Reservoir Calculator

Abstract

The objective is to provide a reservoir calculation device that minimizes the impact on mounting area and manufacturing costs, and suppresses manufacturing variations and fluctuations due to external environmental factors. [Solution] A reservoir calculation device comprising an input unit, an input layer that performs weighting calculations on signals received by the input unit, a reservoir layer connected to the input layer and consisting of a plurality of physical nodes, an output layer connected to at least one physical node and performing weighting calculations on the state output of each physical node, and an output unit that outputs the sum of the weighted signals from the output layer, wherein the physical node is a physical reservoir element that performs a nonlinear transformation based on its own dynamics, and further, a correction mechanism is provided for correcting the output of the physical reservoir element, the correction mechanism is characterized in that it reduces the difference between a predetermined output target value for an input signal and a predetermined output of the physical reservoir element for an input signal.

Description

Technical Field

[0001] This disclosure relates to a reservoir computing device.

Background Art

[0002] Reservoir computing is one of the machine learning in the field of information engineering, and is a method of performing information processing using a non-linear dynamical system. Reservoir computing holds information in a dynamical system called a reservoir, and generates an output by combining that information with input data. The dynamical system in the reservoir is represented as a network of nodes and elements, and exhibits non-linear behavior.

[0003] The concept of reservoir computing was proposed by Jaeger in the early 2000s. It was shown that the reservoir holds the information of the input data and the reservoir generates an output based on the input data because the dynamical system in the reservoir has a random connection structure. This idea is a different approach from the conventional neural network as a recurrent neural network, and has attracted attention as a new information processing method.

[0004] Reservoir computing is applied in various fields such as speech recognition, time series data prediction, and pattern recognition. In particular, reservoir computing exhibits high performance in processing data with strong temporal dependencies and non-linearity. For example, in speech recognition, the waveform of a speech signal is input to the reservoir, and the dynamical system in the reservoir extracts the features of the speech and outputs the recognition result. Also, in time series data prediction, past data is input to the reservoir, and the dynamical system in the reservoir predicts the future state.

[0005] Reservoir computation is characterized by its high generalization ability, which is not significantly affected by the data dimension or the number of samples. This is because the dynamic system within the reservoir exhibits nonlinear behavior, allowing it to capture complex data patterns. Furthermore, reservoir computation is suitable for real-time processing and is effective for applications requiring high-speed information processing. In particular, conventional recurrent neural network (RNN) technologies, including LSTM (Long Short-Term Memory), have faced challenges due to the computational cost of training. In reservoir computation, the connection weights within the reservoir layer responsible for nonlinear transformations can be randomly set fixed values, and only the weights of the output layer are optimized through training, resulting in low computational cost for training. Therefore, the ability to perform online training using inference results, in addition to batch training, is one reason why it is attracting attention.

[0006] Research is progressing on using physical elements to perform reservoir calculations. Instead of using conventional software for reservoir calculations, this approach attempts to perform calculations using physical phenomena such as dynamics involving electricity, magnetism, and polymers. This is called a physical reservoir, or physical reservoir calculation. In order to realize reservoir calculations using physical phenomena, it is necessary to construct a dynamic system using industrially manufactured physical elements. For example, optical elements or electrical elements can be used.

[0007] As an example, in reservoir calculations using optical elements, information processing is performed by utilizing the propagation of light and the nonlinearity of light. The propagation path of light and the intensity of light are represented as information in the reservoir, and the output is generated through interaction with the input data. Optical elements enable high-speed processing and can be used to build large-scale networks. For example, the propagation path of light can be controlled using optical fibers or optical waveguides.

[0008] In reservoir calculations using electrical elements, information processing is performed using electronic circuits and semiconductor devices. The dynamic behavior and nonlinearity of electronic circuits are utilized to hold information within the reservoir, and output is generated through interaction with input data. Electrical elements are becoming smaller, more power-efficient, and offer greater flexibility in implementation, making them suitable for realizing reservoir calculations. For example, such electrical elements can be constructed using transistors, MEMS technology, thin-film fabrication methods, etc.

[0009] Incidentally, when mass-producing physical reservoirs industrially, there are concerns about variations in element characteristics due to manufacturing variability and environmental factors. These variations negatively affect the performance of reservoir calculations. When mass-producing industrially, it is necessary to optimize the manufacturing process to minimize variations in element characteristics. For example, temperature control during manufacturing and material quality control are important. Furthermore, it is necessary to develop robust designs and compensation methods that are resistant to variations in element characteristics. This will stabilize the performance of reservoir calculations.

[0010] Furthermore, establishing methods for evaluating and tuning the performance of physical reservoir calculations is crucial. Performance can be improved by optimizing the parameters and coupling structure of the dynamic system within the reservoir. Additionally, selecting the appropriate model for physical reservoir calculations and considering preprocessing methods for input data are important. These factors will allow for maximizing the performance of physical reservoir calculations.

[0011] In particular, regarding the issues related to variations and fluctuations of electrical elements in physical reservoir calculations, Non-Patent Document 1 describes a method for correcting the results of reservoir calculations using a higher-order polynomial. Non-Patent Document 2 discloses a method for correcting fluctuations due to variations in physical reservoirs using a digital reservoir installed downstream. [Prior art documents] [Non-patent literature]

[0012] [Non-Patent Document 1] X.Guo et al, Micromachines 2023; 14, 161. “Input-Output-Improved Reservoir Computing Based on Duffing Resonator Processing Dynamic Temperature Compensation for MEMS Resonant Accelerometer” Jan. 2023.

[0013] [Non-Patent Document 2] Shohei Tatsumi, Yuki Abe, Kohei Nishida, and Tetsuya Asai, Proposal for a replication technique for physical reservoirs using small-scale digital correction reservoirs, IEICE Technical Report CCS2023-43 (2024-03). [Overview of the project] [Problems that the invention aims to solve]

[0014] The challenges in industrializing electronic devices that perform physical reservoir calculations mainly include variations caused by manufacturing equipment and processes, and the influence of operating environment factors such as temperature on the device's characteristics.

