Storage pool computing device and data processing method

By designing a reserve pool computing device, using mask units and rotating neuron units to process input signals, and periodically switching the neuron circuit array to achieve delay feedback and time-sharing multiplexing, the problems of high overhead and complexity in the prior art are solved. , realizes efficient and low-power storage pool calculations.

CN114418078BActive Publication Date: 2024-11-22TSINGHUA UNIVERSITY
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202210049571.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-17
Publication Date
2024-11-22
Estimated Expiration
2042-01-17

AI Technical Summary

Technical Problem

The existing reserve pool computing architecture based on delay feedback and time-sharing multiplexing has high overhead, complexity and power consumption problems, and lacks parallel computing capabilities, which limits the development and application of hardware RC systems.

Method used

A reserve pool computing device is designed to process input signals through signal input circuit, reserve pool circuit and output layer unit, and to process input signals through periodic switching of neuron circuit arrays and delay feedback and The time-sharing multiplexing function reduces the demand for auxiliary circuits and optimizes operating efficiency.

Benefits of technology

It realizes efficient reserve pool calculation, reduces implementation costs, optimizes operating efficiency, and does not require additional auxiliary circuits, improves the system's parallel computing power and reduces power consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114418078B_ABST
    Figure CN114418078B_ABST
Patent Text Reader

Abstract

The present disclosure provides a reservoir computing device and a data processing method, wherein the reservoir computing device includes a signal input circuit, a reservoir circuit and an output layer unit. The signal input circuit is configured to receive an input signal. The reservoir circuit includes a plurality of reservoir units. Each reservoir unit includes a mask unit and a rotation neuron unit. The mask unit is configured to receive an input signal and perform a first processing on the input signal using a first weight to obtain a first processing result. The rotation neuron unit is configured to receive the first processing result from the mask unit and perform a second processing on the first processing result to obtain a second processing result. The output layer unit is configured to multiply a plurality of second processing results of a plurality of reservoir units by a second weight matrix to obtain a third processing result, and output the third processing result. The reservoir computing device can realize a complete reservoir computing function with a simple structure, optimize the operating efficiency and reduce the implementation cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the present disclosure relate to a storage pool computing device and a data processing method. Background Art

[0002] Reservoir Computing (RC) is an important branch of Recurrent Neural Network (RNN). RC was first proposed in the early 21st century to solve the gradient explosion and gradient vanishing problems in RNN training. As an implementation of RNN, RC network has demonstrated powerful time series signal processing capabilities. In recent years, RC has been widely used in fields such as biological signal processing, chaotic sequence prediction and pattern recognition. Summary of the invention

[0003] At least one embodiment of the present disclosure provides a reservoir computing device, comprising: a signal input circuit, configured to receive an input signal; a reservoir circuit, the reservoir circuit comprising a plurality of reservoir units, wherein each of the plurality of reservoir units comprises: a mask unit, configured to receive the input signal and perform a first processing on the input signal using a first weight to obtain a first processing result, and a rotation neuron unit, configured to receive the first processing result from the mask unit and perform a second processing on the first processing result to perform dimensionality increase, nonlinear operation and recursive connection to obtain a second processing result; an output layer unit, configured to multiply the plurality of second processing results of the plurality of reservoir units by a second weight matrix to obtain a third processing result, and output the third processing result.

[0004] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the mask unit includes: an input end, configured to receive the input signal; an input weight configuration module, configured to receive a control signal about the first weight, perform the first processing on the input signal received by the input end and the first weight, and obtain the first processing result; and an output end, configured to output the first processing result of the input weight configuration module.

[0005] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the input weight configuration module includes: an inverter; a plurality of switches, wherein each of the plurality of switches includes: a first switch input terminal, a second switch input terminal, a switch output terminal, and a switch control terminal, wherein the first switch input terminal is connected to the input terminal of the mask unit to receive the input signal, the inverter is connected to the input terminal of the mask unit to receive the input signal and inverts the input signal to obtain an inverted input signal, the second switch input terminal is connected to the inverter to receive the inverted input signal, the switch output terminal is connected to the output terminal of the mask unit, and the switch control terminal is configured to receive the control signal and output the signal received by the first switch input terminal or the second switch input terminal from the switch output terminal.

[0006] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the control signal is a random control signal.

[0007] For example, in the reserve pool computing device provided by at least one embodiment of the present disclosure, the rotating neuron unit includes: N front neuron rotor circuits, wherein each of the N front neuron rotor circuits includes a first selection signal terminal, a first input terminal, and N first output terminals ranked from 1st to Nth; N rear neuron rotor circuits, wherein the N rear neuron rotors correspond one-to-one to the N front neuron rotor circuits, and each rear neuron rotor circuit includes a second selection signal terminal, N second input terminals ranked from 1st to Nth, and a second output terminal; N neuron circuits, wherein the first terminal of the mth neuron circuit in the N neuron circuits is connected to the mth first output terminal of each of the N front neuron rotor circuits, and the second terminal of the mth neuron circuit in the N neuron circuits is connected to the mth second input terminal of each of the N rear neuron rotor circuits; a timing control circuit connected to the first selection signal terminal of each of the N front neuron rotor circuits. The N neuron rotor circuits are connected to a communication signal terminal, and are connected to a second selection signal terminal of each of the N rear neuron rotor circuits, and are configured to generate a selection signal, thereby applying the selection signal to the N front neuron rotor circuits and the N rear neuron rotor circuits at the same time, wherein the N front neuron rotor circuits are configured as a whole, according to the selection signal, each of the N front neuron rotor circuits selects one of its own N first output terminals, and the serial numbers of the first output terminals selected by the N front neuron rotor circuits are different from each other, and the N rear neuron rotor circuits are configured as a whole, according to the selection signal, each of the N front neuron rotor circuits selects one of its own N second input terminals, and the serial numbers of the second input terminals selected by the N rear neuron rotor circuits are different from each other, and the first output terminal of the front neuron rotor circuit and the second input terminal of the rear neuron rotor circuit connected to the same neuron circuit are selected at the same time, N is a positive integer greater than 1, m=1, 2,…, N.

[0008] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the value of the selection signal varies with time and makes that in N operation cycles: each of the N front neuron rotor circuits selects its own 1st to Nth first output terminals to be selected in sequence; each of the N rear neuron rotor circuits selects its own 1st to Nth second input terminals to be selected in sequence.

[0009] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, each of the multiple neuron circuits includes: a nonlinear activation circuit; an integration circuit; and an attenuation circuit, wherein a first end of the nonlinear activation circuit is connected to an input end of the neuron circuit, a second end of the nonlinear activation circuit is connected to an output end of the neuron circuit, a first end of the integration circuit is connected to the output end of the neuron circuit, a first end of the attenuation circuit is connected to an input end of the neuron circuit, and a second end of the integration circuit is connected to a second end of the attenuation circuit.

