Distributed training air interface feedback method and related device

Abstract

The invention discloses a distributed training air interface feedback method and a related device. In the method, a first node sends first information to at least two second nodes in a plurality of second nodes, the first information being used for indicating at least one first resource and at least one second resource; the first node receives pilots transmitted by the at least two second nodes on the at least one first resource and data transmitted on the at least one second resource; wherein a corresponding relationship exists between the at least one second resource and the at least one model parameter, and the at least one model parameter is a part of or all of the parameters that at least two second nodes respectively utilize local training data to train a model, and a training result needs to be fed back or a gradient needs to be updated; the data transmitted by the second node on the second resource is obtained based on a product of a training result or an update gradient of the model parameter corresponding to the second resource and the pilot frequency. According to the embodiment of the invention, the feedback overhead can be saved and the accuracy of air interface feedback is improved.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a distributed training air interface feedback method and related devices. Background Art

[0002] In the distributed training scenario of artificial intelligence (AI), the central node sends the model to each distributed node, which then trains the distributed model based on local training samples. The training results or update gradients of the local AI model parameters are then fed back to the central node through the air interface. The central node summarizes the training results or update gradients of the parameters for model updating.

[0003] When the number of distributed nodes participating in the training is large, the overhead of each distributed node reporting the training results of the parameters or updating the gradients is large. Therefore, how to reduce the overhead of the training results or updating the gradients is a problem to be solved. Summary of the invention

[0004] The present application provides a distributed training air interface feedback method and related devices, which can reduce the training results of model parameters or the overhead of updating gradients.

[0005] In the first aspect, the embodiment of the present application provides a distributed training air interface feedback method, which can be applied to a first node, or a chip or chip module in the first node, or to a module or unit that can realize all or part of the functions of the first node, etc., and is described below using the first node as an example. In this method, the first node sends first information to at least two second nodes among a plurality of second nodes, and the first information is used to indicate at least one first resource and at least one second resource; the first node receives a pilot transmitted by at least two second nodes on at least one first resource, and data transmitted on at least one second resource; wherein there is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a part or all of a parameter that needs to feedback a training result or an update gradient when at least two second nodes train the model using local training data; the data transmitted by the second node on the second resource is obtained based on the training result or update gradient of the model parameter corresponding to the second resource, and the product of the pilot.

[0006] In the method, since the first node sends the same first information to at least two second nodes, the at least two second nodes can transmit pilot signals on the same first resource, and transmit the training results of the same model parameters obtained through their respective training or data corresponding to the updated gradients on the same second resource. In this way, the pilot signal received by the first node on the first resource is the weighted sum of the pilot signals sent by each second node, and the data signal received on the second resource is the weighted sum of the data sent by each second node. It can be seen that the method utilizes the natural merging effect of each wireless channel to complete the fusion of the training results or updated gradients of the model parameters that need to be fed back, thereby saving channel resources.

[0007] In addition, the present application is also beneficial for the first node to estimate the merged channel of multiple wireless channels between each second node and the first node based on the received pilot signal, and then to equalize the merged data signal (which may be called the merged data signal) based on the estimated merged channel, thereby helping to combat random channel fading, especially small-scale fading with greater randomness, and thus helping to improve the accuracy of air interface feedback.

[0008] In an optional implementation, the first resource indicated by the first information is one, so that at least two second nodes transmit pilot signals on the same first resource, thereby avoiding the problem of excessive pilot resource overhead caused by a large number of second nodes participating in model training.

[0009] In another optional implementation, the first information indicates a plurality of first resources, so that at least two second nodes transmit pilots on the same plurality of first resources. It can be seen that in this implementation, at least two second nodes transmit multiple pilots using multiple first resources, which can improve the signal-to-noise ratio of the pilots, and thus improve the accuracy of channel estimation, compared with using one first resource to transmit a pilot once. For example, when the channel is frequency-selective or time-selective, pilot reception and channel estimation are performed on multiple first resources respectively, and it is possible to estimate how the channel changes over time or frequency, thereby improving the accuracy of channel estimation.

[0010] In an optional implementation, there is a one-to-one correspondence between at least one second resource and at least one model parameter. The at least one model parameter is a parameter for which a training result or an update gradient needs to be fed back. Optionally, the number of second resources indicated by the first information is greater than or equal to the number of model parameters. If the number of second resources is greater than the number of model parameters, some second resources may not need to transmit the training results or update gradients of the corresponding model parameters.

[0011] In an optional implementation, the method further includes: the first node sends second information to at least two second nodes, the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and at least one second resource. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or updated gradients need to be fed back, and the correspondence between these model parameters and the second resources for transmitting the corresponding training results or updated gradients.

[0012] Optionally, among at least two second nodes, the data transmitted by the second node on the second resource is the product of the training result or update gradient of the model parameter corresponding to the second resource and the pilot and the weight corresponding to the second node. The weight corresponding to the second node is determined based on the quantity and quality of the local training data of the second node, and / or the channel information between the second node and the first node, so as to help control the importance of each second node in the merging of the training results. In an optional implementation, the weight is a scalar. In another optional implementation, the weight is a vector.

[0013] In an optional implementation, the method further includes: the first node estimates a combined channel of multiple wireless channels based on the received pilot signal, and the multiple wireless channels are wireless channels between the first node and at least two second nodes; and determines the training result of the model parameter corresponding to the second resource or the combined result of the updated gradient based on the combined channel of the multiple wireless channels and the data signal received on the second resource. The data signal received on the second resource is a combined data signal composed of the product of the transfer function of each wireless channel and the data transmitted by the corresponding second node on the second resource, which are superimposed.

[0014] In an optional implementation, the first node determines the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients based on the merged channel of multiple wireless channels and the data signal received on the second resource, including: performing equalization processing using the merged channel of multiple wireless channels and the data signal received on the second resource to obtain the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients.

[0015] In an optional implementation, equalization processing is performed using a merged channel of multiple wireless channels and a data signal received on a second resource to obtain a training result of a model parameter corresponding to the second resource or a merged result of an updated gradient, including: dividing an inner product between a pilot signal received on at least one first resource and a data signal received on the second resource by an inner product of at least one pilot signal received on the first resource and itself, to obtain a training result of a model parameter corresponding to the second resource after equalization processing or a merged result of an updated gradient.

[0016] It can be seen that in this implementation, the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients are a weighted average of the training results or updated gradients of the model parameters transmitted by each second node on the second resource. In this way, even if the channel is random and unknown, it only affects the weight of the weighted average, and does not directly affect the amplitude and phase of the merged result. Therefore, the impact of the unknown channel on the merged result can be significantly reduced, thereby improving the accuracy of the air interface feedback.

[0017] In an optional implementation, the first information is used to indicate multiple resource blocks, each resource block includes at least one first resource and at least one second resource. Each resource block includes part or all of the frequency domain subcarriers, and / or part or all of the time domain symbols. The frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel. Optionally, all frequency domain subcarriers refer to all frequency domain resources allocated by the first node for the training results of the model parameters or the updated gradient feedback, such as all frequency domain resources indicated by the first information. Optionally, all time domain symbols refer to all time domain symbols allocated by the first node for the training results of the model parameters or the updated gradient feedback, such as all time domain symbols indicated by the first information. It can be seen that this implementation takes into account the frequency selectivity or time selectivity of the channel, divides the channel resources into multiple resource blocks, and configures each resource block with the first resource, so as to effectively overcome the changes in the channel and thus improve the accuracy of the air interface feedback.

[0018] Optionally, for each resource block, the first node receives pilot signals transmitted by at least two second nodes on at least one first resource in the resource block, and data transmitted on at least one second resource; at least one model parameter is a partial parameter of which the at least two second nodes train the model using local training data and need to feed back the training result or update gradient; for each resource block, the data transmitted by the second node on the second resource in the resource block is obtained by multiplying the pilot signal by the training result or update gradient of the model parameter corresponding to the second resource in the resource block.

[0019] Among them, for each resource block, there is a corresponding relationship between at least one second resource of the resource block and at least one model parameter, that is, there is a corresponding relationship between each second resource in all resource blocks and each model parameter for which feedback of training results or updated gradients is required, such as a one-to-one correspondence. Optionally, the model parameter for which feedback of training results or updated gradients is required may be a model parameter for which feedback is required in one training stage of the model, or a model parameter for which feedback is required in all training stages of the model.

[0020] Optionally, the total number of second resources of all resource blocks indicated by the first information is greater than or equal to the number of model parameters. If the number of second resources is greater than the number of model parameters, some second resources may not need to transmit the training results or update gradients of the corresponding model parameters.

[0021] Optionally, the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and each second resource in each resource block. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or updated gradients need to be fed back, and the correspondence between these model parameters and the second resources in each resource block for transmitting the corresponding training results or updated gradients.

[0022] Optionally, for each resource block, the first node estimates a combined channel of multiple wireless channels based on the received pilot signal, where the multiple wireless channels are wireless channels between the first node and at least two second nodes respectively; and determines the training result of the model parameter corresponding to the second resource or the combined result of the updated gradient based on the combined channel of the multiple wireless channels and the data signal received on the second resource in the resource block. The data signal received on the second resource in each resource block is a combined data signal composed of the product of the transfer function of each wireless channel and the data transmitted by the corresponding second node on the second resource, which are superimposed.

[0023] Optionally, the first node may perform equalization processing on each resource block using a merged channel of multiple wireless channels and a data signal received on the second resource to obtain a training result of a model parameter corresponding to the second resource or a merged result of an updated gradient.

[0024] Optionally, the first node may, for each resource block, divide the inner product between a pilot signal received on at least one first resource and a data signal received on a second resource by the inner product of the pilot signal received on at least one first resource and itself to obtain a training result of parameters corresponding to the second resource after equalization or a merged result of updated gradients.

[0025] In the second aspect, the present application also provides a distributed training air interface feedback method, which can be applied to a second node, or a chip or chip module in the second node, or to a module or unit that can realize all or part of the functions of the second node, etc., and the second node is described below as an example. In this method, the second node receives first information from the first node, and the first information is used to indicate at least one first resource and at least one second resource; wherein there is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a part or all of the parameters that need to feedback the training result or update the gradient when the second node trains the model using local training data; the second node determines the data transmitted on the second resource based on the training result of the model parameter corresponding to the second resource or the product between the update gradient and the pilot; and transmits the pilot on at least one first resource, and transmits the determined data on the second resource.

[0026] In the method, the second node transmits a pilot signal on the first resource, and transmits the training results of the corresponding model parameters or the data of the updated gradient on the second resource, so that the first node can obtain the wireless channel between the second node and the data transmitted by the second node on each second resource, thereby facilitating the first node to utilize the natural merging effect of each wireless channel to complete the fusion of the training results of the model parameters or the updated gradients that need to be fed back by each second node, so as to save the channel resources required for the feedback of the training results or the updated gradients.

[0027] Optionally, the data transmitted by the second node on the second resource is the product of the training result or update gradient of the model parameter corresponding to the second resource, the pilot signal, and the weight corresponding to the second node. The weight corresponding to the second node is determined based on the quantity and quality of the local training data of the second node, and / or the channel information between the second node and the first node. In an optional implementation, the weight is a scalar. In another optional implementation, the weight is a vector.

[0028] In an optional implementation, there is a one-to-one correspondence between at least one second resource and at least one model parameter.

[0029] In an optional implementation, the method further includes: the second node receives second information from the first node, the second information is used to indicate at least one model parameter, and the corresponding relationship between at least one model parameter and at least one second resource. This implementation enables the second node to know which model parameters need to feedback training results or update gradients, and transmit the second resources corresponding to each training result or update gradient.

[0030] In an optional implementation, the first information is used to indicate multiple resource blocks, each resource block including at least one first resource and at least one second resource; the frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

[0031] Wherein, for each resource block, there is a corresponding relationship between at least one second resource of the resource block and at least one model parameter, that is, there is a corresponding relationship between each second resource in all resource blocks and each model parameter for which a training result or an update gradient needs to be fed back, such as a one-to-one correspondence. For each resource block, the second node determines the data to be transmitted on the second resource based on the product between the training result or the update gradient of the model parameter corresponding to the second resource and the pilot, and transmits a pilot on at least one first resource, and transmits the determined data on the second resource.

[0032] Optionally, the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and each second resource in each resource block. This implementation enables the second node to know which model parameters need to feedback training results or update gradients, and transmit the second resources of each resource block corresponding to each training result or update gradient.

[0033] Optionally, the optional implementation methods of this aspect can also refer to the relevant content described in the first aspect and will not be described in detail here.

[0034] In the third aspect, the embodiment of the present application also provides a distributed training air interface feedback method, which is explained from the perspective of the interaction between the first node and at least two second nodes among multiple second nodes. In this method, the first node sends first information to at least two second nodes, and the first information is used to indicate at least one first resource and at least one second resource; at least two second nodes receive the first information, and transmit pilot signals on at least one first resource, and transmit determined data on at least one second resource respectively; accordingly, the first node receives the pilot signals transmitted by at least two second nodes on at least one first resource, and the data transmitted on at least one second resource.

[0035] Among them, there is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a parameter in which at least two second nodes train the model using local training data, and feedback of training results or update gradients is required; the data transmitted by the second node on the second resource is obtained based on the product of the training result or update gradient of the model parameter corresponding to the second resource and the pilot.

[0036] In the method, since the first node sends the same first information to at least two second nodes, the at least two second nodes can transmit pilots on the same pilot resources, and transmit the training results of the same model parameters obtained by their respective training or data corresponding to the update gradients on the same second resources, so that the merged channel of multiple wireless channels between each second node and the first node and the merged data signal corresponding to the data transmitted by each second node on each second resource can be obtained. In this way, the natural merging effect of each wireless channel is utilized to complete the fusion of the training results or update gradients of the model parameters that need to be fed back by each second node, thereby saving channel resources.

[0037] In addition, the present application is also conducive to estimating the combined channel based on the pilot, which is used to equalize the combined data signal, thereby helping to combat random channel fading, especially small-scale fading with greater randomness, and further helping to improve the accuracy of air interface feedback.

[0038] In an optional implementation, the first information indicates a first resource and at least one second resource, so that at least two second nodes transmit pilot signals on the same first resource, thereby avoiding the problem of excessive pilot resource overhead caused by a large number of second nodes participating in model training.

