Physical reservoir and information processing device

Abstract

A physical reservoir according to the present embodiment comprises an information processing unit and an output unit. The information processing unit includes a nonlinear processing unit and a linear processing unit respectively connected to the output unit. The nonlinear processing unit and the linear processing unit are given the same input signal. The nonlinear processing unit includes a plurality of physical reservoir elements and nonlinearly converts the input signal. The linear processing unit includes a switch and a plurality of signal holding elements. The switch switches the signal holding element into which to input the input signal from among the plurality of signal holding elements. Each of the plurality of signal holding elements is configured to hold the input signal for a certain time.

Description

Physical reservoir and information processing device 【0001】 This disclosure relates to a physical reservoir and an information processing device. 【0002】 Neuromorphic devices are components that mimic the human brain using neural networks. They artificially mimic the relationships between neurons and synapses in the human brain. 【0003】 A neuromorphic device, for example, has nodes (neurons in the brain) arranged hierarchically and means of communication (synapses in the brain) connecting them. The neuromorphic device improves the accuracy of answering problems by having the means of communication (synapses) learn. The neuromorphic device reflects the learning results in the output by weighting the input data. 【0004】 One type of neural network is the recurrent neural network. Recurrent neural networks can handle nonlinear time-series data. Nonlinear time-series data is data whose values ​​change over time, and stock prices are one example. Recurrent neural networks can process time-series data by feeding back the processing results of neurons in later layers to neurons in earlier layers. 【0005】 Reservoir computing is one method for realizing recurrent neural networks. Reservoir computing performs recursive processing by interacting signals based on internal connections. 【0006】For example, Patent Document 1 discloses an example of reservoir computing. Patent Document 1 discloses that the performance required for reservoir computing includes short-term memory performance and nonlinear performance. Short-term memory performance is a measure of how much past information can be remembered and forgotten. Nonlinear performance is a measure of how much nonlinear transformation is performed on the input data. A reservoir with excellent short-term memory performance can output an estimated solution by taking into account past data of the time series. A reservoir with excellent nonlinear performance has high expressive power and can handle more complex time series data. 【0007】 In recent years, there has been research into replacing some of the mathematical calculations in reservoir computing with physical elements (hereinafter referred to as physical reservoirs). This is because processing reservoir computing using software can result in a high computational load due to the mathematical processing of all calculations, and sufficient processing efficiency may not be achieved. For example, Patent Document 2 discloses a reservoir computing device using an oscillator. 【0008】 International Publication No. 2022 / 024167, Japanese Patent Publication No. 2020-87189 【0009】 There are studies underway to implement reservoir computing functions using physical reservoirs. However, physical reservoirs have difficulty ensuring sufficient short-term memory performance. 【0010】 This disclosure is made in view of the above circumstances and provides a physical reservoir and information processing device with excellent short-term memory performance. 【0011】 A physical reservoir according to the first embodiment comprises an information processing unit and an output unit. The information processing unit has a nonlinear processing unit and a linear processing unit, respectively, connected to the output unit. The same input signal is input to the nonlinear processing unit and the linear processing unit. The nonlinear processing unit has a plurality of physical reservoir elements and performs a nonlinear transformation of the input signal. The linear processing unit has a switch and a plurality of signal holding elements. The switch is configured to switch which of the plurality of signal holding elements receives the input signal. Each of the plurality of signal holding elements is configured to hold the input signal for a certain period of time. 【0012】 The physical reservoir and information processing device according to the above embodiment have excellent short-term memory performance. 【0013】 This is a schematic diagram of a physical reservoir according to the first embodiment. This is a schematic diagram of an example of a nonlinear processing unit according to the first embodiment. This is a schematic diagram of an example of a linear processing unit according to the first embodiment. This is a schematic diagram for explaining the learning process of the physical reservoir according to the first embodiment. This is a schematic block diagram showing a specific example of the functional configuration of the information processing device according to the first embodiment. This is a conceptual diagram of an information processing device according to another example of the information processing device according to the first embodiment. 【0014】 The following description of this embodiment will be given in detail with reference to the drawings as appropriate. The drawings used in the following description may be enlarged for convenience to make the features of this disclosure easier to understand, and the specific configuration of each component may differ from the actual one. The configurations etc. exemplified in the following description are examples, and this disclosure is not limited to them, and can be modified as appropriate to achieve the effects of this disclosure. 