[0010] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the nonlinear activation circuit includes a diode, the cathode of the diode is connected to the output end of the neuron circuit, the anode of the diode is connected to the reference voltage end, the integration circuit includes an integration resistor and a capacitor, the first end of the integration resistor is connected to the input end of the neuron circuit, the second end of the integration resistor is connected to the first end of the capacitor and the output end of the neuron circuit, the second end of the capacitor is connected to the reference voltage end, and the attenuation circuit includes an attenuation resistor, the first end of the attenuation resistor is connected to the output end of the neuron circuit, and the second end of the attenuation resistor is connected to the reference voltage end.

[0011] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, each of the front neuron rotor circuits is a first multiplexer, and each of the rear neuron rotor circuits is a second multiplexer.

[0012] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the timing control circuit includes: a counter configured to generate the selection signal under the control of a clock signal.

[0013] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the output layer unit includes: a multiplication-addition module configured to multiply a plurality of the second processing results with the second weight matrix to obtain the third processing result.

[0014] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the multiplication and addition module includes a memristor array, the memristor array includes a plurality of memristors arranged in an array, and a plurality of conductance values ​​of the plurality of memristors arranged in an array correspond to the values ​​of a plurality of elements of the second weight matrix.

[0015] For example, in the reserve pool computing device provided in at least one embodiment of the present disclosure, the output layer unit further includes: a parameter setting module configured to set the conductance value of the memristor array.

[0016] For example, at least one embodiment of the present disclosure provides a data processing method corresponding to the above-mentioned reserve pool computing device, including: using the reserve pool computing device to perform inference computing operations; or, using the reserve pool computing device to perform training computing operations.

[0017] For example, in the data processing method provided by at least one embodiment of the present disclosure, the inference calculation operation includes: receiving an input signal for the inference calculation operation through the signal input circuit; performing the first processing on the input signal and the first weight through the reserve pool circuit to obtain the first processing result, performing the second processing on the first processing result to perform dimensionality increase, nonlinear operation and recursive connection to obtain multiple second processing results; multiplying the multiple second processing results with the second weight matrix through the output layer unit to obtain the third processing result, and outputting the third processing result.

[0018] For example, in the data processing method provided by at least one embodiment of the present disclosure, the training calculation operation includes: receiving an input signal for the training calculation operation and a label value for the input signal through the signal input circuit; performing the first processing on the input signal and the first weight through the reserve pool circuit to obtain the first processing result, performing the second processing on the first processing result to perform dimension increase, nonlinear operation and recursive connection to obtain multiple second processing results; multiplying the multiple second processing results with the second weight matrix through the output layer unit to obtain the third processing result; calculating the error of the second weight matrix according to the multiple third processing results and the label value of the training input signal to update the second weight matrix; and writing the updated second weight matrix into the output layer unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present disclosure, but are not intended to limit the present disclosure.

[0020] Figure 1 A schematic diagram of a storage pool computing architecture is shown;

[0021] Figure 2 A schematic diagram of a storage pool computing device provided for at least one embodiment of the present disclosure;

[0022] Figure 3 A schematic diagram of a mask unit of a reservoir computing device provided for at least one embodiment of the present disclosure;

[0023] Figure 4AA schematic block diagram of a rotation neuron unit of a reservoir computing device provided for at least one embodiment of the present disclosure;

[0024] Figure 4B A schematic diagram of a rotation neuron unit of a reservoir computing device provided by at least one embodiment of the present disclosure;

[0025] Figure 5A A schematic diagram of the dynamic characteristics of an exemplary neuron circuit provided for at least one embodiment of the present disclosure;

[0026] Figure 5B A schematic diagram of an exemplary neuron circuit provided for at least one embodiment of the present disclosure;

[0027] Figure 6 A schematic diagram of an output layer unit based on a memristor array provided for at least one embodiment of the present disclosure;

[0028] Figure 7 A flowchart of a data processing method provided for at least one embodiment of the present disclosure; and

[0029] Figure 8 A flowchart of another data processing method provided for at least one embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0031] Unless otherwise defined, the technical terms or scientific terms used in the present disclosure should be understood by people with ordinary skills in the field to which the present disclosure belongs. The "first", "second" and similar words used in the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, similar words such as "one", "one" or "the" do not indicate quantity restrictions, but indicate that there is at least one. Similar words such as "include" or "comprise" mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Similar words such as "connect" or "connected" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0032] The present disclosure is described below through several specific embodiments. In order to keep the following description of the embodiments of the present disclosure clear and concise, detailed descriptions of known functions and known components may be omitted. When any component of the embodiments of the present disclosure appears in more than one figure, the component is represented by the same or similar reference numeral in each figure.

[0033] With the full development of RC algorithms, the method of implementing RC with hardware has also received attention in the field of brain-like computing in the past decade. The RC algorithm implemented by software relies on the data storage and processing under the traditional von Neumann architecture. The idea of ​​implementing RC with hardware relies on an architecture that can efficiently mine the physical properties of components and realize the basic functions of RC.

[0034] For example, Figure 1 A schematic diagram of a reserve pool computing architecture is shown. The reserve pool computing architecture consists of three parts: an input layer, an intermediate layer, and an output layer. For example, the input layer can be composed of one or more nodes, which is a type of feed-forward neural network (FNN); the intermediate layer is composed of multiple nodes, which is a type of RNN; and the output layer can be an adder with weights. Specifically, the RC calculation process in the reserve pool computing architecture can be shown as formula (1):

[0035]

[0036] Where i = 1,…,N, x in represents the input signal from the input layer to the middle layer, x i Represents the intermediate processing result of the intermediate layer input to the output layer, y out Represents the output signal of the output layer, Wi Represents the weight matrix of the output layer.

[0037] In RC, the connection between the input layer and the middle layer of the neural network is defined by a randomly generated and fixed matrix, and the complex response of this matrix to the input data maps the data into a high-dimensional and linearly separable feature space. Therefore, training an RC network only requires linear regression of the high-dimensional state matrix of the output layer to implement some functions of RNN, which greatly simplifies the complexity and cost of training.

[0038] As an implementation of RNN, RC network has shown powerful time series signal processing capabilities. In recent years, RC has been widely used in biological signal processing, chaotic sequence prediction, pattern recognition and other fields.

[0039] The use of hardware to implement RC computing has been further developed due to the introduction of time-division multiplexing and delayed feedback loops into the implementation architecture of hardware RC. The operation of time-division multiplexing of input signals reduces the number of neurons required. In addition, the existence of delayed feedback loops gives the network a certain memory capacity, which is the key to the implementation of RC. In the above RC architecture based on delayed feedback and time-division multiplexing, the physical properties of various electronic components, optical components or optoelectronic devices are fully utilized in machine learning related applications and experiments, and good results have been achieved.

