Search

CN-122003820-A - Communication method and electronic equipment

CN122003820ACN 122003820 ACN122003820 ACN 122003820ACN-122003820-A

Abstract

The embodiment of the application provides a communication method and electronic equipment. The method is applied to first equipment and comprises the steps of solving input disturbance in model calibration according to model boundary parameters and confidence requirements output in the model calibration, determining transmitting power and modulation orders according to the input disturbance in the model calibration, modulating first data features based on the transmitting power and the modulation orders to generate modulation symbols, wherein the first data features are used for reasoning by second equipment, and transmitting the modulation symbols. According to the method provided by the embodiment of the application, model calibration can be realized to improve the reasoning efficiency while ensuring the reasoning accuracy.

Inventors

  • HOU QIUSHUO
  • YU GUANDING
  • LI MENGYUAN
  • WANG JIAN

Assignees

  • 华为技术有限公司

Dates

Publication Date
20260508
Application Date
20231010

Claims (20)

  1. A method of communication, the method being applied to a first device, the method comprising: Solving input disturbance during model calibration according to model boundary parameters and confidence coefficient requirements output during model calibration; determining the transmitting power and the modulation order according to the input disturbance during the calibration of the model; Modulating a first data characteristic based on the transmission power and the modulation order, and generating a modulation symbol, wherein the first data characteristic is used for reasoning by a second device; and transmitting the modulation symbol.
  2. The method according to claim 1, wherein the method further comprises: Acquiring reasoning data; And extracting and compressing the features of the reasoning data to generate the first data features.
  3. The method according to claim 1 or 2, characterized in that the method further comprises: Transmitting a second data characteristic to the second device, the second data characteristic being used by the second device to calculate the model boundary parameters; And receiving the model boundary parameters sent by the second equipment and the confidence requirements output during the model calibration.
  4. A method according to any of claims 1-3, characterized in that said determining the transmit power and modulation order from the input disturbances when said model is calibrated comprises: Calculating the transmission bit error rate of the model calibration according to the quantized bit number of the features and the input disturbance of the model calibration; And determining the transmitting power and the modulation order corresponding to the transmission error bit rate in the model calibration according to the transmission error bit rate in the model calibration.
  5. The method according to claim 4, wherein the method further comprises: Receiving channel gain of the first device, which is sent by the second device; and determining the mapping relation between the transmission bit error rate, the transmitting power and the modulation order according to the channel gain of the first equipment.
  6. The method of any of claims 1-5, wherein the solving for the input perturbation at model calibration based on the model boundary parameters and the confidence requirement of the output at model calibration, wherein the solving for the input perturbation at model calibration satisfies: The model output confidence function meets the model calibration requirement; the difference between the input of the neural network of the second device and the output of the neural network of the first device is less than or equal to the variable to be optimized.
  7. A method of communication, the method being applied to a second device, the method comprising: Receiving a modulation symbol sent by a first device, wherein the modulation symbol is a modulation symbol sent by the first device according to the method of any one of claims 1-6; generating recovery data of the first data characteristic according to the modulation symbol; and carrying out reasoning according to the recovery data of the first data characteristic to obtain a reasoning result.
  8. The method of claim 7, wherein prior to receiving the modulation symbols transmitted by the first device, the method further comprises: receiving a second data characteristic sent by the first device; calculating model boundary parameters according to the second data characteristics; And sending the model boundary parameters and the confidence requirements output during model calibration to the first device.
  9. The method according to claim 7 or 8, characterized in that the method further comprises: Estimating a channel gain of the first device, wherein the channel gain of the first device is used for determining a mapping relation between a transmission bit error rate and a transmitting power and a modulation order; and transmitting the channel gain of the first device to the first device.
  10. The method according to any one of claims 7-9, further comprising: calculating a current expected calibration error according to the reasoning result; And restarting the reasoning process when the current expected calibration error does not meet the expected requirement.
  11. A method of communication, the method being applied to a second device, the method comprising: Solving input disturbance during model calibration according to model boundary parameters and confidence coefficient requirements output during model calibration; determining the transmitting power and the modulation order according to the input disturbance during the calibration of the model; and transmitting the transmitting power and the modulation order.
  12. The method of claim 11, wherein the method further comprises: receiving a second data characteristic sent by the first device; And calculating the model boundary parameters according to the second data characteristics.
  13. The method according to claim 11 or 12, wherein said determining the transmit power and modulation order from the input perturbations in the model calibration comprises: Calculating the transmission bit error rate of the model calibration according to the quantized bit number of the features and the input disturbance of the model calibration; And determining the transmitting power and the modulation order corresponding to the transmission error bit rate in the model calibration according to the transmission error bit rate in the model calibration.
  14. The method of claim 13, wherein the method further comprises: estimating a channel gain of the first device; And determining the mapping relation between the transmission bit error rate and the transmitting power and modulation order according to the channel gain of the first equipment.
  15. The method of any of claims 11-14, wherein the solving for the input perturbation at model calibration based on the model boundary parameters and the confidence requirement of the output at model calibration, wherein the solving for the input perturbation at model calibration satisfies: The model output confidence function meets the model calibration requirement; the difference between the input of the neural network of the second device and the output of the neural network of the first device is less than or equal to the variable to be optimized.
  16. The method according to any one of claims 11-15, further comprising: receiving a modulation symbol sent by a first device, wherein the modulation symbol is a modulation symbol generated by the first device for modulating a first data characteristic based on the transmitting power and the modulation order; generating recovery data of the first data feature according to the modulation symbol; And reasoning according to the recovery data of the first data characteristic to obtain a reasoning result.
  17. The method of claim 16, wherein the method further comprises: calculating a current expected calibration error according to the reasoning result; And restarting the reasoning process when the current expected calibration error does not meet the expected requirement.
  18. A method of communication, the method being applied to a first device, the method comprising: Receiving a transmission power and a modulation order transmitted by a second device, wherein the transmission power and the modulation order are generated according to the method of any one of claims 11-17; Modulating a first data characteristic based on the transmission power and the modulation order, and generating a modulation symbol, wherein the first data characteristic is used for reasoning by a second device; and transmitting the modulation symbol.
  19. The method of claim 18, wherein the method further comprises: Acquiring reasoning data; And extracting and compressing the features of the reasoning data to generate the first data features.
  20. The method according to claim 18 or 19, characterized in that the method further comprises: and sending a second data characteristic to the second device, wherein the second data characteristic is used for calculating the model boundary parameters by the second device.