[0015] For example, in a physical reservoir using a MEMS oscillator, the shape and thickness of the vibrating mass will vary depending on the processing accuracy of the deposition equipment and etching equipment. As a result, variations occur in the resonance point and amplitude, causing fluctuations in the output of the physical reservoir. Similar fluctuations can occur in devices that utilize other physical phenomena or electrical behaviors.

[0016] Furthermore, the physical reservoir is also affected by the environment in which it operates. For example, the aforementioned oscillator-type physical reservoir undergoes thermal expansion of the vibrating mass and beams depending on the ambient temperature. Alternatively, it affects the characteristics of the analog circuit used to detect the resonant frequency of the oscillator-type reservoir. As a result, there is a problem of degraded accuracy in physical reservoir calculations. The known technologies shown in Non-Patent Documents 1 and 2 were proposed to address these problems, but all of them require a digital computing infrastructure, which creates challenges such as manufacturing costs and mounting area when commercializing physical reservoir computers. For these reasons, new methods for dealing with these fluctuations are needed in order to industrialize physical reservoirs.

[0017] This disclosure is made in view of the above circumstances and provides a reservoir calculation device that minimizes the impact on mounting area and manufacturing costs, and suppresses manufacturing variations and fluctuations due to external environmental factors. [Means for solving the problem]

[0018] (1) The reservoir computing device according to the first embodiment is a physical reservoir comprising an input layer that receives an input signal from the outside, a physical reservoir layer connected to the input layer and consisting of a plurality of physical nodes, and an output layer connected to at least one of the plurality of physical nodes and outputting the result of weighted calculation of the state of each physical node, wherein the physical node comprises a physical reservoir element that performs a nonlinear transformation based on the input signal and signals from other physical nodes and its own dynamics, and a correction mechanism that corrects the output signal of the physical reservoir element, wherein the correction mechanism corrects the output signal of the physical reservoir element so as to cancel the difference between the output target value of the physical node with respect to the input signal and the output signal of the physical reservoir element.

[0019] (2) In the reservoir calculation device according to the above embodiment, the correction mechanism may be a bias adjustment circuit that adds or subtracts a predetermined amount to the output of the physical reservoir element.

[0020] (3) In the reservoir computing device according to the above aspect, the correction mechanism may be a gain adjustment circuit that enlarges or reduces the dynamic range of the output of the physical reservoir element by a predetermined ratio.

[0021] (4) In the reservoir computing device according to the above aspect, the correction mechanism may be a modulation circuit that non-linearly converts the output of the physical reservoir element.

[0022] (5) In the reservoir computing device according to the above aspect, the signal processing device may include a temperature sensor, the correction mechanism may determine a correction amount using the output of the temperature sensor, and perform a correction process on the output of the physical reservoir element.

[0023] (6) In the reservoir computing device according to the above aspect, the correction mechanism is provided at the physical node where the absolute value of the coupling weight between the physical node of the physical reservoir layer and the output layer is greater than a predetermined value, and the number of physical nodes provided with the correction mechanism is selectively provided so as not to exceed the number of physical nodes coupled to the output layer.

[0024] (7) The reservoir computing device according to the second aspect is composed of a physical reservoir device that electronically realizes only the physical reservoir layer, and a CMOS semiconductor device connected to the physical reservoir device. The CMOS semiconductor device includes a sensor input / output terminal serving as an interface such as a sensor, a sensor control circuit, an input circuit, a reservoir control circuit, a physical reservoir device connection terminal, a correction circuit, an output circuit, and an external input / output terminal. It is a signal processing device characterized by this.

[0025] (8) In the reservoir computing device according to the second embodiment, the physical reservoir device can be fabricated between the wiring layers of a CMOS semiconductor, and the sensor control circuit, the input circuit, the reservoir control circuit, the correction circuit, and the output circuit are fabricated by transistor circuits formed in the wiring and the underlying diffusion layer within the wiring layer of the CMOS semiconductor. Such physical reservoir devices include spintronic devices using magnetic materials such as spin-wave application elements and spin torque oscillation elements. Also, as physical reservoir devices, there are devices that utilize the dynamics of dielectric polarization by dielectric materials such as HfO2, and memristor reservoir devices that utilize non-volatile memristor elements such as ReRAM.

[0026] (9) In the reservoir computing device according to the second embodiment, the CMOS semiconductor device and the physical reservoir device may be integrally encapsulated by a resin mold package. The output of the physical reservoir device is corrected by a correction circuit implemented in the semiconductor device for the output of the physical reservoir element of the physical reservoir device. In this case, the semiconductor device and the physical reservoir device may be fabricated by different manufacturing methods, and each may be integrated using bonding techniques such as wafer bonding and package-on-package, or techniques such as chiplets. Such physical reservoir devices include oscillators, microphones using MEMS, or SAW (surface acoustic wave) devices, piezoelectric resonators, and the like.

Advantages of the Invention

[0027] The reservoir computing device according to the present disclosure has little impact on the mounting area and manufacturing cost, and can suppress manufacturing variations and fluctuations in external environmental factors.

