Improved NW-prioritized individual sequence training

Abstract

Embodiments provide improved NW-first individual sequential training, at least for: training an artificial intelligence decoder using at least an input of a hypothetical artificial intelligence encoder; determining a training dataset, the training dataset comprising an input assuming an artificial intelligence encoder and at least one of: an unquantized predicted channel state information feedback or a dequantized predicted channel state information feedback; transmitting the training data set to the user equipment; and determining reconstructed channel state information using the trained artificial intelligence decoder.

Description

Technical Field

[0001] Example and non-limiting example embodiments relate generally to communications, and more particularly, to NW-first individual sequential training with a raw dataset sharing scheme for AIML-enabled CSI compression. Background Art

[0002] It is known to measure channel state information for use in adapting transmissions in a communication network. BRIEF DESCRIPTION OF THE DRAWINGS

[0003] The foregoing aspects and other features are explained in the following description in conjunction with the accompanying drawings.

[0004] Figure 1 It is a CSI feedback solution that supports AIML (auto encoder) A high-level block diagram of an estimated channel feature vector (“target CSI” in the depicted example).

[0005] Figure 2 It is shown that UE-only training operations are preferred.

[0006] Figure 3 NW-first individual training operation is shown.

[0007] Figure 4 NW-first separate training with NW-side decoder API sharing ("freeze and train") is shown.

[0008] Figure 5 is a flowchart of the NW-first separate training framework with original dataset sharing (Q: quantizer, dQ: dequantizer) described in this paper.

[0009] Figure 6 The NW-first individual training framework with original (unquantized) dataset sharing is shown.

[0010] Figure 7 is a message sequence chart between UE and gNB for over-the-air training dataset sharing scenario.

[0011] Figure 8 UE-side training operations are shown, including AI encoder training and quantizer learning.

[0012] Fig. 9 A UE-side (general) AI encoder model update operation scenario is shown.

[0013] Fig.10 The empirical distribution of (original and quantized) latent vector elements with a TF-based encoder-decoder (each element has uniform SQ with {1,2,3} bits of resolution) is shown.

[0014] Fig.11 Inference results (SGCS) are shown for various training proposals (scheme described in the paper: “sepNwFirstwRawDS” - separate NW-first training with raw dataset sharing; x-axis:

[0015] [52*1 64*1 104*1 52*2 64*3], where latent vector dimension * number of Q bits per element).

[0016] Fig.12 Inference results (SGCS) for various training proposals with a linear y-axis are shown.

[0017] Fig.13 is a block diagram of one possible and non-limiting system in which example embodiments may be practiced.

[0018] Fig.14 is an example apparatus configured to implement the examples described herein.

[0019] Fig.15 A representation of an example of a non-volatile storage medium for storing instructions implementing the examples described herein is shown.

[0020] Fig.16 is an example method based on the examples described herein.

[0021] Fig.17 is an example method based on the examples described herein. DETAILED DESCRIPTION

[0022] 3GPP Release 18 Study Project AIML Air Interface Enhancement

[0023] One of the research items of 3GPP Rel18 is AIML for CSI feedback enhancement. This activity studies a setting in which the CSI feedback processing part on the UE side is replaced by an AI encoder (possibly followed by a quantizer), and its output is used as the CSI feedback information to be signaled to the NW over the air. The format of the AI-encoded CSI feedback does not necessarily follow the current 3GPP-defined format, such as Type I, Type II, enhanced Type II codebook, etc.

[0024] Figure 1 It is a CSI feedback solution that supports AIML (auto encoder) A high-level block diagram of an estimated channel feature vector 105, which in the depicted example is the "target CSI". If the CSI feedback mechanism is lossless, the target CSI can be considered as the version of the CSI that the NW side 111 should retrieve.

[0025] Figure 1A CSI feedback scheme for a general AIML (autoencoder) is depicted in which a joint end-to-end (E2E) model training case is considered. At the UE side 101 (depicted as UE 10), the downlink channel is estimated using the CSI-RS (102), and the channel estimate (e.g., SVD) is subjected to the required preprocessing (104) to obtain channel feature vectors, etc. These channel feature vectors (105), such as in [RAN1#112], where the input CSI precoding matrix is ​​in the spatial frequency domain, are fed to the AI ​​encoder 106. The output of the AI ​​encoder (i.e., the latent vector (108)) will be quantized (110) to reduce the CSI (in this case, AI compressed PMI) feedback overhead. Note that in addition to the channel feature vectors 105, other types of channel information (e.g., the full channel matrix) can also be used as input to the AI ​​encoder 106. At the NW side 111 (depicting the network node 70), which receives the CSI feedback information 120 as input from the UE side 101, reverse operations (i.e., dequantization (112), AI decoding (114)) are performed to reconstruct the original channel information, in this case the channel eigenvectors. If necessary, the reconstructed channel information 115 is further post-processed (e.g., a re-orthogonalization process) at 116, the result of which is referred to as "output CSI" 118 in 3GPP.

[0026] For a typical autoencoder, the AI ​​encoder 106 and the AI ​​decoder 114 are trained by the same training entity in the same training session to derive the best possible parameters at the AI ​​encoder 106 and the AI ​​decoder 114 ([RAN1#110] "Type 1: Joint training with a single training entity"). Note that in this case, the training entity may use the input of the AI ​​encoder (e.g., the AI ​​encoder input (channel feature vector 105) at the UE side 101) and the reconstructed CSI 115 (the AI ​​decoder output at the NW side 111) as parameters of the loss function during training, such as Figure 1 As shown (reference 122). However, this forces disclosure of proprietary AI encoder 106 or AI decoder 114 details (AIML framework, architecture, layer configuration, etc.), so it is impossible for UE and NW vendors to agree on this training option (Type 1). In this sense, "Type 3: Separate Training" has been supported by many UE and NW vendors in 3GPP, because UE vendors and NW vendors do not need to reveal their proprietary AI encoder / decoder related information to other parties. The third option (Type 2: Joint training with two training entities) includes training both the encoder and decoder in the same training session, but with two separate training entities that exchange gradients during training.

