Method and node for training autoencoder in communication network

Abstract

A computer-implemented method in a first node in a communication network for training a first component model of an automatic encoder (AE) machine learning model, the first component model being an encoder or a decoder, and wherein the first component model is a first component model. The first component model is for use in exchanging compressed channel state information (CSI) between a first node, a second node, and a third node in a communication network. The method comprises: i) training (502) the first component model using a first data product obtained from the second node, where the training comprises freezing a first subset of horizontal layers in the first component model during a first reverse pass training phase. The method further comprises ii) initiating (504) a further training of the first component model using a second data product from the third node, where the further training comprises freezing a second subset of horizontal layers in the first component model during a second reverse passing training phase, the first subset of horizontal layers being different from the second subset of horizontal layers.

Description

Technical Field

[0001] The present disclosure relates to methods, nodes, and systems in a communication network. More particularly, but not exclusively, the present disclosure relates to methods and nodes in a communication network for training a first component model of an autoencoder machine learning model. Background Art

[0002] Fifth-generation (5G) mobile wireless communication systems (referred to as New Radio, "NR") use orthogonal frequency division multiplexing (OFDM) with configurable bandwidth and subcarrier spacing to efficiently support a wide variety of use cases and deployment scenarios. NR offers improvements in deployment flexibility, user throughput, latency, and reliability compared to fourth-generation systems (referred to as Long Term Evolution, "LTE"). These throughput performance gains are achieved in part by enhancing support for Multi-User Multiple Input Multiple Output (MU-MIMO) transmission strategies, in which two or more UEs receive data on the same time-frequency resources, i.e., spatially separated transmissions.

[0003] If the network (NW) cannot accurately estimate the complete downlink channel based on the uplink transmission, the active user equipment (UE) needs to report channel information to the NW via the uplink control or data channel. In LTE and NR, the following signaling protocols can be used to perform this feedback: -NW uses N ports to transmit a channel state information reference signal (CSI-RS) over the downlink. - The UE estimates the downlink channel (or an important characteristic thereof) for each of the N ports based on the transmitted CSI-RS. - The UE reports CSI (e.g., Channel Quality Index (CQI), Precoding Matrix Indicator (PMI), Rank Indicator (RI)) to the NW through uplink control and / or data channels. -NW uses UE feedback for downlink user scheduling and MIMO precoding.

[0004] In NR, there are two types of beamforming. Type I selects only one specific beam from a set of beams, while Type 2 selects a set of beams and linearly combines all beams in the same set. Both Type I and Type II reporting are configurable, with the CSI Type II reporting protocol specifically designed to enable MU-MIMO operation based on uplink UE reporting.

[0005] The CSI Type II normal reporting mode is based on the specification of a discrete Fourier transform (DFT) basis function set in the precoder codebook. The UE selects and reports the L DFT vectors from the codebook that best match its channel conditions (just like the classic codebook Precoding Matrix Indicator (PMI) from earlier 3GPP releases). The number of DFT vectors, L, is typically 2 or 4 and is configurable by the network. In addition, the UE reports how the L DFT vectors should be combined with respect to relative amplitude scaling and co-phasing.

[0006] Recently, neural network-based autoencoders (AEs) have shown promising results for compressing downlink MIMO channel estimates for uplink feedback.

[0007] An autoencoder (AE) is a neural network (NN) that can be used to reduce data in a representation space / dimensionality in an unsupervised manner. Figure 1 shows a fully connected (dense) autoencoder. The autoencoder is divided into two parts: an encoder 102 (for compressing input data X) and a decoder 104 (for recovering / reconstructing the input data from the compressed data output by the autoencoder).

[0008] The encoder and decoder are separated by a bottleneck layer 106, which maintains a compressed representation of the input data "X." The compressed representation is labeled "Y" in Figure 1. The variable Y is sometimes called the latent representation of the input X.

[0009] To reduce communication cost, it is desirable that the size of the bottleneck (latent representation) (e.g., the size of Y) is smaller than the size of the input data X. Therefore, the AE encoder compresses the input features X to produce the latent representation Y.

[0010] The decoder part of AE attempts to reverse the compression of the encoder and reconstruct X with minimum error according to some predefined loss function. The decoded or reconstructed X is denoted as ^X.

[0011] In AE-based CSI reporting, the AE encoder is in the UE and the AE decoder is in the NW. UE and NW are typically represented by different vendors (manufacturers), and therefore, AE solutions need to be viewed from a multi-vendor perspective with potential standardization (3GPP) impact (see, for example, 3GPP TSG-RAN WG1, Meeting #109-e, Tdoc R1-2203281, “Evaluation of AI-CSI,” online, May 16-27, 2022; RWS-210448, “Views on studies on AI / ML for PHY,” Huawei-HiSilicon, TSG RAN Rel-18 Workshop, June 28-July 2022).

[0012] To this end, 3GPP 5G networks support uplink physical layer channel decoding (error control decoding) in the following manner: - The UE performs channel coding and the NW performs channel decoding. The channel coder has been specified in 3GPP, which ensures that the UE's behavior is understood by the NW and can be tested. - On the other hand, the channel decoder is left to implementation (vendor proprietary).

[0013] If 3GPP specifies one or more AE-based CSI encoders for use in UEs, the corresponding AE decoder in the NW may be left to be implemented (eg, built in a proprietary manner by training the decoder for the specified AE encoder).

[0014] The perspective on standardization of AE-based CSI reporting can be summarized as follows: Examples where either the AE encoder or the AE decoder or both are normalized: - training within 3GPP (e.g., specifying neural network (NN) architecture, weights, and biases), - training outside 3GPP (e.g., specifying the NN architecture), - specifies the signaling for AE-based CSI reporting / configuration, Example, where the AE encoder and AE decoder are implementation specific (vendor proprietary): - specifies the interface to the AE encoder and AE decoder, -Specifies the signaling for AE-based CSI reporting / configuration. Summary of the Invention

[0015] As noted above in the background section and the references therein, several challenges arise when considering the use of autoencoders for compressing reported CSI I / Q samples across radio channels, particularly when encoding and decoding are performed by different vendors. There are different challenges associated with different levels of model and / or data sharing. Such challenges can be summarized as follows:

[0016] Scenario 1 : The vendor does not want to share its data or models, for example, proprietary data and / or proprietary AE. The challenge in this scenario is how to generate a reciprocal encoder or decoder in the absence of proprietary data and / or proprietary AE.

[0017] A compromise between full sharing is considered to be: Option 1: UE-Vendor(s) and gNB(s) share the encoder Option 2: gNB-Vendor(s) and UE share decoder(s)

[0018] Scenario 2 Sharing common real / synthetic datasets between NW / UE vendors. Challenges associated with this scenario include, but are not limited to, how to train encoders and decoders over the air.

[0019] Scenario 3 : UE manufacturers need to design encoders that are compatible with more than one NW decoder, or NW manufacturers need to design decoders that are compatible with more than one UE encoder. In such scenarios, the UE or NW will need to train more than one encoder or decoder, respectively, and this can impose significant constraints and limitations on the UE or NW, respectively, due to the overhead associated with loading and unloading different models as needed. Due to their size, it is generally not possible to have more than one encoder or decoder occupying the same memory at any given time.

[0020] The disclosure herein addresses some of the aforementioned issues by providing a "global" encoder and / or decoder. For example, in a scenario where there are multiple UEs and a single network node, embodiments herein relate to providing a single decoder that is capable of decoding potential representations of CSI data received from different encoders located on different UEs. In some embodiments, this is provided in a manner that protects the privacy of each UE, for example, the data and encoder used by each of these UEs do not need to be shared. In a scenario where a UE provides CSI data to multiple network nodes, embodiments herein relate to providing a single encoder that is capable of encoding the CSI data in a manner that can be decoded by different decoders on each of these network nodes. In some embodiments, this is provided in a manner that protects the privacy of each network node, for example, the data and decoder used by each of these network nodes do not need to be shared. These and other aspects will be described in more detail below.

[0021] Therefore, in a first aspect, a computer-implemented method for training a first component model of an autoencoder AE machine learning model in a first node in a communication network is provided, the first component model being an encoder or a decoder, and wherein the first component model is for use when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network. The method comprises: i) training the first component model using a first data product obtained from the second node, wherein the training comprises freezing a first subset of horizontal layers in the first component model during a first backward pass training phase; and ii) initiating further training of the first component model using a second data product from the third node, wherein the further training comprises freezing a second subset of horizontal layers in the first component model during a second backward pass training phase, the first subset of horizontal layers being different from the second subset of horizontal layers.

[0022] In this manner, the first component model is trained using data products from the first and second nodes, with different layers in the first component model frozen during training using different data products. This training approach has the technical effect of protecting the learning obtained on each dataset (e.g., learning from the first, second, and third UEs) and preventing "catastrophic forgetting," a phenomenon where previous learning is effectively overwritten by subsequent learning. This creates a balance between the learning obtained from each UE. It further allows the model to learn and retain knowledge from small differences or rare events between the CSI data available at the first, second, and third UEs, resulting in high accuracy.

[0023] In some embodiments, before steps i) and ii), the method further comprises: training a baseline version of the first component model on the first CSI data. The training in step i) may then be performed on the baseline version of the first component model.

[0024] In some embodiments, the method further includes sending the baseline version of the first component model to both the second node and the third node.

[0025] In some embodiments, the method further includes receiving a first data product from the second node, the first data product having been obtained as a result of the second node using CSI data available at the second node to train a second component model to perform a coding operation that is complementary (or opposite / inverse) to a baseline version of the first component model.