【0015】 Figure 1 is a schematic diagram of a physical reservoir 10 according to the first embodiment. The physical reservoir 10 includes, for example, an input unit 1, an information processing unit 2, and an output unit 3. The input unit 1 and the output unit 3 are each connected to the information processing unit 2. 【0016】 The input unit 1 inputs data based on the input signal S1 to the information processing unit 2. The input unit 1 is connected to, for example, a sensor. The input unit 1 propagates the input signal S1 acquired by the sensor to the information processing unit 2. The input unit 1 is, for example, a signal propagation circuit. 【0017】 The input unit 1 may transmit the input signal S1 directly to the information processing unit 2 as input signal S2 without conversion. If the input signal S1 is input directly to the information processing unit 2 without conversion, the input unit 1 may be omitted. 【0018】The input unit 1 may convert the input signal S1 into an input signal S2 and propagate it to the information processing unit 2. For example, the input signal S1 may be weighted to obtain the input signal S2, or the input signal S1 may be converted into a time-division signal to obtain the input signal S2, or the input signal S1 may be digitally converted to obtain the input signal S2, or a combination of these processes may be performed on the input signal S1 to obtain the input signal S2. 【0019】 The input signal S1 may be a scalar or a vector. If the input signal S1 is a vector, each element representing the components of the vector may also be responsible for input signal S1. For example, all the elements representing the components of the vector may be input to the input unit 1 as input signal S1, and each of the elements may be input to the information processing unit 2 as input signal S2. 【0020】 The information processing unit 2 is connected to the input unit 1. The information processing unit 2 receives an input signal S2 from the input unit 1. The information processing unit 2 stores the input signal S2 and converts it into another signal. 【0021】 The information processing unit 2 includes a nonlinear processing unit 4 and a linear processing unit 6. The same input signal S2 is input to both the nonlinear processing unit 4 and the linear processing unit 6. 【0022】 The nonlinear processing unit 4 has multiple physical reservoir elements 5. Each physical reservoir element 5 corresponds to a node in reservoir computing and is equivalent to a neuron in a neural circuit. The number of physical reservoir elements 5 within the nonlinear processing unit 4 is not particularly limited. 【0023】 Each of the multiple physical reservoir elements 5 is connected to other physical reservoir elements 5. The connections between the physical reservoir elements 5 correspond to synapses in a neural circuit. Each physical reservoir element 5 may be coupled to all other physical reservoir elements 5 in the nonlinear processing unit 4, or it may be coupled to some of the physical reservoir elements 5 in the nonlinear processing unit 4. The connections between the physical reservoir elements 5 are random. 【0024】The coupling between physical reservoir elements 5 includes recursive coupling. Recursive coupling is a coupling in which the output returns to the input. For example, a signal output from one physical reservoir element 5 at time t may return to the physical reservoir element 5 that output the signal at time t + n (where n is a natural number). This occurs because the signal output from one physical reservoir element 5 propagates through other physical reservoir elements 5 and returns to the original physical reservoir element 5. This coupling relationship between physical reservoir elements 5, in which the output from one physical reservoir element 5 is input again via another physical reservoir element 5, is called recursive coupling. 【0025】 A coupling coefficient indicating a weight is set between the physical reservoir elements 5. The signal input to the nonlinear processing unit 4 propagates between the physical reservoir elements 5. The signal that has propagated to one physical reservoir element 5 is multiplied according to the coupling coefficient and propagated to the next physical reservoir element 5. 【0026】 The coupling coefficient between the physical reservoir elements 5 can be arbitrarily set, for example, within the range of -1.0 to +1.0. The coupling coefficient between the physical reservoir elements 5 can be set, for example, by a random number. The coupling coefficient between the physical reservoir elements 5 may be set, for example, to maximize the amount of mutual information between the output signal S5 from the physical reservoir 10 and the expected signal (e.g., training data). 【0027】 Each of the physical reservoir elements 5 may be fully coupled to the input unit 1. Fully coupled to the input unit 1 means that each of the physical reservoir elements 5 is connected to the input unit 1. That is, an input signal S2 is input to each of the physical reservoir elements 5 in the nonlinear processing unit 4. The coupling coefficient between the input unit 1 and each of the physical reservoir elements 5 can be set arbitrarily, and a random weight may be multiplied by the input signal S2 input to each of the physical reservoir elements 5. When the input unit 1 and each of the physical reservoir elements 5 are fully coupled, the signals interact in a complex manner in the information processing unit 2. As a result, the nonlinear performance of the information processing unit 2 is enhanced, and the physical reservoir 10 becomes able to handle more complex information. 