[0040] The existing RC architecture based on delayed feedback and time-division multiplexing reduces the cost of hardware implementation, but there are still obvious defects. First, due to the construction of the delayed feedback loop, there is still a large cost and system complexity. For example, in the circuit implementation, the delay signal still requires a digital-to-analog conversion module and a memory; in the optoelectronic RC system, the delay line is usually implemented by a section of optical fiber several kilometers long. These implementation methods all have power consumption or volume problems. In the absence of a delayed feedback loop, the memory capacity and high-dimensional mapping capacity of the RC system will be greatly reduced or lost. Secondly, the time-division multiplexing of the input signal increases the complexity of the system and brings a large number of serial operations. These shortcomings hinder the further development and application of hardware RC systems. In the existing RC systems, in addition to the key physical devices for computing functions, a large number of control units and digital-to-analog conversions are often required to assist the operation of the system, or a bulky optoelectronic system, and an ideal brain-like computing unit implemented by RC should be parallel, low-power, and simple and efficient.

[0041] At least one embodiment of the present disclosure provides a reservoir computing device, which includes a signal input circuit, a reservoir circuit and an output layer unit. The signal input circuit is configured to receive an input signal. The reservoir circuit includes a plurality of reservoir units. Each reservoir unit includes a mask unit and a rotation neuron unit. The mask unit is configured to receive an input signal and perform a first processing on the input signal using a first weight to obtain a first processing result. The rotation neuron unit is configured to receive the first processing result from the mask unit, and perform a second processing on the first processing result to perform dimensionality increase, nonlinear operation and recursive connection to obtain a second processing result. The output layer unit is configured to multiply a plurality of second processing results of a plurality of reservoir units by a second weight matrix to obtain a third processing result, and output the third processing result.

[0042] The reservoir computing device replaces the delayed feedback and time-division multiplexing functions in the existing device by periodically switching the connection between the neuron circuit array and the input and output circuit. The reservoir computing device can realize the complete RC function without other auxiliary circuits, optimizes the operation efficiency and reduces the implementation cost.

[0043] At least one embodiment of the present disclosure further provides a data processing method corresponding to the above-mentioned reserve pool computing device, and the data processing method includes inference computing or training computing.

[0044] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings, but the present disclosure is not limited to these specific embodiments.

[0045] Figure 2 A schematic diagram of a storage pool computing device provided for at least one embodiment of the present disclosure.

[0046] For example, Figure 2 As shown, the reservoir computing device includes a signal input circuit 20, a reservoir circuit 10 and an output layer unit 30. The signal input circuit 20 is configured to receive an input signal. The reservoir circuit 10 includes a plurality of reservoir units. Each reservoir unit includes a mask unit and a rotation neuron unit. The mask unit is configured to receive an input signal and perform a first processing on the input signal using a first weight to obtain a first processing result. The rotation neuron unit is configured to receive the first processing result from the mask unit, and perform a second processing on the first processing result to perform dimensionality increase, nonlinear operation and recursive connection to obtain a second processing result. The output layer unit 30 is configured to multiply a plurality of second processing results of a plurality of reservoir units by a second weight matrix to obtain a third processing result, and output the third processing result.

[0047] For example, Figure 2As shown, the reservoir circuit 10 includes n reservoir units, namely, reservoir unit 11 to reservoir unit 1n, where n is a positive integer, for example, greater than or equal to 2. Each reservoir unit includes a mask unit and a rotation neuron. After receiving the input signal, the signal input circuit 20 outputs the input signal in parallel to n mask units in the n reservoir units, namely, mask unit 101 to mask unit 10n. After each mask unit receives the input signal, it performs a first processing on the input signal using a first weight, obtains a first processing result and inputs it to the corresponding rotation neuron. For example, the input signal includes n input signals 1 to n in the form of unit vectors, and the input signals 1 to n are respectively input to n different reservoir units 11 to 1n. For the same input signal, inputting it in parallel to multiple mask units configured with different first weights can improve the richness of the state, thereby improving the computing performance of the network.

[0048] For example, after each rotation neuron unit receives multiple first processing results from the corresponding mask unit, the rotation neuron unit will play a major role in calculation. The n rotation neuron units 111-11n perform second processing on the multiple first processing results to perform three types of processing: dimension increase, nonlinear operation, and recursive connection. After the operation of the n rotation neuron units 111-11n, multiple second processing results are obtained and input into the output layer unit 30 in parallel.

[0049] For example, the output layer unit 30 includes a multiplication-addition module; for example, the multiplication-addition module is implemented by a storage-computation integrated array, such as a memristor array. The multiplication-addition module stores a trained second weight matrix. The multiplication-addition module is configured to multiply the plurality of second processing results output by the rotation neuron unit with the second weight matrix to obtain a third processing result, and output the third processing result. The second weight matrix in the multiplication-addition module can configure a corresponding target value according to the application scenario to achieve functions such as classification, prediction, and recognition.

[0050] Figure 3 A schematic diagram of a mask unit of a reservoir computing device provided for at least one embodiment of the present disclosure.

[0051] For example, Figure 3As shown, in the i-th reserve pool unit 1i of the reserve pool circuit 10, the mask unit 10i includes an input terminal 10i1, an input weight configuration module 10i2 and an output terminal 10i3, where i is a positive integer greater than or equal to 1 and less than or equal to n. The input terminal 10i1 is configured to receive an input signal i. The input weight configuration module 10i2 is configured to receive a control signal regarding a first weight. The input weight configuration module 10i2 performs a first processing on the input signal i received by the input terminal 10i1 and the first weight to obtain N first processing results. N is a positive integer, for example, greater than or equal to 2. The output terminal 10i3 is configured to output the N first processing results of the input weight configuration module 10i2, thereby completing the first dimensionality increase processing of the input signal.

[0052] For example, Figure 3 As shown, an example of the input weight configuration module 10i2 includes an inverter and N switches S1, S2, ..., S N Each switch includes a first switch input terminal, a second switch input terminal, a switch output terminal and a switch control terminal (not shown in the figure). The first switch input terminal is connected to the input terminal 10i1 to receive an input signal i (the input signal i is transmitted through Figure 2 The signal input circuit 20 in the embodiment is input to the input terminal 10i1). The inverter is connected to the input terminal 10i1 to receive the input signal i and invert the input signal i to obtain the inverted input signal i. The second switch input terminal is connected to the inverter to receive the inverted input signal i. The switch output terminal is connected to the output terminal 10i3. The switch control terminal is configured to receive a control signal and output the signal received by the first switch input terminal or the second switch input terminal from the switch output terminal.

[0053] For example, the specific process of the first processing is that the input weight configuration module 10i2 first divides the input signal i into positive and negative values ​​through an inverter, and then controls N switches S1, S2, ..., S N The input to each neuron is selected as a positive value or a negative value, thereby applying a first weight (here the weight values ​​are randomly distributed 1 and -1). For example, the input signal i first passes through an inverter to generate a signal with the same amplitude and opposite polarity as the original signal (equivalent to multiplying by the weight "-1"). Subsequently, the N control signals are input into the N switches S1, S2, ... S N The switch control terminal of the RC software algorithm is used to select whether the switch output terminal of each switch outputs a positive signal or a negative signal, thereby obtaining N first processing results 1 to N. This process corresponds to multiplying the input signal i in the RC software algorithm by the corresponding first weight.