[0026] In some embodiments, the method further includes receiving a second data product from the third node, the second data product having been obtained as a result of the third node using CSI data available at the third node to train a third component model to perform encoding operations complementary to the baseline version of the first component model.

[0027] In some embodiments, the first data product is a second component model and / or the second data product is a third component model.

[0028] In some embodiments, step i) comprises using the first component model and the second component model against each other during training.

[0029] In some embodiments, the first data product includes a latent representation of the CSI data available at the second node, the latent representation having been obtained by passing the CSI data through the second component model. In some embodiments, the second data product includes a latent representation of the CSI data available at the third node, the latent representation having been obtained by passing the CSI data through the second component model.

[0030] In some embodiments, in step i), the first component model is trained to: decompress a latent representation available at the second node or the third node if the first component model is a decoder; or compress CSI data available at the third node to produce a latent representation if the first component model is an encoder.

[0031] In some embodiments, if the first component model is a decoder, the second component model is an encoder, and if the first component model is an encoder, the second component model is a decoder. In some embodiments, if the first component model is a decoder, the third component model is an encoder, and if the first component model is an encoder, the third component model is a decoder.

[0032] In some embodiments, the further training is performed by the first node.

[0033] In some embodiments, the first node is a first network node, the second node is a first user equipment (UE), the third node is a second UE, and the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE.

[0034] In some embodiments, the method further includes: receiving first compressed CSI data from the first UE and decompressing the first compressed CSI data using the first component model; and / or receiving second compressed CSI data from the first UE and decompressing the second compressed CSI data using the first component model.

[0035] In some embodiments, the first node is a first user equipment (UE), the second node is a first network node, the third node is a second network node, and the first component model is a universal encoder for use by the first user equipment when encoding CSI information, and the CSI information can be decoded by the first network node or the second network node.

[0036] The method may further include compressing the first CSI data using the first component model to obtain compressed first CSI data, and sending the compressed first CSI data to the first network node and / or the second network node.

[0037] In another set of embodiments, the first data product includes a baseline version of the first component model that has been trained by the second node on CSI data available at the second node. The method may then include: using the CSI data available at the second node, training the second component model to perform encoding operations complementary to the baseline version of the first component model; and using the second component model inversely with the first component model to train the first component model in step i).

[0038] In some embodiments, in the training in step i), a baseline version of the first component model may be used as a starting point for the first component model.

[0039] In some embodiments, the first data product further includes a first version of the first component model, the second node having trained the first version of the component model on the CSI data available at the second node by freezing a third subset of horizontal layers in the first component model during a third backward pass training phase, the third subset of horizontal layers being different from the first subset of horizontal layers and the second subset of horizontal layers. In such embodiments, the method may include using the first version of the first component model as a starting point for the first component model in the training in step i).

[0040] In some embodiments, the training in step i) is performed using CSI data available at the first node.

[0041] In some embodiments, step ii) includes sending one or more of the following to a third node to initiate further training on the third node: i) the first component model output from step i); ii) one or more parameters of the first component model output from step i); iii) one or more instructions causing the third node to perform further training.

[0042] In some embodiments, the second data product is CSI data available at the third node.

[0043] In some embodiments, the first node is a first user equipment (UE), the second node is a first network node, the third node is a second UE, and the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE.

[0044] In some embodiments, the method further comprises sending the first component model to the first network node for use in decoding the compressed CSI data from the first UE and / or the second UE.

[0045] In some embodiments, the method further comprises: compressing the first CSI data; and sending the compressed first CSI data to the first network node.

[0046] In some embodiments, the first node is a first network node, the second node is a first user equipment UE, the third node is a second network node, and the first component model is a universal encoder for use by the first UE when encoding CSI information, and the CSI information can be decoded by the first network node or the second network node.

[0047] In some embodiments, the method further comprises compressing the first CSI data using the first component model, and sending the compressed first CSI data to the first network node and / or the second network node.

[0048] In a second aspect, there is a computer-implemented method for training a first component model of an autoencoder (AE) machine learning model in a second node in a communication network, the first component model being an encoder or a decoder, and wherein the first component model is for use in exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network. The method includes: receiving a baseline version of the first component model from the first node regarding CSI data available at the first node; using the CSI data available at the second node, training the second component model to perform encoding operations complementary to the baseline version of the first component model; and sending a first data product based on the training to the first node for use by the first node in further training the first component model.

[0049] In some embodiments, the first data product comprises one or more of: a second component model; and a latent representation of the CSI data available at the second node, the latent representation having been obtained by passing the CSI data through the second component model.

[0050] In some embodiments, if the first component model is a decoder, the second component model is an encoder, and if the first component model is an encoder, the second component model is a decoder.

[0051] In some embodiments, the first node is a first network node, the second node is a first user equipment (UE), and the third node is a second UE; and wherein the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE.

[0052] In some embodiments, the method further comprises: compressing the new CSI data using the second component model; and sending the compressed new CSI data to the first network node.

[0053] In some embodiments, the first node is a first user equipment (UE), the second node is a first network node, and the third node is a second network node, and the first component model is a universal encoder for use by the first user equipment when encoding CSI information, the CSI information being decodable by the first network node or the second network node. In some embodiments, the method further includes: receiving compressed new CSI data from the first UE; and decompressing the compressed new CSI data from the first UE using the second component model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] For a better understanding and to more clearly show how the embodiments herein may be implemented, reference will now be made, by way of example only, to the accompanying drawings, in which: FIG1 shows a prior art diagram of an autoencoder; Figure 2 illustrates an example network node according to some embodiments herein; Figure 3 illustrates an example user device according to some embodiments herein; Figure 4 illustrates various use cases according to embodiments herein; Figure 5 A method in a first node according to some embodiments herein is shown; Figure 6a Illustrate a method of training a first component model according to some embodiments herein; Figure 6b Icon in use Figure 6a The trained model of Figure 7 shows an example autoencoder architecture according to some embodiments herein, where the horizontal layers are split into three subsets; Figure 8 shows a method 800 in a second node according to some embodiments herein; Figure 9 shows an example way of partitioning data according to some embodiments herein; Figure 10 Show Figure 6a An example signal diagram between the nodes shown in FIG; Figure 11 illustrates an example training and execution process according to some embodiments herein; Figure 12 illustrates an example training and execution process according to some embodiments herein; Figure 13 illustrates an example training and execution process according to some embodiments herein; and Figure 14 An example training and execution process according to some embodiments herein is shown. DETAILED DESCRIPTION

[0055] The disclosure herein relates to a communication network (or telecommunication network). The communication network may include any one or any combination of the following: a wired link (e.g., ASDL); or a wireless link, such as a Global System for Mobile Communications (GSM), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), New Radio (NR), WiFi, Bluetooth, or future wireless technologies. The skilled person will appreciate that these are merely examples and that the communication network may include other types of links. The wireless network may be configured to operate according to a specific standard or other type of predefined rules or procedures. Thus, a specific embodiment of the wireless network may implement a communication standard, such as a Global System for Mobile Communications (GSM), a Universal Mobile Telecommunications System (UMTS), a Long Term Evolution (LTE), and / or other suitable 2G, 3G, 4G, or 5G standard; a wireless local area network (WLAN) standard, such as the IEEE 802.11 standard; and / or any other suitable wireless communication standard, such as the Worldwide Interoperability for Microwave Access (WiMax), Bluetooth, Z-Wave, and / or ZigBee standard.

[0056] Generally speaking, embodiments herein relate to nodes in a communication network, such as network nodes and user equipment (UE). Figure 2 An example network node 200 in a communication network according to some embodiments herein is illustrated. In general, the network node 200 may include any component or network function (e.g., any hardware or software module) in the communication network suitable for performing the functions described herein. For example, the network node may include a device capable of, configured to, arranged to, and / or operable to communicate directly or indirectly with a UE (such as a wireless device) and / or with other network nodes or devices in the communication network to enable and / or provide wireless or wired access to the UE and / or perform other functions (e.g., management) in the communication network. Examples of network nodes include, but are not limited to, access points (APs) (e.g., radio access points), base stations (BSs) (e.g., radio base stations, node Bs, evolved node Bs (eNBs), and NR node Bs (gNBs)). Further examples of nodes include, but are not limited to, core network functions, such as, for example, core network functions in a fifth generation core network (5GC).

[0057] The network node 200 may be configured (e.g., adapted, operable, or programmed) to perform any of the embodiments of the methods 500 or 800 described below. It will be appreciated that the network node 200 may include one or more virtual machines running different software and / or processes. Thus, the network node 200 may include one or more servers, switches, and / or storage devices, and / or may include a cloud computing infrastructure or an infrastructure configured to execute software and / or processes in a distributed manner.

[0058] The network node 200 may include a processor (e.g., processing circuitry or logic) 202. The processor 202 may control the operation of the network node 200 in the manner described herein. The processor 202 may include one or more processors, processing units, multi-core processors, or modules configured or programmed to control the network node 200 in the manner described herein. In a particular implementation, the processor 202 may include multiple software and / or hardware modules, each of which is configured to perform or be used to perform a single or multiple steps of the functionality of the network node 200 as described herein.