Brief Description of the Drawings

[0028] [Figure 1] It is a conceptual diagram of the reservoir computing device according to the first embodiment. [Figure 2] It is a configuration diagram of the physical nodes of the physical reservoir layer of the reservoir computing device according to the first embodiment. [Figure 3] This is a configuration diagram showing the details of the physical nodes in the reservoir computing device according to the first embodiment. [Figure 4] This figure shows the learning method of the reservoir calculator according to the first embodiment. [Figure 5] This is a diagram of the correction mechanism of the reservoir calculation device according to the first embodiment. [Figure 6] This figure illustrates an example of another correction mechanism for the reservoir calculator according to the first embodiment. [Figure 7] This is an example of a gain adjustment circuit for a reservoir calculator according to the first embodiment. [Figure 8] This figure shows an example in which a temperature sensor is provided in the reservoir calculator according to the first embodiment. [Figure 9] This figure shows an example in which a temperature sensor is installed within the physical node in the reservoir computing device according to the first embodiment. [Figure 10] This figure shows an example in which a thermistor element is provided within the physical node as a correction mechanism in a reservoir computing device according to the first embodiment. [Figure 11] This figure shows a physical node equipped with a correction mechanism and a physical node without a correction mechanism in a physical reservoir computing device according to the first embodiment. [Figure 12] This figure shows the range of the physical reservoir device according to the second embodiment. [Figure 13] This is a functional block diagram of the reservoir calculator according to the second embodiment. [Figure 14] This is a diagram showing the configuration of the reservoir calculator according to the second embodiment. [Modes for carrying out the invention]

[0029] The following description of this embodiment will be made in detail with reference to the drawings as appropriate. The drawings used in the following description may be enlarged for convenience in order 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.

[0030] (First Embodiment) Figure 1 is a conceptual diagram of a reservoir computing device 1 according to the first embodiment. The reservoir computing device 1 has an input unit 2, an input layer 4, a reservoir layer 6, an output layer 8, and an output unit 10. The input unit 2 is an interface for receiving information from an external device (not shown), and the input layer 4 weights the input signal from the input unit 2 with the weights of the input layer 4 and passes the signal to each physical node 7 of the subsequent reservoir layer 6. In reservoir computing in mathematical models, the nodes that constitute the reservoir layer are called physical nodes in a physical reservoir. Typical external devices include sensors. Alternatively, a current measuring instrument for the motor drive signal in a motor control system can also be used as the input unit 2.

[0031] When the input unit 2 receives a single external input signal, it passes this single input signal as a one-dimensional time-series signal to multiple physical nodes 7 of the reservoir layer 6, weighted by the input weights. When the input unit 2 receives multiple external input signals, each input signal is treated as a multi-dimensional time-series signal and weighted by its respective weight to each physical node 7 of the reservoir layer 6. When the physical reservoir is treated as a mathematical model, the input layer weights 3 are represented as a two-dimensional matrix. The input signal may be passed to all physical nodes 7 constituting the reservoir layer 6, or to only some of the physical nodes 7 of the reservoir layer 6. In the latter case, the input layer weights 3 are represented as a weight matrix where the elements corresponding to the connection weights between the input signal and the physical nodes 7 to which the input signal propagates are zero.

[0032] The reservoir layer 6 of a physical reservoir consists of multiple physical nodes 7. Each physical node 7 may or may not be coupled to one another via coupling weights in the reservoir layer 6. Each physical node 7 has the function of summing the signals received from other physical nodes 7, performing nonlinear transformation and correction processing, and outputting the result. Here, nonlinear transformation processing is a trivial technique commonly used in the field of machine learning. In addition, in a physical reservoir, the physical nodes 7 themselves may have a function equivalent to coupling weights, and there may be no explicit coupling weights between the physical nodes 7. Generally, in reservoir calculations, the coupling between each physical node 7 takes random values, but in recent years, there have been reports that performance can be improved by optimizing the coupling between physical nodes within the reservoir layer 6. In a physical reservoir as well, it is possible to design the coupling between physical nodes 7 to be a value optimized on a mathematical model. For example, in the case of a physical reservoir using MEMS oscillators, such adjustments can be made by adjusting the natural frequency and damping coefficient.

[0033] The mathematical representation of the computation realized by a physical reservoir in the case where there is no coupling to feed back the output of output layer 8 to reservoir layer 6 is shown below.

[0034]

number

[0035]

number

[0036]

number

[0037] Here, k is the time step, Win(i) is the input layer weight 3 that connects the i-th physical node in the reservoir layer 6 to the input, Wres(j,k) is the reservoir layer weight 5 that connects the j-th and k-th physical nodes in the reservoir layer 6, and Wout(m) is the output layer weight 9 that connects the m-th physical node in the reservoir layer 6 to the output. Also, u(k) is the input signal, x represents the time evolution of the reservoir layer 6, and r(k) represents the state vector of the reservoir layer 6 at time k. The output of the physical reservoir is calculated as z(k) by multiplying the state vector r(k) of the reservoir layer 6 by the weight of the output layer 8. Note that α is called the leaking rate and usually takes a fixed value in the range of 0 to 1. The leaking rate determines the update rate of the reservoir layer 6. φ(·) is called the activation function and provides a nonlinear transformation function in the reservoir calculation. Hyperbolic functions such as Tanh, sinc, and wavelet transforms are known.

[0038] Figure 2 shows the physical nodes 7 of the reservoir layer 6 described in this embodiment. Each physical node 7 of the reservoir layer 6 consists of a physical reservoir element 11 and a correction mechanism 12. As will be described later, among the multiple physical nodes 7, there may be physical nodes 7 that do not have a correction mechanism 12. In that case, the physical node 7 and the physical reservoir element 11 are equivalent. The physical reservoir element 11 functions as a node that constitutes the physical reservoir. The physical reservoir element 11 is a major component of the physical reservoir and is known to be achievable with various electronic devices based on MEMS oscillators, ferroelectric elements, spintronic elements, ionic materials, electrochemistry, etc., and is an element with nonlinear time response characteristics. The physical reservoir element 11 receives input signals from the input layer 4 and signals from other physical nodes 7. The physical reservoir element 11 outputs the result of a state transition based on the signal input to the physical reservoir element 11, the current state of the physical reservoir element 11, and the dynamics of the physical reservoir element 11, as the output of the physical reservoir element 11. The correction mechanism 12 is coupled to the output terminal of the physical reservoir element 11 and is a mechanism that converts the output signal of the physical reservoir element 11 to produce an output as a physical node 7.