[0027] Individual training can be divided into two types, namely UE-first training and NW-first training.

[0028] Output-CSI-UE, Input-CSI-NW

[0029] The separate training case does not require that the UE vendor and the NW vendor should share the AIML model details. In the case of UE-first training, where the AI ​​encoder 206 at the UE side 201 is trained first, when the UE-side AI encoder 206 is trained together with the corresponding hypothetical AI decoder 212 ( Figure 2 In stage 1200 in the UE side training process, the UE vendor needs to assume an AI decoder model at the NW side 211. In this UE side model training phase (stage 1200), the output at the assumed AI decoder 212 of the NW side 211 (assumed at the UE side 201) is referred to as "Output-CSI-UE" 215 in 3GPP. Figure 2 Also shown in FIG. 2 is a NW side model training stage 220 .

[0030] In the case of NW-first training, where the AI ​​decoder 314 at the NW side 311 is trained first, when the NW-side AI decoder 314 is trained together with the corresponding hypothetical AI encoder 306 ( Figure 3 In the stage 1300 in UE side 301, the NW vendor needs to assume the AI ​​encoder model 306 at the UE side 301. In this NW side model training stage (stage 1300), the input at the assumed AI encoder 306 at the UE side (assumed at the NW side 311) is called "Input-CSI-NW" 305 in 3GPP. Note that the assumed quantizer 310 is also assumed by the NW side 311 together with the assumed AI encoder model 306. Figure 3 Also shown is the UE-side model training phase 320.

[0031] In CSI compression using a two-sided model use case, separate AI / ML model training (referred to as type 3 AI / ML model training) on ​​the network and UE sides may be beneficial, where the UE-side CSI generation part and the network-side CSI reconstruction part are trained by the UE side and the network side, respectively. The separate training may include sequential training starting from UE-side training or sequential training starting from NW-side training. The evaluation of quantization-aware or non-aware training may be based on quantization-aware training, where quantization or dequantization involves the training process. CSI compression using a two-sided model use case is used when the reconstructed output CSI assumed at the UE or the "input CSI-NW" 305 (the input CSI assumed at the AI ​​encoder assumed by the NW at the UE side) is a precoding matrix, where in one option the precoding matrix is ​​in the space-frequency domain and in another option the precoding matrix is ​​in the angle-delay domain.

[0032] Key trends and key takeaways

[0033] Based on the interpretation of 3GPP Rel18 protocol and submitted TDOC, the following trends and possible convergence points can be observed. Regarding model training strategies for dual-sided model use cases, Type 3 separate training has gained support from many manufacturers, UE vendors, and NW vendors, etc. Type 3 separate training allows independent training on the NW side or UE side based on a training dataset (222, 223, 322, 323) shared with the other party without having to disclose its own model details, such as AIML backbone architecture, model details, etc. In cases where UE-side separate training is preferred (such as Figure 2 As shown in Figure 2, NW vendors need to handle multiple UE-side (encoder) models, while UE vendors can focus on a single (or only a few) encoder models. Figure 3 ), UE vendors need to handle multiple NW-side (decoder) models, while NW vendors can focus on a single (or only a few) decoder models. Due to the above reasons, NW vendors may prefer NW-first separate training as this is relevant to the case where gNB 70 only maintains a small number of models. UE vendors may tend to adopt a NW-first training "freeze and train" scheme. Another major benefit of the NW-first training scheme is that it does not incur a large engineering cost for decoder model updates when UE-side encoder model changes / updates occur. In short, UE model training and updates can occur without involving NW-side training.

[0034] Figure 3 The process of NW-first individual sequential training is shown in FIG. The whole process can be divided into two main stages, namely, the NW-side model training stage 300 and the UE-side model training stage 320 .

[0035] 1. NW-side model training phase (300). NW 70 performs training on its NW-side AI decoder model 314. Figure 3 As depicted in FIG. 3 , the NW side 311 should propose a hypothetical UE side model 306 for generating the UE side expected result (ie, the expected CSI feedback 311). As a result of the NW side training, {the hypothetical input of the AI ​​encoder (Input-CSI-NW: )}305 can be used for its self-supervised learning. Note here that the NW side 311 should have a UE-side assumption model of DL channel measurement and subsequent pre-processing (for example, an assumption model of DL channel measurement 324 and pre-processing 326) and an AI encoder 306. The input of the AIML model on the UE side is the result of passing through the UE-specific signal processing chain before the AIML model, i.e., the RF chain, channel estimation 324, pre-processing 326 (such as SVD), etc. This pre-AIML signal processing operation is highly proprietary to the UE / modem vendor, and if the details are not shared, the NW vendor needs to come up with the assumption model, which is marked with ['] (apostrophe) in this disclosure. Note that sometimes for clarity, the [′] (apostrophe) is omitted when its distinction is not important. Alternatively, the UE vendor may provide pre-processed channel information that takes into account UE vendor specific operations or effects such as channel estimation, SVD, and RF nonlinearity, etc. Upon completion of NW side model training 300 (under its own assumed UE side model 306), the NW may provide the resulting training data set 323 to the UE side, e.g., {(assumed input to AI encoder (Input-CSI-NW: )305, predicted CSI feedback (z′ q ))}311, for subsequent UE-side model training 320.