[0059] The network node 200 may include a memory 204. In some embodiments, the memory 204 of the network node 200 may be configured to store program code or instructions 206 that may be executed by the processor 202 of the network node 200 to perform the functionality described herein. Alternatively or additionally, the memory 204 of the network node 200 may be configured to store any requests, resources, information, data, signals, or the like described herein. The processor 202 of the network node 200 may be configured to control the memory 204 of the network node 200 to store any requests, resources, information, data, signals, or the like described herein.

[0060] It will be appreciated that network node 200 may include additional or alternative Figure 2 . For example, in some embodiments, network node 200 may include a communication interface. The communication interface may be used to communicate with other network nodes (e.g., other physical or virtual nodes) in the communication network. For example, the communication interface may be configured to transmit and / or receive requests, resources, information, data, signals, or the like to other nodes or network functions. The processor 202 of network node 200 may be configured to control such a communication interface to transmit and / or receive requests, resources, information, data, signals, or the like to other nodes or network functions.

[0061] As noted above, some embodiments herein relate to user equipment (UE) or client devices (e.g., such as stations STA) in a wireless network. In more detail, a UE may include a device capable of, configured to, arranged to, and / or operable to wirelessly communicate with a network node and / or other wireless devices. Unless otherwise indicated, the term UE may be used interchangeably with a wireless device (WD) herein. Wireless communication may involve transmitting and / or receiving wireless signals using electromagnetic waves, radio waves, infrared waves, and / or other types of signals suitable for conveying information over the air. In some embodiments, a UE may be configured to transmit and / or receive information without direct human interaction. For example, a UE may be designed to transmit information to a network on a predetermined schedule, when triggered by an internal or external event, or in response to a request from the network. Examples of UEs include, but are not limited to, smartphones, mobile phones, cellular phones, voice-over-IP (VoIP) phones, wireless local loop phones, desktop computers, personal digital assistants (PDAs), wireless cameras, game consoles or devices, music storage devices, rechargers, wearable terminal devices, wireless endpoints, mobile stations, tablets, laptops, laptop embedded equipment (LEEs), laptop mounted equipment (LMEs), smart devices, wireless customer premises equipment (CPEs), in-vehicle wireless terminal devices, and the like. A UE may support device-to-device (D2D) communication, for example, by implementing 3GPP standards for direct link communication, vehicle-to-vehicle (V2V), vehicle-to-infrastructure (V2I), and vehicle-to-everything (V2X), and in such cases may be referred to as a D2D communication device. As another specific example, in an Internet of Things (IoT) scenario, a UE may represent a machine or other device that performs monitoring and / or measurements and transmits the results of such monitoring and / or measurements to another UE and / or a network node. In such a case, the UE may be a machine-to-machine (M2M) device, which in the 3GPP context may be referred to as an MTC device. As a specific example, a UE may be a UE that implements the 3GPP Narrowband Internet of Things (NB-IoT) standard. Specific examples of such machines or devices are sensors, metering devices such as power meters, industrial machinery, or household or personal appliances (e.g., refrigerators, televisions, etc.), personal wearable devices (e.g., watches, fitness trackers, etc.). In other scenarios, a UE may represent a vehicle or other device that is capable of monitoring and / or reporting its operating conditions or other functions associated with its operation. A UE as described above may represent an endpoint of a wireless connection, in which case the device may be referred to as a wireless terminal. Furthermore, a UE as described above may be mobile, in which case it may also be referred to as a mobile device or mobile terminal.

[0062] Figure 3An example UE 300 is shown in accordance with some embodiments herein. The UE 300 includes a processor 302 and a memory 304. In some embodiments, the memory 304 contains instructions 306 executable by the processor 302 to cause the processor to perform the methods and functions described herein.

[0063] The UE 300 may be configured or operable to perform the methods and functions described herein, such as the method 500 or the method 800. The UE 300 may include a processor (or logic) 302. It will be appreciated that the UE 300 may include one or more virtual machines that run different software and / or processes. Thus, the UE 300 may include one or more servers, switches, and / or storage devices, and / or may include a cloud computing infrastructure or an infrastructure configured to run software and / or processes in a distributed manner.

[0064] The processor 302 may control the operation of the UE 300 in the manner described herein. The processor 302 may include one or more processors, processing units, multi-core processors, or modules configured or programmed to control the UE 300 in the manner described herein. In a specific implementation, the processor 302 may include multiple software and / or hardware modules, each of which is configured to perform or be used to perform a single or multiple steps of the functionality of the UE 300 as described herein.

[0065] The UE 300 may include a memory 304. In some embodiments, the memory 304 of the UE 300 may be configured to store program codes or instructions that may be executed by the processor 302 of the UE 300 to perform the functionality described herein. Alternatively or additionally, the memory 304 of the UE 300 may be configured to store any requests, resources, information, data, signals, or the like described herein. The processor 302 of the UE 300 may be configured to control the memory 304 of the UE 300 to store any requests, resources, information, data, signals, or the like described herein.

[0066] It will be appreciated that the UE 300 may include, in addition to or in place of Figure 3 . For example, UE 300 may include a communication interface. The communication interface may be used to communicate with other UEs and / or nodes in the communication network (e.g., such as other physical or virtual nodes, such as node 200 as described above). For example, the communication interface may be configured to transmit and / or receive requests, resources, information, data, signals, or the like to and from a node or network function. The processor 302 of UE 300 may be configured to control such a communication interface to transmit and / or receive requests, resources, information, data, signals, or the like to and from a node or network function.

[0067] As described above, the embodiments herein relate to the use of autoencoders in communication networks, for example, for use in compressing downlink MIMO channel state information (CSI) estimates for uplink feedback. This can be used, for example, in MU-MIMO systems. As described above in the background section, problems may arise when UE manufacturers (operating UEs such as UE 300) and NW manufacturers (operating network nodes such as network node 200) use different AEs in these types of processes. This may result in each UE and / or each NW node needing to store many encoders or decoders in memory at a given time. For example, currently, a UE that sends CSI to more than one NW node may need to use a different encoder for each NW node. Conversely, a NW node that receives compressed CSI data from more than one UE may need a different decoder for the compressed data of each UE. In addition to being inefficient, the AE is also large, and therefore, this is generally not feasible due to memory limitations.

[0068] As also mentioned above, for privacy reasons, a vendor may be reluctant to share its raw CSI data and / or encoders or decoders trained on the raw data with other vendors. This generally means that it is not feasible to pool data for training a single encoder-decoder pair, which can be trained on a single global dataset from all nodes in a traditional manner.

[0069] In short, to this end, this paper proposes a balanced replay incremental learning (BRIL) mechanism to build a universal AE (encoder / decoder) on both sides of the network ((one or more) network vendors or (one or more) UE vendors) using data processing and baseline NW encoder-decoder training. In a scenario where multiple UEs send compressed CSI to a single NW node, a baseline decoder can be trained at the network, and then this baseline can be sent to all UEs from multiple vendors, which train their own encoders by freezing the baseline NW decoder and training their own encoders to encode data available to the corresponding vendor. These encoders are sent to the NW. The NW then uses a layer- and potential-splitting process to train the universal decoder. In this splitting process, the percentage of potential sample size for each vendor can also be addressed.

[0070] Therefore, the embodiments herein propose to solve some of the aforementioned problems by training a global encoder or decoder. For example, a global encoder may be able to encode CSI data, and the CSI data can be decoded by different decoders located on different network nodes, which have been trained on different training data specific to the respective network nodes. In this way, it is not necessary to pass training data sets and / or decoders around the network, improving the privacy of the respective network nodes. As another example, a global decoder may be able to decode CSI data that has been encoded by different encoders located on different UEs, and the different encoders have been trained on different training data sets (e.g., training data sets specific to their respective UEs). In this way, it is not necessary to pass training data sets and / or encoders around the network, improving the privacy of the respective UEs. Using a universal or global encoder or decoder has the advantage of reducing the cost of maintaining multiple autoencoders for each combination of UE (chipset vendor) and NW (network) vendor.

[0071] Figure 4 Some example scenarios are shown in the figure. Figure 4 4. Various use cases are shown to which the proposed incremental learning method described herein can be applied. Use case 400 can be split into two branches: a single network node multiple UE use case in branch 402; and a multiple network node single UE use case in branch 404. These will be discussed in more detail below.

[0072] Now go to Figure 5 , Figure 5 A computer-implemented method 500 according to some embodiments herein is shown. The method 500 may be performed by a first node in a communication network. The method 500 is used to train a first component model of an autoencoder AE machine learning model. The first component model may be an encoder or a decoder. The first component model is for use when exchanging compressed channel state information (CSI) between a first node, a second node, and a third node in a communication network. In short, in a first step 502, the method 500 includes training the first component model using a first data product obtained from the second node, wherein the training includes freezing a first subset of horizontal layers in the first component model during a first backward pass training phase. In a second step 504, the method includes initiating further training of the first component model using a second data product from the third node, wherein the further training includes freezing a second subset of horizontal layers in the first component model during a second backward pass training phase, the first subset of horizontal layers being different from the second subset of horizontal layers.

[0073] The first component model is an encoder or decoder, such as one half of an autoencoder. According to the disclosure herein, the first component model is trained using data products from different nodes, with different subsets of horizontal layers frozen for training using each of the different data products. As will be described in more detail below, the data products can take various forms, for example, the first data product can be: Option 1) the second component model (e.g., trained at a third node), or Option 2) a latent representation of the CSI data available at the third node, obtained by passing the CSI data through the second component model.