[0039] In another optional implementation, the first information indicates multiple first resources and at least one second resource, so that at least two second nodes transmit pilot signals on the same multiple first resources.

[0040] In an optional implementation, there is a one-to-one correspondence between at least one second resource and at least one model parameter. The at least one model parameter is a parameter for which a training result or an update gradient needs to be fed back. Optionally, the number of second resources indicated by the first information is greater than or equal to the number of model parameters. If the number of second resources is greater than the number of model parameters, some second resources may not need to transmit the training results or update gradients of the corresponding model parameters.

[0041] In an optional implementation, the method further includes: the first node sends second information to at least two second nodes, and correspondingly, at least two second nodes receive the second information, wherein the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and at least one second resource. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or updated gradients need to be fed back, and the correspondence between these model parameters and the second resources for transmitting the corresponding training results or updated gradients.

[0042] In an optional embodiment, the method also includes: the first node estimates a merged channel of multiple wireless channels based on the received pilot signal, and the multiple wireless channels are respectively wireless channels between the first node and at least two second nodes; based on the merged channel of the multiple wireless channels and the data signal received on the second resource, determining the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients.

[0043] In an optional implementation, the first information is used to indicate a plurality of resource blocks, each resource block including at least one first resource and at least one second resource. The frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel. This implementation takes into account the frequency selectivity or time selectivity of the channel, divides the channel resources into a plurality of resource blocks, configures each resource block with a first resource, and transmits the pilot and data according to at least one of the above implementations, thereby being able to effectively overcome channel changes.

[0044] Optionally, for each resource block, at least two second nodes transmit a pilot on at least one first resource, and transmit determined data on the second resource; accordingly, the first node receives a pilot transmitted by at least two second nodes on at least one first resource in the resource block, and data transmitted on at least one second resource. Among them, the model parameters corresponding to the transmission on the second resource in each resource block are part of the parameters that need to feedback the training results or update gradients when at least two second nodes train the model using local training data. For each resource block, the data transmitted by the second node on the second resource in the resource block is obtained based on the product of the training results or update gradients of the model parameters corresponding to the second resource in the resource block and the pilot.

[0045] Among them, for each resource block, there is a corresponding relationship between at least one second resource of the resource block and at least one model parameter, that is, there is a corresponding relationship between each second resource in all resource blocks and each model parameter for which feedback of training results or updated gradients is required, such as a one-to-one correspondence. Optionally, the model parameter for which feedback of training results or updated gradients is required may be a model parameter for which feedback is required in one training stage of the model, or a model parameter for which feedback is required in all training stages of the model.

[0046] Optionally, the total number of second resources of all resource blocks indicated by the first information is greater than or equal to the number of model parameters. If the number of second resources is greater than the number of model parameters, some second resources may not need to transmit the training results or update gradients of the corresponding model parameters.

[0047] Optionally, the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and each second resource in each resource block. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or updated gradients need to be fed back, and the correspondence between these model parameters and the second resources in each resource block for transmitting the corresponding training results or updated gradients.

[0048] Optionally, the optional implementation methods of this aspect may also refer to the relevant explanations of the first and second aspects above, which will not be described in detail here. For example, the implementation method of the first node obtaining the training results of the model parameters corresponding to each second resource or the merged results of the updated gradients may refer to the relevant content of the first aspect, which will not be described in detail here; for another example, for each resource block, the first node determines the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients. The implementation method can refer to the relevant content of the first aspect, which will not be described in detail here. For another example, the optional implementation method executed by the second node in this aspect may also refer to the relevant content described in the second aspect, which will not be described in detail here.

[0049] In the fourth aspect, the embodiment of the present application also provides a distributed training air interface feedback method, which can be applied to a first node, or a chip or chip module in the first node, or to a module or unit that can realize all or part of the functions of the first node, etc., and the first node is described below as an example. In this method, the first node sends first information to at least two second nodes among multiple second nodes, and the first information is used to indicate multiple first resources and at least one second resource; wherein, there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter. At least one model parameter is a part or all of the parameters that the second node needs to feedback the training results or update gradients for training the model using local training data; the first node receives the pilot of the transmission layer corresponding to each first resource transmitted by at least two second nodes, and the data transmitted by each transmission layer on at least one second resource; wherein, the data transmitted by at least two second nodes on the transmission layer of the second resource is obtained based on the training results or update gradients of the model parameters corresponding to the second resource and the transmission layer, and the product between the pilot corresponding to the transmission layer.

[0050] In this method, there is a corresponding relationship between the combination of the transmission layer of at least one second resource and the second node and each model parameter. In this way, through multi-layer transmission, at least two second nodes can feedback the training results or update gradients of multiple model parameters on the same second resource, thereby further improving the utilization rate of time-frequency resources.

[0051] In addition, since the first node sends the same first information to at least two second nodes, the at least two second nodes can transmit the pilot on the transmission layer corresponding to the same pilot resource, and transmit the training results of the same model parameters obtained by their respective training or the data corresponding to the update gradient on the same second resource and transmission layer, so that the merged channel of multiple wireless channels between each second node and the first node and the merged data signal corresponding to the data transmitted by each second node on each second resource can be obtained. In this way, the natural merging effect of each wireless channel is utilized to complete the fusion of the training results or update gradients of the model parameters that need to be fed back by each second node, thereby saving channel resources.

[0052] In addition, the present application is also conducive to estimating the combined channel based on the pilot, which is used to equalize the combined data signal, thereby helping to combat random channel fading, especially small-scale fading with greater randomness, and further helping to improve the accuracy of air interface feedback.

[0053] In an optional implementation, there is a one-to-one correspondence between a combination of at least one second resource and a transmission layer of the second node and at least one model parameter.

[0054] In an optional embodiment, the method also includes: the first node sends second information to at least two second nodes, the second information is used to indicate at least one model parameter, and the correspondence between a combination of at least one second resource and a transmission layer of the second node, and at least one model parameter.

[0055] Optionally, the correspondence between the multiple first resources and the transmission layer of the second node includes at least one of a many-to-one or a one-to-one correspondence. The one-to-one correspondence can reduce the overhead of the pilot resource, and the many-to-one correspondence can improve the accuracy of the channel estimation, thereby improving the accuracy of the air interface feedback. Optionally, the first node can send third information to at least two second nodes, and the third information is used to indicate the correspondence between the multiple first resources and the transmission layer.

[0056] In an optional embodiment, the method also includes: the first node estimates a merged channel of multiple wireless channels based on a pilot received on each first resource, where the multiple wireless channels are wireless channels of corresponding transmission layers between the first node and at least two second nodes; and determines the training results of the parameters corresponding to each transmission layer of the second resource or the merged results of the updated gradients based on the merged channel of the multiple wireless channels corresponding to each transmission layer and the data signal received on the second resource.

[0057] The data signal received on the second resource is a combined data signal formed by superimposing the product of the transfer function of the wireless channel of at least two second nodes and each transmission layer and the data signal transmitted by at least two second nodes on the transmission layer on the second resource.

[0058] In an optional implementation, the first node determines the training results of the model parameters corresponding to each transmission layer of the second resource or the merged results of the updated gradients based on the merged channel of the multiple wireless channels corresponding to each transmission layer and the data signal received on the second resource, including: the first node uses the merged channel of the multiple wireless channels corresponding to each transmission layer and the data signal received on the second resource to perform multi-layer equalization processing to obtain the training results of the parameters corresponding to each transmission layer of the second resource or the merged results of the updated gradients.

[0059] In an optional implementation, the first information is used to indicate multiple resource blocks, each resource block including at least one of the above-mentioned first resources and at least one of the above-mentioned second resources; the frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

[0060] Optionally, the total number of second resources in the multiple resource blocks multiplied by the number of transmission layers of the second node is not less than the number of model parameters to be fed back. Optionally, the number of second resources used to transmit training results or update gradients of one or more model parameters is not more than the number of transmission layers of the second node.

[0061] Optionally, for each transport layer, the optional implementation method executed by the first node can also refer to the relevant explanation of the first aspect above, which will not be described in detail here.

[0062] In the fifth aspect, the embodiment of the present application also provides a distributed training air interface feedback method, which can be applied to a second node, or a chip or chip module in the second node, or to a module or unit that can realize all or part of the functions of the second node, etc., and the second node is described below as an example. In this method, the second node receives first information from the first node, and the first information is used to indicate multiple first resources and at least one second resource; wherein, there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is that the second node uses local training data to train the model, and part or all of the parameters of the training results or update gradients need to be fed back; the second node determines the data transmitted by the transmission layer on the second resource based on the training results or update gradients of the model parameters corresponding to the second resource and the transmission layer, and the product between the pilot corresponding to the transmission layer; the second node transmits the pilot of the corresponding transmission layer on the first resource, and transmits the determined data on the transmission layer on the second resource.

[0063] In this method, there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and each model parameter. In this way, through multi-layer transmission, the second node can feedback the training results or update gradients of multiple model parameters on the same second resource, thereby further improving the utilization rate of time-frequency resources.

[0064] In addition, the second node can transmit a pilot signal in the transmission layer corresponding to the pilot resource, as well as the training results or update gradients of the model parameters corresponding to the second resources and the transmission layer, so that the first node can obtain the merged channel of multiple wireless channels between each second node participating in the model training and the first node, and the merged data signal corresponding to the data transmitted by each second node on each second resource. This is beneficial to utilizing the natural merging effect of each wireless channel to complete the fusion of the training results or update gradients of the model parameters that need to be fed back by each second node, thereby saving channel resources.

[0065] In addition, the present application is also beneficial for the first node to estimate the combined channel based on the pilot, which is used to equalize the combined data signal, thereby helping to combat random channel fading, especially small-scale fading with greater randomness, and further helping to improve the accuracy of air interface feedback.

[0066] In an optional implementation, there is a one-to-one correspondence between a combination of at least one second resource and a transmission layer of the second node and at least one model parameter.

[0067] In an optional embodiment, the method also includes: the second node receives second information from the first node, the second information is used to indicate at least one model parameter, and a correspondence between a combination of at least one second resource and a transmission layer of the second node and the at least one model parameter.

[0068] In an optional implementation, the first information is used to indicate multiple resource blocks, each resource block including at least one first resource and at least one second resource; the frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

[0069] Optionally, the relevant contents in this aspect may refer to the optional implementation of the fourth aspect above, which will not be elaborated here. Optionally, for each transport layer, the optional implementation performed by the second node may also refer to the relevant description of the second aspect above, which will not be elaborated here.

[0070] In the sixth aspect, the embodiment of the present application also provides a distributed training air interface feedback method, which is explained from the perspective of interaction between a first node and at least two second nodes among a plurality of second nodes. In the method, the first node sends first information to at least two second nodes among a plurality of second nodes, and correspondingly, at least two second nodes receive the first information. Among them, the first information is used to indicate a plurality of first resources and at least one second resource; wherein, there is a corresponding relationship between a plurality of first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter. At least one model parameter is a part or all of the parameters that the second node uses local training data to train the model, and the training results or update gradients need to be fed back. Furthermore, at least two second nodes determine the data transmitted by the transmission layer on the second resource based on the product between the training results or update gradients of the model parameters corresponding to the second resource and the transmission layer, and the pilot corresponding to the transmission layer; at least two second nodes transmit the pilot of the corresponding transmission layer on the first resource, and transmit the determined data on the transmission layer on the second resource. Correspondingly, the first node receives the pilot of the transmission layer corresponding to each first resource transmitted by at least two second nodes, and the data transmitted by each transmission layer on at least one second resource.

[0071] In this method, there is a corresponding relationship between the combination of the transmission layer of at least one second resource and the second node and each model parameter. In this way, through multi-layer transmission, at least two second nodes can feedback the training results or update gradients of multiple model parameters on the same second resource, thereby further improving the utilization rate of time-frequency resources.

[0072] In addition, since the first node sends the same first information to at least two second nodes, the at least two second nodes can transmit the pilot on the transmission layer corresponding to the same pilot resource, and transmit the training results of the same model parameters obtained by their respective training or the data corresponding to the update gradient on the same second resource and transmission layer, so that the merged channel of multiple wireless channels between each second node and the first node and the merged data signal corresponding to the data transmitted by each second node on each second resource can be obtained. In this way, the natural merging effect of each wireless channel is utilized to complete the fusion of the training results or update gradients of the model parameters that need to be fed back by each second node, thereby saving channel resources.

[0073] In addition, the present application is also conducive to estimating the combined channel based on the pilot, which is used to equalize the combined data signal, thereby helping to combat random channel fading, especially small-scale fading with greater randomness, and further helping to improve the accuracy of air interface feedback.

[0074] In an optional implementation, there is a one-to-one correspondence between a combination of at least one second resource and a transmission layer of the second node and at least one model parameter.

[0075] In an optional embodiment, the method also includes: the first node sends second information to at least two second nodes, the second information is used to indicate at least one model parameter, and the correspondence between a combination of at least one second resource and a transmission layer of the second node, and at least one model parameter.

[0076] In an optional implementation, the first information is used to indicate multiple resource blocks, each resource block including at least one first resource and at least one second resource; the frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

[0077] Optionally, the total number of second resources in the multiple resource blocks multiplied by the number of transmission layers of the second node is not less than the number of model parameters to be fed back. Optionally, the number of second resources used to transmit training results or update gradients of one or more model parameters is not more than the number of transmission layers of the second node.

[0078] Optionally, the optional implementation methods executed by the first node and the second node can also refer to the relevant explanations of the fourth and fifth aspects above, and will not be described in detail here.

[0079] In a seventh aspect, an embodiment of the present application further provides a communication device. The communication device is a first node, or a device of the first node, or a device that can be used in conjunction with the first node. In a possible implementation, the communication device includes a functional module, and the functional module is a hardware circuit, or software, or a combination of a hardware circuit and software.

[0080] In a possible implementation, the communication device includes one or more functional units, such as a communication unit, wherein the communication unit is used to send first information to at least two second nodes among a plurality of second nodes, the first information being used to indicate at least one first resource and at least one second resource; and is also used to receive pilots transmitted by at least two second nodes on at least one first resource, and data transmitted on at least one second resource. There is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a part or all of a parameter for which feedback of training results or update gradients is required when at least two second nodes train a model using local training data; the data transmitted by the second node on the second resource is obtained based on the product of the training results or update gradients of the model parameters corresponding to the second resource and the pilot.

