Patent Application
20220137485
The present invention relates to an optical signal processing apparatus that can be applied to optical reservoir computing.
In recent years, an environment has been constructed to acquire a large amount of data from various sensors via the Internet, and research and business for analyzing the large amount of acquired data and performing highly accurate knowledge processing and future prediction have been actively carried out. In general, because analysis of a large amount of data requires time and incurs costs such as power consumption, computing devices having high speed and high efficiency are required. As a computing scheme for such information processing, an optical computing technique called reservoir computing (RC), which imitates signal processing of the cerebellum, has been proposed. Optical computing devices using a dynamical system called RC are attracting attention because such devices are likely to have both high speed and high efficiency.
In examples of applications of optical RC in the related art, examples of solving a one-dimensional input and output problem such as a chaos approximation problem and NARMA 10 have mainly been reported (for example, see Non Patent Literature 1). Further, it is necessary to improve computing accuracy in order to further widen a range of applications of optical RC.
In RC, it is generally known that the computing accuracy is improved by an increase in the number of nodes of a reservoir layer. In the case of optical RC, because the nodes of the reservoir layer are represented by the number of optical pulses that circulate around a fiber ring, computing processing is performed by time-multiplexing the circulating optical pulses in order to increase the number of nodes and improve the computing accuracy. However, all of tasks and nodes are expanded on a time axis to input data, and thus the higher the number of nodes, the longer the data time to enter into the optical RC, which leads to a problem of reduced throughput.
An object of the present invention is to provide an optical signal processing device capable of improving computing accuracy without increasing the number of nodes of a reservoir layer.
In order to achieve such an object, an aspect of the present invention is an optical signal processing apparatus for converting an input one-dimensional signal to an optical signal to perform signal processing, the optical signal processing apparatus including: an input unit configured to perform linear processing on the input one-dimensional signal to convert the input one-dimensional signal to an optical signal of multi-wavelength; a reservoir unit connected to an output of the input unit and configured to perform linear processing and nonlinear processing on the optical signal; and an output unit connected to an output of the reservoir unit and configured to convert the optical signal to an electrical signal and perform linear processing to output a one-dimensional output.
According to the present invention, by making optical RC that expand a node in the wavelength direction instead of expanding the node in the time axis direction, the throughput of the optical RC can be improved without increasing the input time of data.
Hereinafter, an embodiment of the present disclosure will be described in detail with reference to the drawings.
The optical signal processing apparatus of the present embodiment improves throughput of optical RC by expanding a node in the wavelength direction instead of expanding the node in the time axis direction.
Input Unit
As the multi-wavelength light source 111, an amplified spontaneous emission (ASE) light source, a broadband light source, a plurality of single wavelength light sources, or the like can be used. When the ASE light source is used, the light source can be operated relatively stably because only intensity information is used. When the broadband light source is used, an amount of information can be made twice or more because both intensity information and phase information are used. When the plurality of single wavelength light sources are used, intensity information and phase information can be adjusted and used for each wavelength.
The input unit 11 executes processing indicated in Equation (1).
Math. 1
x
in(λ,t)=win(λ)·u(λ,t) (1)
A weight win(λ) is cumulated to a time series signal task u(λ, t) to generate an input signal xin(λ, t) to the reservoir 12. This weight win(λ) is a value given in advance prior to performing training of the optical RC and is not be changed through training and testing.
The time series signal task u(λ, t) is generated by passing through the time modulation unit 112 connected to the light source 111 (waveform example A in
The time series signal task u(λ, t) is input to the signal modulation unit 113 and is multiplied by the weight win(λ) for each wavelength (waveform example B in
Reservoir Unit
The reservoir unit 12 executes processing indicated in Equation (2).
Math. 2
x
re(λ,t)=cos2[Σλwre(λ)·xin(λ,t−1)+ϕ]×Ω(λ)+xin(λ,t) (2)
The input xin(λ, t−1) from the input unit 11 at time t−1 is multiplied by the weight wre(λ) and signals in the wavelength direction are cumulated. Modulation of the cost function is performed on a signal Ω(λ) weighted in the wavelength direction. Finally, the input xin(λ, t) from the input unit 11 at time t is added to generate the state xre(λ, t) of the reservoir unit 12 at time t. Here, the weight wre(λ) and the signal Ω(λ) are values given in advance before performing the training of the optical RC and are not changed through training and testing. Note that φ is a bias in the time modulation unit described below.
The input xin(λ, t−1) from the input unit 11 input at time t−1 as a first input to the reservoir unit 12 passes through the merging unit 121 and the branch unit 122 and is input to the signal modulation unit 123. In the signal modulation unit 123, the input is multiplied by the weight wre(λ) for each wavelength. A wavelength selective switch, MEMS, or the like can be used as the signal modulation unit 123. When used, the wavelength selective switch can be operated with low power consumption. When the MEMS is used, an extinction ratio can be increased to improve computation performance.
The optical signal that has passed through the signal modulation unit 123 is converted to an electrical signal by a light reception unit 124. Here, rather than providing a light reception unit for each wavelength, a weighted multi-wavelength signal is received by one light reception unit. The electrical signal output from the light reception unit 123 loses information in the wavelength direction, and thus it is possible to perform calculation as if signals in the wavelength direction were added.