[0074] As will be described in more detail below, different subsets of layers may be unfrozen (or updated) during training associated with the second and third nodes. Thus, in this manner, different subsets of layers are updated for training associated with data products from different nodes. If a particular subset of horizontal layers is unfrozen for training associated with a particular node, those horizontal layers may subsequently be frozen for all other training associated with data products from other nodes. In this manner, training or learning from earlier nodes may be “locked” into the AE. This prevents catastrophic forgetting and enables a first component model (e.g., an encoder or decoder) to simultaneously compress or decompress CSI data from the first, second, and third nodes, even when the decompression or compression, respectively, is performed by another decoder or encoder that is specifically trained on data from the corresponding nodes.

[0075] In more detail, in general, the first node may be a network node such as the aforementioned network node 200, a UE such as the aforementioned UE 300, a client device in a Wi-Fi system, or any other node in a communication network.

[0076] In some embodiments, prior to steps i) and ii), a baseline version of the first component model is trained on the first CSI data. In such embodiments, the training in step i) is performed on the baseline version of the first component model. In other words, the baseline version of the first component model can be used to initialize the first component model. If the first component model is an encoder, the baseline version of the first component model is the encoder, and if the first component model is a decoder, the baseline version of the first component model is the decoder. In other words, the baseline version of the first component model is the same "half" of the autoencoder as the first component model.

[0077] The baseline version of the first component model may have been trained at the first node, or alternatively may have been obtained (e.g., received) from another node. The baseline version of the first component model may be one half of a baseline autoencoder (e.g., including an encoder and a decoder). The baseline autoencoder model may be trained using any CSI data, including, for example, but not limited to, CSI data obtained from a repository such as a cloud repository, CSI data available at the first node, and / or synthesized CSI data.

[0078] Appendix I shows the header and some example CSI data. The examples are shown using CSI filtered using L3 (not L1). Those skilled in the art will appreciate that this is merely an example, and that additional or different headers may be present in the CSI data. For example, the CSI data may further include fields including, but not limited to, SINR, delay, number of users, and / or bandwidth.

[0079] In other examples, a newly initialized version of the first component model (eg, with arbitrary weights and biases) may be used in step 502 to replace the baseline version of the first component model.

[0080] Step 502 may include obtaining a first data product from a second node. The first data product may have been obtained as a result of the second node using CSI data available at the second node to train a second component model to perform encoding operations complementary to a baseline version of the first component model (e.g., opposite or inverse encoding / decoding operations).

[0081] This is Figure 6a In the figure, Figure 6a An embodiment of method 500 is shown. In this embodiment, the first node is a network node 600, the second node is a first UE 608a, and the second node is a second UE 608b. In this embodiment, the first component model is a decoder. The decoder is used to decode encoded CSI from the first UE and the second UE. In this embodiment, the first UE and the second UE are encoding the CSI using different encoders, each of which has been trained on data available to the first UE and the second UE, respectively. Therefore, the first component model is a universal decoder capable of decoding compressed CSI from two different encoders on the first UE and the second UE.

[0082] Figure 6a The method for training a global decoder as described in the previous paragraph is illustrated. In this embodiment, the steps outlined in stage 1 are performed at a first node 600, which is a network node; the steps in stage 2 are performed on a first UE 608a, a second UE 608b, and a third UE 608c. The steps in stage 3 can be performed by the first node 600 (e.g., a network node).

[0083] In Phase 1, step 602, a baseline autoencoder is trained on CSI data, which in this example is a Cloud-Based Dataset (CDS). Figure 6a In the example, training is performed at the network vendor node to produce a baseline (BL) encoder 604 ( Figure 6a BL-Enc-NW) and BL decoder 606 ( Figure 6b BL-Dec-NW in ).

[0084] Those skilled in the art will be familiar with the training of autoencoders, however, a training tutorial titled "Intro to Autoecoders" is available from, for example, TensorFlow. This tutorial is currently available at https: / / www.tensorflow.org / tutorials / generative / autoencoder. A baseline autoencoder can be trained in a known manner using a training dataset of CSI data (e.g., from CDS).

[0085] In summary, Phase-1 involves training the baseline (BL) encoder and decoder at the NW or UE via a cloud-based public CSI dataset (CDS). After training both the BL encoder and decoder parts using CDS, the decoder is sent to each UE for individual training.

[0086] In phase 2, the baseline decoder 606BL-Dec-NW is sent to the first UE 608a, the second UE 608b, and the third UE 608c. It will be appreciated that these are merely examples and that the method can be extended to more than three UEs.

[0087] The first UE 608a then trains a second component model (which in this example is the encoder 610a) to encode the data available at the first UE in a manner that can be decoded by the baseline decoder BL-Dec-NW 606. During the training in phase 2, the baseline decoder 606 is "frozen" in the following sense: during the backpropagation phase, the weights and biases of the baseline decoder 606 are not updated in the training, only the weights and biases of the encoder 610a are updated.

[0088] Freezing prevents the weights of a neural network layer from being modified during the backward pass of training. You gradually "lock" the weights used for each layer to reduce the amount of computation in the backward pass and reduce training time. When a parameter is frozen, its partial derivative is not calculated when backpropagation is performed, and therefore it is "skipped."

[0089] Note that you can unfreeze a certain layer level if you decide to continue training - an example of this is transfer learning: starting from a pretrained model, unfreezing the weights, and then continuing training on a different dataset.

[0090] In step 612a, the first data product is then sent to the first node 600. In this example, the first data product may be: option 1) the second component model, such as encoder 610a, or option 2) a latent representation 614 of the CSI data available at a third node, which has been obtained by passing the CSI data through the second component model, such as compressed CSI data output by the second component model.

[0091] In option 1), the encoder 610a trained at the first UE is sent directly to the first node 600 (e.g., a network node). In option 2), the output of the encoder 610a is sent to the first node 600. As noted above, these outputs are referred to herein as "latents," i.e., latent representations of the CSI data available at the first UE. In other words, the latent representation of the CSI data is a compressed version of the CSI data (e.g., the output of the encoder 610a when the CSI data is provided as input). The latent representation 614 is a pictorial representation of FIG. 1 and FIG. Figure 7 Option 2) may be performed, for example, in a scenario where it is not desirable for the first UE 608a to send the encoder 610a directly to the first node 600 for privacy reasons (or technical reasons such as a desire to reduce signaling overhead).

[0092] The second UE 608b performs the same steps as the first UE 608a in stage 2 and, using the CSI data available at the second UE, trains a third component model (encoder 610b) to compress the CSI data available at the second UE in a manner that can be decoded by the baseline decoder 606. The second UE 608b then sends the output of the training to the first node 600 in 612b. As noted above, once trained, the encoder 610b can be sent to the first node 600, or a potential representation of the CSI data available at the second UE can be sent to the first node 600, according to options 1) and 2) as described above.

[0093] The third UE 608c also performs the same steps as the first UE 608a in stage 2 and, using the CSI data available at the third UE, trains a fourth component model (encoder 610c) to compress the CSI data available at the second UE in a manner that can be decoded by the baseline decoder 606. The third UE 608c then sends the output of the training to the first node 600 in 612c. As noted above, once trained, the encoder 610c can be sent to the first node 600, or a potential representation of the CSI data available at the second UE can be sent to the first node 600, according to options 1) and 2) as described above.

[0094] So, in summary, Phase 2 is about training the respective UE vendor encoder on each UE side (on the data available at the respective UE). In this example, given a frozen BL decoder sent from the first node 600 (NW), each UE 608a, 608b, 608c trains the encoder using its own dataset. After the respective UE encoders are trained on the UE side, they are sent to the first node 600 (which can be a gNB) for the actual BRIL training of the universal decoder.

[0095] Will understand, Figure 6a This is merely an example, and the fourth and / or subsequent UEs may perform Phase 2 in a manner equivalent to that described above.

[0096] Executed by a first node (eg, a network node) Figure 6a In phase 3, the first node 600 performs steps 502 and 504 of the method 500 described above. In step 502, the first node trains a first component model using the first data product obtained from the second node. The training includes freezing a first subset of horizontal layers in the first component model during a first backward pass training phase.

[0097] As used in this article, horizontal layers refer to the routes that data can take from the input layer to the output layer through the AE. Figure 7 In the figure, Figure 7 An autoencoder 700 is shown, comprising an encoder 702 and a decoder 704. In this example, each circle represents a neuron (or graph node) in the autoencoder. Figure 7 , which is illustrated in

[0045] as three graph nodes labeled "1." In other words, a horizontal layer as defined herein is a sequence of graph nodes through a decoder through which data may pass during a forward pass through the network.

[0098] exist Figure 7In the example of , the decoder 704 has been split into three subsets of horizontal layers, a first subset of horizontal layers labeled 1, a second subset of horizontal layers labeled 2, and a third subset of horizontal layers labeled 3. It will be appreciated that Figure 7 The three subsets indicated in are only examples, and an encoder or decoder may include Figure 7 Furthermore, the first subset of layers, the second subset of layers, and the third subset of layers may include layers corresponding to the number of horizontal layers shown in FIG. Figure 7 The different numbers of horizontal layers are shown in FIG.

[0099] Return to Figure 6a In stage-3 of the present invention, in step 502, a first subset of horizontal layers is frozen during a first backward pass training phase, and a first component model is trained using the first data product. The forward pass through the network proceeds as usual, but during the backward pass phase, the first subset of horizontal layers is frozen, or remains unchanged.

[0100] In this sense, 1. given a batch of inputs, the multiplication of data with neurons occurs in a so-called feed-forward (FF) pass…

[0101] 2. Compute the loss function, i.e., the output of this FF pass with respect to the background facts of the labels.

