CN-121981179-A - Circuit for realizing cyclic neural network and construction method thereof

CN121981179ACN 121981179 ACN121981179 ACN 121981179ACN-121981179-A

Abstract

The invention relates to the technical field of electronic circuits and signal processing, in particular to a circuit for realizing a cyclic neural network and a construction method thereof. The circuit for realizing the cyclic neural network comprises a plurality of processing units and a plurality of resistors, wherein each processing unit comprises a signal input end and a signal output end, each two processing units are connected through a resistor, one processing unit is used as an input processing unit, the signal input end of the input processing unit is used for receiving an input signal, at least one processing unit different from the input processing unit is used as an output processing unit, and the signal output end of the output processing unit is used for transmitting an output signal. Compared with the traditional neural network computing architecture, the information processing power consumption and delay are reduced, and the efficiency of processing time sequence information is improved.

Inventors

Jiang Tianxi
ZHOU ZIXIN
YANG MENGLONG
ZHANG SHIWU
HE QINGBO

Assignees

中国科学技术大学

Dates

Publication Date: 20260505
Application Date: 20251225

Claims (10)

1. A circuit for implementing a recurrent neural network, comprising a plurality of processing units and a plurality of resistors, each of said processing units comprising a signal input and a signal output; Each two processing units are connected with each other through one resistor; At least one processing unit different from the input processing unit is used as an output processing unit, and the signal output end of the output processing unit is used for transmitting an output signal.
2. A circuit for implementing a recurrent neural network as defined in claim 1, further comprising: The signal input end of each processing unit is connected with at least one resistor, and the resistor is connected with the signal input ends of other processing units.
3. The circuit for implementing a recurrent neural network of claim 1, wherein the processing unit comprises a first resistor, a first frequency-dependent negative resistance element, and a second frequency-dependent negative resistance element; the first end of the first resistor is connected with the signal input end, and the second end of the first resistor is respectively connected with the signal output end and the input end of the second frequency-dependent negative resistance element; The input end of the first frequency-variable negative resistance element is connected with the signal input end; the grounding end of the first frequency-variable negative resistance element is used for grounding; the grounding end of the second frequency-dependent negative resistance element is used for grounding.
4. A circuit for implementing a recurrent neural network as defined in claim 3, wherein the first frequency-dependent negative resistance element comprises a first capacitor, a second resistor, a third resistor, a fourth resistor, a fifth capacitor, a first amplifier, and a second amplifier, and the first capacitor, the second resistor, the third resistor, the fourth resistor, and the fifth capacitor are sequentially connected in series; The positive input end of the first amplifier is connected to one end of the first capacitor which is not connected with the second resistor, the negative input end of the first amplifier is connected to a node between the second resistor and the third resistor, and the output end of the first amplifier is connected to a node between the third resistor and the fourth resistor; The positive input end of the second amplifier is connected to the node between the fourth resistor and the fifth capacitor, the negative input end of the second amplifier is connected to the node between the second resistor and the third resistor, and the output end of the second amplifier is connected to the node between the first capacitor and the second resistor.
5. The circuit for implementing a recurrent neural network of claim 4, wherein the first amplifier and the second amplifier are operational amplifiers.
6. A construction method of a circuit for realizing a cyclic neural network comprises the following steps: Constructing a mechanical resonance network model; training the mechanical resonance network model based on training data of a target task, wherein the mechanical resonance network model comprises a plurality of mechanical resonance units and mechanical coupling elements connected with the mechanical resonance units; Mapping the trained mechanical resonance unit into a processing unit, mapping the trained mechanical coupling element into a resistor, and constructing the circuit for realizing the cyclic neural network according to any one of claims 1 to 5.
7. The method for constructing a circuit for implementing a recurrent neural network as defined in claim 6, wherein training the mechanical resonance network model based on training data of a target task includes optimizing intrinsic parameters of the mechanical resonance unit and coupling parameters of the mechanical coupling element based on training data of the target task, so that the trained mechanical resonance network model can generate a dynamic response corresponding to the target task according to an input mechanical excitation signal.
8. The method for constructing a circuit for implementing a recurrent neural network as defined in claim 6, further comprising: Selecting one processing unit as an input processing unit, wherein a signal input end of the input processing unit is used for receiving an input signal, and the input signal comprises a current signal; At least one processing unit different from the input processing unit is selected as an output processing unit, and a signal output end of the output processing unit is used for transmitting an output signal.
9. The method for constructing a circuit for implementing a recurrent neural network as defined in claim 6, further comprising: And mapping each mechanical parameter in the trained mechanical resonance network model into a corresponding circuit parameter.
10. The method for constructing a circuit for implementing a recurrent neural network according to claim 9, wherein mapping each mechanical parameter in the trained mechanical resonance network model to a corresponding circuit parameter comprises: the force is mapped to the current in the circuit, the displacement is mapped to the voltage in the circuit, the viscous damping is mapped to the capacitance value in the circuit, the mass is mapped to the frequency-dependent negative resistance, and the compliance is mapped to the resistance value.