【0028】Furthermore, each of the physical reservoir elements 5 may be fully coupled to the output unit 3. Fully coupled to the output unit 3 means that each of the physical reservoir elements 5 is connected to the output unit 3. By fully coupling the output unit 3 and each of the physical reservoir elements 5, more signals are input to the output unit 3. By obtaining an estimated solution using a large amount of data, the accuracy (correct response rate) of the output from the physical reservoir 10 can be improved. 【0029】 Figure 2 is a schematic diagram of an example of a nonlinear processing unit 4 according to this embodiment. The nonlinear processing unit 4 is composed of a combination of physical elements. For example, the nonlinear processing unit 4 shown in Figure 2 is a coupled oscillator. Each oscillator corresponds to a physical reservoir element 5. Each physical reservoir element 5 vibrates due to a spring 5A and interacts with each other due to a spring 5B. For example, an input signal S2 input to one physical reservoir element 5 causes the spring 5A of the physical reservoir element 5 to vibrate. The input signal S2 propagates to other physical reservoir elements 5 via the spring 5B. The input signal S2 propagates into the nonlinear processing unit 4 via the spring 5B and is nonlinearly transformed. The coupling coefficient between the physical reservoir elements 5 can be changed by changing the spring constant of the spring 5B. 【0030】 Here, we have shown an example where the nonlinear processing unit 4 is a coupled oscillator, but the nonlinear processing unit 4 is not limited to this example. The nonlinear processing unit 4 can be anything that converts the input signal into vibration, electromagnetic field, magnetic field, voltage, current, spin wave, etc. 【0031】 For example, the nonlinear processing unit 4 may be a MEMS (Micro Electronics Mechanical System) microphone array in which multiple MEMS microphones are arranged and electrically connected to each other. A MEMS microphone converts sound waves into electrical signals, for example. The vibration of the diaphragm of each MEMS microphone affects the diaphragms of other MEMS microphones, and the vibrations of the diaphragms of each MEMS microphone interact with each other. The input signal, which is input to the MEMS microphone array as a sound wave, is nonlinearly converted and output as an electrical signal from each MEMS microphone. 【0032】Alternatively, the nonlinear processing unit 4 may be a SAW (Surface Acoustic Wave) filter. In a SAW filter, the elastic wave input as the input signal interacts with the piezoelectric material. As a result, the input signal, which was input as an elastic wave, is nonlinearly transformed by the piezoelectric material and output as an electrical signal from the piezoelectric material. 【0033】 Alternatively, for example, the nonlinear processing unit 4 may be a spin-torque oscillator array in which spin-torque oscillators are arranged in an array. The magnetic fields generated by the magnetization of each spin-torque oscillator interact with each other, and the input signal is nonlinearly transformed. 【0034】 For example, the nonlinear processing unit 4 may be a combination of multiple RLC circuits. If there are RLC circuits with different time constants, a phase difference will occur, and the input signal will be nonlinearly transformed. 【0035】 The linear processing unit 6 comprises a switch 7 and a plurality of signal holding elements 8. 【0036】 Switch 7 is connected to the input unit 1. Switch 7 is configured to switch its connection to each of the multiple signal-holding elements 8. Switch 7 is configured to switch which of the multiple signal-holding elements 8 the input signal is input to. For example, switch 7 can be connected to any of the signal-holding elements 8, and at a certain timing, it is connected to a specific signal-holding element 8. For example, if there are M signal-holding elements 8 (M is a natural number), switch 7 switches the connection between switch 7 and the signal-holding elements 8 sequentially from the 1st signal-holding element 8 to the Mth signal-holding element 8. After reaching the Mth signal-holding element 8, switch 7 is connected to the 1st signal-holding element 8, and the switching of connections from the 1st signal-holding element 8 to the Mth signal-holding element 8 is repeated again. Switch 7 repeats the switching operation between switch 7 and the signal-holding elements 8 at regular intervals. A known switch 7 can be used. The switching timing of switch 7 can be set arbitrarily. 【0037】The signal holding element 8 is configured to hold the input signal S2 for a certain period of time. The certain period of time is arbitrary, for example, it is the time from the disconnection of the connection by the switch 7 from the first signal holding element 8 to the Mth signal holding element 8. 【0038】 The number of signal holding elements 8 in the linear processing unit 6 is arbitrary. The larger the number of signal holding elements 8 in the linear processing unit 6, the more information of the old input signal S2 is stored, and thus the short-term memory performance of the physical reservoir 10 is enhanced. 【0039】 The signal holding element 8 is, for example, a vibrator. The vibrator is not limited to a vibrator composed of a spring and a weight, and may be other vibrators represented by the same mathematical model, such as a vibrator using a cantilever beam with one-sided support or a vibrator using a beam supported at both ends. FIG. 3 is a schematic diagram of an example of the linear processing unit 6 according to the present embodiment. Each of the signal holding elements 8 is independent. For example, when the switch 7 is connected to a certain signal holding element 8, the spring 8A of the signal holding element 8 vibrates due to the input signal S2. The signal holding elements 8 are not connected to each other, and the spring 8A continues to vibrate even after the connection between the switch 7 and the signal holding element 8 is disconnected. That is, the signal holding element 8 converts the input signal S2 into the vibration of the spring 8A and holds the input signal S2 for a certain period of time as the vibration of the spring 8A. 