[0102] 3. Based on this loss, backpropagation occurs across all neurons: 3.1 Does not affect frozen neurons 3.2 Update the unfrozen neurons based on the loss gradient with respect to the neuron weights.

[0103] So, in this way, in the case of a frozen parameter, its partial derivative is not calculated during backpropagation and is therefore "skipped". In other words, the loss function is agnostic to frozen / unfrozen layers - it only considers the output. The frozen layers (or neurons) are there, but the backpropagation process does not affect them. Therefore, they contribute, but they are never learned / updated. Therefore, their values ​​(or weights) remain constant and therefore have an impact on the loss.

[0104] like Figure 6a As illustrated in , the first subset of horizontal layers may include layers 616b and 616c. Thus, when training the first model using the first data product, only layer 616a may be updated during the backward pass through the network.

[0105] It is further noted that the input layer 616d and / or the output layer 616e may also be frozen during training.

[0106] In option 1) above, where the first data product is a first component model (e.g., first encoder 610a as trained by first UE 608a), training is performed by freezing the first encoder and a first subset of horizontal layers during the backpropagation phase. In other words, only the unfrozen layers 616a in the first component model (e.g., decoder 606) are updated. Thus, the (remaining) unfrozen layers are trained to decode the CSI data compressed by encoder 610a, which was trained by the first UE on the CSI data available at the first UE.

[0107] In scenario b) above, where the first data product is a latent representation 614 of the CSI data available at the first UE 608a, the latent representation is fed as input to the decoder 606, and the decoder is trained to reconstruct the CSI data for each latent representation. Thus, both the latent representation and the original CSI data may be sent to the first node in step 614a.

[0108] According to step 504 of the above method 500, after training the first component model using the first data product obtained from the first UE 608a, the process is repeated using the second data product from the second UE 608b. In step 504, the second subset of horizontal layers is frozen to the first subset of horizontal layers. For example, with respect to Figure 6a , layers 616a and 616c may be frozen during the second reverse pass phase. Layer 616b may be unfrozen and updated during the second reverse pass. However, it will be appreciated that Figure 6a The layers indicated in are merely examples, and other layers and / or other combinations of layers may be frozen in the second backward pass.

[0109] After training the first component model using the second data product obtained from the second UE 608b, the process may be repeated using data products from other UEs. For example, the training may be repeated for a third UE 608c and / or subsequent UEs. For example, during training of the first component model using the third data product from the third UE, the third subset of horizontal layers may be frozen during a third backward pass. The third subset of layers may be different from the first subset of horizontal layers and / or the second subset of horizontal layers. For example, with respect to Figure 6a , layers 616a and 616b may be frozen during the third reverse pass phase. Layer 616c may be unfrozen and updated during the third reverse pass.

[0110] However, it will be appreciated that Figure 6aThe layers indicated in are merely examples, and other layers and / or other combinations of layers may be frozen in the second reverse pass. As an example, the horizontal layers may be divided between a first subset of layers, a second subset of layers, and / or a third and / or subsequent subsets of layers based on the amount of CSI data available at each corresponding UE. For example, if the first UE 608a has more available CSI data than the second UE 608b, more horizontal layers may be unfrozen in step 502 (e.g., when training using the first data product) compared to step 504 (e.g., when training using the second data product). As another example, the horizontal layers may be divided between a first subset of layers, a second subset of layers, and / or a third and / or subsequent subsets of layers based on the relative proportions of CSI data exchanged between the first UE 608a and the first node and between the second UE 608b and the first node. For example, if the first UE 608a sends more compressed CSI data to the first node than the second UE 608b sends to the first node, more horizontal layers may be unfrozen in step 502 (e.g., when training using the first data product) than in step 504 (e.g., when training using the second data product). It will be appreciated that these are merely examples, however, and that the layers may be divided between the first subset of horizontal layers, the second subset of horizontal layers, and / or the third and / or subsequent subset(s) of horizontal layers in any other manner according to any criteria other than those described herein.

[0111] It will be further appreciated that other layers may be frozen during training. For example, in some embodiments, the input layer and / or the output layer may be frozen (e.g., frozen compared to a baseline version of the first component model).

[0112] In other words, Phase 3 is about incremental learning of the balanced playback buffer for the proposed construction of a universal decoder. For this step, the following components can be considered: 1) individually trained UE encoders (610a, 610b, 610c), 2) potential outputs of each UE + common encoder, 3) multi-layer decoder 608a (each group of layers 616a, 616b, 616c represents a virtual focus for each UE manufacturer, and in addition, input layer 616d and output layer 616e represent common learning). As mentioned above, Phase 3 can be trained via two options:

[0113] Option-1 Consider the following case where all respective encoders and decoders are placed at the NW (e.g., UEs 608a, 608b, 608c send their encoders to the first node) and the input is taken from the Common Channel Data Service (CDS). The UE's encoders 610a, 610b, 610c are frozen (non-trainable parameters) and are not considered in the back-propagation phase of training.

[0114] Option-2 Consider the following scenario where encoders 610a, 610b, 610c are not sent to the first node and reside at UEs 608a, 608b, 608c. In this option, UEs 608a, 608b, 608c send their corresponding potential output (e.g., CSI data compressed / output by one of encoders 610a, 610b, 610c) to the first node (e.g., to the NW), where the decoder resides. Inputs to the UE vendor are used from a UE-specific dataset. The UE's encoder is frozen and not considered during the backpropagation phase of training.

[0115] In both options, the first node (at NW) incrementally trains the decoder by partitioning (e.g., splitting into groups or subsets of layers) its layers into multiple segments (layer partitioning) and common layers (616e), where the multiple segments represent each UE vendor (differently shaded in the figure: 616a for vendor A1, 616b for vendor A2, 616c for vendor A3, etc.). When training for a certain UE vendor, its corresponding segment is unfrozen, while other segments for other UE vendors are frozen. The corresponding input to the decoder will depend on the layer group trained in this step. If the layer group represents vendor A1, most of the potential data output comes from the encoder 610a, while smaller segments of potential data are sampled from other UE / common vendors (this is called data partitioning and potential partitioning). After training for each potential segment for all UE vendors, the decoder is expected to converge to a stable loss value.

[0116] In addition to segmenting the layers of the first component model, the data used to train it can also be segmented.

[0117] As an example, in one embodiment, method 500 includes splitting the layers of the first component model and the CSI data (used to train it) into:

[0118] - NN segments (e.g. comprising subsets of layers) per vendor (or group of vendors, depending on the clustering algorithm proposed below that uses similarities between their latent spaces). In other words, layers are assigned to different subsets of layers according to the vendor.

[0119] - At each iteration, split the CSI data or latent representations available for training (if option 2) using the above split to include the main chunks from the current UE vendor (or vendor group, based on clustering based on similarity of their latent samples), while including smaller chunks of data / latent representations from other vendors (and public) into the current training for the current vendor. This can be called a "balanced replay buffer" because the data for each subset of training levels is selected to reflect the specific representation / distribution of the data. This is in Figure 9In the figure, Figure 9 Four options are shown for how to put together a dataset comprising CSI data segments from different vendors (e.g., different UEs) for training purposes. Each segment represents data from a single vendor, and the segments can be of similar size or have different segment lengths. As an example, when training is performed in step 502, the data used for the training step can be taken equally from UEs 608a, 608b, 608c, or the dataset can be assembled using different proportions of data from different individual sources. For example, the majority of the training data can be taken from the respective UE, while a smaller proportion of data can be taken from other UEs. Alternatively, the data can be split proportionally to the number of CSI reports sent by each respective UE.

[0120] In this example, method 500 may be written in pseudocode as follows:

[0121] In the pseudo-code above, step 502 of method 500 corresponds to, for example, a first iteration of the loop, and step 504 corresponds to a second iteration of the loop.

[0122] Turning now to various methods for segmenting the data, in one example, a large percentage of the latent representations is segmented from the encoder portion belonging to a specific vendor or BL encoder. Here, we identify two methods that can be used to identify the percentage value of the latent representation segmentation for each vendor in each iteration. Assume that there is a loop over all vendors, and in each iteration a single vendor is targeted, which in this example is called the target vendor. Method for prioritizing by target manufacturers:

[0123] In this approach, the percentage of each potential segment for each manufacturer is calculated using two components:

[0124] The normalized difference between the target manufacturer (tv) potential sample size and the minimum sampled manufacturer potential sample size. We call the normalized output of this difference Lat diff (tv)∈[0, 1]. A value of zero means that the target firm’s potential size is the smallest sampled firm.

[0125] The aggregate or average difference in the (normalized) Kullback-Leibler (KL) divergence between the target vendor (tv) latent space and the latent spaces of all other vendors, which we refer to as Lat KL (tv)∈[0,1].

[0126] The percentage of the target manufacturer's potential segment size (referred to as: % STV ) will be the following function:

[0127] Where v is the number of manufacturers, w diff and w KL is a weighting scale that gives more or less value to each component (normalized difference in latent space size and mean KL divergence) separately. norm is the normalized value of STV. norm This can be found by inserting the minimum or maximum parameter into the original formula, as shown below: Uniform method:

[0128] This is simply to have the same percentage for all manufacturers (target and others), so:

[0129] Thus, in the manner described above, method 400 can be used to obtain a global first component model (encoder or decoder). Jón Rúnar Baldvinsson, in his paper titled "Rare Event Learning in URLLC Wireless Networking Environment Using GANs" (2021), describes freezing different layers during training on different datasets to allow the algorithm to update different parts of a generative adversarial network (GAN). The inventors have recognized that the techniques described in this paper can be equally applied to training autoencoders for compressing and decompressing CSI data.