[0081] Optionally, in this implementation, the first information, the second information and other related contents and possible implementations of the communication device can be found in the relevant description of the first aspect and will not be described in detail here.

[0082] In another possible implementation, in the communication device, the communication unit is used to send first information to at least two second nodes among a plurality of second nodes, the first information being used to indicate a plurality of first resources and at least one second resource; wherein there is a corresponding relationship between the plurality of first resources and the transmission layer of the second node; there is a corresponding relationship between a combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is a part or all of a parameter for which the second node trains the model using local training data and needs to feed back the training result or update the gradient when training; the communication unit is also used to receive a pilot of the transmission layer corresponding to each first resource transmitted by at least two second nodes, and data transmitted by each transmission layer on the at least one second resource; wherein the data transmitted by the transmission layer of at least two second nodes on the second resource is obtained based on the product between the training result or update gradient of the model parameters corresponding to the second resource and the transmission layer and the pilot corresponding to the transmission layer.

[0083] Optionally, in this implementation, the first information, the second information and other related contents and possible implementations of the communication device can be found in the relevant description of the fourth aspect and will not be described in detail here.

[0084] In an eighth aspect, an embodiment of the present application further provides a communication device. The communication device is a second node, or a device of the second node, or a device that can be used in conjunction with the second node. In a possible implementation, the communication device includes a functional module, which is a hardware circuit, or software, or a combination of a hardware circuit and software.

[0085] In one possible implementation, the communication device includes one or more functional units, such as a communication unit and a processing unit, wherein the communication unit is used to receive first information from a first node, the first information is used to indicate multiple first resources and at least one second resource; wherein there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is a part or all of a parameter that the second node uses local training data to train the model and needs to feedback the training result or update the gradient; the processing unit, the processing unit is used to determine the data transmitted by the transmission layer on the second resource based on the product between the training result or update gradient of the model parameter corresponding to the second resource and the transmission layer and the pilot corresponding to the transmission layer; the communication unit is also used to transmit the pilot of the corresponding transmission layer on the first resource, and transmit the determined data on the transmission layer on the second resource.

[0086] Optionally, in this implementation, the first information, the second information and other related contents and possible implementations of the communication device can be found in the relevant description of the second aspect and will not be described in detail here.

[0087] In another possible implementation, in the communication device, the communication unit is used to receive first information from the first node, the first information is used to indicate multiple first resources and at least one second resource; wherein there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is a part or all of a parameter that the second node trains the model using local training data and needs to feed back the training results or update the gradient; the processing unit is used to determine the data transmitted by the transmission layer on the second resource based on the product between the training results or update gradients of the model parameters corresponding to the second resource and the transmission layer and the pilot corresponding to the transmission layer; the pilot of the transmission layer corresponding to the first resource transmission, and the data determined by the transmission layer transmission on the second resource.

[0088] Optionally, in this implementation, the first information, the second information and other related contents and possible implementations of the communication device can be found in the relevant description of the fifth aspect and will not be described in detail here.

[0089] For the seventh and eighth aspects, as examples, the processing unit can be a processing unit or can be embodied as a processing circuit or a logic circuit; the transceiver unit can be an input / output interface, interface circuit, output circuit, input circuit, pin or related circuit on the chip or chip system.

[0090] During the implementation process, the processor can be used to perform, for example, but not limited to, baseband related processing, and the transceiver or communication interface can be used to perform, for example, but not limited to, radio frequency transceiver. The above-mentioned devices can be arranged on independent chips, or at least partially or completely on the same chip. For example, the processor can be further divided into an analog baseband processor and a digital baseband processor. Among them, the analog baseband processor can be integrated with the transceiver (or communication interface) on the same chip, and the digital baseband processor can be arranged on an independent chip. With the continuous development of integrated circuit technology, more and more devices can be integrated on the same chip. For example, a digital baseband processor can be integrated with a variety of application processors (such as but not limited to a graphics processor, a multimedia processor, etc.) on the same chip. Such a chip can be called a system on a chip (System on a Chip, SoC). Whether each device is independently arranged on different chips or integrated on one or more chips often depends on the needs of product design. The embodiment of the present application does not limit the implementation form of the above-mentioned devices.

[0091] In the ninth aspect, the embodiment of the present application also provides a processor for executing the method of any one of the first aspect, the second aspect, the fourth aspect and the fifth aspect, or the method of any possible implementation of any one of the first aspect, the second aspect, the fourth aspect and the fifth aspect. In the process of executing these methods, the process of sending the above-mentioned signal and receiving the above-mentioned signal can be understood as the process of outputting the above-mentioned signal by the processor, and the process of the above-mentioned signal input by the processor. When outputting the above-mentioned signal, the processor outputs the above-mentioned signal to the transceiver so that it is transmitted by the transceiver (or communication interface). After the above-mentioned signal is output by the processor, it may also need to perform other processing before it reaches the transceiver (or communication interface). Similarly, when the processor receives the above-mentioned signal input, the transceiver (or communication interface) receives the above-mentioned signal and inputs it into the processor. Further, after the transceiver (or communication interface) receives the above-mentioned signal, the above-mentioned signal may need to perform other processing before it enters the processor.

[0092] For the sending and receiving operations involved in the processor, unless otherwise specified, or unless they conflict with their actual function or internal logic in the relevant description, they can be more generally understood as processor output, reception, input and other operations, rather than sending and receiving operations performed directly by the RF circuit and antenna.

[0093] In the implementation process, the processor may be a processor specifically used to execute these methods, or a processor that executes computer instructions in a memory to execute these methods, such as a general-purpose processor. The memory may be a non-transitory memory, such as a read-only memory (ROM), which may be integrated with the processor on the same chip or may be separately arranged on different chips. The embodiment of the present application does not limit the type of memory and the arrangement of the memory and the processor.

[0094] In a tenth aspect, an embodiment of the present application further provides a communication device, comprising: a processor, configured to call a computer program stored in a memory, and through a transceiver, enable the communication device to implement the method of any one of the first, second, fourth and fifth aspects or any possible implementation method. Optionally, the communication device further comprises a memory, and the processor is coupled to the memory.

[0095] In the eleventh aspect, the present application further provides a communication system, the system comprising the first node executing the first aspect or any optional implementation method in the first aspect and at least two second nodes executing the second aspect or any possible implementation method in the second aspect; and / or, the system comprises the first node executing the fourth aspect or any optional implementation method in the fourth aspect and at least two second nodes executing the fifth aspect or any possible implementation method in the fifth aspect. In another possible design, the system may also include other devices that interact with the first device and / or the second device in the solution provided in the embodiment of the present application.

[0096] In the twelfth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed, the computer device implements the method of any aspect or any possible implementation method of the above-mentioned first aspect, second aspect, fourth aspect and fifth aspect through a transceiver and a processor.

[0097] In the thirteenth aspect, the present application also provides a computer program product comprising instructions, the computer program product comprising: computer program code, when the computer program code is run in parallel, enables a computer device to implement the method of any aspect or any possible implementation of the above-mentioned first aspect, second aspect, fourth aspect and fifth aspect through a transceiver and a processor.

[0098] In a fourteenth aspect, the present application provides a chip system, which includes a processor and an interface, the interface is used to obtain a program or instruction, the processor is used to call the program or instruction, and the interface is used to implement the method of any aspect or any possible implementation of the above-mentioned first aspect, second aspect, fourth aspect and fifth aspect. In a possible design, the chip system also includes a memory, which is used to store program instructions and data necessary for the terminal. The chip system can be composed of chips, or it can include chips and other discrete devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 It is a flow chart of a distributed training air interface feedback method provided in an embodiment of the present application;

[0100] Figure 2 It is a scenario diagram of a distributed training air interface feedback method provided in an embodiment of the present application;

[0101] Figure 3 It is a flowchart of another distributed training air interface feedback method provided in an embodiment of the present application;

[0102] Figure 4 It is a scenario diagram of another distributed training air interface feedback method provided in an embodiment of the present application;

[0103] Figure 5 is a schematic diagram of multiple resource blocks provided in an embodiment of the present application;

[0104] Figure 6 is a schematic diagram of a model training method provided in an embodiment of the present application;

[0105] Figure 7 is a structural diagram of a communication device provided in an embodiment of the present application;

[0106] Figure 8 It is a structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0107] The embodiments of the present application can be applied to communication systems of various radio access technologies (RAT), such as narrowband Internet of things (NB-IoT), long term evolution (LTE) communication systems, 5G (or new radio, NR) communication systems, and of course future communication systems, such as the sixth generation (6G) or even the seventh generation (7G) systems. The embodiments of the present application can also be applied to non-terrestrial networks (NTN), vehicle to everything (V2X), long term evolution-vehicle (LTE-V), vehicle to vehicle (V2V), machine type communications (MTC), Internet of things (IoT), long term evolution-machine to machine (LTE-M), machine to machine (M2M), or future mobile communication systems. The embodiments of the present application may also be applied to other wireless networks that adopt various standards or protocols, such as Bluetooth, wireless local area network (WLAN), or other wireless networks that are currently known or developed later.

[0108] In the embodiment of the present application, the first node is used to send at least one pilot resource and at least one data resource to at least two second nodes among the plurality of second nodes, and at least two second nodes can send pilots on the pilot resources and transmit corresponding data on at least one data resource, wherein the data transmitted by each second node is used to feed back the training results or update gradients of the model parameters obtained by each second node through model training. Thus, the natural merging effect of each wireless channel is utilized to complete the fusion of the training results or update gradients of the model parameters to be fed back by at least two second nodes, thereby saving channel resources.

[0109] Optionally, the first node may be a central node for maintaining an artificial intelligence model or a machine learning model for a specific function. The first node may be a network device or a terminal device, etc. The first node may send the model parameters of the model at the current moment to at least two second nodes through a downlink channel, and at least two second nodes may perform a training once using local training data to obtain the training results or update gradients of each model parameter. Furthermore, the distributed training air interface feedback method provided in the embodiment of the present application may be used to feedback the training results or update gradients of each model parameter. Optionally, the second node is a distributed node for participating in the distributed training of an artificial intelligence model or a machine learning model, and the second node may be a terminal device or a network device. In the case where the first node is a terminal device and the second node is also a terminal device, the first node may send the initial model parameters at the current moment to other terminal devices participating in the distributed training through a side link, and then the terminal device may perform the relevant operations of the first node in the embodiment of the present application, and the other terminal devices participating in the distributed training may perform the relevant operations of the second node in the embodiment of the present application.

[0110] Optionally, the network device is an access network (AN) device, such as a base station, that a terminal accesses to a mobile communication system in a wireless manner. The network device may also refer to a device that communicates with a terminal device at an air interface. The network device may include an evolved NodeB (eNodeB or eNB) transmission reception point (TRP) in a long term evolution (LTE) system or long term evolution-advanced (LTE-A), a next generation NodeB (gNB) in a fifth generation (5G) mobile communication system, an access network device in an open radio access network (O-RAN or open RAN), a next generation base station in a sixth generation (6G) mobile communication system, or a base station in a future mobile communication system, or an access node in a wireless fidelity (WiFi) system, etc.; or, the network device may be a relay station, a vehicle-mounted device, a future evolved public land mobile network (PLMN) device, a device in a device-to-device (D2D) network, a device in an M2M network, a device in an IoT network, or a network device in a public land mobile network (PLMN), etc.

[0111] Optionally, the network device takes a base station as an example. The base station can communicate with the terminal device or communicate with the terminal device through a relay station. The terminal can communicate with multiple base stations in different access technologies. The network device can be a module or unit that completes some functions of the base station, for example, it can be a centralized unit (CU), a distributed unit (DU), a centralized unit control plane (CU control plane, CU-CP) module, or a centralized unit user plane (CU user plane, CU-UP) module. Multiple DUs can be centrally controlled by one CU. CU and DU can be divided according to the protocol layer functions of the wireless network they possess, for example, the functions of the packet data convergence protocol (PDCP) layer and the above protocol layers are set in the CU, and the protocol layers below the PDCP, such as the radio link control (RLC) layer and the medium access control (MAC) layer, are set in the DU. It should be noted that this division of the protocol layer is only an example, and it can also be divided at other protocol layers. The radio frequency device can be remote and not placed in the DU, or it can be integrated in the DU, or partly remote and partly integrated in the DU, and the embodiments of the present application do not impose any restrictions. In addition, in some embodiments, the control plane (control plan, CP) and user plane (user plan, UP) of the CU can also be separated and implemented by different entities, namely the control plane CU entity (CU-CP entity) and the user plane CU entity (CU-UP entity). In this network architecture, the signaling generated by the CU can be sent to the terminal device through the DU, or the signaling generated by the terminal device can be sent to the CU through the DU. The DU can directly encapsulate the signaling through the protocol layer and transparently transmit it to the terminal device or CU without parsing it. In this network architecture, the CU is divided into a network device on the radio access network (radio access network, RAN) side. In addition, the CU can also be divided as a network device on the core network (core network, CN) side, and the present application does not impose any restrictions on this. Optionally, the access network device can be a server, etc. For example, the network device in the vehicle V2X technology can be a road side unit (road side unit, RSU). Optionally, the network device may also be various types of devices constituting an access node, such as an active antenna unit (AAU), a baseband unit (BBU), and the like.

[0112] Optionally, the network device is a network device in the NTN system, and can be deployed on a high-altitude platform or a satellite, such as a satellite or a satellite base station.

[0113] Optionally, the network device may be a macro base station (also known as a large station), a micro base station or an indoor station (also known as a small station), or a relay node or a donor node, etc. The specific technology and specific device form adopted by the access network device are not limited in this application. Optionally, the communication device used to implement the function of the network device may be a network device, or a device that can support the network device to implement the function, such as a chip system, which may be installed in the network device.