[0039]

number

[0040]

number

[0041] The output of the physical reservoir element 11, corrected by the correction mechanism 12, becomes the output of the physical node 7. The output of the physical node 7 is passed to other physical nodes 7 or to the output layer 8.

[0042] Output layer 8 takes the sum of the signals after applying the output layer 8 weights to the output signals from the physical nodes 7 of the reservoir layer 6, and outputs this sum to the outside as the output signal. Alternatively, the output signal may be generated by taking only the outputs of some of the physical nodes 7 of the reservoir layer 6 and applying the weights of those outputs to output layer 8. Note that not all physical nodes of the reservoir layer 6 need to be coupled to the output signal; only some physical nodes may be coupled to the output signal.

[0043] Although the output unit 10 is described here as a single signal, it may contain multiple signals. Figure 3 is a conceptual diagram of the signal processing device described in this embodiment, incorporating the detailed parts of the physical node 7 in Figure 2.

[0044] Next, we will explain the learning of the output layer 8 of the physical reservoir. Generally, in reservoir computation, the weights of the input layer 4 and the reservoir layer 6 are set randomly and are not subject to learning update. Since only the weights of the output layer 8 are optimized as learning targets, we will show an example here in which the weights of the output layer 8 are updated by learning. In recent research, techniques for optimizing the weights of the reservoir layer 6 and the input layer 4 have been proposed. For example, it is also possible to optimize the coupling weights between the physical nodes 7 of the reservoir layer 6 using the mutual information between the output of the physical node 7 of the reservoir layer 6 and the correct label which functions as a training signal. Even in such cases, as already mentioned, it is possible to design the physical reservoir element 11 so that mathematically predetermined weights are reflected in its characteristics, and even in such cases, this embodiment functions effectively against variations due to manufacturing variations and external environmental factors of the physical reservoir element 11.

[0045] The learning process in reservoir computation will be explained using Figure 4. As shown in Figure 4, first, a learning input signal S1 is applied to the input unit 2. The learning input signal S1 is propagated to each physical node 7 of the reservoir layer 6 after being weighted by the input layer 4. The internal state of the reservoir layer 6 changes according to the learning input signal S1. The output layer 8 weights the outputs of the physical nodes 7 of the reservoir layer 6, and the sum of these weights is output as an output signal from the output unit 10. The output signal is compared with the teacher signal S2, which is paired with the learning input signal S1, by the comparator 13, and the difference signal is passed to the learning algorithm calculation unit 14. The learning algorithm calculation unit 14 calculates the output layer weight update amount S3 of the output layer 8 so that the output signal approaches the teacher signal S2. The weight update unit 15 subtracts or adds to each weight of the output layer 8 according to the output layer weight update amount S3 calculated by the learning algorithm calculation unit 14. A learning method in which the learning input signal S1 and the teacher signal S2 are provided together and learned at once is called batch learning. On the other hand, a learning method in which the weights of the output layer 8 are updated sequentially using the inference results is called online learning.

[0046] This section describes the application of batch learning to a physical reservoir. In batch learning for a physical reservoir, a mathematical equivalent model of the physical reservoir is created in advance using a computer with high computational resolution, and training data is input to this model to obtain the calculated output of the physical reservoir element 11. The weights of the output layer 8 are optimized by batch learning using this result and the training signal S2. In batch learning, the weights of the output layer 8 are obtained by optimization calculations, including inverse matrix calculations and regularization effects. These output layer 8 weights are implemented according to the implementation method of the output layer 8. When the output layer 8 is implemented as a microcontroller switch, ASIC, FPGA, or other digital circuit, the obtained output layer weights are stored in the memory of the microcontroller, ASIC, or FPGA and used to calculate the output value. If necessary, they may be adjusted to the arithmetic word length of the computational base used for quantization, etc.

[0047] If the output of physical node 7 of a physical reservoir is a voltage-modulated signal, it is also possible to use an array of memristor elements as the output layer 8 of the physical reservoir. A memristor element is an array of elements fabricated using spintronics, ferroelectrics, phase-change memory, etc., arranged in a two-dimensional array. For example, if multiple inputs are applied from the direction corresponding to the rows of the two-dimensional array of memristor elements, the current corresponding to the resistance value of each memristor element becomes the output signal of the memristor. The sum of these output signals can be obtained by OR wiring them. This is equivalent to multiplying the input signal (voltage) by the conductance (reciprocal of resistance) = current, and then summing them up by OR wiring. In other words, it becomes possible to construct the output layer 8 of a physical reservoir using a memristor element array.

[0048] Although not explained in detail here, even when applying online learning to a physical reservoir, a certain level of learning effect can be obtained by adjusting the correction mechanism of this embodiment. Online learning is an algorithm that sequentially learns and updates the weights of the output layer 8 based on the inference results output during actual use, and methods such as sequential least squares, FORCE learning, and Kalman filters are used.

[0049] In this embodiment, both batch learning and online learning algorithms are applicable, and as described later, it is also possible to implement the comparator 13 and the learning algorithm calculation unit 14 in the digital computing HW that constitutes the output layer 8 and the output unit 10. Furthermore, this embodiment can also be applied to edge devices dedicated to inference without a learning mechanism.

[0050] Furthermore, a learning method without a learning input signal S1 is also acceptable. For example, when using a physical reservoir in a device anomaly detection system, the sound and vibration of the device to be monitored are acquired in advance by sensors, and the physical reservoir is trained. In some cases, this training may be performed by modeling the physical reservoir on a computer or similar device. In that case, the weights of the output layer 8 optimized by the training are reflected in the output layer 8 of the physical reservoir before monitoring the device's status. If an anomaly occurs in the device, the internal state of the physical reservoir deviates from normal, allowing it to output an anomaly detection signal.