[0036] 2. UE side model training phase (320). The UE side 321 can be based on the training data set 323 provided by the NW (for example, {(AI encoder's assumed input (Input-CSI-NW: )305, predict CSI feedback (z′ q ) 311}) to start training its AI encoder 328. Note that the assumed input to the AI ​​encoder provided by the NW (Input-CSI-NW 305) is the output of the previously assumed UE-side operation (i.e., DL channel measurement and pre-processing), which may be different from the actual UE-side operation (324, 326) in use (i.e., UE 10-specific) unless it is directly provided by a specific UE vendor. For UE-side model training 320, an appropriate loss function needs to be defined. One option is to use CSI feedback 330 (the result of the actual UE-side model considering the quantization effect: z q 332) and the expected CSI feedback 311 (the result of the hypothetical UE-side model trained according to the assumptions made by the NW vendor during the NW-side model training, taking into account the quantization effect: z′ q 334). Note again that the quantizer 336 at the UE side 321 and the dequantizer 338 at the NW side should be aligned to eliminate ambiguity in interpreting the CSI feedback bit sequence to facilitate z q Correct retrieval of 332.

[0037] Note here that Figure 3 As shown (see the south-west section), the quantized CSI feedback 311 may be coarse-grained, e.g., only 4 different levels per potential vector element when SQ is used with a 2-bit / element allocation for the depicted example. This may degrade the model training performance at the UE-side model training process 320, especially at low quantization resolutions, when the loss function takes the quantized CSI feedback 311 as its input parameter.

[0038] UE priority individual sequential training ( Figure 2 ) does not suffer from this problem because the training loss calculated when training the NW side model is based on the reconstructed CSI with a finer granularity 225. For a better understanding, refer to Figure 2 .

[0039] It has been assumed that the quantization and dequantization (Q-dQ) operations (230, 232, 340, 342) may also be subject to training. The Q-dQ operation itself may be learned during the model training process of the first training session, i.e., NW-side decoder training 300 in case of NW-first training, and UE-side 201 encoder training in case of UE-first training. After the first phase of the training session, the Q-dQ operation and its associated generated codebook (if any) are fixed. In the subsequent operation, i.e., the second phase of the training session, the quantizer itself is frozen, but the model is being trained with the appropriate Q-dQ. Therefore, with respect to Figure 2 and Figure 3 All cases explained in the discussion can be considered as quantization-aware training.

[0040] Unlike UE-first separate sequential training where the training loss calculated when training the NW-side model is based on the output (or reconstructed) CSI225 with a finer granularity, the model training performance may be degraded in the case of NW-first separate sequential training. For example, a ~0.4dB loss has been reported based on SGCS in the case of the “Type 3: gNB-first” training case compared to the “Type 3: UE-first” training case or the “Type 1: Joint Baseline with VQ” case. This performance gap may seem insignificant at first glance, which can be used as a parameter for the NW-first training scheme, as the performance degradation can be added whenever other possible sources of degradation are allowed (e.g., quantization resolution limitations, channel estimation errors, channel time evolution, etc.). It is always good to have a “tool set” ready for possible performance enhancement schemes, just in case.

[0041] Furthermore, the performance degradation level of NW-priority separate training (350) can be as high as ~2dB (>1.8dB) when using low resolution (1.2dB at CR (compression ratio) 16, >1.8dB at CR8, both with 1-bit SQ) SQ with uniform SQ (scalar quantization) compared to UE-priority separate training (250) or joint E2E training (150).

[0042] The main reason for the performance gap between the NW-first scheme 350 and the UE-first scheme 250 is that the training loss calculated when training (320) the UE-side model for the NW-first training use case 350 is based on (quantized) CSI feedback (311, 330, 332) with a coarse granularity.

[0043] To address this issue, a "Gradient Exchange-Based NW-First Type 2 Training" (also known as: "Freeze and Train") approach (400) may be used, which has been shown to improve encoder training performance by facilitating reconstructed CSI for UE-side encoder training sessions. However, this approach may not be preferred because it requires the NW vendor to provide an API 402 of the trained NW-side decoder model. Although this does not require explicit disclosure of the NW-side decoder model details, there are still privacy issues from the NW side in the sense that reverse engineering practices may reveal enough information about the trained NW-side model (e.g., its capabilities, behavior on various inputs, etc.).

[0044] A NW-first individual training scheme is described herein that provides the closest performance to competing schemes, such as UE-first individual training (250), NW-first training with a "freeze and train" option (400), while not disclosing NW vendor proprietary information.

[0045] This paper describes a model training framework for the NW-first separate sequential training use case, which aims to address the possible UE-side model training performance degradation caused by the coarse / low granularity of the input parameters of the loss function without simultaneously revealing the NW-side decoder model information.

[0046] This document also describes exemplary UE-gNB signaling essentials, i.e. possible messages and their contents, focusing on over-the-air training dataset sharing scenario.

[0047] The scheme described in this article maintains the benefits of separate training, which is that both sides, namely the UE side and the NW side, do not need to disclose their own model details, such as backbone architecture, model layer details, etc., and the model training of each party can be done via a shared training dataset, which allows separate training.

[0048] At the same time, the benefits of maintaining the NW-first training program from the perspective of NW suppliers are (1-2):

[0049] 1. No need for the NW side to maintain multiple models corresponding to each individual UE model. Regarding the NW side decoder model at each vendor, a single (or only a few) models should be sufficient.