Description

Circuit for realizing cyclic neural network and construction method thereof Technical Field The invention relates to the technical field of electronic circuits and signal processing, in particular to a circuit for realizing a cyclic neural network and a construction method thereof. Background With the rapid development of artificial intelligence technology, the increasing application demands of the artificial intelligence technology are increasingly in conflict with the current computing system in terms of energy consumption and efficiency, which promotes the development of a novel computing paradigm of a physical neural network. Physical neural networks, which use the inherent properties of physical systems for information processing, are considered to be a very potential specialized computing architecture. Among them, the physical neural network technology based on the wave system exhibits excellent timing information processing potential. In a wave physical system, a cyclic relationship exists in a dynamic equation of a physical body, and physical waves naturally have memory and calculation capacity for time sequence information when transmitted in space. Based on the theory, in the related technology, a mechanical resonance circulating neural network based on a local resonance elastic metamaterial (namely a mechanical resonance network model corresponding to the invention) is provided, and the mapping relation between a dynamic equation of a spring vibrator system and the circulating neural network is disclosed in the field of low-frequency elastic waves. The related art proposes that the local resonance elastic metamaterial is composed of resonance units arranged in a5×5 matrix, the units are connected and coupled by springs with a stiffness coefficient of k c, and four corner units are subjected to fixing treatment. The resonance unit is used as a cell for constructing the metamaterial, the mass of an outer spherical shell is M, the mass of an inner spherical body is M, and the outer spherical shell and the inner spherical body are connected by a spring with the rigidity coefficient of k n. And inputting vibration excitation to a specified position of the trained spring vibrator system, measuring the vibration response of the ball in the output cell, and comparing the energy of the vibration response, so that the classification and discrimination of input signals can be realized. However, the related art researches or the real physical model is too complex, so that a simplified theoretical framework can be constructed, or the time sequence information containing complex frequency spectrum and low frequency components is difficult to process due to the limitation of the volume and the scale of hardware. And the output result is mostly non-electrical signals such as mechanical quantity, and the like, and the output result still needs to depend on additional sensors and analog-digital conversion units to be integrated with other electronic systems, so that the overall complexity of the system is increased. The related physical neural network technology has insufficient capability of processing complex time sequence information due to theoretical simplification and hardware limitation, and the system integration complexity is increased. Disclosure of Invention The invention aims to solve the technical problems of providing a circuit for realizing a cyclic neural network and a construction method thereof, which realize the rapid processing of complex time sequence information, output of electrical signals and easy integration. In order to solve the technical problems, the invention adopts a technical scheme that: a circuit for implementing a recurrent neural network, comprising a plurality of processing units and a plurality of resistors, each of said processing units comprising a signal input and a signal output; Each two processing units are connected with each other through one resistor; At least one processing unit different from the input processing unit is used as an output processing unit, and the signal output end of the output processing unit is used for transmitting an output signal. In order to solve the technical problems, the invention adopts another technical scheme that: a construction method of a circuit for realizing a cyclic neural network comprises the following steps: Constructing a mechanical resonance network model; training the mechanical resonance network model based on training data of a target task; Mapping the parameters of each mechanical element in the trained mechanical resonance network model into parameters of corresponding circuit elements, wherein the mechanical resonance network model comprises a plurality of mechanical resonance units and mechanical coupling elements connected with the mechanical resonance units; Mapping the mechanical resonance unit into a processing unit, and mapping the mechanical elastic coupling element into a resistor; and constructing a circuit fo