[0114] Optionally, a terminal, also known as a terminal device (terminal), user equipment (UE), mobile station (MS), mobile terminal (MT), etc., is a device with wireless transceiver function, which can send signals to network devices or receive signals from network devices. The terminal device may include user equipment (UE), sometimes also called terminal, access station, UE station, remote station, wireless communication equipment, or user device, etc. The terminal device is used to connect people, objects, machines, etc., and can be widely used in various scenarios, such as but not limited to the following scenarios: cellular communication, D2D, V2X, M2M / MTC, IoT, virtual reality (VR), augmented reality (AR), industrial control, self driving, remote medical, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, drone, robot and other scenarios of terminal equipment. For example, the terminal can be a mobile phone, a tablet computer, a computer with wireless transceiver function, a VR terminal, an AR terminal, a wireless terminal in industrial control, a wireless terminal in unmanned driving, a smart speaker in an IoT network, a wireless terminal device in telemedicine, a wireless terminal device in a smart grid, a wireless terminal device in transportation safety, a wireless terminal device in a smart city, or a wireless terminal device in a smart home, etc. As an example but not a limitation, the terminal can also be a wearable device, which can also be called a wearable smart device or a smart wearable device, etc., which is a general term for the intelligent design and development of wearable devices for daily wear using wearable technology, such as glasses, gloves, watches, clothing and shoes, etc. The various terminals introduced above, if located on a vehicle (for example, placed in a vehicle or installed in a vehicle), can be considered as vehicle-mounted terminal devices, which are also called on-board units (OBUs). The terminal may also be a vehicle-mounted module, vehicle-mounted module, vehicle-mounted component, vehicle-mounted chip or vehicle-mounted unit that is built into the vehicle as one or more components or units. The vehicle may implement the methods described in the embodiments of the present application through the built-in vehicle-mounted module, vehicle-mounted module, vehicle-mounted component, vehicle-mounted chip or vehicle-mounted unit.

[0115] The network equipment and / or terminal equipment can be fixed or movable. The network equipment and / or terminal equipment can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; or can be deployed on the water surface; or can be deployed on aircraft, balloons and artificial satellites in the air. This application does not limit the environment / scenario in which the network equipment and terminal equipment are located. The network equipment and terminal equipment can be deployed in the same or different environments / scenarios, for example, the network equipment and terminal equipment are deployed on land at the same time; or, the network equipment is deployed on land and the terminal equipment is deployed on the water surface, etc., and examples are not given one by one.

[0116] In the distributed training scenario of artificial intelligence (AI), the central node sends the model to each distributed node, and the distributed node performs distributed model training in combination with local training samples, and then feeds back the training results or update gradients of the local artificial intelligence model parameters to the central node through the air interface, and the central node summarizes the training results or update gradients of the parameters for model update. When the number of distributed nodes participating in the training is large, the overhead of each distributed node reporting the training results or update gradients of the parameters is large. Therefore, how to reduce the overhead of training results or update gradients is a problem to be solved. In addition, it was found in the study that since the wireless channel is random, the channel information mastered by the central node or distributed node is erroneous. Therefore, in the process of using the natural merging effect of each wireless channel to complete the training results of the model parameters or update the feedback of the gradient, the influence of the channel is uncontrollable, which will cause a large gap between the received signal and the actual merging result.

[0117] In the present application, the second nodes participating in the feedback can share the same pilot resources (or share the same pilot resources and pilots), and the data sent by each second node is obtained based on the training results of the model parameters to be fed back or the product between the update gradient and the corresponding pilot. This pilot air merging, and the pilot and model parameter training results or update gradient superposition transmission method can effectively combat random channel resources and improve the accuracy of air interface feedback. It can be understood that in the embodiment of the present application, the total number of distributed nodes (i.e., second nodes) can be K' (K' is an integer greater than 2), and some of the second nodes (such as K, K is greater than or equal to 2 and less than or equal to K') can be selected to adopt the scheme of the embodiment of the present application, and other second nodes can adopt other methods, such as other methods that do not adopt the pilot air merging and pilot superposition transmission methods described in the present application. The embodiments of the present application are not limited to this. In conjunction with the accompanying drawings, the following embodiments included but not limited to the present application are elaborated.

[0118] The embodiment of the present application provides a distributed training air interface feedback method, in which a first node sends a first message to at least two second nodes among a plurality of second nodes, and at least two second nodes receive the first message. The first information is used to indicate at least one first resource and at least one second resource. There is a corresponding relationship between at least one second resource and at least one model parameter. At least one model parameter is a parameter that at least two second nodes use local training data to train the model, and part or all of the parameters of the training result or update gradient need to be fed back. At least two second nodes determine the data transmitted on the second resource based on the training result of the model parameter corresponding to the second resource or the product between the update gradient and the pilot, transmit the pilot on at least one first resource, and transmit the determined data on the second resource. Accordingly, the first node receives the pilot transmitted by at least two second nodes on at least one first resource, and the data transmitted on at least one second resource. Since the first node sends the same first information to at least two second nodes, at least two second nodes can transmit the pilot on the same pilot resource, and transmit the training result or update gradient corresponding to the same model parameter obtained by their respective training on the same second resource. It can be seen that this method utilizes the natural merging effect of each wireless channel to complete the training results of the model parameters that need to be fed back or the fusion of the updated gradients, thereby saving channel resources.

[0119] See also Figure 1 , Figure 1 is a flow chart of a distributed training air interface feedback method provided in an embodiment of the present application, such as Figure 1 As shown, the method is described by taking the case where at least two of the plurality of (eg, K') second nodes are K second nodes as an example, where K is an integer greater than or equal to 2 and less than or equal to K'. Figure 1 As shown, the method includes but is not limited to the following steps:

[0120] S101. A first node sends first information to K second nodes; correspondingly, the K second nodes receive the first information; the first information is used to indicate at least one first resource and at least one second resource.

[0121] The first resource is used to transmit a pilot signal, and the second resource is used to transmit a training result or an update gradient of a model parameter obtained by the second node performing model training. Optionally, the first resource may be referred to as a pilot signal resource, and the second resource may be referred to as a data resource, and both may include certain time domain resources, such as a number of symbols, time slots or sub-time slots, and certain frequency domain resources, such as a number of subcarriers, resource blocks or resource block groups.

[0122] Optionally, the first node may send the first information to the K second nodes in a broadcast manner, and correspondingly, the K second nodes receive the first information in a broadcast manner.

[0123] In an optional implementation, the first resource indicated by the first information is one, so that at least two second nodes transmit pilot signals on the same first resource, thereby avoiding the problem of excessive pilot resource overhead caused by a large number of second nodes participating in model training.

[0124] In another optional implementation, the first information indicates a plurality of first resources, so that at least two second nodes transmit pilots on the same plurality of first resources. It can be seen that in this implementation, at least two second nodes transmit multiple pilots using multiple first resources, which can improve the signal-to-noise ratio of the pilots compared to using one first resource to transmit a pilot once, thereby improving the accuracy of channel estimation. For example, in the case where the channel has frequency selectivity or time selectivity, pilot reception and channel estimation are performed on multiple first resources respectively, and the change of the channel over time or frequency can be estimated, thereby improving the accuracy of channel estimation. Optionally, the first information indicates a plurality of first resources. If the pilots corresponding to the plurality of first resources are the same, the first node can average the received pilot signals on the plurality of first resources. If the pilots corresponding to the plurality of first resources are different, the first node can receive pilot signals on the plurality of first resources respectively, but the ratio of the pilots on the plurality of first resources between different second nodes must be the same. For details, please refer to the relevant description of formula (3) below.

[0125] There is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a part or all of the parameters that require feedback of training results or updated gradients for at least two second nodes to train the model using local training data. Optionally, the model maintained by the first node may be an artificial intelligence model or a machine learning model, such as a distributed training scenario of distributed artificial intelligence in wireless communications such as a neural network model for wireless resource management functions or channel state information feedback functions. The K second nodes are distributed nodes that participate in model training in one of the one or more training stages. Optionally, the distributed training of the model may include but is not limited to the initial training stage, the training stage after the initial training stage, and the final training stage. The distributed training air interface feedback method described in the embodiment of the present application may be applicable to one or more of the training stages. Optionally, before or after S101, the first node may also transmit the initial model parameters of the model at the current moment to the K second nodes through the downlink channel. The K second nodes, based on the initial model parameters received, respectively use local training data to train the model to obtain the training results of the model parameters or update the gradients.

[0126] In an optional implementation, there is a one-to-one correspondence between at least one second resource and at least one model parameter. Among them, the at least one model parameter is a parameter for which a training result or an updated gradient needs to be fed back. Optionally, the number of second resources indicated by the first information is greater than or equal to the number of model parameters. If the number of second resources is greater than the number of model parameters, some second resources may not need to transmit the training results or updated gradients of the corresponding model parameters. In an optional implementation, the first information indicates one second resource, and there is a corresponding relationship between one second resource and one model parameter, so that at least two second nodes respectively feed back the training results or updated gradients of the one model parameter on the same second resource. In another optional implementation, the first information indicates multiple second resources, and there is a one-to-one correspondence between multiple second resources and multiple model parameters. For each second resource in the multiple second resources, at least two second nodes can respectively feed back the training results or updated gradients of the model parameters corresponding to the second resource on the second resource, thereby realizing feedback on multiple model parameters.

[0127] In an optional implementation, the method may further include: the first node sends second information to K second nodes, the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and at least one second resource. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or updated gradients need to be fed back, and the correspondence between these model parameters and the second resources for transmitting the corresponding training results or updated gradients.

[0128] S102. The K second nodes determine, for a second resource in at least one second resource, data to be transmitted on the second resource based on a training result or an update gradient of a model parameter corresponding to the second resource multiplied by a pilot signal.

[0129] Optionally, the first node may further indicate the pilot p to the K second nodes respectively. Optionally, the K pilots of the K second nodes may be the same or different. Optionally, the K pilots may be designed to have a constant amplitude and a random phase jump, thereby improving the effectiveness of the combined data signal received by the first node against random channel fading. Optionally, in the case where there are multiple first resources indicated by the first information, the pilots corresponding to the multiple first resources of a second node may be the same or different from the pilots corresponding to the multiple first resources of other second nodes, but the pilots corresponding to the multiple first resources of the same second node need to be different.

[0130] Optionally, the data transmitted by the second node on the second resource is the product of the training result or update gradient of the model parameter corresponding to the second resource and the pilot and the weight corresponding to the second node. The weight corresponding to the second node is determined based on the quantity and quality of the local training data of the second node, and / or the channel information between the second node and the first node. In an optional implementation, the weight is a scalar. In another optional implementation, the weight is a vector.

[0131] S103. K second nodes send a pilot on each first resource and transmit certain data on at least one second resource; accordingly, the first node receives pilots transmitted by at least two second nodes on at least one first resource and data transmitted on at least one second resource.

[0132] It can be seen that in this method, since the first node sends the same first information to at least two second nodes, the at least two second nodes can transmit the pilot on the same pilot resource, and transmit the training results of the same model parameters obtained by their respective training or the data corresponding to the updated gradient on the same second resource. It can be seen that this method utilizes the natural merging effect of each wireless channel to complete the fusion of the training results of the model parameters that need to be fed back or the updated gradient, thereby saving channel resources.

[0133] In addition, the present application is also beneficial for the first node to obtain the merged channel of multiple wireless channels between each second node and the first node and the merged data signal corresponding to the data transmitted by each second node on each second resource, which is further beneficial to estimate the merged channel based on the pilot signal and equalize the merged data signal, which is beneficial to combat random channel fading, especially small-scale fading with greater randomness, and is further beneficial to improving the accuracy of air interface feedback.

[0134] The following combination Figure 2 ,right Figure 1 The distributed training air interface feedback method is described by taking an example. Figure 2 is a scenario diagram of a distributed training air interface feedback method provided in an embodiment of the present application, such as Figure 2 As shown in FIG. 1 , the network device is used as the first node; K terminal devices are used as the second node, where K is an integer greater than or equal to 2, to participate in the distributed training of the model. During the distributed training, the network device sets the initial model parameter w of the current model to 0 Transmitted to K terminal devices through the downlink channel, where the model parameter w 0 A vector containing M model parameters to be fed back, i.e., model parameter w 0 is an M-dimensional vector. The K terminal devices receive the initial model parameters w 0Based on this, the model is trained using local training data to obtain the training results of the model parameters or update the gradients, such as Figure 2 As shown, the model parameters w obtained by training the terminal device 1 are 0 The training result: w 1 =w 0 +Δw 1 , the model parameters w obtained by training the terminal device 2 0 The training result: w 2 =w 0 +Δw 2 ,…, the model parameters w obtained by training the terminal device k 0 The training result: w k =w 0 +Δw k ,…, the model parameters w obtained by training the terminal device K 0 The training result: w K =w 0 +Δw k . Among them, Δw k is the model parameter w obtained by training the terminal device k 0 The updated gradient of 1 To W K , and Δw 1 To Δw K It is also an M-dimensional vector.

[0135] like Figure 2 As shown, taking M second resources as an example, where M is an integer greater than or equal to 1, ( Figure 2 In the scenario shown, the second resource may be an uplink channel resource). On the M second resources, the terminal device 1 sends data x 1 =f(p 1 Δw 1 ), terminal device 2 sends data x 2 =f(p 2 Δw 2 ), …, terminal device k sends data x k =f(p k Δw k ), …, data x sent by terminal device K K =f(p K Δw K ). Among them, x 1 to x K are M-dimensional vectors respectively.

[0136] Optionally, the data transmitted by the second node k on a second resource is the training result or update gradient of the model parameter corresponding to the second resource, and ... k The weight a corresponding to the second nodek The product between. Among them, the weight a k The weight is determined according to the quantity and quality of the local training data of the second node k and the channel information between the second node and the first node. In an optional implementation, the weight is a scalar. In another optional implementation, the weight is a vector.

[0137] For example, Figure 2 For example, K terminal devices simultaneously send certain data on M second resources, where the data x sent by terminal device k on second resource m is k,m , corresponding to the M-dimensional vector w fed back by the terminal device k to the network device k One of the components is w k,m′ , or the M-dimensional vector Δw that the terminal device k feeds back to the network device k One of the components, namely Δw k,m′ The second resource m corresponds to the model parameter m' among the M model parameters. That is, the training result of the model parameter m' fed back by the terminal device k in the second resource m can be expressed as w k,m′ Or the updated gradient of the model parameter m' can be expressed as Δw k,m′ The data x sent by the terminal device k on the second resource m k,m It can be expressed as:

[0138] x k,m =p k a k w k,m ',or,

[0139] x k,m =p k a k Δw k,m ′;

[0140] Among them, x k,m It can be the aforementioned M-dimensional vector x k The data corresponding to the second resource m in p k . represents the pilot signal sent by terminal device k, a k is the weight corresponding to terminal device k.