Description

Communication method and electronic equipment Technical Field The present application relates to the field of computer technologies, and in particular, to a communication method and an electronic device. Background The neural network (Neural network, NN) is a specific model in machine learning technology. According to the general approximation theorem, the neural network can theoretically approximate any continuous function, so that the neural network has the capability of learning any mapping. The neural network model is widely applied in the field of wireless communication by virtue of the advantage that the complexity of the traditional mathematical optimization algorithm can be reduced. While the conventional communication system needs to design a communication module by means of abundant expert knowledge, the deep learning communication system based on the neural network can automatically discover an implicit mode structure from a large number of data sets, establish a mapping relation between data and obtain performance superior to that of the conventional modeling method. However, for some application scenarios with high safety requirements, such as unmanned or industrial manufacturing, the NN model output needs to have high accuracy and confidence level indicating the current result. Thus, the output of the NN model requires confidence in the accuracy of the model, i.e., model calibration. Generally, NN models based on conventional training with minimized loss function tend to be uncalibrated because NN models have better learning ability, especially if there is less training data, they may be overfitted to produce an overly confident output, i.e., an output with a confidence greater than the model accuracy. Therefore, a method is needed to reduce the confidence of the NN model output while ensuring the NN model accuracy. Disclosure of Invention Aiming at the problem of how to reduce the confidence of NN model output on the premise of ensuring NN model accuracy, the application provides a communication method and electronic equipment, and also provides a computer readable storage medium. The embodiment of the application adopts the following technical scheme: In a first aspect, the present application provides a communication method, the method being applied to a first device, the method comprising: Acquiring a first data characteristic, wherein the first data characteristic is used for reasoning by a second device; Solving input disturbance during model calibration according to model boundary parameters and confidence coefficient requirements output during model calibration; Determining the transmitting power and the modulation order according to the input disturbance during model calibration; Modulating a first data characteristic based on the transmission power and the modulation order, and generating a modulation symbol, wherein the first data characteristic is used for reasoning by a second device; the modulation symbols are transmitted. According to the method of the first aspect, the inference performance (comprising the inference accuracy and the model calibration) and the channel state are jointly considered, and the transmission power and the modulation order used for the terminal equipment to transmit the inference feature (the inference feature is used for the base station equipment to infer) are adaptively determined, so that the model calibration is realized to improve the inference efficiency while the inference accuracy is ensured. According to the method of the first aspect, the self-adaptive transmission mechanism of the first equipment side for the neural network model calibration can improve the calibration capability of the model on the premise of ensuring the model reasoning accuracy. According to the method of the first aspect, the confidence level of the model output is calibrated by utilizing the input disturbance caused by the channel noise, so that the model calibration capability can be ensured while the model reasoning accuracy is met, and the problem that the model calibration cannot be performed due to insufficient computing resources is effectively solved. In one implementation manner of the first aspect, acquiring the data feature includes: Acquiring reasoning data; And extracting and compressing the features of the reasoning data to generate a first data feature. In an implementation manner of the first aspect, before acquiring the data feature, the method further includes: Transmitting second data characteristics to the second device, the second data characteristics being used by the second device to calculate model boundary parameters; And receiving the model boundary parameters sent by the second equipment and the confidence requirements output during model calibration. In one implementation manner of the first aspect, determining the transmission power and the modulation order according to the input disturbance in the model calibration includes: Calculating the