Search

CN-122024789-A - Evolution method, system and device of intelligent perception memory

CN122024789ACN 122024789 ACN122024789 ACN 122024789ACN-122024789-A

Abstract

The invention discloses an evolution method, system and device of an intelligent perception memory, the method comprises the steps of receiving a multi-mode physical signal sent by an upper computer, modulating the multi-mode physical signal to obtain a conductance distribution parameter, performing controlled relaxation on the conductance distribution parameter to obtain steady-state current distribution data, performing sparse sampling on the steady-state current distribution data according to a preset threshold to obtain a sparse event stream, uploading the sparse event stream to the upper computer, receiving a target hardware gene parameter fed back by the upper computer according to the sparse event stream, updating a hardware control parameter according to the target hardware gene parameter, and re-executing the conductance distribution parameter obtained by modulating the multi-mode physical signal. The invention improves scene adaptability and calculation adaptability.

Inventors

  • Chen Kuangxiang
  • GAO WEI

Assignees

  • 联和存储科技(江苏)有限公司

Dates

Publication Date
20260512
Application Date
20260122

Claims (10)

  1. 1. An evolution method of an intelligent perception memory is characterized by comprising the following steps: Receiving a multi-mode physical signal sent by an upper computer, and modulating the multi-mode physical signal to obtain a conductivity distribution parameter; Performing controlled relaxation on the conductivity distribution parameters to obtain steady-state current distribution data; Sparse sampling is carried out on the steady-state current distribution data according to a preset threshold value to obtain a sparse event stream; Uploading the sparse event stream to the upper computer, and receiving target hardware gene parameters fed back by the upper computer according to the sparse event stream; Updating the control parameters of the hardware according to the target hardware gene parameters, and re-executing the conductance distribution parameters obtained by modulating the multi-mode physical signals.
  2. 2. The method for evolving an intelligent sensing memory according to claim 1, wherein the hardware comprises a memristor array, a programmable switch array and a parallel median filter circuit, and the method further comprises, before receiving the multi-mode physical signal sent by the upper computer: receiving a configuration instruction of the multi-mode task sent by the upper computer; Analyzing configuration instructions of the multi-mode task to obtain initial hardware gene parameters and initial topology configuration parameters, wherein the initial hardware gene parameters comprise initial global pulse voltage, initial dynamic threshold coefficients, initial relaxation time constants and initial network connection preference weights; configuring control parameters for the memristor array, the programmable switch array and the parallel median filtering circuit according to the initial hardware genetic parameters; And controlling the programmable switch array to reconstruct the topological structure of the memristor array into a target topological structure matched with the multi-mode task based on the initial topological configuration parameters.
  3. 3. The method for evolution of intelligent sensing memory according to claim 2, wherein the conductance distribution parameters obtained by modulating the multi-modal physical signal include: Sending the multi-modal physical signal to the memristor array; Collecting macroscopic conductance variation of the memristor array when receiving the multi-mode physical signal; fitting the macroscopic conductance variation based on a preset central limit theorem to obtain Gaussian distribution parameters, wherein the Gaussian distribution parameters comprise a mean value and a variance; And taking the Gaussian distribution parameter as the conductivity distribution parameter.
  4. 4. The method of claim 3, wherein said subjecting said conductance profile parameters to controlled relaxation to obtain steady state current profile data comprises: configuring a noise intensity matrix according to the conductivity distribution parameters, and acquiring a preset relaxation coefficient and a preset random noise item; Generating a current dynamics equation according to the noise intensity matrix, the preset relaxation coefficient and the preset random noise item; Feeding back a sensing ready signal to the upper computer, and receiving a global pulse excitation instruction issued by the upper computer according to the sensing ready signal; Applying global pulse excitation to the memristor array according to the global pulse excitation instruction, and controlling the memristor array to start controlled relaxation calculation based on the current dynamics equation; Judging whether the calculated duration of the controlled relaxation calculation reaches the initial relaxation time constant or not; And stopping controlled relaxation calculation when the calculation time length reaches the initial relaxation time constant, and collecting steady-state currents of all nodes in the memristor array to obtain the steady-state current distribution data.
  5. 5. The method of claim 4, wherein the memristor array includes a plurality of nodes, and wherein generating the current kinetic equation from the noise strength matrix, the preset relaxation coefficient, and the preset random noise term includes: ; Wherein, for any node in the memristor array, Indicating the current time rate of change of the node, Representing the preset relaxation coefficient, I representing the instantaneous current of the node at time t, Representing the matrix of the intensity of the noise, Representing the element-by-element product operator, And representing the preset random noise item of the node at the time t.
  6. 6. The method of claim 5, wherein after performing controlled relaxation on the conductance profile to obtain steady-state current profile data, further comprising: Controlling the parallel median filter circuit to perform median calculation on the steady-state current distribution data to obtain a current median; And calculating the product of the initial dynamic threshold coefficient and the current median as the preset threshold.
  7. 7. The method for evolving an intelligent sensing memory according to claim 6, wherein sparse sampling the steady-state current distribution data according to a preset threshold to obtain a sparse event stream comprises: traversing all nodes of the memristor array, and judging whether steady-state current of each node is larger than the preset threshold value; when the steady-state current of the node is larger than the preset threshold value, recording the node address and the time stamp of the node; and ordering the node addresses and the time stamps of all the nodes according to the time sequence to obtain the sparse event stream in the node address-time stamp key value pair format.
  8. 8. The method of claim 7, wherein the target hardware genetic parameter comprises a target global pulse voltage, a target dynamic threshold coefficient, a target relaxation time constant, and a target network connection preference weight, and wherein updating the control parameter of the hardware according to the target hardware genetic parameter comprises: Updating the initial global pulse voltage according to the target global pulse voltage; Updating the initial dynamic threshold coefficient according to the target dynamic threshold coefficient; updating the initial relaxation time constant according to the target relaxation time constant; And updating the initial network connection preference weight according to the target network connection preference weight.
  9. 9. An evolution system of an intelligent perception memory, which is characterized by comprising the intelligent perception memory and an upper computer, wherein the intelligent perception memory comprises a controller, a memristor array, a programmable switch array and a parallel median filter circuit, and the controller can execute the evolution method of the intelligent perception memory according to any one of claims 1 to 8.
  10. 10. An evolution device of an intelligent perception memory, comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores instructions for execution by the at least one processor to enable the at least one processor to perform the method of evolution of the smart aware memory of any one of claims 1 to 8.