[0141] The K pilot signals (or pilot signal sequences) of K terminal devices can be expressed as: 1 , p 1 ,…,p k ,…,p K , taking the first information indicating a first resource as an example, K terminal devices transmit their respective corresponding pilot signals on the first resource simultaneously, and the pilot signal q received by the network device on the first resource can be expressed as shown in the following formula (1):

[0142] The wireless channels between the K terminal devices and the network device (for example, in the order of the index size of the terminal devices) can be represented as h 1 ,h 2 ,…,h k ,…,h K , such as h k represents the transfer function of the wireless channel between the terminal device k and the network device, and these wireless channels are respectively an N-dimensional vector; n (q) represents the noise on the first resource (ie, the pilot resource); q is an N-dimensional vector, wherein N is the number of antennas of the first node, and N is an integer greater than or equal to 1.

[0143] K terminal devices transmit certain data simultaneously on M second resources, such as Figure 2 The M-dimensional vector x shown 1 to x K .by Figure 2 The data x transmitted by the terminal device k on the second resource m k,m =p k a k Δw k,m ′ is taken as an example (i.e., taking a data signal on a data resource and the feedback is an update gradient of a model parameter m′ as an example), for simplicity, the subscript m or m′ is omitted, and the data signal y received by the network device on the second resource m can be expressed as shown in the following formula (2):

[0144] Where y is an N-dimensional vector; n (y) represents the noise on the second resource.

[0145] In the embodiment of the present application, the correspondence between the second resource and the model parameter is independent of the index of the second node, but the first node and each second node must agree on the correspondence between each second resource and each model parameter, which can be indicated by the second information as mentioned above. It can be seen that, as formula (2) utilizes the natural superposition effect of the wireless channel, the data signal y received by the network device on the second resource is equivalent to the weighted sum of the update gradient or training result of the same model parameter. In addition, in the embodiment of the present application, the update gradient or training result of the model parameter is multiplied by the pilot, which is equivalent to the equivalent channel through which the data sent by the second node passes is the product of the real channel and the pilot; in this way, the data sent by multiple second nodes at the same time passes through the merged equivalent channel, which is equivalent to the merger of the product of the channel and the pilot of each second node. Since the pilot received by the first node is also the merger of the product of the channel and the pilot of each second node, it is the same as the merged equivalent channel of the data and the pilot, which is beneficial to balance the merged data signal, thereby helping to combat random channel fading, especially small-scale fading with large randomness, and thus helping to improve the accuracy of air interface feedback.

[0146] In the embodiment of the present application, the at least one second resource can be shared between the second nodes, and the second node multiplies the training result or update gradient of the model parameter to be fed back by the pilot, determines the data, and maps the data to each second resource according to the corresponding relationship (the data is not quantized and digitally modulated). Therefore, in the embodiment of the present application, the number of second resources consumed by the feedback of the training result or update gradient of the model parameter is determined by the number of model parameters to be fed back, and each second node adopts the traditional digital coding modulation, and then maps the modulated data to different uplink channel resources. The number of data transmission resources consumed is proportional to K times the number of model parameters to be fed back. In particular, when the number K of second nodes participating in model training and the number M of model parameters to be fed back are large, the uplink resource overhead brought by the traditional feedback method may be too large to bear, and the distributed training air interface feedback method described in the embodiment of the present application can greatly save feedback overhead.

[0147] Since the first node sends the same first information to at least two second nodes, the at least two second nodes can transmit the pilot on the same pilot resource, and the merged channel of multiple wireless channels between each second node and the first node can be obtained, which is beneficial to estimate the merged channel based on the pilot, and is used to equalize the merged data signal, which is beneficial to combat random channel fading, especially small-scale fading with greater randomness, and further helps to improve the accuracy of air interface feedback.

[0148] Optional, Figure 1 The distributed training air interface feedback method shown may also include the following steps:

[0149] S104. The first node estimates a merged channel of multiple wireless channels based on the received pilot signal; and determines the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients based on the merged channel of the multiple wireless channels and the data signal received on the second resource.

[0150] The data signal received on the second resource, as shown in formula (2), is a combined data signal formed by superimposing the product of the transfer function of each wireless channel and the data transmitted by the corresponding second node on the second resource.

[0151] In an optional implementation, the first node determines the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients based on the merged channel of multiple wireless channels and the data signal received on the second resource, including: performing equalization processing using the merged channel of multiple wireless channels and the data signal received on the second resource to obtain the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients.

[0152] In an optional implementation, equalization processing is performed using a merged channel of multiple wireless channels and a data signal received on a second resource to obtain a training result of a model parameter corresponding to the second resource or a merged result of an updated gradient, including: dividing the inner product between a pilot signal received on at least one first resource and a data signal received on the second resource by the inner product of at least one first resource received and the pilot signal itself, to obtain a training result of a parameter corresponding to the second resource after equalization processing or a merged result of an updated gradient.

[0153] It can be seen that in this implementation, the training results of the model parameters corresponding to the second resource or the merged results of the updated gradients are a weighted average of the training results or updated gradients of the model parameters transmitted by each second node on the second resource. In this way, even if the channel is random and unknown, it only affects the weight of the weighted average, and does not directly affect the amplitude and phase of the merged result. Therefore, the impact of the unknown channel on the merged result can be significantly reduced, thereby improving the accuracy of the air interface feedback.

[0154] For example, Figure 2 For example, the inner product between the pilot signal q (as shown in formula (1)) received by the network device on the first resource and the data signal y (as shown in formula (2)) received on the second resource can be expressed as the following formula (3) (for simplicity, the noise n is ignored below) (q) and n (y) ):

[0155]

[0156] Where * represents the conjugate of a complex number, Indicates p kThe conjugate of Indicates p k′ The conjugation of.

[0157] The inner product of the pilot signal q received by the network device on the first resource and itself can be expressed as the following formula (4):

[0158]

[0159] Where H represents the conjugate transpose, q H represents the conjugate transposed matrix of the received pilot signal q.

[0160] The training result of the model parameter corresponding to the second resource or the merged result of the updated gradient can be expressed as the following formula (5):

[0161]

[0162] It can be seen that the calculated z can be regarded as the a of each terminal device. k Δw k is a weighted average of , which is a scalar, where z is mainly determined by the first term in formula (5), and the first term is a positive real number, so a k Δw k The superposition is in-phase superposition, and the vector inner product The operation of the inner product in high-dimensional space is For the case where the vector inner product is large, the first node can control the corresponding second node not to participate in the same stage of training at the same time, so as to send pilots in different time domains or frequency domains. In this way, even if the channel h k is random and unknown, and will only affect the weight of the weighted average, and will not directly affect the amplitude and phase of the merging result z. Therefore, the embodiment of the present application can significantly reduce the impact of the unknown channel on the merging result, and solve the problem that the channel impact is uncontrollable during the air merging process, resulting in a large gap between the actual merging result of the merged data signal and the feedback model parameters. Optionally, the K pilots can be designed to have a constant amplitude and a random phase jump, which is used to whiten (or randomize) the cross term in formula (5) This further reduces the impact of channel randomness on the merging results.

[0163] In the embodiment of the present application, the first information indicates a first resource that is multiple, such as the first information indicates multiple (such as N p The pilot signal q[i] received by the first node on the first resource i can be expressed as shown in the following formula (6):

[0164]

[0165] Among them, the symbols q[i], p k[i]、n (q) [i] represents the received pilot, sent pilot and received noise on the first resource i, i=1,…,N p The pilot signal sent by each second node on the first resource i satisfies That is, each second node in N p N sent on the first resource p The long pilot sequence is composed of different scalars of each second node The same spreading sequence c[i] as that of each second node, i=1,…,N p generated.

[0166] Accordingly, the data transmitted by each second node on the second resource is based on the update gradient or training result of the model parameters and In this way, when the first node receives the pilot signal, it uses As the “average” received pilot signal on multiple first resources (* denotes the conjugate of a complex number), replace the formula q in the equation can be equalized. p =1, the above formula naturally degenerates into the case where there is only one pilot resource mentioned above.

[0167] Optionally, the distributed training air interface feedback method may further include: the network device updates the maintained model using the merged result of the model parameters. Optionally, in the embodiment of the present application, the model parameters that the second node needs to feedback may be M model parameters, and the M model parameters may be vectors of M neural network parameters.

[0168] The following Figure 1 to Figure 2 In the embodiment, the first information indicates that the first resource is one or two, and the second resource is one or two, and further examples are given:

[0169] Example 1, assuming that the first node sends the first information to two second nodes, such as the second node 1 and the second node 2. The first information indicates a first resource, such as the first resource 1, and indicates a second resource, such as the second resource 1, and the second resource 1 corresponds to the model parameter 1. In addition, the first node allocates the pilot 1-1 transmitted on the first resource to the second node 1, and allocates the pilot 1-2 transmitted on the first resource 1 to the second node 2. The pilot 1-1 and the pilot 1-2 may be the same or different. In this way, the second node 1 determines the data 1-1 transmitted by the second resource 1 according to the training result of the model parameter 1 obtained by its own training or the product of the updated gradient and the pilot 1-1; and the second node 2 determines the data 1-2 transmitted on the second resource 1 according to the training result of the model parameter 1 obtained by its own training or the product of the updated gradient and the pilot 1-2. The second node 1 sends a pilot 1-1 on the first resource 1, the second node 2 sends a pilot 1-2 on the first resource 1, and the second node 1 sends data 1-1 on the second resource 1, and the second node 2 sends data 1-2 on the second resource 1. It can be seen that the second node 1 and the second node 2 can feed back the training results or update gradients obtained by training the model parameter 1 on the same second resource 1, so that the data signal received by the first node on the second resource 1 is the fusion of the training results or update gradients of the model parameter 1 fed back by the first node 1 and the second node 2, thereby saving channel resources. In addition, the first node can estimate the combined channel of the wireless channels between the first node and the second node 1 and the second node 2 respectively based on the pilot signal received on the first resource 1 (i.e., the average received pilot of the pilot 1-1 and the pilot 1-2), and then balance the data signal received on the second resource 1 based on the combined channel, so as to be beneficial to combat random channel fading, especially small-scale fading with large randomness, and thus be beneficial to improve the accuracy of air interface feedback.

[0170] Example 2 is different from Example 1 in that the first information indicates two first resources, such as first resource 1 and first resource 2. Optionally, the first node allocates pilot 1-1 transmitted on the first resource 1 and pilot 2-1 transmitted on the first resource 2 to the second node 1; the first node allocates pilot 1-2 transmitted on the first resource 1 and pilot 2-2 transmitted on the first resource 2 to the second node 2. In addition, it is also different from Example 1 in that, as described in formula (6), the pilot 1-1 and pilot 2-1 of the second node 1 are based on the same scalar Therefore, the second node 1 determines that the data 1-1 on the second resource 1 is based on the training result or update gradient of the model parameter 1, and the scalar Similarly, the second node 2 determines that the data 1-2 on the second resource 1 is based on the training result or update gradient of the model parameter 1, and the scalar The product between .

[0171] Example 3 is different from Example 2 in that the first information indicates two second resources, such as second resource 1 and second resource 2. Second resource 1 also corresponds to model parameter 1, and second resource 2 corresponds to model parameter 2. Second node 1 obtains the training result of model parameter 1 or updates the gradient and scalar according to its own training. The product of determines the data 1-1 transmitted by the second resource 1; and the training result of the model parameter 2 obtained by its own training or the updated gradient and scalar The second node 2 determines the data 2-1 transmitted by the second resource 2 according to the training result of the model parameter 1 obtained by its own training or updates the gradient and scalar The product of determines the data 1-2 transmitted on the second resource 1; and the training result of the model parameter 2 obtained by its own training or updates the gradient and scalar The product of determines the data 2-2 transmitted on the second resource 2.

[0172] Example 4 is different from the above-mentioned Example 1 in that the first information indicates two second resources, such as second resource 1 and second resource 2, and second resource 1 also corresponds to model parameter 1, and second resource 2 corresponds to model parameter 2. The second node 1 determines the data 1-1 transmitted by the second resource 1 according to the training result of the model parameter 1 obtained by its own training or the product of the updated gradient and the pilot 1-1; and determines the data 2-1 transmitted by the second resource 2 according to the training result of the model parameter 2 obtained by its own training or the product of the updated gradient and the pilot 1-1. The second node 2 determines the data 1-2 transmitted on the second resource 1 according to the training result of the model parameter 1 obtained by its own training or the product of the updated gradient and the pilot 1-2; and determines the data 2-2 transmitted on the second resource 2 according to the training result of the model parameter 2 obtained by its own training or the product of the updated gradient and the pilot 1-2.

[0173] The embodiment of the present application also provides a distributed training air interface feedback method, which uses multiple-input multiple-output (MIMO) space division multiplexing to extend the pilot air merging described in the above embodiment, and the pilot and model parameter training results or update gradient superposition transmission air merging mechanism to multi-stream scenarios, so as to effectively combat random channel fading while improving resource utilization. In this method, a first node sends a first message to at least two second nodes among a plurality of second nodes, and correspondingly, at least two second nodes receive the first message. Among them, the first information is used to indicate a plurality of first resources and at least one second resource. There is a corresponding relationship between the plurality of first resources and the transmission layer of the second node. There is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter. Among them, at least one model parameter is a part or all of the parameters that the second node needs to feedback the training results or update the gradient when training the model using local training data. At least two second nodes determine the data transmitted by the transmission layer on the second resource based on the product between the training results or update gradients of the model parameters corresponding to the second resource and the transmission layer and the pilot corresponding to the transmission layer, and then transmit the pilot of the transmission layer corresponding to the first resource, and the determined data transmitted by the transmission layer on the second resource. Correspondingly, the first node receives the pilot of the transmission layer corresponding to each first resource transmitted by at least two second nodes, and the data transmitted by each transmission layer on at least one second resource. It can be seen that this method can feedback the training results or update gradients of multiple model parameters on the same second resource by means of multi-layer transmission (that is, by means of multi-layer transmission, the data corresponding to multiple model parameters can be mapped to the same second resource for transmission), thereby further improving the utilization rate of time-frequency resources.