[0130] To address the privacy concerns of malicious vendors sending corrupted data or corrupted encoders to the NW, the following steps can be performed: - Avoid UE-localized training of the decoder, so the impact of a malicious UE will be only on the encoder or input data - only use input data from CDS (i.e., frequently updated and tested datasets) (for encoders frozen by the manufacturer). In this case, the only possibility of causing harm from a malicious UE is to send a corrupted encoder, which can be done via Ways to avoid: - Train the decoder at the gNB and freeze the malicious encoder, -Towards% STV_tv Adding terms to reduce the potential samples generated by malicious encoders, it becomes: in, Among them, Ver iTr(tv) is a factor that measures the degree of distrust between the gNB and this particular tv vendor (the current potential sample is generated from its encoder). The more the gNB trusts the vendor, the lower this factor is, and the less the gNB trusts the vendor, the lower the Ver iTr (tv) will be higher.

[0131] We point out that the above process is applied when the trust level between gNB and UE vendors is established. iTr (tv)), we calculate the ratio (STV Norm ), the ratio is used to filter out a certain percentage of UEs with a limited degree of trust. Therefore, in short, identifying the degree of trust between the NW and the UE manufacturer is not within the scope of the present invention, but it is assumed that this degree of trust has been obtained.

[0132] Training the decoder using method 500 preserves the learning obtained on each dataset (e.g., learning from the first UE, the second UE, and the third UE) and prevents "catastrophic forgetting," a phenomenon whereby previous learning is effectively overwritten by subsequent learning. This creates a balance between the learning obtained from each UE. It further allows the model to learn and retain knowledge from small differences or rare events between the CSI data available at the first UE, the second UE, and the third UE. It has been further appreciated that this method can be applied to CSI data because the distribution of CSI datasets is sufficiently similar across different UEs to allow convergence. In summary, the freezing process described herein enables training a single "global" decoder that is fine-tuned to accurately decode the compressed CSI output from three different encoders that were trained to compress the CSI based on the three different datasets.

[0133] Figure 6b The trained model is shown in use. After training, the decoder 608c (eg, the first component) output from the training process can be used to decode the compressed CSI from the encoder 610a, 610b, or 610c.

[0134] The first UE may perform a reciprocal process with the first node. For example, Figure 8A computer-implemented method for training a first component model of an autoencoder AE machine learning model in a second node in a communication network is shown, wherein the first component model is an encoder or a decoder, and wherein the first component model is used when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network. In a first step 802, the method includes receiving a baseline version of the first component model about the CSI data available at the first node from the first node. In a second step 804, the method includes using the CSI data available at the second node to train the second component model to perform a coding operation that is complementary (e.g., inverse) to the baseline version of the first component model. In a third step, the method includes sending a first data product based on the training to the first node for use by the first node in further training the first component model.

[0135] As mentioned above Figure 5 As described with method 500, the first data product may include one or more of the following: a second component model; and a latent representation of the CSI data available at the second node, the latent representation having been obtained by passing the CSI data through the second component model. Thus, in the embodiment described with respect to FIG. 6 , the method may include the second node sending an encoder trained on data available at the second node as opposed to a baseline decoder. Alternatively or additionally, the second node may send a compressed representation of the data that has been passed through such an encoder.

[0136] More generally, if the first component model is a decoder, the second component model will be an encoder, and if the first component model is an encoder, the second component model will be a decoder. In other words, the second component model will be the opposite half (or inverse half) of the full encoder to the first component model.

[0137] In some embodiments, as shown in the embodiment of FIG6 , the first node is a first network node, the second node is a first user equipment UE, and the third node is a second UE. As noted above, in this example, the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE. This corresponds to Figure 4 Use Case-2 (block 412) in , and is also summarized in Figure 12 middle.

[0138] In use, the second component model may compress the new CSI data using the second component model and send the compressed new CSI data to the first network node. When the UE vendor sends its potential channel to the BRIL generic decoder at the gNB for decoding, this execution phase of BRIL ( Figure 6b Scenario-2) in the above example occurs.

[0139] It will be appreciated that the second node and subsequent nodes may all execute method 800 and send data products to the first node for use by the first node in training the first component model according to method 500 .

[0140] Now go to Figure 10 , Figure 10 Shown in Figure 6a A signal diagram between different nodes in an embodiment of the present invention describes a BRIL for developing a universal decoder for multiple chipset vendors and single network vendors. Figure 10 The signals in are as follows: 1002: Manufacturer request to join the network for the first time or to rejoin the network: UEs 608a, 608b from different manufacturers (1 to N) join the network. 1004: Configure UE with CSI-MeasConfig: The network node configures all UEs with CSI-MeasConfig (via RRCConfiguration and / or RRCReconfiguration). 1006: Configure UE with CSI-ReportConfig: The network node configures all UEs with CSI-ReportConfig (via RRCConfiguration and / or RRCReconfiguration). 1008: Request UE vendor to send AI capabilities, including the ability to support AI-related procedures: The network requests AI capabilities (general AI-related and specific to CSI compression) from the UE, such as processing, ML model and data quality capabilities. 1010: Report AI and data information, such as (processing, CPU, energy, data bias, drift, etc.): UE responds to the network with its computing capabilities and data quality Some operations are related to network training and data operations for baseline CSI and / or models. Note that any of the following messages can be sent via RRC reconfiguration messages or MAC-CE messages. In the sequence diagrams, brackets <> are used to indicate that this signal / message or operation is optional. 1012: Cluster UEs into groups based on the AI / data capability set sent by the UE vendor: The network runs a clustering algorithm to find UE vendors that are suitable for training together in a common AE. For example, some requirements for UE vendors to be in the same cluster are the distance between their CSI distributions. 1014: Notify UE IDs in the cluster: The network notifies all vendor IDs in the same incremental learning cluster to UE nodes belonging to different vendors. Operations related to network training and data processing for the CSI AE baseline 1016: Request CSI or synthesis parameters and / or trained baseline model: <The network requests CSI or synthesized CSI parameters and / or a trained baseline from the UE> 1018: Respond to the network with the requested CSI data: <The UE vendor responds to the network with the requested CSI data> 1020: Store all data from the UE or present in the CDS (Common Cloud for CSI datasets): Store all data sets from the UE or present in the CDS (Common Cloud for CSI datasets). Operations related to generating a baseline CSI model called BL-AE-NW 1022: Pseudocode for the BP to train BL-AE-NW: The network runs backpropagation to train BL-AE-NW (pseudocode for backpropagation) 1024: Send the generated baseline model (BL-Dec-NW): The network sends the generated baseline model (BL-Dec-NW) to all UE vendors Operations related to the incremental learning preparation process 1026 / 1028: UE (belonging to different vendors) 1. Receive: 1.a <Receive vendor group or clustering> 1.b. <Receive / Send the trained baseline BL-Dec-NW> 2. <Discover neighboring UE vendors>, 3. <Project / align the shared baseline data or model into the existing model or local data>. Generate a baseline UE CSI model called BL-Enc-UE 1030 / 1032: Train BL-Enc-UE using the agreed data (specific UE vendor or CDS) and BL-Dec-NW (frozen): The UE vendor uses the agreed data (from a specific UE vendor or CDS) and BL-Dec-NW (frozen) to train BL-Enc-UE. 1034 / 1036: Pseudocode for the BP to train BL-Enc-UE: The UE runs BP to train BL-Enc-UE (pseudocode for backpropagation) Operations related to the incremental learning of the general decoder · Initialize the learning rate, batch size, epochs, etc Option-1: If the target scenario is BRIL general decoder training at the NW node · 1038: Announce BL-Enc-UE: The UE vendor sends a message to the network including BL-Enc-UE. 1040: Looping on UE vendor 'u' Network initialization: o BRIL_Enc = BL-Enc of UE vendor 'u' (and freeze it) Freeze the first and last layers of BL_Dec_'u'_NW οBRIL_Dec_'u'=BL_Dec_'u'_NW ο Layer segmentation stage on BRIL_Dec_'u' o BRIL_Dec=BRIL_Dec_'u' while freezing all other parts not selected via the layer segmentation stage. o A data segmentation process (i.e., balanced replay buffer) is applied to treat the current vendor latent representation as the majority of the training data, plus small portions of other vendors and the common latent space. Loop on each round o loop over each batch 1042: The network runs a feed-forward pass to calculate y = f(BRIL_Enc + BRIL_Dec) = f(BRIL_AE) 1044: Calculate the reconstruction loss of AE_'u' |y'-y|: The network calculates the reconstruction loss of AE_'u' |y'-y| 1046: Compute backpropagation across all BRIL_Enc and BRIL_Dec (except the last layer of BL_Dec and all non-'u' vendor parts): The network computes backpropagation across all BRIL_Enc and BRIL_Dec (except the last layer of BL_Dec and all non-'u' vendor parts). 1048: Update unfrozen weights: The network updates all weights (except frozen weights) Option-2: If the case is incremental learning of a general decoder over the air Loop on each UE vendor 'u' 1050: Network initialization: a.Out_BRIL_Enc = Transmission of the output of BL-Enc of manufacturer 'u' b. Freeze the first and last layers of BL_Dec_'u'_NW c. Implicit freezing of BL-Enc-'u', since there is no transmission of gradients back to the UE d.BRIL_Dec_'u'=BL_Dec_'u'_NW e. Layer segmentation stage on BRIL_Dec_'u' f. A data segmentation process (i.e., balanced replay buffer) is applied to treat the current vendor latent representation as the majority of the training data, plus small portions of other vendors and the common latent space. g. BRIL_Dec = BRIL_Dec_'u' while freezing all other parts not selected via the layer segmentation stage. Loop on each round o loop over each batch 1052 and 1054: UE vendor runs a feed-forward pass to calculate y_latent = f(BRIL_Enc) 1056: UE manufacturer sends a message with y_latent to the network 1058: The network runs a feed-forward pass to compute y = f(BRIL_Enc(y_enc)) 1060: The network calculates the reconstruction loss of AE_'u' |y'-y| 1062: The network computes the backpropagation across all BRIL_Dec (except the last layer of BL_Dec and all non-'u' vendor parts). 1064: No backpropagation across BRIL_Enc, i.e. no message passing of weight gradients to UE

[0141] Figure 6a The training process is described for a general decoder, however, an equivalent process can also be applied to the training of a general encoder.