[0054] For example, in the above example, the first weight may be the inverter and N switches S1, S2, ... NThe random distribution of N -1 and 1 is realized. In other examples, the first weight may also be N random numbers between 0 and 1 realized by other methods, or may be other numerical distributions that can realize signal dimensionality increase classification processing.

[0055] For example, the control signal is a random control signal, and N switches S1, S2, ... N The switch control terminal controls each switch individually.

[0056] For example, the input signal may be a continuous analog input signal. For example, the inverter may be an inverting operational amplifier circuit. The embodiments of the present disclosure do not limit these specific implementations.

[0057] It should be noted that after the debugging of the reserve pool computing device is completed or during its operation, the n groups of switches controlling the first weights in all n mask units must remain unchanged to ensure that similar responses are generated to similar input signals during use for processing by the next level of rotation neuron units.

[0058] Figure 4A A schematic block diagram of a rotation neuron unit of a reservoir computing device provided for at least one embodiment of the present disclosure; Figure 4B A schematic diagram of an example of a rotating neuron unit of a reservoir computing device provided for at least one embodiment of the present disclosure.

[0059] For example, Figure 4A As shown, in the i-th reservoir unit 1i of the reservoir circuit 10, the rotating neuron unit 11i includes N front neuron rotor circuits 11i1, N neuron circuits 11i2, N rear neuron rotor circuits 11i3 and a timing control circuit 11i4 (such as Figure 4B Here, N is a positive integer, for example, greater than or equal to 2.

[0060] For example, Figure 4B As shown, in one example, in the i-th rotation neuron unit, the N front neuron rotor circuits 11i1 include neuron rotor circuits 11i11-11i1N (labeled as m1, m2, ..., m N ), N neuron circuits 11i2 include neuron circuits 11i21-11i2N, and N post-neuron rotor circuits 11i3 include post-neuron rotor circuits 11i31-11i3N (labeled as m1', m2', ..., m N ').

[0061] For example, each pre-neuron rotor circuit includes a first selection signal terminal, a first input terminal, and N first output terminals (in1, in2, ..., in N). The order of the N first output terminals of adjacent pre-neuron rotor circuits is different; for example, the N first output terminals of the pre-neuron rotor circuit 11i11 labeled as m1 are in1, in2, and in3. N ,…,in2, the N first output terminals of the pre-neuron rotor circuit 11i12 marked as m2 are in2, in1,…, in N , ..., for the N The N first output terminals of the front neuron rotor circuit 11i1N are respectively in N ,in N-1 ,…,in1.

[0062] For example, Figure 4B As shown, the N post-neuron rotors 11i3 correspond to the N pre-neuron rotor circuits 11i1 one by one; each post-neuron rotor circuit includes a second selection signal terminal, N second input terminals (out1, out2, ..., out N ) and the second output terminal. The order of the N second input terminals of adjacent post-neuron rotor circuits is different; for example, the N second input terminals of the post-neuron rotor circuit 11i31 marked as m1' are out1, out N ,…,out2, the N second input terminals of the post-neuron rotor circuit 11i32 marked as m2' are out2, out1,…,out N , ..., for the N The N second input terminals of the post-neuron rotor circuit 11iN are out N ,out N-1 ,…,out1.

[0063] For example, Figure 4B As shown, the first end of the neuron circuit 11i2m is connected to the mth first output end in of each front neuron rotor circuit. m , that is, the N first output terminals in of the N front neuron rotor circuits 11i11 to 11i1N m The second end of the neuron circuit 11i2m is connected to the mth second input end out of each post neuron rotor circuit. m , that is, the N first output terminals out of the N post-neuron rotor circuits 11i21 to 11i2N m Here m = 1, 2, ..., N; therefore, the neuron circuit 11i2m is simultaneously connected to the N first output terminals in of the N front neuron rotor circuits 11i1 m connected, and at the same time connected to the N second input terminals out of the N post-neuron rotor circuits 11i3 m connect.

[0064] For example, each pre-neuron rotor circuit 110 may be a first multiplexer, that is, a N-to-one multiplexer, at N first output terminals (in1, in2, ..., in N ) is selected to be connected to the first input terminal; for example, each post-neuron rotor circuit 130 can be a second multiplexer, that is, an N-to-one multiplexer, at the N second input terminals (out1, out2, ..., out N ) is selected to be connected to the second output terminal.

[0065] For example, the timing control circuit 11i4 includes a counter configured to generate a selection signal under the control of a clock signal. The timing control circuit 11i4 is connected to the first selection signal terminal of each front neuron rotor circuit and to the second selection signal terminal of each rear neuron rotor circuit. Thus, the selection signal is applied to the N front neuron rotor circuits 11i1 and the N rear neuron rotor circuits 11i3 at the same time, that is, the first selection signal terminal and the second selection signal terminal receive the same selection signal.

[0066] It should be noted that the first selection signal terminals of the N front neuron rotor circuits 11i1 and the second selection signal terminals of the N rear neuron rotor circuits 11i3 can receive the same selection signal or different selection signals.

[0067] For example, the N pre-neuron rotor circuits 11i1 are configured as a whole, and each pre-neuron rotor circuit selects its own N first output terminals (in1, in2, ..., in N ), and the serial numbers of the first output terminals selected by the N front neuron rotor circuits 11i1 are different from each other. The N rear neuron rotor circuits 11i3 are configured as a whole, according to the selection signal, each rear neuron rotor circuit selects its own N second input terminals (out1, out2, ..., out N ), and the serial numbers of the second input terminals selected by the N rear neuron rotor circuits 11i3 are different from each other. The first output terminal of the front neuron rotor circuit and the second input terminal of the rear neuron rotor circuit connected to the same neuron circuit are selected at the same time, thereby obtaining a path from the first input terminal to the second output terminal.

[0068] For example, the value of the selection signal varies with time and makes it possible that in N operation cycles: each pre-neuron rotor circuit selects its first to Nth first output terminals (in1, in2, ..., in N ) are selected in sequence; each post-neuron rotor circuit selects its own 1st to Nth second input terminals (out1, out2, ..., out N ) are selected in sequence.

[0069] For example, Figure 4B As shown, the timing control circuit 11i4 may include a counter, for example, a log2N-bit counter, for timing output, for example, an increasing numerical signal. Each pre-neuron rotor circuit 110 may be a first multiplexer (m1, m2, ..., m N ), N first multiplexers (m1, m2, ..., m N ) are connected to the output end of the counter. Each post-neuron rotor circuit 130 can be a second multiplexer (m1', m2', ..., m1') of N-select. N '), N second multiplexers (m1', m2', ..., m N ') are also connected to the output end of the counter. When the input end of the counter receives a driving pulse signal of a specific frequency, the counter simultaneously outputs a selection signal to the first selection signal end and the second selection signal end, and the first output end (in1, in2, ..., in1) of the first multiplexer of the N front neuron rotor circuit 11i1 is connected to the output end of the counter. N ) and the N second input terminals (out1, out2, ..., out N ) are selected in sequence.