[0174] In the method, the first node allocates pilot resources (or allocates pilot resources and pilots) to each transmission layer. The pilot resources allocated to any two different transmission layers are different, or the allocated pilot resources are the same but the pilot sequences are orthogonal. Different second nodes send the pilots corresponding to the same transmission layer on the pilot resources corresponding to the same transmission layer, and the data sent on the same transmission layer and the same data resource are obtained based on the product between the model parameters corresponding to the transmission layer and the data resource and the pilot corresponding to the transmission layer.

[0175] Optionally, the second node supports multi-stream transmission. If the number of transmission layers is L, where L is an integer greater than 1, the first node configures at least L pilot resources for each transmission layer (i.e., L first resources, or configures at least L pilot resources and pilots), so that each transmission layer is allocated different pilot resources. In this way, for each of the at least two second nodes, the second node sends the pilot corresponding to the second node in the transmission layer on the pilot resources corresponding to each layer.

[0176] See also Figure 3 , Figure 3 is a flow chart of another distributed training air interface feedback method provided in an embodiment of the present application, such as Figure 3 As shown, the method is explained by taking the example that at least two second nodes among the plurality of second nodes are K second nodes, where K is an integer greater than or equal to 1. Figure 3 The distributed training air interface feedback method and Figure 1 The difference is that Figure 3 In the distributed training air interface feedback method, at least two of the plurality of second nodes support the capability of multi-stream transmission, and both transmit multiple layers of data on the same second resource. Figure 3 As shown, the distributed training air interface feedback method may include but is not limited to the following steps:

[0177] S201. A first node sends first information to K second nodes, and correspondingly, the K second nodes receive the first information, wherein the first information is used to indicate a plurality of first resources and at least one second resource.

[0178] Optionally, the related explanations of the first resource and the second resource can also be found in the previous text and will not be described in detail here.

[0179] There is a correspondence between the multiple first resources and the transmission layer of the second node, such as at least one of a multiple-to-one or a one-to-one correspondence. The number of first resources is greater than or equal to the number of transmission layers, so that each layer can be allocated with different first resources, so that the second node can send a corresponding pilot on the first resource corresponding to each transmission layer. The number of antennas of the K second nodes must be greater than or equal to the number of transmission layers.

[0180] In an optional implementation, one first resource corresponds to one transmission layer. In this way, the problem of excessive pilot resource overhead caused by a large number of second nodes participating in model training can be avoided. In another optional implementation, multiple first resources correspond to one transmission layer. In this way, the second node transmits a pilot on the same multiple first resources for each transmission layer, which can improve the signal-to-noise ratio of the pilot, thereby improving the accuracy of channel estimation.

[0181] Among them, there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter. For the relevant description of at least one model parameter, please refer to the previous text and will not be described in detail here. Optionally, the corresponding relationship is a one-to-one corresponding relationship. For example, taking the number of transmission layers as L (L is an integer greater than 1) and the number of second resources as M (M is an integer greater than or equal to 1) as an example, the one-to-one corresponding relationship can be expressed as: {(second resource i, transmission layer l), model parameter m′}, where i=1, 2,…, M; l=1, 2,…, L. It can be seen that in the embodiment of the present application, the second node can send multiple layers of data on the same second resource, and the second node can send M*L data on M second resources, and can correspondingly feedback M*L model coefficients, thereby greatly improving resource utilization.

[0182] In an optional implementation, the method further includes: the first node sends second information to K second nodes, the second information is used to indicate at least one model parameter, and the correspondence between a combination of at least one second resource and a transmission layer of the second node and at least one model parameter. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or gradient updates are required to be fed back, and the correspondence between these model parameters and the second resource and the transmission layer.

[0183] Optionally, the first node may further indicate the pilot signals p corresponding to the first resources to the K second nodes respectively. Optionally, the pilot signals corresponding to the first resources are different.

[0184] S202. The K second nodes determine, for a second resource in at least one second resource, data corresponding to the second resource and the transmission layer based on the training result of the model parameter corresponding to the combination of the second resource and the transmission layer or the product between the update gradient and the pilot corresponding to the transmission layer;

[0185] Optionally, the data transmitted by the second node on the second resource and transmission layer is the product of the training result or update gradient of the model parameters corresponding to the second resource and transmission layer, the pilot corresponding to the transmission layer and the weight corresponding to the second node. The weight corresponding to the second node can be found in the above text. Figure 1 , Figure 2 The description of the embodiments will not be described in detail here.

[0186] S203.K second nodes respectively transmit the pilot signals of the corresponding transmission layer on each first resource, and transmit the determined corresponding data on the transmission layer of the second resource; correspondingly, the first node receives the pilot signals of the corresponding transmission layer transmitted by the K second nodes on each first resource, and the data transmitted on the transmission layer of the second resource.

[0187] It can be seen that, through the multi-layer transmission method, at least two second nodes can feed back the training results or update gradients of multiple model parameters on the same second resource, thereby further improving the utilization rate of time-frequency resources.

[0188] Optional, Figure 3 The method shown also includes the following steps:

[0189] S204. The first node estimates a merged channel of multiple wireless channels based on the pilot signals received on each first resource, where the multiple wireless channels are wireless channels of corresponding transmission layers between the first node and at least two second nodes; and determines the training results of the parameters corresponding to each transmission layer of the second resource or the merged results of the updated gradients based on the merged channel of the multiple wireless channels corresponding to each transmission layer and the data signals received on the second resources.

[0190] The data signal received on the second resource is a combined data signal formed by superimposing the product of the transfer function of the wireless channel of at least two second nodes and each transmission layer and the data signal transmitted by at least two second nodes on the transmission layer on the second resource. For each transmission layer, the first node estimates a combined channel of multiple wireless channels based on the pilot signals received on one or more first resources corresponding to the transmission layer, and the multiple wireless channels are the wireless channels of the corresponding transmission layers between the first node and each second node.

[0191] In an embodiment of the present application, the pilot signals received by each transmission layer on the corresponding first resource are arranged into a column vector to obtain a pilot column vector, and the number of elements of the pilot column vector is the number of antennas of the first node; the column vectors corresponding to multiple transmission layers are arranged in order of layer numbers into a receiving pilot matrix, and the number of columns of the receiving pilot matrix is ​​equal to the number of first resources. The pseudo-inverse matrix of the receiving pilot matrix is ​​calculated as an equalization matrix. The data signals received on the second resource are arranged into a column vector, and the number of elements of the data column vector is the number of antennas of the first node. The first node calculates the product of the equalization matrix and the data column vector to obtain the equalization result of the second resource, and each element of the equalization result is, in turn, the training results of the parameters corresponding to each transmission layer on the second resource arranged in order of layer numbers or the merged results of the updated gradients.

[0192] The following combination Figure 4 ,right Figure 3 The distributed training air interface feedback method is described by taking an example. Figure 4 is a scenario diagram of a distributed training air interface feedback method provided in an embodiment of the present application, wherein: Figure 4 Take the air interface feedback of one of the terminal devices as an example for explanation. Figure 4 As shown, taking the number of transmission layers as 2, the number of first resources as 2, each transmission layer corresponding to one first resource, and the number of second resources as 6 as an example, Figure 4 The black filled squares in the middle represent the first resources, which transmit the pilots of the corresponding transmission layers respectively; Figure 4 The square filled with medium gray indicates the second resource; the square filled with white indicates that the transmission layer does not transmit signals on the resource indicated by the square. For example, the terminal device transmits the update gradient of the model parameter with an even index on the transmission layer (layer) 1 and each second resource, and transmits the update gradient of the model parameter with an odd index on the transmission layer 2 and each second resource. Figure 4 As shown, the update gradient of the model parameters corresponding to the transmission layer 1 and each second resource can be recorded as: Δw k [0],Δw k [2],Δw k [4],Δw k [6],Δw k [8],Δw k

[10] ; The update gradient of the model parameters corresponding to the transport layer 2 and each second resource is denoted as: Δw k [1],Δw k [3],Δw k [5],Δw k [7],Δw k [9],Δw k

[11] .

[0193] In an optional implementation, each transmission layer corresponds to a first resource, and the number of transmission layers of the second node is L. Then, the pilot signal received by the first node on the first resource corresponding to the transmission layer l can be expressed as shown in the following formula (7):

[0194]

[0195] Among them, p k,l represents the pilot signal transmitted by the second node k on the first resource corresponding to the transmission layer l, h k,l represents the transfer function of the wireless channel between the transmission layer l of the second node k and the first node, Represents the noise on the first resource corresponding to transmission layer 1.

[0196] In this case, multi-layer transmission requires the addition of a weight vector. Each second node and each transmission layer uses a weight vector to transmit pilot and data signals respectively. The number of elements in the weight vector is equal to the number of transmission channels of the second node. For example, j in formula (7) k,l is the equivalent channel corresponding to the transmission layer l of the second node k, which is equivalent to the product of the channel matrix of the second node k and the transmission weight of the transmission layer l of the second node k:

[0197] h k,l =H k t k,l

[0198] Among them, H k is the channel matrix of the second node k, with dimension N×N t , N is the number of antennas of the first node, N t is the number of transmit antennas of the second node k. k,l is the transmission weight of the second node k at the transmission layer l, which is an N t dimensional vector.

[0199] The pilot matrix received by the first node on the first resources corresponding to the L transmission layers can be expressed as shown in the following formula (8):

[0200] Q=[q 1 ,…,q L ]

[0201] Assuming that the data signal received by the first node on a second resource is y, the data signal y can be expressed as:

[0202]

[0203] Among them, the weight a k,l It is determined according to the quantity and quality of the local training data of the second node k, and / or the channel information of the transmission layer l between the second node and the first node. k,l It represents the update gradient of the model parameters corresponding to the second node k on the transmission layer l and the second resource according to the aforementioned mapping relationship.

[0204] Then, according to the data signal received on the second resource, the training result of the parameter corresponding to each transmission layer of the second resource or the merged result of the updated gradient is determined, which can be expressed as shown in the following formula (9):

[0205]

[0206] Among them, z is an L-dimensional vector, and the lth element of z is the training result of the model parameters corresponding to the lth layer (or called transmission layer l) on the second resource or the merged result of the updated gradient.

[0207] In the case where each transmission layer corresponds to one or more first resources, assuming that the number of transmission layers of the second node is L and each transmission layer corresponds to N p first resource, then the pilot signal received by the first node on the first resource i corresponding to the transmission layer l can be expressed as shown in the following formula (10):

[0208]

[0209] Among them, p k,l[i] represents the pilot signal transmitted by the second node k on the first resource i corresponding to the transmission layer l, n (q) [i] is the noise on the first resource i. The pilot signal sent by each distributed node satisfies That is, each node and each layer in N p N sent on resources p The long pilot sequence is a scalar that can be different for each node and each layer. The spreading sequence c is the same for all nodes l [i], spreading sequence c of each layer l [i] Orthogonality must be satisfied, that is, when l = l' When l≠l′ The data transmitted by each distributed node at each layer on the data resource is When receiving, use As the "average" received pilot on multiple pilot resources of layer l, replace the above formula Q = [q 1 ,…,q L ] l Just perform equalization.

[0210] The implementation method of each transmission layer corresponding to one first resource described in formula (7) is a case described in formula (10). For example, for formula (10), let i = l when c l [i]=1,i≠l l [i]=0, then

[0211]

[0212] This is equivalent to the form shown in the above formula (7).

[0213] The following Figure 3 to Figure 4 In the embodiment, the case where the second resource indicated by the first information is one is further illustrated by an example: it is assumed that the first node sends the first information to two second nodes, such as the second node 1 and the second node 2. The first information indicates two first resources, such as the first resource 1 corresponding to the transmission layer 1 and the first resource 2 corresponding to the transmission layer 2, and indicates one second resource, such as the model parameter 1 corresponding to the second resource 1 and the transmission layer 1, and the model parameter 2 corresponding to the second resource 1 and the transmission layer 2.

[0214] In addition, the first node allocates the first resource 1 and the pilot 1-1 transmitted on the transmission layer 1, and the first resource 2 and the pilot 1-2 transmitted on the transmission layer 2 to the second node 1. The first node allocates the first resource 1 and the pilot 2-1 transmitted on the transmission layer 1, and the first resource 2 and the pilot 2-2 transmitted on the transmission layer 2 to the second node 2.

[0215] The second resource 1 determined by the second node 1 and the data 1-1-1 transmitted by the transmission layer 1 are obtained based on the training result of the model parameter 1 or the product between the update gradient and the pilot 1-1. The second resource 1 determined by the second node 1 and the data 1-1-2 transmitted by the transmission layer 2 are obtained based on the training result of the model parameter 2 or the product between the update gradient and the pilot 1-2.

[0216] The second resource 1 determined by the second node 2 and the data 2-1-1 transmitted by the transmission layer 1 are obtained based on the training result of the model parameter 1 or the product between the update gradient and the pilot 2-1. The second resource 1 determined by the second node 2 and the data 2-1-2 transmitted by the transmission layer 2 are obtained based on the training result of the model parameter 2 or the product between the update gradient and the pilot 2-2.

[0217] In this way, the second node 1 transmits pilot 1-1 on the first resource 1 and transmission layer 1, and the second node 2 transmits pilot 2-1 on the first resource 1 and transmission layer 1. The second node 1 transmits pilot 1-2 on the first resource 1 and transmission layer 2, and the second node 2 transmits pilot 2-2 on the first resource 1 and transmission layer 2.

[0218] Second node 1 transmits data 1-1-1 on second resource 1 and transport layer 1, and data 1-1-2 on second resource 1 and transport layer 2. Second node 2 transmits data 2-1-1 on second resource 1 and transport layer 1, and data 2-1-2 on second resource 1 and transport layer 2.

[0219] It can be seen that in this example 1, the first information indicates a second resource, which can realize the feedback of two model parameters, thereby further saving feedback overhead.

[0220] In the case where the first information indicates multiple second resources, the method described in the embodiments of the present application can be adopted for each second resource to transmit multi-stream data on one second resource, thereby saving feedback overhead to a certain extent when the amount of data of the model parameters to be fed back is large and multiple second resources need to be allocated.

[0221] For the case where the first information indicates multiple first resources and each transmission layer corresponds to multiple first resources, the implementation method of each transmission layer corresponding to multiple first resources is the same as the above Figure 1 to Figure 2 In the embodiment, the first information indicates a plurality of first resources in a similar manner, as can be seen from the relevant contents described in the above-mentioned examples 2 and 3, except that the scalar used by the second node to determine the data It is related to the transmission layer, that is, the pilot of the first resource of different transmission layers can use different scalars For details, please refer to the content described in the above formula (10).