[0051] The correction mechanism 12 in this embodiment will now be described. For example, Figure 5 shows a correction mechanism 12a that provides an offset to the output signal of the physical reservoir element 11. The mathematical formula is as follows.

[0052]

number

[0053] If the output of the physical reservoir element 11 is a voltage, the above-mentioned ΔOffset will be a DC voltage. Similarly, if the output of the physical reservoir element 11 is a current, it will be a DC current. For example, if the output value of the fabricated physical reservoir element 11 is shifted to the negative side compared to the expected output value when a reference input is applied due to manufacturing variations, the physical node 7 can output the correct value by applying a bias with the correction mechanism 12a. Conversely, if it is larger than the expected output value, the physical node 7 can output the correct value by having the correction mechanism 12a suppress the output of the physical reservoir element 11. For example, this can be adjusted by providing multiple resistors outside the physical reservoir element 11 and selecting the appropriate resistor from among them based on measurements such as those taken during manufacturing and shipping tests. Such external resistors can also be fabricated using a CMOS circuit, for example.

[0054] The correction mechanism 12 in this embodiment may be an amplification circuit that uses a gain adjuster to expand or decrease the output range of the physical reservoir element 11, as shown in the correction mechanism 12b of Figure 6. In that case, it is expressed by the following equation.

[0055]

number

[0056] The amplification circuit may also be a transistor circuit using amplifier 16 as shown in Figure 7. Figure 7(a) shows the case of a non-inverting amplifier, and Figure 7(b) shows an example of an inverting amplifier. In the non-inverting amplifier shown in Figure 7(a), the amplification factor of the output voltage Vout with respect to the input voltage Vin is expressed by the following equation. In equations (8) and (9) below, resistor R1 corresponds to resistor 17, and resistor R2 corresponds to resistor 18.

[0057]

number

[0058] On the other hand, in the inverting amplifier shown in Figure 7(b), the amplification factor of the output voltage Vout with respect to the input voltage Vin is expressed by the following formula.

[0059]

number

[0060] The amplification factor is determined based on the output value of the physical reservoir element 11, which has been measured in advance. The resistance values ​​of resistors 17 and 18 may be pre-built into the circuit based on resistance values ​​calculated during the design phase. If the amplification factor is to be adjustable, multiple resistors may be provided and determined by wiring. For example, it may be set individually based on the measurement results of the physical reservoir element 11 measured during the manufacturing test of the physical reservoir.

[0061] Furthermore, the correction mechanism 12 may also provide a nonlinear effect. A nonlinear effect could be, for example, an exponential modulation of the input signal. Such an effect can be expected by using a diode as the correction mechanism 12. The nonlinear effect could also be a modulation method such as a hyperbolic function, where the output converges to a constant value. While the physical reservoir element 11 is originally designed on the premise of nonlinear operation and memory capabilities based on the physical principles and electrical circuit theory it utilizes, it may not satisfy the required nonlinearity due to, for example, manufacturing variations, equipment constraints, or structural constraints of the device. In such cases, such a correction mechanism 12 is expected to improve the expressiveness of the physical reservoir and improve its performance.

[0062] Figure 8 shows a configuration diagram in which a temperature sensor 19 is provided outside the reservoir layer 6, and the degree of action of the correction mechanism 12 is determined based on its output signal. As shown in Figure 8, the temperature sensor 19 may be provided in the physical reservoir, and the action of the correction mechanism 12 may be adjusted according to the output value of the temperature sensor 19. The electrical characteristics of electronic devices fabricated by semiconductor processes, magnetic thin films, MEMS methods, etc., are generally temperature-dependent. As a result, when the operating temperature of the element changes, the output of the physical reservoir element 11 fluctuates. Consequently, even if the same input signal is applied, the output of the physical reservoir fluctuates, and the correct output cannot be obtained. Therefore, the output fluctuation of the reservoir calculator can be suppressed by a correction mechanism 12 that suppresses fluctuations in response to the output fluctuation of the physical reservoir element 11 due to temperature fluctuations. Note that the temperature sensor 19 and the physical reservoir may be sealed in the same package. Figure 9 shows a case in which the temperature sensor 19 is provided inside the physical reservoir node. The temperature sensor 19 is known to be a thermistor or a CMOS circuit.

[0063] Furthermore, when the output of the physical reservoir element 11 is a voltage, an element whose resistance changes with temperature, such as a thermistor, can itself serve as a correction mechanism 12. For example, as shown in Figure 10, by arranging a thermistor element 20 and a reference resistor 21 in series in the output stage of the physical reservoir element 11 and taking the midpoint potential as the output, control circuits and the like become unnecessary. For example, consider a physical reservoir element 11 in which the output potential decreases when the temperature rises. In that case, by using an NTC thermistor as thermistor element 20, it is expected that the resistance will decrease with increasing temperature. As a result, the midpoint potential will rise, making it possible to mitigate the temperature dependence of the physical reservoir element 11.

[0064] Furthermore, the correction mechanism 12 does not need to be provided for all physical reservoir elements 11 that constitute the physical reservoir computer. Each physical node 7 is weighted by the weights of the output layer 8 to produce an output signal, and the larger the weight of the output layer 8, the greater the impact on the output. Therefore, the correction mechanism 12 can be effective even if it is only provided for the physical reservoir elements 11 connected to the output unit 10 with large coupling weights. This will be explained using Figure 11. Physical node 70 is the i-th physical node, and its output layer 8 coupling weight, |Wout(i)|, is large. On the other hand, physical node 71 is the j-th physical node, and its output layer 8 coupling weight, |Wout(j)|, is small. In this way, it is also possible to select N (N is a natural number) nodes with large absolute values ​​of output layer 8 coupling weights and provide the correction mechanism 12 only for the physical reservoir elements 11 of the physical nodes connected to the selected weights. By concentrating such large coupling weight physical reservoir elements 11 in specific regions within the physical reservoir, it becomes possible to optimize the wiring to the correction mechanism 12 and the circuit of the correction mechanism 12.