[0050] 2. The UE side needs to cope with multiple NW-side models associated with various vendors by developing multiple dedicated UE models corresponding to each individual NW-side model, or by developing a (limited number of) generic UE-side encoder models that can generalize themselves to multiple NW-side decoder models.

[0051] Although it can improve the UE-side model as well as the overall E2E model training performance by providing raw (unquantized) CSI feedback (raw latent vector; z e ) instead of its quantized version z as in the usual case q The first training entity (NW side) of the UE-side training is used to enhance the granularity of the input parameters (i.e., CSI feedback) of the training loss function for UE-side training. For the loss function parameters, the UE-side encoder training can be performed on the original / unquantized CSI feedback, which is shown to improve the encoder performance by up to ~2dB.

[0052] After its own NW-side quantization-aware training is completed, the NW side is assumed to provide {(hypothetical input to the AI ​​encoder (Input-CSI-NW: ), (raw / unquantized) predicted CSI feedback (z′ e ))} as the UE side training data set.

[0053] This means that quantized CSI feedback, e.g., a unique predefined level (e.g., 1, 2, or 3 bits of information) per latent vector element or codeword index referring to a predefined codebook, is replaced with raw CSI feedback, e.g., a high resolution (e.g., 32-bit floating point) value per latent vector element.

[0054] Note that Figure 6 As shown, the quantizer and dequantizer operations are applicable to the NW side decoder training phase and the UE side encoder training phase. The model parameters are trained with the quantization effect taken into account.

[0055] The NW side is responsible for quantizer training, and the NW side model together with the associated quantizer-dequantizer operation is not tied to the model or quantization scheme of a specific UE vendor. Once the Q-dQ operation is trained, its operation rules or associated codebooks are fixed and frozen at the UE side training process. The task of model generalization (if considered beneficial) is on the UE side, which means that the UE side may need to develop a generic UE side encoder model (including associated quantizers) that can handle possible multiple NW side decoder models to avoid multiple one-to-one matching UE models.

[0056] The whole process can be completed in 500 seconds. Figure 5 , where certain aspects are labeled as blocks 505 (including operations 510, 520), 515 (including operations 540, 550), and 525 (including operation 560).

[0057] The main advantages and technical effects of the scheme described here are (1-4):

[0058] 1. Simple but very effective. The entire traditional NW-first training operation can remain intact. Up to ~2dB SGCS performance improvement is observed relative to the conventional case (see the "Method Validation" section for details). The gap with the upper reference case (i.e., the joint end-to-end case) is less than ~0.1dB

[0059] 2. The cost is very low. No complexity added on the NW side, no complexity added on the UE side, the only foreseeable cost: increased dataset size means this should not be important in the case of offline IODT stage training, it may be important in the case of over-the-air dataset sharing use case.

[0060] 3. Further possible improvements to facilitate UE-side training, better quantization learning (if needed), and expected better general encoder model training performance.

[0061] 4. High level of detectability of third party use via inspection of shared data sets.

[0062] The main use case of the examples described in this article is offline vendor-to-vendor IODT phase training. The basic assumption about the main mode of operation of the scheme described in this article is that the whole process can be done offline, for example in the vendor-to-vendor IODT phase. This may also be preferred. This means that no runtime gradient sharing is required at all. Moreover, unlike the "freeze and train" scheme, the NW side does not need to provide any NW vendor's proprietary information related to its decoder model to the UE vendor (no privacy issues for the NW vendor).

[0063] There may be another use case for the solution described herein, namely, an over-the-air training dataset sharing scenario. Considering this use case, Figure 7 The signaling operations between UE 10 and gNB 70 are shown in FIG. The messages and their possible contents may be standardized in upcoming 3GPP specifications. Note that this process is also designed to reduce over-the-air signaling overhead based on observations of our numerical results.

[0064] The NW-first individual training framework described in this paper consists of two stages, as used for conventional quantization-aware NW-first individual training, i.e., NW-side decoder training together with quantizer-dequantizer training, quantizer-dequantizer training followed by training dataset generation, and UE-side encoder training based on a shared training dataset. The shared dataset is a set of tuples (e.g., Input-CSI-NW (channel feature vector), original (unquantized) latent vector, quantized latent vector, i.e. ). Here, the last term (the quantized latent vector, z q ) may be optional and need not be provided to the UE side when the trained / finalized Q-dQ rule and / or the resulting codebook and associated mapping scheme are shared by the NW side (first training entity). In the case where the Q-dQ rule is not provided, the dataset {(z e , z q )} can be used for Q-dQ learning. The point here is to share the original (unquantized) latent vector (z) for the NW side (first training entity) e ) to improve encoder training on the UE side by providing rich and fine-grained training data (refer to Figure 6 and Fig.10 , about the collection of z e and z q Here, Uniform SQ is used as the quantization scheme.

[0065] A potentially important use case scenario for NW-first individual training is over-the-air training dataset sharing for over-the-air UE and gNB model training and updates. Since dataset sharing needs to be done over the air interface in this case, reducing over-the-air signaling traffic is critical. This article describes a general framework (exemplary signaling between UE and gNB) that may allow for reduced over-the-air traffic. This has a direct impact on the 3GPP specifications.

[0066] This section also includes possible implementation examples of UE-side training operations focusing on AI encoder training and quantizer learning, as well as UE-side (general) AI encoder model update operation scenarios.