Description

Evolution method, system and device of intelligent perception memory Technical Field The application relates to the technical field of data storage, in particular to an evolution method, system and device of an intelligent perception memory. Background With the high-speed development of technologies such as edge intelligence and body intelligence, the market has made stringent demands on the data handling overhead of memories, and memories integrated with the memory technology can complete calculation in the memories, thus relieving the data handling overhead. At present, the existing memory integrating calculation generally uses fixed parameters to calculate, self-learning and self-optimization cannot be performed according to application scenes, physical connection and functions of a calculating unit are fixed, and dynamic data flow requirements of multi-mode tasks cannot be adapted. Therefore, the conventional memory integrated with memory is liable to cause problems of poor scene adaptability and poor computation adaptability. Disclosure of Invention The application mainly aims to provide an evolution method, system and device of an intelligent perception memory, and aims to solve the problems of weak scene adaptability and poor calculation adaptability. In order to achieve the above object, the present application provides an evolution method of an intelligent sensing memory, the evolution method of the intelligent sensing memory includes: Receiving a multi-mode physical signal sent by an upper computer, and modulating the multi-mode physical signal to obtain a conductivity distribution parameter; Performing controlled relaxation on the conductivity distribution parameters to obtain steady-state current distribution data; Sparse sampling is carried out on the steady-state current distribution data according to a preset threshold value to obtain a sparse event stream; Uploading the sparse event stream to the upper computer, and receiving target hardware gene parameters fed back by the upper computer according to the sparse event stream; Updating the control parameters of the hardware according to the target hardware gene parameters, and re-executing the conductance distribution parameters obtained by modulating the multi-mode physical signals. In some embodiments, the hardware comprises a memristor array, a programmable switch array and a parallel median filter circuit, and before receiving the multi-mode physical signal sent by the upper computer, the method further comprises: receiving a configuration instruction of the multi-mode task sent by the upper computer; Analyzing configuration instructions of the multi-mode task to obtain initial hardware gene parameters and initial topology configuration parameters, wherein the initial hardware gene parameters comprise initial global pulse voltage, initial dynamic threshold coefficients, initial relaxation time constants and initial network connection preference weights; configuring control parameters for the memristor array, the programmable switch array and the parallel median filtering circuit according to the initial hardware genetic parameters; And controlling the programmable switch array to reconstruct the topological structure of the memristor array into a target topological structure matched with the multi-mode task based on the initial topological configuration parameters. In some embodiments, the conductance distribution parameter obtained by modulating the multi-mode physical signal includes: Sending the multi-modal physical signal to the memristor array; Collecting macroscopic conductance variation of the memristor array when receiving the multi-mode physical signal; fitting the macroscopic conductance variation based on a preset central limit theorem to obtain Gaussian distribution parameters, wherein the Gaussian distribution parameters comprise a mean value and a variance; And taking the Gaussian distribution parameter as the conductivity distribution parameter. In some embodiments, the controlled relaxation of the conductance profile parameter yields steady-state current profile data comprising: configuring a noise intensity matrix according to the conductivity distribution parameters, and acquiring a preset relaxation coefficient and a preset random noise item; Generating a current dynamics equation according to the noise intensity matrix, the preset relaxation coefficient and the preset random noise item; Feeding back a sensing ready signal to the upper computer, and receiving a global pulse excitation instruction issued by the upper computer according to the sensing ready signal; Applying global pulse excitation to the memristor array according to the global pulse excitation instruction, and controlling the memristor array to start controlled relaxation calculation based on the current dynamics equation; Judging whether the calculated duration of the controlled relaxation calculation reaches the initial relaxation time constant or not; A