[0065] (Second Embodiment) In the second embodiment described below, a device in which only the reservoir layer 6 in Figure 1 is realized using electronic devices such as MEMS oscillators, resonators, ferroelectrics, and spintronics is referred to as a physical reservoir device.

[0066] The physical reservoir device 301 is explained using Figure 12. The physical reservoir device 301 is an electronic device that implements only the function of the reservoir layer in reservoir calculations. The physical reservoir device 301 has an input interface 302, a physical node 7, an output interface 303, and a control interface 304. The input interface 302 includes one or more terminals 305 and receives signals from an externally provided input layer via the terminals 305 of the input interface 302. The output interface 303 also includes one or more terminals 305 and outputs the state of the physical node 7 to an externally provided output layer via the terminals 305 of the output interface 303. The control interface 304 includes one or more terminals 305 and communicates with an externally provided control circuit via the terminals 305 of the control interface 304. For example, when resetting the physical node 7, a reset signal is sent from the externally provided control circuit via the terminals 305 of the control interface 304. The control interface 304 is also used to notify the physical reservoir device 301 of the communication timing of the input and output interfaces. The physical reservoir device 301 has internal circuits for resetting each physical node 7 or sampling the state of each physical node 7 as needed.

[0067] In the second embodiment, an example of applying this correction mechanism to a physical reservoir device will be described. For the sake of explanation, in the following description, a plate-shaped semiconductor (for example, silicon) component on which circuit elements and the like are stacked will be referred to as a substrate.

[0068] Figure 13 is a circuit block diagram of the reservoir calculation device shown in the second embodiment. The reservoir calculation device 400 consists of a semiconductor device 300 and a physical reservoir device 301. The semiconductor device 300 consists of a sensor input / output terminal 101, a reservoir control block 200, a physical reservoir device connection terminal 104, and an external input / output terminal 108.

[0069] The reservoir control block 200 is an integrated circuit in which a sensor control circuit 103, an input circuit 102, a reservoir control circuit 105, a correction circuit 106, and an output circuit 107 are stacked on a CMOS substrate. Specifically, the sensor control circuit 103 is a circuit block for controlling the sensor connected via the sensor input / output terminal 101 and for receiving sensor signals. The reservoir control circuit 105 performs various controls for generating and transmitting input signals to the physical reservoir device 301, which is connected via the physical reservoir device connection terminal 104, and for receiving output (status) signals from the physical reservoir device 301. The input circuit 102 is a functional block corresponding to the input layer 4, and is a circuit that weights the signal from the sensor using the weights of the input layer 4 and generates an input signal to the physical reservoir device 301. The correction circuit 106 is a circuit that receives a signal from the physical reservoir element 11 of the physical reservoir device 301 via the physical reservoir device connection terminal 104 and corrects that signal in a predetermined manner. The output circuit 107 is a functional block corresponding to the output layer 8. It performs signal conversion processing, such as weighting calculations, on the corrected signal received from the correction circuit 106 using the weights of the output layer 8, and outputs it to an external computer or other device via the external input / output terminal 108. Although a computer is used as an example of an external device here, it may also be connected to a microcontroller, an FPGA circuit, or other devices via wired or wireless communication. The physical reservoir device connection terminal 104 is connected to the physical reservoir element 11 of the physical reservoir device 301.

[0070] The second embodiment will be described in more detail. External sensors such as inertial sensors for detecting acceleration and angular velocity, gas sensors for detecting carbon dioxide and other gases, temperature and humidity sensors, pressure sensors, and MEMS microphones can be connected. Furthermore, it is possible to use sensor modules with multiple sensor elements sealed in the same package, or to connect multiple sensors to create a sensor fusion system.

[0071] The input / output terminals are hardware (HW) that connects the sensor to the physical reservoir control circuit. The sensor control unit performs specific control depending on the connected sensor. For example, if the sensor's output signal is an analog signal, the sensor control unit incorporates an ADC circuit to sample the sensor's analog signal and input it as a digital signal into the sensor control unit's data buffer. On the other hand, if the sensor's output signal is a digital signal, the sensor control unit includes circuits necessary for receiving the digital signal or for communication control related to that. Such communication control methods include SPI, I2C, and I3C. Note that the type of sensor connected is not specified here, so the detailed circuit functions of the sensor control unit will not be explained.

[0072] The input circuit 102 performs multiplication of the sensor signal acquired by the sensor control circuit 103 with the input layer weight 3. Specifically, it multiplies the sensor signal stored in the data buffer (not shown) of the sensor control circuit 103 by the input layer weight 3 corresponding to each of the physical reservoir elements 11 inside the physical reservoir device 301. The input layer weight 3 may be stored in a memory or storage device provided in the sensor control circuit 103, or it may be stored in an external non-volatile memory and configured to be rewritable by an external computer. Furthermore, if the input layer weight 3 is designed to be expressed as a power of 2, the operation can be performed by bit shifting instead of multiplication. For example, if the input layer weight 3 is +8, it is sufficient to shift only 3 bits towards the MSB.

[0073] The reservoir control circuit 105 is a circuit block for controlling the physical reservoir device 301. The physical reservoir device 301 is composed of physical reservoir elements 11 realized by electronic devices and is, in principle, an analog system. However, depending on the use of the physical reservoir device 301, it may have a digital interface by incorporating functions such as an ADC or DAC internally. Each case is described below.