[0070] For example, at a certain (the first) moment, the timing control circuit 11i4 outputs a selection signal to the first selection signal terminal and the second selection signal terminal, and the selection signal is used to select the first output terminal of a predetermined number in the first multiplexer, and to select the second input terminal of the same predetermined number in the second multiplexer.

[0071] For example, at this moment, the strobe signal is used to select Figure 4B The first first output terminal of the first multiplexer and the first second input terminal of the second multiplexer are selected, then the first output terminal in1 of the first multiplexer m1 is selected, and the first output terminal in2 of the first multiplexer m2 is selected, and the first output terminal in3 of the first multiplexer m3 is selected...; correspondingly, the second input terminal out1 of the second multiplexer m1' is selected, and the second input terminal out2 of the second multiplexer m2' is selected, and the second input terminal out3 of the second multiplexer m3' is selected...

[0072] In this case, the neuron circuit 1 is connected to the first output terminals in1 of all the first multiplexers and to the second input terminals out1 of all the second multiplexers, thereby obtaining a path from the first input terminal of the first multiplexer m1, the first output terminal in1 of the first multiplexer m1, the neuron circuit 1, the second input terminal out1 of the second multiplexer m1', and the second output terminal of the second multiplexer m1'; similarly, the neuron circuit j is connected to the first output terminals in1 of all the first multiplexers. j connected to the second input terminal out of all the second multiplexers j connection, thereby obtaining from the first multiplexer m j The first input terminal of the first multiplexer m j The first output terminal in1, the neuron circuit j, the second multiplexer m j 'The second input terminal out j , the second multiplexer m j Here, j is greater than 1 and less than or equal to N.

[0073] For example, at the next (second) moment, the strobe signal is used to select Figure 4B The second first output terminal of the first multiplexer and the second second input terminal of the second multiplexer in the first multiplexer are connected to the first output terminal in the first multiplexer m1. N is selected, and the first output terminal in1 of the first multiplexer m2 is selected, the first output terminal in2 of the first multiplexer m3 is selected...; correspondingly, the second input terminal out N is selected, and the second input terminal out1 of the second multiplexer m2' is selected, the second input terminal out2 of the second multiplexer m3' is selected...

[0074] In this case, the neuron circuit 1 is connected to the first output terminals in1 of all the first multiplexers and to the second input terminals out1 of all the second multiplexers, thereby obtaining a path from the first input terminal of the first multiplexer m2, the first output terminal in1 of the first multiplexer m2, the neuron circuit 1, the second input terminal out1 of the second multiplexer m2', and the second output terminal of the second multiplexer m2'; similarly, the neuron circuit j is connected to the first output terminals in1 of all the first multiplexers. j connected to the second input terminal out of all the second multiplexers j connection, thereby obtaining from the first multiplexer m j+1 The first input terminal of the first multiplexer m j+1 The first output terminal inj , neuron circuit j, second multiplexer m j+1 'The second input terminal out j , the second multiplexer m j+1 Here, j is greater than 1 and less than N. The neuron circuit N is connected to the first output terminal of all the first multiplexers in N connected to the second input terminal out of all the second multiplexers N The first input terminal of the first multiplexer m1 and the first output terminal of the first multiplexer m1 are connected. N , neuron circuit N, the second input terminal out of the second multiplexer m1' N , a path of the second output terminal of the second multiplexer m1'.

[0075] …

[0076] For example, at N times, the strobe signal is used to select Figure 4B The Nth first output terminal of the first multiplexer and the Nth second input terminal of the second multiplexer are selected, then the first output terminal in2 of the first multiplexer m1 is selected, and the first output terminal in3 of the first multiplexer m2 is selected, and the first output terminal in4 of the first multiplexer m3 is selected...; correspondingly, the second input terminal out2 of the second multiplexer m1' is selected, and the second input terminal out3 of the second multiplexer m2' is selected, and the second input terminal out4 of the second multiplexer m3' is selected...

[0077] In this case, the neuron circuit 1 is connected to the first output terminals in1 of all the first multiplexers and to the second input terminals out1 of all the second multiplexers, thereby obtaining the first multiplexers m N The first input terminal of the first multiplexer m N The first output terminal in1, the neuron circuit 1, the second multiplexer m N 'The second input terminal out1, the second multiplexer m N 'The second output terminal of the path; similarly, the neuron circuit j due to the first output terminal of all the first multiplexers in j connected to the second input terminal out of all the second multiplexers j connection, thereby obtaining from the first multiplexer m j-1 The first input terminal of the first multiplexer m j-1 The first output terminal in j , neuron circuit j, second multiplexer m j-1 'The second input terminal outj , the second multiplexer m j-1 Here, j is greater than 1 and less than or equal to N.

[0078] It can be seen from the above that, at each moment, a neuron circuit can only connect to one first multiplexer and one second multiplexer, and it is not possible to connect to multiple first multiplexers or multiple second multiplexers at the same time.

[0079] In particular, therefore, from the 1st moment to the Nth moment, the second multiplexer m j The second output terminal of ' sequentially outputs the signals processed by neuron circuit j, neuron circuit j-1, ..., neuron circuit 1, neuron circuit N, neuron circuit N-1, ..., neuron circuit j+1. Here, j is greater than or equal to 1 and less than or equal to N.

[0080] For example, Figure 4B As shown, in the i-th reservoir unit 1i of the reservoir circuit 10, the specific process of the rotation neuron unit 11i performing the second processing on the N first processing results is that the mask unit 10i outputs the N first processing results 1 to N to the rotation neuron unit 11i, that is, the first multiplexer m j The first input terminal receives the jth switch S of the mask unit j The first processing result j is output from the switch output terminal of the first multiplexer m. j After the first input terminal of is input, from the first moment to the Nth moment, it is processed by neuron circuit j, neuron circuit j-1, ..., neuron circuit 1, neuron circuit N, neuron circuit N-1, ..., neuron circuit j+1 in sequence, and is output from the second multiplexer m j The second output terminal of ' is outputted as the second processing result j in sequence. Here, i is greater than or equal to 1 and less than or equal to n, and j is greater than or equal to 1 and less than or equal to N.

[0081] For example, in the rotating neuron unit of the reserve pool computing device provided in the above embodiment, each front neuron rotor circuit and each rear neuron rotor circuit sequentially poll different neuron circuits, thereby achieving the effect of "rotation" of the neuron circuit and realizing the dimensionality increase processing of the signal. The process of circuit switching is equivalent to the matrix multiplication of the ring RC in the software algorithm.

[0082] Figure 5A A schematic diagram of the dynamic characteristics of an exemplary neuron circuit provided in at least one embodiment of the present disclosure, Figure 5B A schematic diagram of an exemplary neuron circuit provided for at least one embodiment of the present disclosure.