[0142] Such an embodiment is Figure 14 is illustrated in , and corresponds to Figure 4 Use Case-4 in (Block 422). Figure 14 In the embodiment, the first node is a first user equipment UE 1400, the second node is a first network node 1402a, the third node is a second network node 1402b, and the first component model 1404 is a universal encoder for the first user equipment 1400 to use when encoding CSI information, and the CSI information can be decoded by the first network node 1402a or the second network node 1402b.

[0143] In use ( Figure 14 ), the UE 1400 uses the trained first component model (generic encoder) 1400 to encode CSI, which can be decoded by each of the corresponding decoders on the network nodes 1402a, 1402b.

[0144] Turning now to other embodiments, the above embodiments describe the proposed procedure (standardized approach) when UE manufacturers agree to exchange AI-related signaling. However, the present invention can exist independently as a proprietary solution, although this may compromise the quality of the algorithm execution, but it is still possible. In order to achieve the independent behavior of the proposed algorithm, the following changes can be made to the above procedure, as shown below (reflected in Figure 2 middle): 1. In Phase-1, we allow the UE to train its own decoder instead of sending the BL decoder from the NW to the UE, so signaling is now required. 2. In stage 3, in this embodiment, the CSI latent space can be sent by each vendor to train the NW BRIL decoder, rather than requiring the UE to send their trained encoder to the NW and use it to locally generate the latent space for each vendor. Although this step may involve sending the CSI latent representation, no additional signaling is required.

[0145] Therefore, in this embodiment, Figure 6a , transferring the first component model between different UE nodes for training the encoder of the corresponding UE.

[0146] In some embodiments of method 500, the first data product received from the second node includes a baseline version of the first component model, which has been trained by the second node on CSI data available at the second node. In other words, the second node performs step 502 and forwards the resulting partially trained model to the first node.

[0147] The first node then trains the second component model to perform an encoding operation that is inverse (e.g., opposite or complementary) to the baseline version of the first component model using the CSI data available at the second node. In this sense, if the first component model is an encoder, the inverse encoding operation or the complementary encoding operation is, for example, a decoding operation, or if the first component model is a decoder, the inverse encoding operation or the complementary encoding operation is an encoding operation. Figure 6a In the embodiment of , this results in encoders 610a, 610b, and 610c. The method then includes using the second component model inversely with the first component model to train the first component model in step i). The training in step i) may be performed on a baseline version of the first model (if this is the first iteration of training). In other words, method 500 may further include using the baseline version of the first component model as a starting point for the training of the first component model in step i).

[0148] Thus, in summary, after the second component model has been trained against the baseline component model, the baseline component model is then retrained by the first node itself on the CSI data at the first node, freezing the first subset of layers as described above. This training may be performed using the CSI data available at the first node.

[0149] The first node then initiates further training in step ii). In this embodiment, step ii) includes sending one or more of the following to the third node to initiate further training on the third node: i) the first component model output from step i); ii) one or more parameters of the first component model output from step i); or iii) one or more instructions causing the third node to perform further training.

[0150] The third node then repeats the method 500 as described above. From the perspective of the third node executing the method 500, the first node sends a data product comprising a first version of a first component model, the first node having trained the first version of the component model on the CSI data available at the second node by freezing a first subset of horizontal layers in the first component model during a first backward pass training phase, and the method then includes using the first version of the first component model as a starting point for training the first component model in step i).

[0151] In one example, if Figure 11 , the first node is a first user equipment UE 1102a, the second node is a first network node 1100, and the third node is a second UE 1102b, and wherein the first component model is a general decoder 1104 for use by the first network node 1000 when decoding compressed CSI information from the first UE 1102a or the second UE 1102b (and any other UE 1102N). In this embodiment, the second component model is an encoder 1106a, and during the training phase, this is trained against a baseline version of the first component model 1108.

[0152] In use ( Figure 11 The method 500 may further include: sending a first component model to the first network node for use in decoding the compressed CSI data from the first UE and / or the second UE. The method may further include compressing the first CSI data and sending it to the first network node (shown as Figure 11 The first CSI data compressed by the arrow 1110 in FIG. 1 is used for decompression by the first component model. Figure 11 Corresponding to Use Case-1( Figure 4 406 in FIG.

[0153] Figure 12 Show Figure 4This is an overview of Use Case-2 (block 412) in the previous section. Figure 6a and Figure 6b described and in Figure 12 Provided for reference and comparison Figure 11 、 Figure 13 and Figure 14 .

[0154] In another example, Figure 13 , the first node is a first network node 1302a, the second node is a first user equipment UE 1300, the third node is a second network node 1302b, and the first component model is a general encoder 1304 for use by the first UE 1300 in encoding CSI information, which can be decoded by the first network node 1302a or the second network node 1302b (and any other network node 1302N). In use ( Figure 13 The method may thus further comprise compressing the first CSI data using the first component model and sending it to the first network node and / or the second network node (shown as Figure 13 The first CSI data compressed by the arrow 1310 in the figure is decompressed by a corresponding decoder trained on the data at each of the first node, the second node and / or the subsequent node, for example. Figure 13 Corresponding to Use Case-3( Figure 4 In another embodiment, a computer program product is provided that includes a computer-readable medium having computer-readable code embodied therein, the computer-readable code being configured to, when executed by a suitable computer or processor, cause the computer or processor to perform one or more of the methods described herein.

[0155] It will thus be appreciated that the present disclosure also applies to computer programs, in particular computer programs on or in a carrier, suitable for putting the embodiments into practice. The program may be in the form of source code, object code, a code intermediate source and object code such as in partially compiled form, or in any other form suitable for use in the implementation of the method according to the embodiments described herein.

[0156] It will also be appreciated that such a program can have many different architectural designs. For example, the program code that implements the functionality of the method or system can be subdivided into one or more subroutines. Many different ways of distributing functionality among these subroutines will be apparent to those skilled in the art. The subroutines can be stored together in an executable file to form a self-contained program. Such an executable file can include computer-executable instructions, for example, processor instructions and / or interpreter instructions (e.g., Java interpreter instructions). Alternatively, one or more or all of the subroutines can be stored in at least one external library file and linked statically or dynamically (e.g., at runtime) to the main program. The main program contains at least one call to at least one of the subroutines. The subroutines can also include function calls between each other.

[0157] The carrier of a computer program can be any entity or device capable of carrying the program. For example, the carrier may include a data storage device, such as a ROM (for example, a CD ROM or a semiconductor ROM) or a magnetic recording medium (for example, a hard disk). In addition, the carrier may be a transmissible carrier, such as an electrical signal or an optical signal, which may be communicated via an electrical cable or an optical cable or by radio or other means. When the program is implemented in such a signal, the carrier may be composed of such a cable or other device or means. Alternatively, the carrier may be an integrated circuit in which a program is embedded, which is suitable for executing the related method or for the execution of the related method.

[0158] Those skilled in the art, in practicing the claimed invention, can understand and implement changes to the disclosed embodiments by studying the drawings, the present disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "a" does not exclude a plurality. A single processor or other unit may perform the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. The computer program may be stored / distributed on a suitable medium (such as an optical storage medium or solid-state medium supplied together with other hardware or as part of other hardware), but the computer program may also be distributed in other forms (such as via the Internet or other wired or wireless telecommunications systems). Any reference signs in the claims should not be interpreted as limiting the scope. Appendix 1 Table 1 Table 1 (continued) Table 1 (continued) Table 1 (continued)

Claims

1. A computer-implemented method for training a first component model of an autoencoder (AE) machine learning model in a first node in a communication network, the first component model being an encoder or a decoder, and wherein: The first component model is for use when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network, the method comprising: i) training (502) the first component model using a first data product obtained from the second node, wherein the training comprises freezing a first subset of horizontal layers in the first component model during a first backward pass training phase; and ii) initiating (504) further training of the first component model using the second data product from the third node, wherein the further training comprises freezing a second subset of horizontal layers in the first component model during a second backward pass training phase, the first subset of horizontal layers being different from the second subset of horizontal layers.

2. The method according to claim 1, wherein Before steps i) and ii), the method further comprises: A baseline version of the first component model is trained on first CSI data; and wherein: The training in step i) is performed on the baseline version of the first component model.

3. The method of claim 2, further comprising: The baseline version of the first component model is sent to both the second node and the third node.