[0067] Extensive model training and inference testing have been performed to validate the ideas described herein, and the results can be found in the latter part of this disclosure. An intermediate KPI, namely SGCS, is used for performance measurement. The scheme described herein is consistent with other competing alternatives, including the upper reference case (joint E2E training), UE-first training, and NW-first training with freeze and train options, with a performance gap of less than 0.1dB. It has been shown that the scheme described herein can improve the inference performance by up to ~2dB compared to the traditional NW-first training approach. It confirms that the ideas described herein can be used as a promising alternative for the NW-first separate sequential training case without any risk of disclosing the details of the NW-side decoder model.

[0068] NW-first individual training framework with original dataset sharing

[0069] To better understand the following concept description, you can refer to Figure 5 (Flowchart) and Figure 6 (System model block diagram).

[0070] 1(510). Perform NW side model training with the following configuration. ( Figure 6 1601 in the process.

[0071] A quantizer 606 is placed (enabled) at the UE side 602, and a corresponding dequantizer 616 is placed (enabled) at the NW side 612. The methods described herein are transparent to the specific quantization scheme (i.e., SQ or VQ) in use. Note that the Q-dQ operation itself can be configured to be trained. SQ: uniform (each element has a predefined fixed level) and non-uniform (quantization levels are subject to Q-dQ training). VQ: The overall latent vector, i.e., 64 dimensions, can be grouped into multiple sub-vectors, each with a smaller manageable dimension, e.g., 2 or 4. Therefore, Q-dQ training produces a dedicated codebook. The codebook can be generated at various levels of generality, e.g., dedicated to each sub-vector, dedicated to each layer (but sub-vectors are universal), or universal to each layer (assuming sub-vectors are universal).

[0072] The E2E autoencoder (UE: (hypothetical) AI encoder 604, NW: AI decoder 614) training dataset 611 is a set of Input-CSI-NW 603, which is the input to the hypothetical AIML encoder 604 at the UE side 602. Depending on the (3GPP or vendor-to-vendor) protocol, it is a channel feature vector or a full channel matrix (or any other form of channel information). Note that in actual scenarios, in order for the UE 10 to obtain the Input-CSI-NW 603, the UE 10 needs to perform channel parameter estimation based on the CSI-RS (or other reference signals, if available), and perform pre-processing such as SVD, etc. For this reason, the UE vendor may also want to provide the Input-CSI-NW 603 for the AIML-enabled CSI compression use case to the NW vendor to better reflect the details of the UE vendor chipset.

[0073] The loss function for E2E autoencoder training is based on the output CSI 618 (reconstructed channel feature vector, e.g., at the NW side 612). In this practice, the main subject of training is the AI ​​decoder 614 at the NW side 612 and the quantizer 606 and dequantizer 616 operations (if configured). Note that this is quantization-aware training since the quantization effect is taken into account during the training process.

[0074] 2(520). Once the NW side 612 model training is completed, it is necessary to generate a UE side training dataset 630 on the trained NW side decoder model 614 (as well as the trained hypothetical UE side encoder model 604 and trained Q-dQ operation, if configured).

[0075] 3(530). The NW side 612 provides the training data set for UE-side training 622 to the UE side 624: x′: ['] (apostrophe) is used to indicate the values ​​from the assumed pre-AIML model 604 at the UE side 602, i.e., channel estimation, preprocessing (such as SVD), etc., which may be different from the actual model. When its indicative value is obvious or does not play a key role in its interpretation, for simplicity, ['] (apostrophe) is sometimes omitted in this disclosure. {(Assumed input of AI encoder (Input-CSI-NW: )603, (raw / unquantized) predicted CSI feedback (z′ e )605, (optional) predict CSI feedback (z′ q )632}. Original predicted CSI feedback (z e )605 has a higher q )617Richer fine-grained structure. Figure 6An empirical histogram 613 of the original / unquantized CSI feedback (latent vector) 605 and an empirical histogram 621 of the quantized CSI feedback 617 are shown. The dimension of the latent vector 605 is 64, and each latent vector element is depicted as having a different color (only selected indexes are shown for clarity). A scalar uniform quantizer with 2 bits / element (4 different levels) is used for the quantization scheme. If the Q-dQ operation configuration details (meaning The mapping rules and / or the codebook set generated in the case of VQ) will be provided by the first training entity (the NW side 612 in the case of NW-first training) to the second training entity (the UE side 624 in the case of NW-first training), and the UE-side training does not need to share z q 617. If not, the UE side 624 also needs to learn Mapping rules (step 5 (550)). In this case, the UE side 624 can e , z q )} for Q-dQ learning.

[0076] 4(540). Perform UE-side 624 model training with the following configuration. Figure 6 The trained quantizer 606 is placed (enabled) on the UE side 624. Note that the Q-dQ operation in this stage is frozen and does not undergo changes. Using the training data set 630 provided by the NW side, i.e., {(assumed input of the AI ​​encoder (Input-CSI-NW: )603, (raw / unquantized) predicted CSI feedback (z′ e )605, (optional) predict CSI feedback (z′ q )632}. The training is done via appropriate APIs without the NW side decoder 614 or its runtime image. The loss function for AI encoder 634 training is based on the (raw / unquantized) CSI feedback 605, which is the output of AI encoder 604 at UE side 602. In this practice, the main subject of training is AI encoder 634 at UE side 624. AI encoder parameter updates should be done at z e Completed on the domain, z e The domain has a ratio of q The domain is richer and finer granularity. Since the quantization effect is still observed during the training process, this is quantization-aware training.

[0077] 5(550). (Optional) Perform quantizer learning and / or codebook generation (in the case of VQ), based on {((original / unquantized) predicted CSI feedback (z′) e )605, predict CSI feedback (z′ q)632)}. This can be done in a separate and independent manner relative to the UE side model training. Note that at this stage, the Q-dQ operation itself is fixed and frozen (no changes are allowed). The main purpose of this step is to learn the trained Q-dQ operation on the UE side without sharing the Q-dQ rules and / or associated codebook.