[0083] For example, Figure 5AAs shown, each neuron circuit (in Figure 4B Taking the neuron circuit 11i2m in the figure as an example) it includes a nonlinear activation circuit im1, an integration circuit im2 and an attenuation circuit im3. A first end of the nonlinear activation circuit im1 is connected to an input end im01 of the neuron circuit 11i2m, and a second end of the nonlinear activation circuit im1 is connected to an output end im02 of the neuron circuit 11i2m; a first end of the integration circuit im2 is connected to an output end im02 of the neuron circuit 11i2m, a first end of the attenuation circuit im3 is connected to an input end im01 of the neuron circuit 11i2m, and a second end of the integration circuit im2 is connected to a second end of the attenuation circuit im3.

[0084] For example, Figure 5A As shown, the nonlinear activation circuit im1 can correspond to a nonlinear function f, and the integration circuit im2 can correspond to the integral Z -1 , the attenuation circuit im3 can correspond to the attenuation d in the integration process. After the first processing result m is input from the input terminal im01 to the neuron circuit 11i2m, the nonlinear activation circuit im1 performs nonlinear processing on the first processing result m through the nonlinear function f; at the same time, the attenuation circuit im3 attenuates the first processing result m by d, and the integration circuit im2 integrates the attenuated first processing result Z -1 Finally, a second processing result m after nonlinear and recursive processing is obtained and output from the output terminal im02.

[0085] For example, Figure 5B As shown, in one example, the nonlinear activation circuit im1 includes a diode D ReLU , diode D ReLU The cathode of the diode D is connected to the output terminal im02. ReLU The positive terminal and the reference voltage terminal V ref The integration circuit im2 includes an integration resistor R int and capacitor C int , integral resistance R int The first end is connected to the input terminal im01, and the integral resistor R int The second end of the capacitor C int The first end and the output end im02 are connected, the capacitor C int The second end is connected to the reference voltage terminal V ref The attenuation circuit im3 includes an attenuation resistor R leakage , attenuation resistance R leakage The first end is connected to the output terminal im02, and the attenuation resistor R leakage The second end is connected to the reference voltage terminal V ref For example, in Figure 5B In the circuit shown, the nonlinear equation f is controlled by the diode DReLU Provide, points Z -1 By R int and C int Provided, attenuation d is provided by resistor R leakage Provided. Resistor R leakage The resistance is greater than the integral resistance R int resistance value.

[0086] It should be noted that the nonlinear activation circuit im1 may also include other components that can provide a nonlinear processing function; the integration circuit im2 may also include other components that can provide an integration processing function; and the attenuation circuit im3 may also include other components that can provide an attenuation processing function. The embodiments of the present disclosure do not limit the specific configuration of these circuits.

[0087] For example, in the neuron circuit provided in the above embodiment, due to the integration of the first processing result input at the past moment by the integration circuit, the second processing result output by the neuron circuit at each moment includes the information of the first processing result at the current moment and the information of the first processing result at the past moment. Therefore, the memory characteristics of the neuron circuit itself can be used to maintain the input information at the past moment, so that the reserve pool computing device does not need additional memory during the simulation computing process, thereby optimizing the operating efficiency and reducing the implementation cost.

[0088] Figure 6 A schematic diagram of an output layer unit 30 based on a memristor array according to at least one embodiment of the present disclosure.

[0089] For example, Figure 6 As shown, the output layer unit 30 includes a multiplication and addition module and a parameter setting module (not shown in the figure). The multiplication and addition module includes a storage-computation integrated array, which may include a memristor array, configured to multiply the plurality of second processing results with the second weight matrix to obtain a third processing result. The parameter setting module is configured to set the conductance value of the memristor array.

[0090] For example, the memristor array of the multiplication-addition module includes a plurality of memristors arranged in an array, and a plurality of conductance values ​​of the plurality of memristors arranged in an array correspond to the values ​​of a plurality of elements of the second weight matrix.

[0091] For example, Figure 6 As shown, the memristor array is composed of a plurality of memristor units, which form an array of M rows and R columns, where M and R are both positive integers. Each memristor unit includes a switch element (eg, a transistor) and one or more memristors. Figure 6 BL <1> BL <2> ...BL <r>They represent the bit lines of the first column, the second column, ..., the Rth column, respectively. The memristors in the memristor units of each column are connected to the bit lines corresponding to the column; SL <1> , SL <2> ...SL <m>Respectively represent the source lines of the first row, the second row, ... the Mth row, the source of the switch element in each row of the memristor unit (for example, the source of the transistor) is connected to the source line corresponding to the row; in addition, the control electrode of the switch element in each row of the memristor unit (for example, the gate of the transistor) is connected to the word line corresponding to the row (not shown in the figure). According to Kirchhoff's law, by setting the state of the memristor unit (for example, the resistance value or the conductance value) and applying the corresponding word line signal and bit line signal to the word line and the bit line, the above memristor array can complete the multiplication and accumulation calculation in parallel.

[0092] For example, each rotation neuron unit outputs N second processing results, and the n rotation neuron units in the n reserve pool units output n×N second processing results in total, that is, R=n×N.

[0093] For example, R second processing results are voltage outputs (V1, V2...V R ), can be directly used as a high-dimensional feature vector for multiplication and addition of the output layer unit 30. For example, the multiple conductance values ​​G of the memristors arranged in M ​​rows and R columns are t =[G 11 ,G 12 …G 1R ; G 21 ,G 22 …G 2R ;…;G M1 ,G M2 …G MR ] corresponds to the values ​​of multiple elements of the second weight matrix. Multiple conductance values ​​of the memristors arranged in M ​​rows and R columns are set by the parameter setting module.

[0094] For example, on the bit line BL <1> BL <2> ...BL <r>The R second processing results (for example, the input voltage vector V t =[V1,V2,…,V R ]), the R second processing results and the second weight matrix with M rows and R columns (for example, the conductivity matrix G t ) are multiplied to obtain M third processing results (for example, the output current vector I t =[I1,I2,…,I M ]) and on the source line SL <1> , SL <2> ...SL <m>Output M third processing results. For example, the specific calculation process is shown in formula (2).

[0095]

[0096] For example, for Figure 2 The storage pool computing device shown in the figure outputs R (R = n × N) second processing results to the storage pool computing device. Figure 6 The output layer unit 30 shown multiplies the R second processing results by the second weight matrix in the output layer unit 30, and finally outputs n×M third processing results.

[0097] For example, Figure 6 The memristor cells in the memristor array may have, for example, a 1T1R structure or a 2T2R structure, wherein the memristor cell of the 1T1R structure includes a transistor and a memristor, and the memristor cell of the 2T2R structure includes two transistors and two memristors. The memristor array may further include word lines (not shown in the figure) for each transistor, each word line being used to control the conduction or cutoff of transistors in, for example, a row of memristor cells. For example, the memristor is a non-volatile memristor, including but not limited to RRAM, PCRAM, ECRAM, flash memory (Flash), etc. The present disclosure does not limit the structure of the memristor cell, and memristor cells of other structural forms that can implement multiplication and accumulation operations may also be used.