[0074] As mentioned earlier, there are various types of physical reservoir elements 11 that constitute a physical reservoir. However, all of them are essentially mechanisms that utilize dynamics in continuous time, and their state is observed as an analog signal. On the other hand, there are both analog and digital interface cases for connection to external devices. First, as an example of a physical reservoir device with an analog interface, we will explain a physical reservoir device using a MEMS oscillator as an example. Such a physical reservoir device may involve modulating and superimposing the input signal onto the drive signal, and detecting the resulting change in the resonant frequency of the oscillator as a change in the capacitance of the oscillator's comb electrodes. In that case, the reservoir control unit requires a drive signal generation circuit, an input signal modulation circuit, and a DAC to drive the oscillator. Furthermore, an ADC circuit to detect the response (output) of the physical reservoir device, and in some cases a QV conversion circuit to convert charge quantity to voltage value are implemented. In addition, a mechanism to detect the envelope of the resonant waveform may be incorporated into the circuit as a signal detection mechanism for the physical reservoir. These detection mechanisms can be provided in one for each physical reservoir element 11, or they can be acquired sequentially using time-division multiplexing. In systems where the simultaneity of all physical elements is strictly required, a circuit is needed to sample and hold each physical node 7 simultaneously and acquire the data sequentially into a data buffer.

[0075] On the other hand, a physical reservoir device with a digital interface implements the drive signal generation circuit and ADC circuits necessary for the operation of the physical reservoir elements inside the physical reservoir device, and provides only a digital interface externally. In the case of a physical reservoir device equipped with a digital interface, the input signal weighted by the input circuit 102 is transmitted to the physical reservoir device using a digital communication function. In addition, in order to receive the output of the physical reservoir device, the state of all physical reservoir elements 11 inside the physical reservoir device is acquired using a digital communication function. Similar to sensors, SPI, I2C, I3C, and other communication interfaces can be used.

[0076] The correction circuit 106 is a mechanism for correcting the output of the physical reservoir element 11 of the physical reservoir, as described in the first embodiment. It performs correction processing on the output of the physical reservoir element 11 received from the physical reservoir device 301. Signal conversions such as bias processing, gain processing, and nonlinear conversion processing, as described in the first embodiment, can be used for the correction processing. Through signal conversion, even if the physical reservoir device 301 has output fluctuations due to manufacturing variations, a signal close to the highly accurate ideal output calculated using a computer can be obtained. Furthermore, by providing appropriate sensors for environmental changes such as operating temperature, output fluctuations can be suppressed, and the performance as a reservoir calculator can be maintained. The correction circuit 106 can also be implemented as a switch by incorporating a digital circuit or a microcontroller inside the reservoir control block 200.

[0077] Furthermore, the correction circuit 106 does not perform calculations; it could also be a signal processing mechanism such as a filter. For example, if you want to suppress the noise present in the physical reservoir element 11, you could use an LPF (low-pass filter). An RC circuit would also work. Not only LPFs, but also BPFs (band-pass filters) and HPFs (high-pass filters) can be used as filters. In particular, when using reservoir calculations for audio signal processing, these filter functions greatly affect the feature generation results of the reservoir calculations. Therefore, correction processing using filter functions is useful.

[0078] In the reservoir calculator 400 according to the above embodiment, the reservoir calculator 400 is equipped with a non-volatile memory, and it is also possible to store the parameters necessary to adjust the output of the correction circuit in the non-volatile memory. Non-volatile memory refers to information storage devices such as flash memory, MRAM, FeRAM, EEPROM, or memory embedded in a semiconductor.

[0079] In the reservoir calculation device 400 according to the above embodiment, the correction amount of the correction circuit 106 may be adjusted from the difference between the state of the physical reservoir element 11 when a predetermined reference input signal is applied to the input circuit 102 and the ideal output value of the physical reservoir element 11 with respect to the predetermined reference input signal.

[0080] Next, another example of the second embodiment will be described in Figure 14. The reservoir calculation device 400 is shown in a state where it is integrally molded in a resin mold package. Inside the resin mold package, a physical reservoir device 301 and a semiconductor device 300 are mounted on a wiring layer 402. The wiring layer 402 is laminated on the substrate 401. Wiring is formed on the wiring layer 402, and the physical reservoir device 301 and the semiconductor device 300 are joined by solder balls, micro-solder balls, etc. Furthermore, the wiring layer 402 is connected to an external electrode 404 via a wire 403. The semiconductor device 300 has the circuit block shown in Figure 12 mounted on it, and various physical reservoir devices used in the second embodiment can be used for the physical reservoir device 301. A temperature sensor or the like may be mounted on the substrate 401 and connected to the semiconductor device 300, and the correction amount of the correction circuit 106 in the semiconductor device 300 may be determined using the output of the temperature sensor. In addition to temperature sensors, other sensors capable of sensing environmental factors that affect the operation of the physical reservoir device may also be provided. Specifically, examples include magnetic sensors and inertial sensors that detect acceleration and angular velocity.

[0081] With the reservoir calculation device according to the first and second embodiments described above, it becomes possible to appropriately correct the state output of the physical reservoir layer even in physical reservoirs that include manufacturing variations, and it can demonstrate performance equivalent to that of reservoir calculations expressed by mathematical models.

[0082] Furthermore, according to the signal processing device of the first and second embodiments described above, even when operating in an environment where external environmental factors such as temperature fluctuate, it is possible to correct the state of the physical reservoir at an appropriate timing, and it is possible to maintain performance equivalent to that of reservoir calculations expressed by mathematical models.

[0083] Furthermore, the signal processing devices according to the first and second embodiments described above do not require a new digital computing infrastructure to correct the output of the physical reservoir element.

[0084] Furthermore, there is a method called virtual node for physical reservoirs. In this method, instead of arranging many physical reservoir elements 11 as physical nodes 7 in the reservoir layer 6, multiple observations are obtained for a single physical element with nonlinear dynamics, including time delays, thereby virtually configuring the same number of physical nodes 7. Examples of such methods include optical modulation type physical reservoirs using optical elements and methods that observe vibration dynamics while sampling, such as MEMS oscillators. This embodiment also functions effectively in such virtual node type physical reservoirs.