[0078] 6(560). Perform E2E model (with quantizer / dequantizer as appropriate) inference testing to evaluate E2E model performance.

[0079] The above process should be repeated for each CR (compression ratio) and quantization bit resolution combination in use to obtain a training parameter set associated with each CR and bit resolution combination. Fig.10 As shown, when quantization (1002) is bypassed, the base original latent vector (z e ) distribution is very close to a Gaussian distribution (except for a few outliers). However, the “insertion” of the Q-dQ operation (for Fig.10 The case depicted in the figure uses uniform SQ) which significantly changes z e distribution, and it shows very different shapes for each quantization resolution ( Fig.10 The number of bits per element in ; refers to z as the quantization bit resolution increases, including after 1-bit Q-aware training 1004, after 2-bit Q-aware training 1006, and after 3-bit Q-aware training 1008. e “Evolution” of the distribution).

[0080] Over-the-air training data sharing scenarios: potential impact on 3GPP standards

[0081] The main use case of the scheme described in this paper is NW-first training during the IODT phase. In this case, no over-the-air dataset sharing is required as the gNB device or its equivalent decoder model and the UE device or its equivalent encoder model. The associated channel estimation, pre-processing schemes can be provided with a dataset that can be uploaded to or downloaded from a designated server using a wired connection (no need to share over the air interface). However, once these trained UE devices are up and running along with the trained NW devices after the IODT is completed, there may be cases where over-the-air training dataset sharing is unavoidable due to the need to retrain model parameter updates.

[0082] Conceivable situations may include, but are not limited to: (1-2):

[0083] 1. When UE-side model retraining is required due to the following reasons: UE-side AI encoder model update, such as changes in the AIML model architecture, model expansion (more / fewer layers, parameter resolution changes, etc.) and / or UE-side non-AIML signal processing chain update, such as different RF components (and therefore different nonlinear behaviors), front-end signal processing algorithm updates, baseband channel estimation and preprocessing (SVD) algorithm updates, etc.

[0084] 2. When the gNB wants to initiate the latest channel data collection for model update to enable: i) update of gNB decoder as well as UE encoder (possibly Q-dQ operation as well) due to performance degradation detected by the gNB of the previously trained model, ii) development of cell-specific and / or day-specific time of training models to adaptively tailor the gNB-UE model for specific cell traffic scenarios and / or specific time of day (at night, in the morning or during busy daytime, etc.).

[0085] In the over-the-air training dataset sharing scenario, it is important to reduce the signaling overhead at the air interface. This section presents an example interaction between the UE and the gNB in ​​terms of signaling and associated operations for this scenario. Figure 7 for corresponding explanation.

[0086] 1(701). As part of UE capability reporting, the UE 10 may report to the gNB 70 its supported training schemes (e.g., UE first, NW first, NW first with raw data set sharing, etc.) and its capabilities for the over-the-air data set (DS) sharing feature via function-based or model ID-based LCM identification.

[0087] 2(702). In the context of NW-priority separate training reported via UE (capability), the gNB knows the training scheme of air DS sharing supported by the UE to enable the CSI compression function of AIML, and commands to enable CSI compression of AIML with air DS sharing feature support in use. From this point on, the gNB and UE should know that air DS sharing can be triggered based on the CSI performance monitoring results of enabling AIML on the gNB side. There can be multiple format options for encoding the real CSI (e.g., channel feature vectors), and the format in use can also be configured at this stage.

[0088] 3(703).Following AIML-enabled CSI operation. The gNB may apply performance monitoring in parallel as part of LCM, for example by monitoring the BLER of PDSCH.

[0089] 4 (704). Events defined in 3GPP or gNB vendor proprietary trigger the gNB to initiate over-the-air DS sharing for model retraining and updating. The gNB needs to determine the CR and quantization bit resolution combination based on the selected (CR, Q resolution) combination, and the original (unquantized) latent vector (z e ) or quantized latent vector (z q ). For example, Fig.11 As depicted in , the model inference performance converges at certain combinations with high quantization resolution. In this case, sharing the quantized latent vectors should be sufficient while saving signaling overhead over the air. Otherwise, sharing the original latent vectors is required for better / acceptable training performance at the expense of air interface traffic.

[0090] 5(705).gNB requests UE to measure downlink channel and report ground truth CSI. If necessary, a specific format for reference truth CSI can be configured. gNB can inform UE of the configuration for upcoming model retraining, i.e., UE-first or NW-first training, necessity of Q-dQ learning, etc.

[0091] 6(706).UE measures DL channel upon request of gNB, generates and reports true CSI. Since UE knows the upcoming model retraining process in advance via gNB signaling, it should keep the CSI data sets (they should be saved in UE internal memory).

[0092] 7(707).UE sends the ground truth CSI dataset to gNB. Due to its expected large size, a data channel, i.e. PUSCH, can be used.

[0093] 8(708).gNB performs NW-side decoder training based on the received true CSI. Once the training is completed, a training dataset needs to be generated. If the gNB does not intend to provide the UE with the (updated) Q-dQ rules (if any), the training dataset to be shared for UE-side retraining should contain z e or q (depending on the gNB's decision in step 4), or optionally e With z q In the latter case, the UE needs to e , z q )}Learn the Q-dQ rule and / or the associated codebook.

[0094] 9(709).gNB sends the UE training data set for retraining the UE side model. If the gNB also decides to share the updated Q-dQ rules, relevant information signaling can be performed at this stage.