【0040】 The Q value (quality factor) of the vibrator constituting the signal holding element 8 is preferably 10 times or more the Q value of the vibrator constituting the physical reservoir element 5. Here, when there is variation in the Q value of the vibrator constituting the physical reservoir element 5, the average value of the Q values of each physical reservoir is taken as the Q value of the vibrator constituting the physical reservoir element 5. Also, the Q value of the vibrator constituting the signal holding element 8 is preferably 100 or more. 【0041】 The Q value is a dimensionless quantity representing the state of vibration. The Q value is the value obtained by dividing the energy stored in the system by the energy dissipated from the system during one cycle. The larger the Q value, the more difficult it is for the vibration to decay. When the Q value is sufficiently large, the vibration of the spring 8A is stabilized and the input signal S2 can be held for a long time. 【0042】Here, an example is disclosed in which the signal holding element 8 is an oscillator using a spring 8A, but the signal holding element 8 is not limited to this example as long as it can hold an input signal for a certain period of time. 【0043】 For example, the signal holding element 8 may include an LC resonant circuit. An LC resonant circuit is an electrical circuit composed of a coil and a capacitor. An LC resonant circuit can store the amplitude of a voltage or current oscillating at the resonant frequency between the coil and the capacitor for a certain period of time (since power does not oscillate). The LC resonant circuit converts the input signal S2 into an oscillation of voltage or current and holds the input signal S2 as an oscillation of voltage or current for a certain period of time. The LC resonant circuit differs from an oscillator using a spring 8A, where the displacement of the weight oscillates, in that the object that oscillates is voltage or current. In the case of an LC resonant circuit, the Q value is preferably 10 times or more the Q value of the oscillator constituting the physical reservoir element 5, and preferably 100 or more, similar to the case of the spring 8A. 【0044】 Alternatively, the signal holding element 8 may be a memristor. A non-volatile memristor is preferred. The memristor is, for example, a magnetic domain wall moving element that can store data in analog form at the position of the magnetic domain wall. The memristor holds the input signal S2 as data. 【0045】 Each of the signal-holding elements 8 is connected, for example, to the output unit 3. By fully connecting the output unit 3 and each of the signal-holding elements 8, more signals are input to the output unit 3. By obtaining an estimated solution using a large amount of data, the accuracy (correct response rate) of the output from the physical reservoir 10 can be improved. 【0046】 The output unit 3 is connected to the nonlinear processing unit 4 and the linear processing unit 6 of the information processing unit 2, respectively. The output unit 3 receives a nonlinear signal S3 from the nonlinear processing unit 4 and a linear signal S4 from the linear processing unit 6. The nonlinear signal S3 is a signal input from each of the physical reservoir elements 5. The linear signal S4 is a signal input from each of the signal holding elements 8. 【0047】The output unit 3 is configured to be able to apply weights to each of the non-linear signals S3 and each of the linear signals S4. For example, the output unit 3 has a product operation element that performs a product operation on each of the non-linear signals S3 and each of the linear signals S4 with a signal serving as a weight. 【0048】 The weights applied to each of the non-linear signals S3 may be constant or different. The weights applied to each of the non-linear signals S3 are determined by learning. 【0049】 The weights applied to each of the linear signals S4 may be constant or different. The weights applied to each of the linear signals S4 are determined by learning. The weights applied to each of the linear signals S4 may be set according to, for example, the number of steps after the switch 7 and the signal holding element 8 are connected. 【0050】 For example, a case where there are M signal holding elements 8 and the Nth signal holding element 8 is connected to the switch 7 will be described as an example. The Nth signal holding element 8 is currently connected to the switch 7 and corresponds to 0 steps. For example, a weight w ０ is applied to the Nth signal holding element 8. The (N - 1)th signal holding element 8 was connected to the switch 7 one step before. For example, a weight w １ is applied to the Nth signal holding element 8. According to the same rule, the first signal holding element 8 was connected to the switch 7 (N - 1) steps before. For example, a weight w Ｎ－１ is applied to the first signal holding element 8. 【0051】 After one moment, the (N + 1)th signal holding element 8 is connected to the switch 7. At this time, a weight w ０ is applied to the (N + 1)th signal holding element 8, a weight w １ is applied to the Nth signal holding element 8, a weight w ２ is applied to the (N - 1)th signal holding element 8, and a weight w Ｎ－２ is applied to the first signal holding element 8. 【0052】 Weights w ０ to w Ｍ－１This value is set through learning and changes depending on the task required of the physical reservoir 10. For example, if the task required of the physical reservoir 10 places importance on past information, the weight w will be set. Ｍ－１ The weight lol ０ If it gets bigger and you want to prioritize the most recent information, then the weight lol ０ The weight lol Ｍ－１ It will get bigger. 【0053】 Next, the operation of the physical reservoir 10 will be explained. Figure 4 is a schematic diagram illustrating the learning process of the physical reservoir 10 according to this embodiment. The physical reservoir 10 performs learning and estimation processes. 