On the other hand, a signal light input from a multi-wavelength light source 125 is input to a signal modulation unit 126. As the multi-wavelength light source 125, an ASE light source, a broadband light source, a plurality of single wavelength light sources, or the like can be used. When the ASE light source is used, the light source can be operated relatively stably because only intensity information is used. When the broadband light source is used, an amount of information can be made twice or more because both intensity information and phase information are used. When the plurality of single wavelength light sources are used, intensity information and phase information can be adjusted and used for each wavelength.
In the signal modulation unit 126, the weight Ω(λ) for wavelength is cumulated. The electrical signal for modulation is input from a personal computer, an FPGA, or the like. When the personal computer is used, data can be easily rewritten. When the FPGA is used, a high speed electrical signal can be input. A wavelength selective switch, MEMS, or the like can be used as the signal modulation unit 126. When used, the wavelength selective switch can be operated with low power consumption. When the MEMS is used, an extinction ratio can be increased to improve computation performance.
The weighted signal light Ω(λ) and the electrical signal converted by the light reception unit 124 are input to a time modulation unit 127 to newly generate a modulated signal. As the time modulation unit 127, an optical attenuator such as an LN modulator or an optical amplifier such as a semiconductor optical amplifier can be used. When the optical attenuator is used, it is possible to shorten a computing time because modulation can be performed at high speed. When the optical amplifier is used, it is possible to suppress deterioration of computing capability due to a loss because a signal can be amplified.
The signal light that has passed through the time modulation unit 127 is input to the merging unit 121 as a second input to the reservoir unit 12 and added to the input xin(λ, t) from the input unit at time t. As the merging unit 121, a space optical system such as a beam splitter, a fiber optical system such as an optical coupler, or a planar optical system such as a PLC can be used. When the spatial optical system is used, polarization of light is not easily changed, and thus it is possible to increase computation performance. When the fiber optical system is used, configuration of a device can be relatively easily changed by changing optical fiber connection. When the planar optical system is used, loss in an optical component can be suppressed, and thus it is possible to increase computation performance.
The signal light xre(λ, t) that has passed through the merging unit 122 is input to the branch unit 122 to be branched into two paths, that is, the reservoir unit 12 and the output unit 13. As the branch unit 122, a space optical system such as a beam splitter, a fiber optical system such as an optical coupler, a planar optical system such as a PLC, or the like can be used. When the spatial optical system is used, polarization of light is not easily changed, and thus it is possible to increase computation performance. When the fiber optical system is used, configuration of a device can be relatively easily changed by changing optical fiber connection. When the planar optical system is used, loss in an optical component can be suppressed, and thus it is possible to increase computation performance.
Output Unit
The output unit 13 executes processing indicated in Equation (3).
Math. 3
x
out(t)=Σλwout(λ)·xre(λ,t) (3)
The input xre(λ, t) from the reservoir unit 12 at time t is multiplied by the weight wout(λ) and data in the wavelength direction is added to generate an output signal. Here, the weight wout(λ) is a variable function. The weight wout(λ) is determined so as to output a desired state T(t) for the state xre(λ, t) of the reservoir unit 12, in accordance with Penrose pseudo-inverse matrix. Compared to a backpropagation method, there is no need to repeat update of weight, and thus it is possible to perform computation at high speed. The computation of the weight wout(λ) is performed by a personal computer, an FPGA, or the like. When the personal computer is used, a state during calculation is easily monitored. When the FPGA is used, computation can be performed at high speed.
At time t, the input signal xre(λ, t) from the reservoir unit 12 is input to the signal modulation unit 131. In the signal modulation unit 131, the input is multiplied by the weight wout(λ) for each wavelength. A wavelength selective switch, MEMS, or the like can be used as the signal modulation unit 131. When used, the wavelength selective switch can be operated with low power consumption. When the MEMS is used, an extinction ratio can be increased to improve computation performance.
The optical signal that has passed through the signal modulation unit 131 is input to the light reception unit 132 and converted into an electrical signal. Here, rather than providing a light reception unit for each wavelength, a weighted multi-wavelength signal is received by one light reception unit. An electrical signal output from the light reception unit 132 loses information in the wavelength direction, and thus it is possible to perform calculation as if signals in the wavelength direction were added.
The computation time of the optical signal processing apparatus of the present embodiment is determined approximately by “modulation speed of light pulse×number of data for task×(number of nodes/number of wavelength multiplexing)”. A value in the parentheses indicates the number of nodes expanded in the time direction. The number of wavelength multiplexing has been 1 for the optical RC in the related art, and thus all nodes have been expanded on the time axis. In the present embodiment, the number of nodes expanded on the time axis can be reduced as the number of multiplexed wavelengths increases. This indicates that a throughput is improved by the reciprocal of the number of multiplexed wavelengths compared to the optical RC in the related art.
As described above, according to the present embodiment, when the optical RC that expands a node in the wavelength direction instead of expanding the node in the time axis direction is made, the throughput of the optical RC can be improved without increasing the number of nodes in the reservoir layer.
Note that it is also possible to further improve the throughput of the optical RC by expanding a node in the time axis direction at the same time as expanding the node in the wavelength direction even when the number of nodes is greater than the number of wavelengths.
With reference to
| Number | Date | Country | Kind |
|---|---|---|---|
| 2019-027369 | Feb 2019 | JP | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2020/004955 | 2/7/2020 | WO | 00 |