[0095] 10(710). The UE receives {z e , (z q )} triggers UE-side AI encoder model training. If necessary, quantizer learning can also be initiated. Note here that at this stage, the UE should still have access to the baseline truth CSI It should also be noted that The order of arrangement should be consistent with the training dataset provided for gNB, that is, {z e , (z q )} are matched exactly in order for UE training exercises. At this stage, the UE may have the following possible combinations of training data sets, namely, or A more detailed description can be found in the subsequent subsections below.

[0096] The NW of the original dataset sharing solution is trained separately first; the UE-side AI encoder model is trained separately at z e Retraining on the domain

[0097] Conventional NW is trained separately first; UE-side AI encoder model is trained on domain z q retraining.

[0098] The NW is trained separately first using the original dataset sharing scheme; the UE-side AI encoder model is based on {(z e , z q )} in z e Retraining on domain,Q-dQ rule learning.

[0099] 11(711).When the UE completes the UE-side model retraining (possibly with Q-dQ learning), it may send an indication of the completion of the UE-side model training to the gNB.

[0100] Note here that steps 4 (704) and 5 (705) may be changed to reflect the UE-initiated situation, which means that the UE may request to trigger over-the-air DS sharing and model training upon identifying a UE-side model update.

[0101] UE-side training operation example: AI encoder training and quantizer learning

[0102] This subsection is dedicated to explaining the UE-side training operation example in more detail ( Figure 7 Step 10 (710)). Figure 8 Provide explanation.

[0103] Assume that UE 10 needs to support multiple gNB decoder models. Due to the limited computational power at the UE, it is expected that UE vendors develop a general AI encoder (802, 822) that can handle multiple gNB decoder models. The quantized latent vector (z q The mapping of ) (806, 826) to the bit sequence (808, 828) and its reverse operation rules (810, 812, 830, 832) are shared between UE vendors and gNB vendors and can be subject to 3GPP standardization. Q-dQ (quantizer-dequantizer) operation rules: shared from z e (804, 824) to z q The exact mapping of (806, 826) is and the associated training / generated codebooks can be subject to vendor-to-vendor alignment and this information can be shared between UE vendor and gNB vendor, or remain dedicated to the first training entity (gNB vendor in case of NW-first training scheme).

[0104] UE-side training scenario: Generic AI encoder (802, 822). When the Q-dQ rules (and the associated codebook in the case of VQ) are provided by the gNB vendor, the gNB vendor may only provide {z e}(804,824) or {z q}(806, 826), assuming that the UE supplier should already have (801, 821). Utilizing per-gNB vendor specific quantizers (805-1, 805-2, 805-3) [Top diagram]: A dedicated per-gNB vendor specific quantizer may be implemented at the UE 10. The UE vendor may implement a dedicated per-gNB vendor specific quantizer based on the quantizers it receives from multiple gNB vendors. (801,804) or (801, 806) to train the general AI encoder 802. No Q-dQ rule learning is required. When the gNB vendor does not provide the Q-dQ rule (and the associated codebook in the case of VQ), this means that the gNB vendor should provide {(z e , z q )}(804, 806, 824, 826). Using a common quantizer (possibly NN-based) [bottom figure]: UE vendors can use multiple gNB vendor sources to (821, 824, 826) Develop and train a general quantizer 825 on top of the general AI encoder 822. Alternatively ( Figure 8 ), the UE vendor can be trained to learn a dedicated quantizer specific to each gNB vendor.

[0105] UE-side (general) AI encoder model update operation scenario

[0106] This subsection describes the general AI encoder model training and update features ( Fig. 9 ) general operation scenario. One of the main benefits of NW-first separate training is that the AI ​​encoder model update 910 at the UE side (903 to 904) will not cause the AI ​​decoder (912, 914) at the gNB side at gNB70-1 and gNB 70-2 to be retrained. Fig. 9 As shown in 904, the previously shared data set (e.g., 918 and 920) without involving gNB-side retraining. Therefore, whenever the associated UE model 916 is updated, there is no need to involve all gNB vendors to update the gNB-side decoder models (912, 914).

[0107] Note that for the generic AI encoder training use case, the approach described in this paper can be implemented by providing a hypothetical AI encoder output {z e} to improve the general model training performance. e} of finer granularity (refer to Fig.10 To observe the collected z e (left side) and z q (Right side), when uniform SQ is used as the Q-dQ rule), better AI encoder model training can distinguish the signature of each gNB vendor, and more capable universal AI encoder training becomes possible.

[0108] Fig.10 The distribution z is shown as the quantization bit resolution increases, including after 1-bit Q-aware training 1014, after 2-bit Q-aware training 1016, and after 3-bit Q-aware training 1018. q Evolution from top to bottom.

[0109] Validation of the method

[0110] The performance of the method described in this paper has been evaluated through model training and inference testing. A 630,000 sample training dataset has been collected for the configurations defined in Table 1. The data type of the ground truth CSI is a normalized channel feature vector (decomposed into real and imaginary parts) with dimension N sb ×2N tx :where N sb : the number of subbands (13), and N tx : The number of TxAnt ports (32). Only the main channel eigenvectors have been considered for training and inference testing of this exercise.

[0111] AIML backbone model: TF (5 layers); matching encoder-decoder

[0112] Training objective function: SGCS, MSE (when the reconstructed CSI is not available during the training phase)

[0113] Number of epochs used for model training: 200

[0114] Quantization: Uniform SQ with predefined levels (quantization is not trained).