【0054】 First, let's explain the learning process of the physical reservoir 10. In the learning process, the information to be learned is input to the input unit 1 as input signal S1. 【0055】 Input signal S1 is, for example, input signal S2 and input to the information processing unit 2. Input signal S2 is input to both the nonlinear processing unit 4 and the linear processing unit 6. 【0056】 In the nonlinear processing unit 4, the input signal S2 is nonlinearly transformed via the physical reservoir element 5. The nonlinearly transformed signal is output to the output unit 3 as a nonlinear signal S3. In the output unit 3, the nonlinear signal S3 from each physical reservoir element 5 is multiplied by a weight w. 【0057】 In the linear processing unit 6, the input signal S2 reaches the signal holding element 8 via the switch 7. By switching the signal holding element 8 connected to the switch 7, past input signals S2 can be stored in the signal holding element 8. For example, when the Nth signal holding element 8 is connected to the switch 7, the (N-1)th signal holding element 8 stores the input signal S2 from one time point ago, the (N-2)th signal holding element 8 stores the input signal S2 from two time points ago, and the 1st signal holding element 8 stores the input signal S2 from N-1 time points ago. 【0058】 The signals held by the signal-holding element 8 are output to the output unit 3 as linear signals S4. In the output unit 3, the linear signals S4 from each signal-holding element 8 are multiplied by a weight w. 【0059】In the output unit 3, an output signal is generated based on the result of multiplying the nonlinear signal S3 by a weight and the result of multiplying the linear signal S4 by a weight. For example, the sum of the result of multiplying the nonlinear signal S3 by a weight and the result of multiplying the linear signal S4 by a weight is substituted into the activation function to obtain the output signal. 【0060】 Next, the output unit 3 compares the output signal with the training data D. The output unit 3 adjusts the weights applied to each of the nonlinear signals S3 and each of the linear signals S4 according to the comparison result. For example, the weights applied to each linear signal S4 are set according to the number of steps since the switch 7 and the signal holding element 8 were connected. 【0061】 During the learning process, these weights are adjusted repeatedly until the mutual information between the output signal and the training data D exceeds a certain level. When the mutual information between the output signal and the training data D exceeds a certain level, the accuracy (correctness) of the estimated solution output from the physical reservoir 10 during the estimation process exceeds the desired value. During the learning process, the physical reservoir 10 determines the weights to be applied to each of the nonlinear signals S3 and the weights to be applied to each of the linear signals S4. This learning may be online learning, where the data is updated each time, or batch learning, where learning is performed using a fixed amount of training data and output. 【0062】 Next, the estimation process for the physical reservoir 10 will be explained. The estimation process for the physical reservoir 10 will be explained based on Figure 1. 【0063】 In the estimation process, the information to be estimated is input to the input unit 1 as input signal S1. 【0064】 Input signal S1 is input to the information processing unit 2 as input signal S2. Input signal S2 is input to the nonlinear processing unit 4 and the linear processing unit 6, respectively. 【0065】 In the nonlinear processing unit 4, the input signal S2 is nonlinearly transformed via the physical reservoir element 5. The nonlinearly transformed signal is output to the output unit 3 as a nonlinear signal S3. The nonlinear signals S3 from each physical reservoir element 5 are multiplied by weights in the output unit 3. These weights are determined during the learning process. 【0066】 In the linear processing unit 6, the input signal S2 reaches the signal holding element 8 via the switch 7. By switching the signal holding element 8 connected to the switch 7, past input signals S2 can be stored in the signal holding element 8. The timing of switching the switch 7 is the same as during the learning process. 【0067】 The signals held by the signal-holding elements 8 are output to the output unit 3 as linear signals S4. Each linear signal S4 from the signal-holding elements 8 is multiplied by a weight w at the output unit. This weight is determined during the learning process. 【0068】 In the output unit 3, an estimated solution is obtained based on the result of multiplying the nonlinear signal S3 by a weight and the result of multiplying the linear signal S4 by a weight. For example, the sum of the result of multiplying the nonlinear signal S3 by a weight and the result of multiplying the linear signal S4 by a weight is substituted into the activation function to obtain the estimated solution. The estimated solution is output from the output unit 3 as output signal S5. Output signal S5 is the answer to the task required of the physical reservoir 10. 【0069】 The physical reservoir 10 according to this embodiment has a linear processing unit 6, and therefore can obtain an estimated solution based on the input signal S2 of a past time. The physical reservoir 10 has excellent short-term memory performance because it can retain past information for a certain period of time. 【0070】 Figure 5 is a schematic block diagram showing a specific example of the functional configuration of the information processing device 100 according to this embodiment. The information processing device 100 is composed of, for example, a personal computer or a server device. The information processing device 100 includes, for example, a physical reservoir 10, an input unit 20, an output unit 30, a storage unit 40, and a control unit 50. 