[0222] In another embodiment, Figure 1 or Figure 2 The difference between the embodiments is that, in this embodiment, the first information is used to indicate multiple resource blocks, each resource block includes at least one first resource described in the above embodiment and at least one second resource described above; the frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resource of each resource block is less than or equal to the coherence time of the channel. The distributed training air interface feedback method takes into account the frequency selectivity or time selectivity of the channel, divides the channel resources into multiple resource blocks, assumes that the channels in each resource block are the same, and configures the first resource for each resource block, and transmits the pilot and data according to at least one implementation method in the above embodiments, thereby effectively overcoming the channel changes.

[0223] For example, Figure 5 As shown, the resources used for feedback of the training results of module parameters or updating gradients, such as resources composed of frequency domain subcarriers and time domain symbols, are divided into several resource blocks, such as Figure 5 As shown, a resource block includes time domain resources represented by 7 small grids in the time domain and frequency domain resources represented by 4 small grids in the frequency domain. Each resource block includes part or all of the frequency domain subcarriers and part or all of the time domain symbols. The frequency domain bandwidth occupied by each resource block is less than or equal to the channel coherence bandwidth, and the time domain time occupied by each resource block is less than or equal to the channel coherence time. Therefore, it can be approximately considered that the channel is unchanged within a resource block. Consider a scenario of one-layer transmission, where a time-frequency unit is used as a pilot resource in each resource block (such as Figure 5 The black filled squares in each resource block shown in the figure represent resources), and the time-frequency unit is configured with a pilot. When the data on the second resource in each resource block is sent, it is multiplied with the pilot in the resource block before being sent, that is, the data on the second resource in each resource block is obtained by multiplying the training result or update gradient of the model parameter corresponding to the second resource in each resource block with the pilot corresponding to the first resource in the resource block. The pilots between different resource blocks can be the same or different.

[0224] Optionally, the resource used for the training results of the feedback module parameters or the updated gradient is called a resource block, then the resource block described in this article can also be called a resource sub-block, that is, the resource used for the training results of the feedback module parameters or the updated gradient is divided into multiple sub-blocks. For ease of description, the resource block in this article is a resource used for the training results of the feedback module parameters or the updated gradient is divided into several parts.

[0225] In this embodiment, for each resource block, the first node and the second node may perform the first resource and the second resource in the resource block in an optional implementation manner as shown in FIG. Figures 1 to 4For example, for each resource block, the first node receives pilot signals transmitted by at least two second nodes on at least one first resource in the resource block, and data transmitted on at least one second resource; for each resource block, the data transmitted by the second node on the second resource in the resource block is obtained based on the training result or update gradient of the model parameter corresponding to the second resource in the resource block, multiplied by the pilot signal.

[0226] For another example, for each resource block, there is a corresponding relationship between at least one second resource of the resource block and at least one model parameter, that is, there is a corresponding relationship between each second resource in all resource blocks and each model parameter for which feedback of training results or updated gradients is required, such as a one-to-one correspondence. Optionally, the model parameter for which feedback of training results or updated gradients is required may be a model parameter for which feedback is required in one training phase of the model, or a model parameter for which feedback is required in all training phases of the model.

[0227] For another example, the total number of second resources of all resource blocks indicated by the first information is greater than or equal to the number of model parameters. If the number of second resources is greater than the number of model parameters, some second resources may not need to transmit the training results or update gradients of the corresponding model parameters.

[0228] For another example, the second information is used to indicate at least one model parameter, and the correspondence between at least one model parameter and each second resource in each resource block. It can be seen that this implementation can inform at least two second nodes of the model parameters for which training results or updated gradients need to be fed back, and the correspondence between these model parameters and the second resources in each resource block for transmitting the corresponding training results or updated gradients.

[0229] In this embodiment, for each resource block, other optional implementations can be found in Figures 1 to 4 The relevant contents in the described embodiments will not be described in detail here.

[0230] In another embodiment, in combination Figure 3 or Figure 4 The embodiment of space division multiplexing using MIMO is different in that, in this embodiment, the first information is used to indicate multiple resource blocks, each resource block includes at least one first resource mentioned above and at least one second resource mentioned above. Among them, the relevant description of the resource blocks can be found in the relevant description of the above embodiment, which will not be described in detail here. In this embodiment, the total number of second resources in the multiple resource blocks multiplied by the number of transmission layers of the second node is not less than the number of model parameters to be fed back. Optionally, the number of second resources used to transmit training results or update gradients of one or more model parameters is not more than the number of transmission layers of the second node.

[0231] In this embodiment, for each first resource and each second resource in each resource block, the first node and the second node may Figure 3 or Figure 4 The embodiment of space division multiplexing using MIMO describes an optional implementation method of the distributed training air interface feedback method for each transmission layer, which will not be described in detail here. For example, the pilot signal received by each transmission layer on the corresponding first resource is arranged into a column vector to obtain a pilot column vector, and the number of elements of the pilot column vector is the number of antennas of the first node; the column vectors corresponding to multiple transmission layers are arranged in order of layer numbers into a receiving pilot matrix, and the number of columns of the receiving pilot matrix is ​​equal to the number of first resources in the resource block to which it belongs. Calculate the pseudo-inverse matrix of the receiving pilot matrix as the equalization matrix of the resource block. The data signal received on the second resource in the resource block is arranged into a column vector, and the number of elements of the data column vector is the number of antennas of the first node. The first node calculates the product of the equalization matrix and the data column vector to obtain the equalization result of the second resource, and each element of the equalization result is, in turn, the training result of the parameters corresponding to each transmission layer on the second resource arranged in order of layer numbers or the merged result of the updated gradient.

[0232] See also Figure 6 , Figure 6 is a schematic diagram of a model training provided in an embodiment of the present application. Optionally, Figure 6 There are a large number of second nodes within the coverage of the network device that supports MIMO, among which some second nodes support 2-stream transmission and some nodes do not support 2-stream transmission. There are a large number of second nodes, and the model parameters of the distributed artificial intelligence model are large. Assuming that the number of second resources used for training result feedback in each training cycle is limited, such as half of the number of model parameters, the method described in this application can be used to improve resource utilization and training efficiency. Figure 6 As shown, the model training is divided into three stages:

[0233] Initial training phase (e.g. Figure 6 In order to train and converge the model parameters as quickly as possible and obtain a set of available model parameters as soon as possible, the purpose of this training stage is to iterate as quickly as possible to accelerate convergence. In this way, in the feedback stage of the initial training stage, each terminal device can improve the efficiency of time-frequency resource utilization as much as possible (for every doubling of the efficiency of time-frequency resource utilization, the amount of data that can be fed back in the same resource can be doubled, and the number of training iterations that can be completed in the same time can be doubled), but the accuracy requirements for training result feedback can be relaxed. Therefore, this training stage can adopt the method in this application. Figure 3 The embodiments described, such as Figure 4The air interface feedback method executed by the second node adopts a multi-layer transmission method, allowing all second nodes that support uplink multi-layer transmission to participate in training, and second nodes that do not support uplink multi-layer transmission do not participate in the training of this stage. In each training cycle in the initial training stage, each second node feeds back all model parameters.

[0234] For example, Figure 6 As shown, the terminal devices participating in the feedback support the transmission of transport layer 1 (Layer 1) and transport layer 2 (Layer 2), and adopt the following Figure 4 The first resource and the second resource shown send a pilot and corresponding model parameters respectively.

[0235] The training phase after the initial training phase (e.g. Figure 6 In the training phase 2 shown in the figure, after the initial training phase reaches training convergence, in order to further improve the generalization performance of the model, so that the training results can characterize the data distribution of all second nodes, then all second nodes must participate in the training, so the air interface feedback method can be adopted Figure 1 The single-layer transmission method described above is used for feedback, such as Figure 6 As shown in the above Figure 2 In the embodiment described, the terminal device participating in the feedback can send a pilot signal on the black filled resources, and send the corresponding model parameters on the 7 gray filled resources. For example, the terminal device k sends w on the corresponding resources in sequence. k [0],w k [1],w k [2],w k [3],w k [4],w k [5],w k [6]. Optionally, since the spectrum efficiency is relatively reduced by half during the training phase, half of the model parameters can be updated in each training cycle, and one round of training is completed every two training cycles.

[0236] The final training phase (e.g. Figure 6 In the training phase 3 shown in FIG. 1 , the online learning phase can be entered to perform fine online adjustments. The training cycle can be extended. At this time, the overhead of updating model parameters or finer gradients is not the main bottleneck. Each second node can use the traditional method (digital modulation instead of air interface merging) to load the feedback model parameter training results or update gradients on the uplink channel and accurately upload them to the first node. Figure 6 The terminal devices participating in the feedback shown in the figure use the traditional feedback method (such as digital modulation, rather than air interface merging) to provide air interface feedback to the network device. In this way, the training results of a small number of second nodes can be fed back in each training cycle, which, combined with the digital coding modulation technology, can ensure the accuracy of the training results of the model parameters or the updated gradient feedback.

[0237] In the embodiments provided by the present application above, the scheme of the air interface feedback method provided by the embodiment of the present application is introduced from the perspective of each node itself and from the perspective of the interaction between each node. It is understandable that each node, such as the first node, the second node, etc., in order to realize the above functions, includes a hardware structure and / or software unit corresponding to each function. It should be easily appreciated by those skilled in the art that the present application can be implemented in the form of hardware or a combination of hardware and computer software in combination with the units and algorithm steps of each example described in the embodiments disclosed in the present application. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0238] See also Figure 7 , Figure 7 A schematic diagram of the structure of a communication device provided in an embodiment of the present application. Figure 7 The communication device shown includes a transceiver module 701 and a processing module 702 .

[0239] In one design of this embodiment, the communication device is the first node or a related device in the first node:

[0240] Exemplarily, the transceiver module 701 is used to send first information to at least two second nodes among multiple second nodes, where the first information is used to indicate at least one first resource and at least one second resource; receive pilot signals transmitted by at least two second nodes on at least one first resource, and data transmitted on at least one second resource; wherein there is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a part or all of a parameter for which feedback of training results or updated gradients is required when the at least two second nodes train the model using local training data; the data transmitted by at least two second nodes on the second resource is obtained by multiplying the training results or updated gradients of the model parameters corresponding to the second resource by the pilot signal.

[0241] Another exemplary embodiment, the transceiver module 701 is used to send first information to at least two second nodes among multiple second nodes, where the first information is used to indicate multiple first resources and at least one second resource; wherein there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is a part or all of a parameter that the second node trains the model using local training data and needs to feed back the training result or update the gradient; and is also used to receive the pilot of the transmission layer corresponding to each first resource transmitted by at least two second nodes, and the data transmitted by each transmission layer on at least one second resource; wherein the data transmitted by the transmission layer of at least two second nodes on the second resource is obtained based on the product between the training result or update gradient of the model parameters corresponding to the second resource and the transmission layer, and the pilot corresponding to the transmission layer.

[0242] Optionally, when the communication device is the first node or a device related to the first node, it is used to implement Figures 1 to 6 The functions and optional implementations of the first node in the illustrated embodiment.

[0243] In one design, the communication device is the second node or a related device in the second node:

[0244] Exemplarily, the transceiver module 701 is used to receive first information from a first node, where the first information is used to indicate at least one first resource and at least one second resource; wherein there is a corresponding relationship between at least one second resource and at least one model parameter, and at least one model parameter is a parameter in which the second node trains the model using local training data and needs to feed back the training result or update the gradient; the processing module 702 is used to determine, for a second resource in at least one second resource, the data to be transmitted on the second resource based on the training result of the model parameter corresponding to the second resource or the product between the updated gradient and the pilot; the transceiver module 701 is also used to transmit a pilot on at least one first resource, and to transmit determined data on each second resource.

[0245] Another exemplary embodiment is that the transceiver module 701 is used to receive first information from a first node, where the first information is used to indicate multiple first resources and at least one second resource; wherein there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is a part or all of a parameter that the second node trains the model using local training data and needs to feed back the training results or update the gradient; the processing module 702 is used to determine the data transmitted by the transmission layer on the second resource based on the product between the training results or update gradients of the model parameters corresponding to the second resource and the transmission layer and the pilot corresponding to the transmission layer; the transceiver module 701 is also used to transmit the pilot of the transmission layer corresponding to the first resource, and to transmit the determined data on the second resource.

[0246] Optionally, when the communication device is a second node, it is used to implement Figures 1 to 6 The functions and optional implementations of the second device or the sensing device in the illustrated embodiment.

[0247] See also Figure 8 , Figure 8 It is a structural diagram of another communication device provided in an embodiment of the present application. Figure 8 The communication device shown includes at least one processor 801 and a memory 802, and optionally, may further include a transceiver 803. The specific connection medium between the processor 801 and the memory 802 is not limited in the embodiment of the present application. Figure 8 In the figure, the memory 802 and the processor 801 are connected via the bus 804 as an example. The bus 804 is represented by a thick line in the figure. The connection between other components is only for schematic illustration and is not limited to this. The bus 804 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0248] The processor 801 may have a data transceiver function and may communicate with other devices. Figure 8 In the device shown, an independent data transceiver module, such as a transceiver 803, may also be provided for transmitting and receiving data; when the processor 801 communicates with other devices, data transmission may be performed through the transceiver 803.

[0249] In one example, when the first node adopts Figure 8 When the form shown is Figure 8 The processor 801 in the memory 802 may call the computer execution instruction stored in the memory 802 to make the first node execute Figures 1 to 6 A method executed by the first node in any one of the embodiments.

[0250] In one example, when the second node adopts Figure 8 When the form shown is Figure 8 The processor 801 in the memory 802 may call the computer execution instruction stored in the memory 802 to make the second node execute Figures 1 to 6 A method executed by the second node in any of the embodiments.

[0251] The present application also provides a communication system, which may include Figures 1 to 6 The first node and at least two second nodes in the method may be specifically referred to the method embodiment described above.