4. The method of claim 3, further comprising: The first data product is received from the second node, the first data product having been obtained as a result of the second node training a second component model to perform encoding operations complementary to the baseline version of the first component model using CSI data available at the second node.

5. The method according to claim 3 or 4, further comprising: The second data product is received from the third node, the second data product having been obtained as a result of the third node training a third component model to perform encoding operations complementary to the baseline version of the first component model using CSI data available at the third node.

6. The method according to claim 4 or 5, wherein: The first data product is the second component model and / or the second data product is the third component model.

7. The method of claim 4, 5 or 6, wherein: Step i) comprises using said first component model and said second component model against each other during said training.

8. The method according to claim 4 or 5, wherein: the first data product comprising a latent representation of the CSI data available at the second node, the latent representation having been obtained by passing the CSI data through the second component model; and / or The second data product comprises a latent representation of the CSI data available at the third node, the latent representation having been obtained by passing the CSI data through the second component model.

9. The method of claim 8, wherein: In step i), the first component model is trained to: If the first component model is a decoder, decompressing the latent representation available at the second node or the third node; or If the first component model is an encoder, the CSI data available at the third node is compressed to produce the latent representation.

10. The method according to any one of claims 4 to 9, wherein If the first component model is a decoder, then the second component model is an encoder, and if the first component model is an encoder, then the second component model is a decoder; and wherein If the first component model is a decoder, the third component model is an encoder, and if the first component model is an encoder, the third component model is a decoder.

11. The method according to any one of claims 1 to 10, wherein: The further training is performed by the first node.

12. The method according to any one of claims 1 to 11, wherein The first node is a first network node, the second node is a first user equipment UE, and the third node is a second UE; and wherein the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE.

13. The method of claim 12, further comprising: receiving first compressed CSI data from the first UE, and decompressing the first compressed CSI data using the first component model; and / or Second compressed CSI data is received from the first UE, and the second compressed CSI data is decompressed using the first component model.

14. The method according to any one of claims 1 to 11, wherein: The first node is a first user equipment (UE), the second node is a first network node, and the third node is a second network node, and the first component model is a universal encoder for use by the first user equipment when encoding CSI information, and the CSI information can be decoded by the first network node or the second network node.

15. The method of claim 14, further comprising: compressing first CSI data using the first component model to obtain compressed first CSI data; as well as The compressed first CSI data is sent to the first network node and / or the second network node.

16. The method of claim 1, wherein: The first data product comprises a baseline version of the first component model that has been trained by the second node on CSI data available at the second node; And wherein the method further comprises: training a second component model to perform encoding operations complementary to the baseline version of the first component model using CSI data available at the second node; and The second component model is used in conjunction with the first component model in order to train the first component model in step i).

17. The method of claim 16, further comprising: In the training in step i), the baseline version of the first component model is used as a starting point for the first component model.

18. The method of claim 16, wherein: the first data product further comprising a first version of the first component model, the second node having trained the first version of the component model on the CSI data available at the second node by freezing a third subset of horizontal layers in the first component model during a third backward pass training phase, the third subset of horizontal layers being different from the first subset of horizontal layers and the second subset of horizontal layers; as well as In the training in step i), the first version of the first component model is used as a starting point for the first component model.

19. The method of claim 16, 17 or 18, wherein Said training in step i) is performed using CSI data available at said first node.

20. The method of any one of claims 16 to 19, wherein Step ii) comprises sending one or more of the following to the third node to initiate the further training on the third node: i) the first component model output from step i); ii) one or more parameters of the first component model output from step i); iii) causing the third node to perform one or more instructions of the further training.

21. The method of any one of claims 16 to 20, wherein: The second data product is CSI data available at the third node.

22. The method of any one of claims 16 to 21, wherein: The first node is a first user equipment UE, the second node is a first network node, and the third node is a second UE; and wherein the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE.

23. The method of claim 22, further comprising: The first component model is sent to the first network node for use in decoding compressed CSI data from the first UE and / or the second UE.

24. The method of claim 22 or 23, wherein: The method further comprises: compressing the first CSI data; and The compressed first CSI data is sent to the first network node.

25. The method of any one of claims 16 to 21, wherein: The first node is a first network node, the second node is a first user equipment UE, and the third node is a second network node; and wherein the first component model is a universal encoder for use by the first UE when encoding CSI information, and the CSI information can be decoded by the first network node or the second network node.

26. The method of claim 25, further comprising: compressing first CSI data using the first component model; as well as The compressed first CSI data is sent to the first network node and / or the second network node.

27. A computer-implemented method for training a first component model of an autoencoder (AE) machine learning model in a second node in a communication network, the first component model being an encoder or a decoder, and wherein: The first component model is for use when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network, the method comprising: receiving, from a first node, a baseline version of a first component model for CSI data available at the first node; training a second component model to perform encoding operations complementary to the baseline version of the first component model using CSI data available at the second node; and A first data product based on the training is sent to the first node for use by the first node in further training the first component model.

28. The method of claim 27, wherein: The first data product includes one or more of the following: the second component model; and A latent representation of the CSI data is available at the second node, the latent representation having been obtained by passing the CSI data through the second component model.

29. The method of claim 27 or 28, wherein If the first component model is a decoder, then the second component model is an encoder, and if the first component model is an encoder, then the second component model is a decoder.

30. The method of claim 27, 28 or 29, wherein The first node is a first network node, the second node is a first user equipment UE, and the third node is a second UE; and wherein the first component model is a general decoder for use by the first network node when decoding compressed CSI information from the first UE or the second UE.

31. The method of claim 30, further comprising: compressing new CSI data using the second component model; as well as The compressed new CSI data is sent to the first network node.

32. The method of claim 27, 28 or 29, wherein The first node is a first user equipment (UE), the second node is a first network node, and the third node is a second network node, and the first component model is a universal encoder for use by the first user equipment when encoding CSI information, and the CSI information can be decoded by the first network node or the second network node.

33. The method of claim 32, further comprising: receiving compressed new CSI data from the first UE; as well as The compressed new CSI data from the first UE is decompressed using the second component model.

34. A first node in a communication network for training a first component model of an autoencoder (AE) machine learning model, wherein the first component model is an encoder or a decoder, and wherein: The first component model is for use when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network, the first node comprising: a memory including instruction data representing an instruction set; and a processor configured to communicate with the memory and to execute the set of instructions, wherein the set of instructions, when executed by the processor, causes the processor to: i) training the first component model using a first data product obtained from the second node, wherein the training comprises freezing a first subset of horizontal layers in the first component model during a first backward pass training phase; and ii) initiating further training of the first component model using the second data product from the third node, wherein the further training comprises freezing a second subset of horizontal layers in the first component model during a second backward pass training phase, the first subset of horizontal layers being different from the second subset of horizontal layers.

35. The first node of claim 34, wherein: The set of instructions, when executed by the processor, further causes the processor to perform the method of any one of claims 2 to 26.

36. A first node in a communication network for training a first component model of an autoencoder (AE) machine learning model, wherein the first component model is an encoder or a decoder, and wherein: The first component model is for use when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network, wherein the first node is configured to: i) training the first component model using a first data product obtained from the second node, wherein the training comprises freezing a first subset of horizontal layers in the first component model during a first backward pass training phase; and ii) initiating further training of the first component model using the second data product from the third node, wherein the further training comprises freezing a second subset of horizontal layers in the first component model during a second backward pass training phase, the first subset of horizontal layers being different from the second subset of horizontal layers.

37. The first node of claim 36, further configured to perform the method of any one of claims 2 to 26.

38. A second node in a communication network for training a first component model of an autoencoder (AE) machine learning model, wherein the first component model is an encoder or a decoder, and wherein: The first component model is for use when exchanging compressed channel state information (CSI) between the first node, the second node, and the third node in the communication network, the second node comprising: a memory including instruction data representing an instruction set; and a processor configured to communicate with the memory and to execute the set of instructions, wherein the set of instructions, when executed by the processor, causes the processor to: receiving, from a first node, a baseline version of a first component model for CSI data available at the first node; training a second component model to perform encoding operations complementary to the baseline version of the first component model using CSI data available at the second node; and A first data product based on the training is sent to the first node for use by the first node in further training the first component model.

39. The second node according to claim 38, wherein: The set of instructions, when executed by the processor, further causes the processor to perform the method of any one of claims 28 to 33.

40. A second node in a communication network for training a first component model of an autoencoder (AE) machine learning model, wherein the first component model is an encoder or a decoder, and wherein: The first component model is for use when exchanging compressed channel state information (CSI) between the first node, a second node, and a third node in the communication network, wherein the second node is configured to: receiving, from a first node, a baseline version of a first component model for CSI data available at the first node; training a second component model to perform encoding operations complementary to the baseline version of the first component model using CSI data available at the second node; and A first data product based on the training is sent to the first node for use by the first node in further training the first component model.

41. The second node of claim 40, further configured to perform the method of any one of claims 28 to 33.

42. A system in a communication network, the system comprising: a first node according to claim 34, 35, 36 or 37; as well as A second node according to claim 38, 39, 40 or 41.

43. A computer program comprising instructions which, when executed on at least one processor, cause the at least one processor to carry out the method according to any one of claims 1 to 33.

44. A carrier comprising a computer program according to claim 43, wherein The carrier includes one of an electronic signal, an optical signal, a radio signal or a computer-readable storage medium.

45. A computer program product comprising a non-transitory computer readable medium having stored thereon a computer program according to claim 43.

Images