【0071】 The input unit 20 is the part to which the input signal S1 is input. The input unit 20 is connected to the input unit 1 of the physical reservoir 10. The input unit 20 is an interface that handles connections to, for example, sensors. The input signal S1 input to the input unit 20 is not limited to information from passive elements such as sensors, but may also be image, audio, or text information selected by the user. 【0072】The output unit 30 outputs information in a format that the user can recognize. The output unit 30 is connected to the output unit 3 of the physical reservoir 10. The output unit 30 outputs the estimated solution of the physical reservoir 10. The output unit 30 may be an image display device such as a liquid crystal display or an organic EL (Electro Luminescence) display. The output unit 30 may also be an interface for connecting the image display device to the information processing device 100. In this case, the output unit 30 generates a video signal for displaying image data and outputs the video signal to the image display device connected to it. The output unit 30 may also be a device that outputs sound, such as a speaker. The output unit 30 may also be an interface for connecting an audio output device such as a speaker or headphones to the information processing device 100. In this case, the output unit 30 generates an audio signal for playing back audio data and outputs the audio signal to the audio output device connected to it. The output unit 30 may also be configured as a touch panel integrated with the input unit 20. 【0073】 The storage unit 40 is configured using a storage medium such as a magnetic hard disk drive or a semiconductor storage device. The storage unit 40 stores data used for controlling the physical reservoir 10. The storage unit 40 also stores data necessary for the control unit 50 to perform processing. 【0074】 The storage unit 40 includes, for example, a training data storage unit 41 and a coupling coefficient storage unit 42. 【0075】 The training data storage unit 41 stores pre-input training data D. The coupling coefficient storage unit 42 stores, for example, the weights to be applied to each of the nonlinear signals S3 and the weights to be applied to each of the linear signals S4. The coupling coefficient storage unit 42 may update the stored weights w after each learning process. 【0076】The control unit 50 is composed of a processor 51 such as a CPU (Central Processing Unit) and a memory 52. ​​The control unit 50 functions when the processor 51 executes a program. Note that all or part of the functions of the control unit 50 may be implemented using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). 【0077】 The program is stored in memory 52. ​​Memory 52 may be a computer-readable recording medium. Computer-readable recording media include, for example, portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, and semiconductor memory devices (e.g., SSDs: Solid State Drives), as well as storage devices such as hard disks and semiconductor memory devices built into a computer system. The above program may be transmitted to the information processing device 100 via a telecommunications line. 【0078】 The memory 52 includes a learning program that performs learning processing for the physical reservoir 10 and an inference program that performs inference processing. 【0079】 The information processing device 100 according to this embodiment has the above-described physical reservoir 10 and therefore has excellent short-term memory performance. 【0080】 Figure 6 is a schematic diagram of another example of the information processing device according to this embodiment. The information processing device 101 includes a terminal device 60 and a storage unit 40. The terminal device 60 and the storage unit 40 are communicated with each other via a network 70. The network 70 may be a wireless communication network or a wired communication network. The network 70 may be configured using, for example, the Internet or a local area network (LAN). The network 70 may be configured by combining multiple networks. 【0081】The information processing device 101 differs from the information processing device 100 in that its storage unit 40 is located on an external server. The terminal device 60 has the aforementioned physical reservoir 10, input unit 20, output unit 30, and control unit 50. The terminal device 60 is configured using information devices such as smartphones, tablets, personal computers, and dedicated devices. 【0082】 The terminal device 60 and the storage unit 40 may each have a communication unit. The communication unit may be configured, for example, as a network interface. The communication unit communicates data with other devices via the network 70 in accordance with the control of the control unit 50. The communication unit may be a wireless communication device or a wired communication device. 【0083】 The information processing device 101 is simply an externally formed storage unit 40, and achieves the same effects as the information processing device 100. 【0084】 Although this embodiment has been described in detail above with reference to the drawings, the configurations and their combinations in this embodiment are merely examples, and additions, omissions, substitutions, and other modifications to the configurations are possible without departing from the spirit of the present invention. 【0085】 1 Input unit 2 Information processing unit 3 Output unit 4 Nonlinear processing unit 5 Physical reservoir element 6 Linear processing unit 7 Switch 8 Signal holding element 10 Physical reservoir 20 Input unit 30 Output unit 40 Storage unit 41 Training data storage unit 42 Coupling coefficient storage unit 50 Control unit 51 Processor 52 Memory 60 Terminal device 70 Network 100, 101 Information processing device D Training data S1, S2 Input signals S3 Nonlinear signals S4 Linear signals S5 Output signals

