third embodiment is different from the information processing system 2S of the second embodiment in that the feedback control system includes the photodetector 108, the clock extractor 701, the phase shifter 702, and the optical phase modulator 703. Have. The reservoir of FIG. 7 has a configuration called reproduction mode synchronization. A part of the optical signal branched from the optical coupler 501 is input to the photodetector 108. A part of the electric signal output from the photodetector 108 is input to the clock extractor 701, and the clock frequency is extracted. When the light phase of the loop structure L2 changes due to a change in temperature or the like, it appears as a change in the clock frequency. By inputting the electric signal output from the clock extractor 701 to the optical phase modulator 703 through the phase shifter 702, the phase change generated in the loop structure L2 is corrected and the optical signal generated in the loop structure L2 is stabilized. . After that, by inputting the teacher data and optimizing the coupling coefficient as in the second embodiment, the voice data can be classified.

  As described above, according to the present embodiment, by utilizing the nonlinear dynamics of the optical amplifier and the loop structure of the optical fiber and performing the stabilization by the reproduction mode synchronization, it is possible to efficiently realize the time series data processing. it can.

  FIG. 8 is a diagram showing a configuration example of a computer that realizes a computer of the information processing system according to each embodiment of the present invention. A computer 5000 that implements the computer 101 of the information processing systems 1S to 3S according to the first to third embodiments includes a CPU (Central Processing Unit) 5300, a memory 5400 such as a RAM (Random Access Memory), an input device 5600 (for example, a keyboard, a mouse, A touch panel or the like) and an output device 5700 (eg, a video graphics card connected to an external display monitor) are interconnected through a memory controller 5500. In the computer 5000, a predetermined program is read from the external storage device 5800 such as SSD or HDD via the I / O (Input / Output) controller 5200, and executed by the cooperation of the CPU 5300 and the memory 5400. A computer 101 of the information processing system 1S to 3S is realized. Alternatively, the program for realizing the computer 101 of the information processing systems 1S to 3S may be acquired from an external computer by communication via the network interface 5100.

  The present invention is not limited to the above-mentioned embodiments, but includes various modifications. For example, the above-described embodiments have been described in detail for the purpose of explaining the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of a certain embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of a certain embodiment. Further, it is possible to add / delete / replace other configurations with respect to a part of the configurations of the respective embodiments.

1S, 2S, 3S ... Information processing system, 1 ... Signal input section, 2 ... Signal conversion section, 3 ... Signal output section, L2 ... Loop structure, FC ... Feedback control system, 101 ... Calculator (calculation section), 102 ... D / A converter, 103 ... Light source, 104 ... Light intensity modulator, 105 ... Electric wire, 106 ... Optical fiber, 107 ... Optical amplifier, 108 ... Photodetector, 109 ... A / D conversion, 201 ... Electro-optic crystal, 202 Input optical signal, 203 ... Output optical signal, 204 ... Input electric signal, 301 ... Impurity doped optical fiber, 302 ... Wavelength division multiplexing optical coupler, 303 ... Wavelength division multiplexing optical coupler, 304 ... Excitation light source, 305 ... Excitation Light, 306 ... Input optical signal, 307 ... Excitation light, 308 ... Output optical signal, 501 ... Optical coupler, 502 ... Optical isolator, 503 ... Optical filter, 601 ... Input electric signal to reservoir, 6 2 ... output electrical signals from the reservoir, 701 ... clock extractor, 702 ... phase shifter, 703 ... optical phase modulator.

Claims (6)

A signal conversion unit that holds at least a part of a signal input in the past as an internal state and that converts the input signal into a non-linear state in the internal state,
A signal input unit for inputting the input signal to the signal conversion unit,
A signal output unit for reading an output signal which is a signal converted by the signal conversion unit;
And a calculation unit that calculates a linear combination of the read output signals and an optimum combination strength,
The information processing system, wherein the signal conversion unit includes an optical amplifier that nonlinearly converts the input signal that is an optical signal into the output signal that is an optical signal. The information processing system according to claim 1, wherein the non-linear conversion from the input signal to the output signal is reproducible. The information processing system according to claim 2, wherein the optical amplifier is incorporated in a loop configured by an optical fiber. The optical amplifier and the optical phase modulator are incorporated in a loop composed of an optical fiber,
A clock extractor for outputting an electric signal indicating a clock frequency extracted from the optical signal output from the loop,
A phase shifter for inputting the electric signal output from the clock extractor to the optical phase modulator,
Have
The information processing system according to claim 2, wherein the optical phase modulator corrects a phase change generated in the loop based on the input electric signal. The signal conversion unit, the signal input unit, and the signal output unit form a loop,
The information processing system according to claim 1, wherein at least a part of the output signal affects a signal that is next input to the signal conversion unit. An information processing method executed by an information processing system,
The information processing system,
A signal conversion unit that holds at least a part of a signal input in the past as an internal state and that converts the input signal into a non-linear state in the internal state,
A signal input unit for inputting the input signal to the signal conversion unit,
A signal output unit for reading an output signal which is a signal converted by the signal conversion unit;
And a calculation unit that calculates a linear combination of the read output signals and an optimum combination strength,
The signal conversion unit includes an optical amplifier,
The information processing method in an information processing system, wherein the optical amplifier nonlinearly converts the input signal that is an optical signal into the output signal that is an optical signal.

Images