[0098] It should be noted that the transistors used in the embodiments of the present disclosure may be thin film transistors or field effect transistors (such as MOS field effect transistors) or other switching devices with the same characteristics. The source and drain of the transistor used here may be symmetrical in structure, so the source and drain may have no difference in structure.

[0099] The reserve pool computing device provided by the above embodiment applies the non-volatile memristor array to the output layer unit to realize efficient multiplication and addition. In the above device, the physical properties of the components are fully exploited for RC calculation, and the system cost and operating power consumption are reduced within the acceptable error range.

[0100] Figure 7 A flowchart of a data processing method provided for at least one embodiment of the present disclosure; Figure 8 A flowchart of another data processing method provided for at least one embodiment of the present disclosure.

[0101] For example, the data processing method for the above-mentioned storage pool computing device includes an inference computing operation or a training computing operation. The first weight matrix and the second weight matrix in the inference computing operation can be target values ​​obtained by training according to the application scenario in the training computing operation to achieve functions such as classification, prediction, and recognition of the input signal.

[0102] For example, Figure 7 As shown, in one example, the inference computing operation includes the following steps S101-S103.

[0103] Step S101 : receiving an input signal for an inference calculation operation through the signal input circuit 20 .

[0104] Step S102: performing a first processing on the input signal and the first weight through the reserve pool circuit 10 to obtain a first processing result, and performing a second processing on the first processing result to perform dimension increase, nonlinear operation and recursive connection to obtain multiple second processing results;

[0105] Step S103: multiply the plurality of second processing results by the second weight matrix through the output layer unit 30 to obtain a third processing result, and output the third processing result.

[0106] For example, Figure 7 As shown, in Figure 2 When performing an inference calculation operation in the storage pool computing device shown in the figure, an input signal is first received through the signal input circuit 20. Then, the signal input circuit 20 inputs the input signal to n, for example Figure 3 The input terminals 1011-10n1 of the mask units 101-10n are as follows (for example, the input signal includes n input signals 1-n in the form of unit vectors, and the input signals 1-n are respectively input into n different reservoir units 11-1n). In the mask unit 10i, the input signal i is first processed with the first weight corresponding to the input weight configuration module 10i2 to obtain N first processing results 1-N. Subsequently, the mask unit 10i outputs the N first processing results 1-N to, for example Figure 4B The rotating neuron unit 11i shown in the figure simultaneously inputs a pulse frequency matching the time constant and timing signal of the neuron circuit to the timing control circuit 11i4, and performs the second processing of the N first processing results 1 to N in the rotating neuron unit 11i to perform dimension increase, nonlinear operation and recursive connection, and obtains N second processing results 1 to N at each moment. Finally, for example Figure 6 The output layer unit 30 shown collects R (R=n×N) second processing results of n rotated neuron units 111~11n, processes the R second processing results through the output layer unit 30, for example, multiplying them with the second weight matrix mapped in the memristor array, to obtain n×M third processing results, and outputs the n×M third processing results for subsequent processing.

[0107] For example, Figure 8 As shown, in one example, the training calculation operation includes the following steps S201-S205.

[0108] Step S201 : receiving an input signal for a training calculation operation and a label value for the input signal through the signal input circuit 20 .

[0109] Step S202: The input signal and the first weight are subjected to a first processing through the reserve pool circuit 10 to obtain a first processing result, and the first processing result is subjected to a second processing to perform dimension increase, nonlinear operation and recursive connection to obtain multiple second processing results.

[0110] Step S203: multiplying the plurality of second processing results by the second weight matrix through the output layer unit 30 to obtain a third processing result.

[0111] Step S204: Calculate the error of the second weight matrix according to the plurality of third processing results and the label value of the training input signal to update the second weight matrix.

[0112] Step S205: Write the updated second weight matrix into the output layer unit 30.

[0113] For example, Figure 8 As shown, in Figure 2 When performing a training calculation operation in the storage pool computing device shown in the figure, the input signal and the label value corresponding to the input signal are first received through the signal input circuit 20. Then, the signal input circuit 20 inputs the input signal to n, for example Figure 3 The input terminals 1011-10n1 of the mask units 101-10n are as follows (for example, the input signal includes n input signals 1-n in the form of unit vectors, and the input signals 1-n are respectively input into n different reservoir units 11-1n). In the mask unit 10i, the input signal i is first processed with the first weight corresponding to the input weight configuration module 10i2 to obtain N first processing results 1-N. Subsequently, the mask unit 10i outputs the N first processing results 1-N to, for example Figure 4B The rotating neuron unit 11i shown in the figure simultaneously inputs a pulse frequency matching the time constant and timing signal of the neuron circuit to the timing control circuit 11i4, and the N first processing results 1 to N are subjected to the second processing in the rotating neuron unit 11i to perform dimension increase, nonlinear operation and recursive connection, and N second processing results 1 to N are obtained at each moment. Then, for example Figure 6 The output layer unit 30 shown collects R (R = n × N) second processing results of n rotation neuron units 111 ~ 11n, and processes the R second processing results through the output layer unit 30, for example, multiplying them with the second weight matrix mapped in the memristor array to obtain n × M third processing results. Subsequently, the error of the second weight matrix is ​​calculated based on the n × M third processing results and the label value of the training input signal, for example, the error is used to calculate the loss value, and the loss value is used for back propagation to update the second weight matrix; finally, the updated second weight matrix is ​​written into, for example Figure 6 In the output layer unit 30 shown, for example, the conductance value of each memristor unit in the memristor array as a storage-computation-in-one array is reset.

[0114] For example, in step S204, the second weight matrix may be updated using an algorithm such as linear regression, which is not limited in the embodiments of the present disclosure. During the training process, the R second processing results output by the reserve pool circuit 10 at each moment may be used as variables in the linear regression algorithm, and the desired third processing result may be set as the target value of the linear regression algorithm, thereby calculating a set of second weight matrices. Figure 6 The parameter setting module in sets the conductance values ​​of the M-row and R-column memristors in the memristor array according to the second weight matrix obtained through training.

[0115] There are a few points to note about this disclosure:

[0116] (1) In the drawings of the embodiments of the present disclosure, only the structures related to the embodiments of the present disclosure are involved, and other structures can refer to the general design.

[0117] (2) In the absence of conflict, features in the same embodiment or in different embodiments of the present disclosure may be combined with each other.