[0252] The scheme described in the present application can be implemented in various ways. For example, these technologies can be implemented in a combination of hardware, software or hardware. For hardware implementation, the processing module for executing these technologies at a communication device (e.g., a base station, a terminal, a network entity or a chip) can be implemented in one or more general-purpose processors, digital signal processors (DSP), digital signal processing devices, application-specific integrated circuits (ASIC), programmable logic devices, field programmable gate arrays (FPGA), or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, or any combination thereof. The general-purpose processor can be a microprocessor, and optionally, the general-purpose processor can also be any traditional processor, controller, microcontroller or state machine. The processor can also be implemented by a combination of computing devices, such as a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a digital signal processor core, or any other similar configuration.

[0253] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0254] The present application also provides a computer-readable medium on which instructions are stored. When the instructions are executed by a computer, the computer device implements the functions of any of the above method embodiments through a transceiver and a processor.

[0255] The present application also provides a computer program product. When the computer program product is executed by a computer, the computer device implements the functions of any of the above method embodiments through a transceiver and a processor.

[0256] The present application also provides a processing device, which can be a chip device or other products, and the processing device is used to determine Figures 1 to 6The relevant information required in the embodiments described in any of the diagrams. For example, the processing device is used to determine the first resource and the second resource indicated by the first information, the first resource is used for the transmitted pilot, and at least one second resource is used for transmitting data; wherein there is a corresponding relationship between the at least one second resource and at least one model parameter, and at least one model parameter is a part or all of the parameters that need to feedback the training result or update gradient when the at least two second nodes train the model using local training data; the data transmitted by the second node on the second resource is obtained based on the training result or update gradient of the model parameter corresponding to the second resource, and the product of the pilot. For another example, the processing device is used to determine multiple first resources and at least one second resource indicated by the first information; wherein, there is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of at least one second resource and the transmission layer of the second node and at least one model parameter, and at least one model parameter is a part or all of a parameter that the second node trains the model using local training data and needs to feed back the training result or update the gradient; each first resource is used to transmit the pilot of the corresponding transmission layer, and at least one second resource is used to transmit the data of each transmission layer; wherein, the data transmitted by at least two second nodes on the transmission layer of the second resource is obtained based on the product between the training result or update gradient of the model parameter corresponding to the second resource and the transmission layer, and the pilot corresponding to the transmission layer.

[0257] Optionally, the processing device may be a baseband processing module, and the information determined by the processing device may be sent out through the radio frequency processing module, for example, the first information or the second information. Alternatively, the radio frequency processing module receives information sent by other devices, and the processing device determines other information based on the information, for example, determining a first resource for transmitting a pilot and a second resource for transmitting data based on the first information.

[0258] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a high-density digital video disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)).

[0259] It is understandable that some optional features in the embodiments of the present application may be implemented independently in certain scenarios without relying on other features, such as the solution on which they are currently based, to solve corresponding technical problems and achieve corresponding effects, or may be combined with other features according to needs in certain scenarios. Accordingly, the devices provided in the embodiments of the present application may also realize these features or functions accordingly, which will not be elaborated here.

[0260] Those skilled in the art can also understand that the various illustrative logical blocks and steps listed in the embodiments of the present application can be implemented by electronic hardware, computer software, or a combination of the two. Whether such functions are implemented by hardware or software depends on the specific application and the design requirements of the entire system. Those skilled in the art can use various methods to implement the functions for corresponding applications, but such implementation should not be understood as exceeding the scope of protection of the embodiments of the present application.

[0261] It is understood that the "embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, the various embodiments in the entire specification do not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It is understood that in various embodiments of the present application, the size of the sequence number of the above-mentioned processes does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0262] It can be understood that in the present application, "when", "if" and "if" all mean that the device will take corresponding actions under certain objective circumstances, but do not limit the time, nor do they require the device to make judgments when implementing it, nor do they mean that there are other limitations.

[0263] In this application, elements expressed in the singular are intended to mean "one or more" rather than "one and only one", unless otherwise specified. In this application, unless otherwise specified, "at least one" is intended to mean "one or more", and "plurality" is intended to mean "two or more". In the text description of this application, "including at least one of A, B and C" may mean: including A; including B; including C; including A and B; including A and C; including B and C; including A, B and C. In the text description of this application, "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may mean: A exists alone, A and B exist at the same time, and B exists alone. A may be singular or plural, and B may be singular or plural.

[0264] It is to be understood that the various numbers involved in the embodiments of the present application are only for the convenience of description and are not intended to limit the scope of the embodiments of the present application. The size of the sequence number of the above-mentioned processes does not mean the order of execution, and the order of execution of each process should be determined by its function and inherent logic. In addition, the terms "system" and "network" are often used interchangeably in this article.

[0265] The predefined in the present application may be understood as defined, predefined, stored, pre-stored, pre-negotiated, pre-configured, solidified, or pre-burned.

[0266] Those skilled in the art will appreciate that, for the sake of convenience and brevity of description, the specific working processes of the systems, devices and units described above may refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0267] The same or similar parts between the various embodiments in this application can refer to each other. In the various embodiments in this application, and the various implementation methods / implementation methods / implementation methods in each embodiment, if there is no special explanation and logical conflict, the terms and / or descriptions between different embodiments and the various implementation methods / implementation methods / implementation methods in each embodiment are consistent and can be referenced to each other. The technical features in different embodiments and the various implementation methods / implementation methods / implementation methods in each embodiment can be combined to form new embodiments, implementation methods, implementation methods, or implementation methods according to their inherent logical relationships. The above-described implementation methods of this application do not constitute a limitation on the scope of protection of this application.

[0268] The above description is only a specific implementation manner of the present application, but the protection scope of the present application 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 application, which should be included in the protection scope of the present application.

Claims

1. A distributed training air interface feedback method, characterized in that: Applied to the first node, the method comprises: Sending first information to at least two second nodes among the plurality of second nodes, where the first information is used to indicate at least one first resource and at least one second resource; receiving pilots transmitted by the at least two second nodes on the at least one first resource and data transmitted on the at least one second resource; There is a corresponding relationship between the at least one second resource and at least one model parameter, and the at least one model parameter is a part or all of a parameter for which the at least two second nodes respectively train the model using local training data and need to feed back the training result or update the gradient; The data transmitted by the second node on the second resource is obtained by multiplying the training result or update gradient of the model parameter corresponding to the second resource by the pilot.

2. The method according to claim 1, characterized in that The method further comprises: Second information is sent to the at least two second nodes, where the second information is used to indicate the at least one model parameter and a corresponding relationship between the at least one model parameter and the at least one second resource.

3. The method according to claim 1 or 2, characterized in that: There is a one-to-one correspondence between the at least one second resource and the at least one model parameter.

4. The method according to any one of claims 1 to 3, characterized in that The method further comprises: estimating a combined channel of a plurality of wireless channels according to the received pilot signal, the plurality of wireless channels being wireless channels between the first node and the at least two second nodes respectively; According to the combined channel of the multiple wireless channels and the data signal received on the second resource, the training results of the model parameters corresponding to the second resource or the combined results of the updated gradients are determined.

5. The method according to claim 4, characterized in that The data signal received on the second resource is a combined data signal formed by superimposing the product of the transfer function of each wireless channel and the data transmitted by the corresponding second node on the second resource.

6. The method according to claim 4 or 5, characterized in that: The determining, according to the combined channel of the multiple wireless channels and the data signal received on the second resource, the training result of the model parameter corresponding to the second resource or the combined result of the updated gradient includes: Equalization processing is performed using a merged channel of the multiple wireless channels and a data signal received on the second resource to obtain a training result of a model parameter corresponding to the second resource or a merged result of an updated gradient.

7. The method according to claim 6, characterized in that The step of performing equalization processing by using a combined channel of the multiple wireless channels and a data signal received on the second resource to obtain a training result of a model parameter corresponding to the second resource or a combined result of an updated gradient includes: The inner product between the pilot signal received on the at least one first resource and the data signal received on the second resource is divided by the inner product of the pilot signal received on the at least one first resource and itself to obtain the training result of the parameters corresponding to the second resource after equalization processing or the merged result of the updated gradient.

8. The method according to any one of claims 1 to 7, characterized in that The first information is used to indicate a plurality of resource blocks, each resource block including the at least one first resource and the at least one second resource; The frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

9. A distributed training air interface feedback method, characterized in that: Applied to the second node, the method comprises: receiving first information from a first node, wherein the first information is used to indicate at least one first resource and at least one second resource; There is a corresponding relationship between the at least one second resource and at least one model parameter, and the at least one model parameter is a part or all of a parameter for which the second node trains the model using local training data and needs to feed back a training result or update a gradient; For a second resource among the at least one second resource, determining data to be transmitted on the second resource based on a training result of a model parameter corresponding to the second resource or a product between an update gradient and a pilot; The pilot is transmitted on the at least one first resource, and the determined data is transmitted on the second resource.

10. The method according to claim 9, characterized in that The method further comprises: Second information is received from the first node, where the second information is used to indicate the at least one model parameter and a corresponding relationship between the at least one model parameter and the at least one second resource.

11. The method according to claim 9 or 10, characterized in that: There is a one-to-one correspondence between the at least one second resource and the at least one model parameter.

12. The method according to any one of claims 9 to 11, characterized in that: The first information is used to indicate a plurality of resource blocks, each resource block including the at least one first resource and the at least one second resource; The frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

13. A distributed training air interface feedback method, characterized in that: Applied to the first node, the method comprises: Sending first information to at least two second nodes among the plurality of second nodes, where the first information is used to indicate the plurality of first resources and at least one second resource; There is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of the at least one second resource and the transmission layer of the second node and at least one model parameter, and the at least one model parameter is a part or all of the parameters that need to feedback the training results or update the gradient when the second node trains the model using local training data; Receiving a pilot of a transmission layer corresponding to each first resource transmitted by the at least two second nodes, and data transmitted by each transmission layer on the at least one second resource; The data transmitted by the second node on the second resource in the transmission layer is obtained based on the product of the training result or update gradient of the model parameters corresponding to the second resource and the transmission layer and the pilot corresponding to the transmission layer.

14. The method according to claim 13, characterized in that The method further comprises: Second information is sent to the at least two second nodes, where the second information is used to indicate the at least one model parameter, and a corresponding relationship between a combination of the at least one second resource and a transmission layer of the second node, and the at least one model parameter.

15. The method according to claim 13 or 14, characterized in that There is a one-to-one correspondence between the combination of the at least one second resource and the transmission layer of the second node and the at least one model parameter.

16. The method according to any one of claims 13 to 15, characterized in that The method further comprises: estimating a combined channel of a plurality of wireless channels according to the pilot received on each of the first resources, wherein the plurality of wireless channels are wireless channels of corresponding transmission layers between the first node and the at least two second nodes; According to the combined channel of the multiple wireless channels corresponding to each transmission layer and the data signal received on the second resource, the training results of the parameters corresponding to each transmission layer of the second resource or the combined results of the updated gradients are determined.

17. The method according to claim 16, characterized in that The data signal received on the second resource is a combined data signal formed by superimposing the product of the transfer function of the wireless channel of the at least two second nodes and each transmission layer and the data signal transmitted by the at least two second nodes on the second resource in the transmission layer.

18. The method according to claim 16 or 17, characterized in that The determining, according to the combined channel of the multiple wireless channels corresponding to each transmission layer and the data signal received on the second resource, the training result or the combined result of the updated gradient of the model parameter corresponding to each transmission layer of the second resource comprises: Multi-layer equalization processing is performed using the merged channel of multiple wireless channels corresponding to each transmission layer and the data signal received on the second resource to obtain the training results of the parameters corresponding to each transmission layer of the second resource or the merged results of the updated gradients.

19. The method according to any one of claims 13 to 18, characterized in that The first information is used to indicate a plurality of resource blocks, each resource block including the at least one first resource and the at least one second resource; The frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

20. A distributed training air interface feedback method, characterized in that: Applied to the second node, the method comprises: Receiving first information from a first node, where the first information is used to indicate a plurality of first resources and at least one second resource; There is a corresponding relationship between the multiple first resources and the transmission layer of the second node; there is a corresponding relationship between the combination of the at least one second resource and the transmission layer of the second node and at least one model parameter, and the at least one model parameter is a part or all of the parameters that need to feedback the training results or update the gradient when the second node trains the model using local training data; For a second resource among the at least one second resource, determining data transmitted by the transmission layer on the second resource based on a product of the training result or update gradient of the model parameter corresponding to the second resource and the transmission layer and a pilot corresponding to the transmission layer; A pilot of a corresponding transmission layer is transmitted on the first resource, and the transmission layer transmits the determined data on the second resource.

21. The method according to claim 20, characterized in that The method further comprises: Second information is received from the first node, where the second information is used to indicate the at least one model parameter and a corresponding relationship between a combination of the at least one second resource and a transmission layer of the second node and the at least one model parameter.

22. The method according to claim 20 or 21, characterized in that There is a one-to-one correspondence between a combination of the at least one second resource and the transmission layer of the second node and at least one model parameter.

23. The method according to any one of claims 20 to 22, characterized in that The first information is used to indicate a plurality of resource blocks, each resource block including the at least one first resource and the at least one second resource; The frequency domain bandwidth of each resource block is less than or equal to the coherence bandwidth of the channel, and / or the duration of the time domain resources of each resource block is less than or equal to the coherence time of the channel.

24. A communication system, characterized in that: The system comprises: A first node for performing the method of any one of claims 1 to 8; and At least two second nodes for executing the method according to any one of claims 9 to 12.

25. A communication system, characterized in that: The system comprises: A first node for performing the method of any one of claims 13 to 19; and At least two second nodes for executing the method of any one of claims 20 to 23.

26. A communication device, characterized in that: The method comprises one or more functional units, wherein the one or more functional units are used to execute the method according to any one of claims 1 to 8, or to execute the method according to any one of claims 9 to 12, or to execute the method according to any one of claims 13 to 19, or to execute the method according to any one of claims 20 to 23.

27. A communication device, characterized in that: The method comprises a processor and a transceiver, wherein the processor calls a computer program stored in a memory through the transceiver, so that the communication device implements the method according to any one of claims 1 to 8, or implements the method according to any one of claims 9 to 12, or implements the method according to any one of claims 13 to 19, or implements the method according to any one of claims 20 to 23.

28. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the computer device implements the method according to any one of claims 1 to 8, or the method according to any one of claims 9 to 12, or the method according to any one of claims 13 to 19, or the method according to any one of claims 20 to 23 through a transceiver and a processor.

Images