Claims

1. A physical reservoir comprising an information processing unit and an output unit, wherein the information processing unit has a nonlinear processing unit and a linear processing unit connected to the output unit, respectively, the same input signal is input to the nonlinear processing unit and the linear processing unit, the nonlinear processing unit has a plurality of physical reservoir elements and performs a nonlinear transformation of the input signal, the linear processing unit has a switch and a plurality of signal holding elements, the switch is configured to switch which of the plurality of signal holding elements receives the input signal, and each of the plurality of signal holding elements is configured to hold the input signal for a certain period of time.

2. The physical reservoir according to claim 1, wherein the output unit is configured to apply weights to the linear signals output from each of the plurality of signal-holding elements, and the weights are set according to the number of steps since the switch and the signal-holding element were connected.

3. The physical reservoir according to claim 1, wherein at least one of the plurality of signal-holding elements includes an oscillator.

4. The physical reservoir according to claim 3, wherein each of the plurality of physical reservoir elements includes an oscillator, and the Q value of each oscillator of the plurality of signal-holding elements is 10 times or more the average value of the Q values ​​of each oscillator of the plurality of physical reservoir elements.

5. The physical reservoir according to claim 3, wherein the Q value of each oscillator of the plurality of signal-holding elements is 100 or more.

6. The physical reservoir according to claim 1, wherein at least one of the plurality of signal-holding elements includes a memristor.

7. The physical reservoir according to claim 1, wherein at least one of the plurality of signal-holding elements includes an LC resonant circuit.

8. An information processing apparatus comprising the physical reservoir described in claim 1.

Images