[0118] The above are only specific embodiments of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present disclosure, which should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.< / m> < / r> < / m> < / r>

Claims

1. A storage pool computing device, comprising: a signal input circuit configured to receive an input signal; A reservoir circuit, the reservoir circuit comprising a plurality of reservoir units, wherein each of the plurality of reservoir units comprises: a mask unit configured to receive the input signal and perform a first process on the input signal using a first weight to obtain a first processing result, and a rotation neuron unit configured to receive the first processing result from the mask unit, and perform a second processing on the first processing result to perform dimension increase, nonlinear operation and recursive connection to obtain a second processing result; and an output layer unit, configured to multiply the plurality of second processing results of the plurality of storage pool units by the second weight matrix to obtain a third processing result, and output the third processing result, Wherein, the rotating neuron unit comprises: N pre-neuron rotor circuits, wherein each of the N pre-neuron rotor circuits comprises a first selection signal terminal, a first input terminal, and N first output terminals sequenced from 1st to Nth; N post-neuron rotor circuits, wherein the N post-neuron rotors correspond one-to-one to the N pre-neuron rotor circuits, and each post-neuron rotor circuit comprises a second selection signal terminal, N second input terminals ranked from 1st to Nth, and a second output terminal; N neuron circuits, wherein a first end of an m-th neuron circuit among the N neuron circuits is connected to an m-th first output end of each of the N front neuron rotor circuits, and a second end of the m-th neuron circuit among the N neuron circuits is connected to an m-th second input end of each of the N rear neuron rotor circuits; a timing control circuit connected to a first selection signal terminal of each of the N front neuron rotor circuits and to a second selection signal terminal of each of the N rear neuron rotor circuits, and configured to generate a selection signal, thereby applying the selection signal to the N front neuron rotor circuits and the N rear neuron rotor circuits simultaneously, The N pre-neuron rotor circuits are configured as a whole such that, according to the selection signal, each of the N pre-neuron rotor circuits selects one of its N first output terminals, and the sequence numbers of the first output terminals selected by the N pre-neuron rotor circuits are different from each other. The N post-neuron rotor circuits are configured as a whole such that, according to the selection signal, each of the N post-neuron rotor circuits selects one of its N second input terminals, and the sequence numbers of the second input terminals selected by the N post-neuron rotor circuits are different from each other. The first output terminal of the front neuron rotor circuit and the second input terminal of the rear neuron rotor circuit connected to the same neuron circuit are simultaneously selected. N is a positive integer greater than 1, m=1,2,…,N.

2. The reserve pool computing device according to claim 1, wherein: The mask unit comprises: An input terminal configured to receive the input signal; an input weight configuration module, configured to receive a control signal regarding the first weight, perform the first processing on the input signal received by the input end and the first weight, and obtain the first processing result; and The output end is configured to output the first processing result of the input weight configuration module.

3. The reserve pool computing device according to claim 2, wherein: The input weight configuration module includes: Inverter; A plurality of switches, wherein each switch of the plurality of switches comprises: The first switch input terminal, The second switch input terminal, Switching outputs, and Switch control terminal, Among them, the first switch input terminal is connected to the input terminal of the mask unit to receive the input signal, the inverter is connected to the input terminal of the mask unit to receive the input signal and inverts the input signal to obtain an inverted input signal, the second switch input terminal is connected to the inverter to receive the inverted input signal, the switch output terminal is connected to the output terminal of the mask unit, and the switch control terminal is configured to receive the control signal and output the signal received by the first switch input terminal or the second switch input terminal from the switch output terminal.

4. The reserve pool computing device according to claim 3, wherein: The control signal is a random control signal.

5. The reserve pool computing device according to claim 1, wherein: The value of the strobe signal varies with time and causes in N operation cycles: Each of the N front neuron rotor circuits selects its own first to Nth first output terminals to be selected in sequence; Each of the N post-neuron rotor circuits enables its 1st to Nth second input terminals to be enabled in sequence.

6. The reserve pool computing device according to claim 1, wherein: Each of the plurality of neuron circuits comprises: Non-linear activation circuits; an integrator circuit; and Attenuation circuit, Among them, the first end of the nonlinear activation circuit is connected to the input end of the neuron circuit, the second end of the nonlinear activation circuit is connected to the output end of the neuron circuit, the first end of the integration circuit is connected to the output end of the neuron circuit, the first end of the attenuation circuit is connected to the input end of the neuron circuit, and the second end of the integration circuit is connected to the second end of the attenuation circuit.

7. The reserve pool computing device according to claim 6, wherein: The nonlinear activation circuit includes a diode, the cathode of the diode is connected to the output end of the neuron circuit, and the anode of the diode is connected to the reference voltage end. The integration circuit comprises an integration resistor and a capacitor, wherein a first end of the integration resistor is connected to an input end of the neuron circuit, a second end of the integration resistor is connected to a first end of the capacitor and an output end of the neuron circuit, and a second end of the capacitor is connected to the reference voltage end. The attenuation circuit includes an attenuation resistor, a first end of the attenuation resistor is connected to the output end of the neuron circuit, and a second end of the attenuation resistor is connected to the reference voltage end.

8. The reserve pool computing device according to claim 1, wherein: Each of the front neuron rotor circuits is a first multiplexer, Each post-neuron rotor circuit is a second multiplexer.

9. The reserve pool computing device according to claim 1, wherein: The timing control circuit comprises: The counter is configured to generate the strobe signal under the control of a clock signal.

10. The reserve pool computing device according to any one of claims 1 to 9, wherein: The output layer unit comprises: The multiplication and addition module is configured to multiply a plurality of the second processing results by the second weight matrix to obtain the third processing result.

11. The reserve pool computing device according to claim 10, wherein: The multiplication and addition module includes a memristor array, The memristor array includes a plurality of memristors arranged in an array, and a plurality of conductance values ​​of the plurality of memristors arranged in an array correspond to values ​​of a plurality of elements of the second weight matrix.

12. The reserve pool computing device according to claim 11, wherein: The output layer unit also includes: The parameter setting module is configured to set the conductance value of the memristor array.

13. A data processing method, used in the reserve pool computing device according to any one of claims 1 to 12, comprising: Using the reserve pool computing device to perform inference computing operations; or The reserve pool computing device is used to perform training computing operations.

14. The data processing method according to claim 13, wherein: The inference calculation operation includes: receiving an input signal for the inference computing operation through the signal input circuit; The input signal and the first weight are subjected to the first processing by the reserve pool circuit to obtain the first processing result, and the first processing result is subjected to the second processing to perform dimension increase, nonlinear operation and recursive connection to obtain multiple second processing results; The output layer unit multiplies the plurality of second processing results by the second weight matrix to obtain the third processing result, and outputs the third processing result.

15. The data processing method according to claim 13, wherein: The training calculation operation includes: receiving, through the signal input circuit, an input signal for the training calculation operation and a label value for the input signal; The input signal and the first weight are subjected to the first processing by the reserve pool circuit to obtain the first processing result, and the first processing result is subjected to the second processing to perform dimension increase, nonlinear operation and recursive connection to obtain multiple second processing results; Multiplying the plurality of second processing results by the second weight matrix through the output layer unit to obtain the third processing result; Calculating an error of the second weight matrix according to the plurality of third processing results and the label value of the input signal to update the second weight matrix; The updated second weight matrix is ​​written into the output layer unit.

Citation Information

Patent Citations

  • Computer system incorporating an adaptive model, and methods for training the adaptive model

    CN108369662A

  • Reservoir element and neuromorphic element

    CN110895952A

  • Reserve pool computing system based on dynamic memristor

    CN112488308A

  • Spin element and reservoir element

    CN113632239A

  • Convolutional neural network on-chip learning system based on non-volatile memory

    US20200342301A1