[0085] Furthermore, one type of physical reservoir involves providing multiple pairs of electrodes to an element composed of aggregated particles, and utilizing the differences in resistance and time constants due to differences in bonding between particles within the element and particle arrangement in the current path between the pairs of electrodes. This embodiment also functions effectively in such a physical reservoir.

[0086] Furthermore, depending on the type of physical reservoir, the boundaries of physical nodes may not be clear. This embodiment functions effectively even in such cases. For example, let's explain using a physical reservoir that utilizes resonance and interference by oscillators as an example, where multiple physical reservoir elements each correspond to nodes in the reservoir layer of the mathematical model. The vibration state of each oscillator is detected as a change in charge amount by a dielectric element provided outside the oscillator. The change in charge amount is converted into the output of the physical reservoir element by using an integrating circuit such as a QV conversion circuit. The correction mechanism may be implemented by directly acting on the vibration state of such oscillators, for example, by applying a bias voltage to the oscillator itself. The correction mechanism may be a mechanism that acts on the polarization state of the dielectric element in the detection unit, or a mechanism that provides a bias or gain to the voltage at the stage where it is detected as a voltage converted from the charge amount. The vibration state of the oscillator can also be detected as a change in resistance using an element such as a piezoresistive. The same approach is possible in that case as well.

[0087] The following describes the application of this method to physical reservoirs that utilize a medium. For example, a physical reservoir has been proposed that detects interference phenomena caused by spin waves as spatiotemporal signals using multiple electrodes placed on a magnetic surface. Spin waves can be generated using microwaves from a microstrip antenna, spin transfer torque from spin-polarized current injection, or visible light pulses. In such physical reservoirs, correction mechanisms such as locally applying a magnetic field to affect the characteristics of the spin wave itself, such as the damping term, can be considered. Similar methods can also be applied to physical reservoirs that utilize skyrmions.

[0088] While the first and second embodiments relating to this disclosure have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications are possible without departing from the spirit of the invention. Furthermore, these embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of Symbols]

[0089] 1 Reservoir Calculator 2 Input section 3. Input layer weights 4 Input Layers 5. Reservoir layer weights 6 Reservoir Layer 7 physical nodes 8 Output Layers 9 Output layer weights 10 Output section 11 Physical Reservoir Elements 12 Correction mechanism 12a Correction mechanism (bias adjustment mechanism) 12b Correction mechanism (gain adjustment mechanism) 13 Comparator 14. Learning Algorithm Calculation Unit 15. Weight update section 16 Amplifiers 17 Resistors 18 resistors 19. Temperature sensor 20 Thermistor elements 21 Reference resistor 70 physical nodes (high output layer connection weights) 71 physical nodes (with small output layer connection weights) 101 Sensor Input / Output Terminal 102 Input Circuit 103 Sensor control circuit 104 Physical Reservoir Device Connection Terminal 105 Reservoir control circuit 106 Correction Circuit 107 Output Circuit 108 External input / output terminal 200 Reservoir control block 300 semiconductor equipment 301 Physical Reservoir Devices 302 Input Interface 303 Output Interface 304 Control Interface 305 terminal 400 Reservoir Calculator 401 circuit board 402 Wiring layer 403 Wire 404 External electrode S1 Learning input signal S2 teacher signal S3 Output layer weight update amount Δ

Claims

1. An input section that receives input signals from an external source, The input layer performs weight calculations on the signals received by the input unit, A reservoir layer connected to the aforementioned input layer and consisting of multiple physical nodes, An output layer connected to at least one of the physical nodes, which performs weighted calculations on the state output of each physical node, An output unit that outputs the sum of weighted signals from the output layer, Equipped with, The physical node includes a physical reservoir element that performs a nonlinear transformation based on its own dynamics, and a correction mechanism for correcting the output of the physical reservoir element. The reservoir calculation device is characterized by reducing the difference between a predetermined output target value for a predetermined input signal and the output of the physical reservoir element for a predetermined input signal.

2. The reservoir calculation device according to claim 1, characterized in that the correction mechanism is a bias adjustment circuit that adds or subtracts a predetermined amount to the output value of the physical reservoir element.

3. The reservoir calculation device according to claim 1, characterized in that the correction mechanism is a gain adjustment circuit that multiplies the dynamic range of the output value of the physical reservoir element by a predetermined amount.

4. The reservoir calculation device according to claim 1, wherein the correction mechanism is a modulation circuit that nonlinearly transforms the output value of the physical reservoir element.

5. A reservoir calculation device according to any one of claims 1 to 4, comprising a temperature sensor and performing correction processing on the output value of the physical reservoir element using the output of the temperature sensor.

6. The reservoir calculator according to claim 5, wherein the correction mechanism is provided at the physical node connected to the weight of the output layer whose absolute value is higher than a predetermined value, and the number of physical nodes on which the correction mechanism is provided does not exceed the number of physical nodes connected to the output layer.

7. A physical reservoir device and a physical reservoir control circuit, A signal processing device comprising a semiconductor device including, Previously, semiconductor devices were fabricated using semiconductor substrates. Sensor input / output terminals, Physical reservoir control circuit, Physical reservoir device connection terminal, It consists of external input / output terminals, The physical reservoir control circuit is, A control circuit for controlling the sensor and acquiring sensor data, An input circuit that calculates weights for sensor data, A reservoir control circuit for controlling a physical reservoir device and obtaining the response of the physical reservoir device, A correction circuit for correcting the signal from the physical reservoir device received from the reservoir control circuit, It consists of an output unit that performs weighted calculations on the output value from the correction circuit, A reservoir computing device in which the semiconductor device and the physical reservoir device are sealed within the same package.

Images