[0115] Table 1: SLS parameters for the two-sided model training dataset

[0116]

[0117] The two-sided model training scheme used for comparative evaluation is (1-5):

[0118] 1. Joint End-to-End (“JointE2E”): Type 1 training (see Figure 1 ); Upper reference situation

[0119] 2. UE priority separate training ("sepUeFirst"): Type 3 UE priority training (see Figure 2 )

[0120] 3. NW priority separate training ("sepNwFirst"): Type 3 NW priority training; with z q Shared routine training (see Figure 3 ).

[0121] 4. NW First Separate Training with Raw Dataset Sharing (“sepNwFirstwRawDS”): The scheme described in this paper; with z e Shared Type 3 NW Priority Training (see Figure 6 ).

[0122] 5.a. NW-first separate training with freezing and training (“sepNwFirstwFnT”); Type 2 NW-first; also known as gradient exchange-based (see Figure 4 )

[0123] The selected (CR [latent vector dimension], bit allocation per element) combinations are (16

[52] , 1), (16

[52] , 2), (13

[64] , 1), (13

[64] , 2), (13

[64] , 3), (8

[104] , 1).

[0124] The reasoning results in terms of SGCS can be found in Fig.11 (y-axis, in dB) and Fig.12 (y-axis, in linear units). The collected results have been sorted bit by bit in increasing order of total CSI feedback overhead, i.e., [52*164*1104*152*264*264*3], where N dim *n q :Where:N dim : latent vector dimension, n q : The number of quantization bits per latent vector element. The (104*1) case has been shifted forward by 0.2 for better visibility relative to the (52*2) case.

[0125] Observations. As expected, the higher the CSI feedback overhead bits, the lower the SGCS in dB. The differences between the various bilateral model training options are small (less than 0.1 dB) as long as the quantization bit allocation for each element is greater than 1 bit. When the quantization resolution is coarse (1 bit / element for the simulation case), conventional NW-first separate training shows a significant performance degradation (from ~1 dB up to ~2 dB) compared to other cases including the scheme described in this paper. The scheme described in this paper (NW-first separate training with original (unquantized) dataset sharing) brings up to ~2 dB performance enhancement compared to the conventional conventional scheme, while the performance is significantly improved compared to the upper reference case (joint E2E), UE-first training and Figure 4 Compared with other methods shown in , it has a high risk of leaking the details of the NW side decoder model (frozen and trained).

[0126] The method described herein is numerically analyzed using uniform SQ and matched encoder-decoder cases. The mismatched encoder-decoder case and / or VQ in place (i.e., codebook-based quantization) may provide another aspect of the performance behavior of the method described herein compared to other competing schemes. In any case, it is expected that the method described herein will show superior performance relative to the traditional NW-first single training scheme.

[0127] The scheme described in this paper (i.e., NW-first separate training with original dataset sharing) can be used as a very promising and capable candidate for the NW-first separate training option, which may be superior to other options such as UE-first training or NW-first training with freeze and train for NW vendors.

[0128] The experimental results also indicate that, depending on the operating area (in terms of overall CSI feedback overhead and / or quantization resolution), a training strategy can be developed in a certain area (usually the low overhead case) where traditional NW-first individual training is simply not a viable option due to its poor performance. In this area, the approach described in this article can be a good alternative. The sharing of the original latent vector does not pose any challenges for the offline IODT phase training use case, but may be burdensome for the over-the-air training dataset sharing use case. In the case of the over-the-air DS sharing use case, it may be beneficial to switch to a lower CR with a higher bit resolution (e.g., CR8 with 1 bit / element means CR16 with 2 bits / element) with equal overall CSI feedback overhead. Sharing quantized latent vectors can save the amount of over-the-air traffic while guaranteeing model training / inference performance.

[0129] Fig.10 The empirical distribution of (original and quantized) latent vector elements with a TF-based encoder-decoder (each element has uniform SQ with {1,2,3} bits of resolution) is shown.

[0130] Fig.11 Inference results (SGCS) are shown for various training proposals (scheme described in the paper: “sepNwFirstwRawDS” - separate NW-first training shared with the original dataset; x-axis: [52*164*1104*152*264*3], where latent vector dimension * number of Q-bits per element).

[0131] Fig.12 Inference results (SGCS) for various training proposals with a linear y-axis are shown.

[0132] The method described herein may be related to Rel-18 and may be used by both UE side and NW side for NW-first separate training scenarios to enable the CSI enhancement feature of AIML.

[0133] The methods described herein may be specified in 3GPP specifications (expected: TS38.214) as part of the procedures for CSI processing. There is a possibility that this level of training collaboration details is left to vendor-to-vendor alignment, and the NW vendor may implement the scheme described herein for NW-first training scenarios at the IODT. In this case, the corresponding UE vendor may adjust its solution accordingly. The scheme described herein requires a shared training dataset. As Fig.10 It is clearly shown in e and z q The distribution is clearly different.

[0134] The UE-gNB signaling operations described herein for sharing of over-the-air training datasets as a whole or in part may be adopted by 3GPP (expected: TS 38.214) as one of the expected use cases.

[0135] Fig.13 A block diagram of one possible and non-limiting example of a cellular network 1 connected to a user equipment (UE) 10 is shown. Fig.13 A number of network elements are shown in the cellular network of FIG: a base station 70 ; and a core network 90 .

[0136] exist Fig.13 , a user equipment (UE) 10 wirelessly communicates with a base station 70 of a cellular network 1 via a radio link 11. The UE 10 is a wireless communication device, such as a mobile device, configured to access a cellular network. The UE 10 is shown with one or more antennas 28. The ellipsis 2 indicates that there may be multiple UEs 10 wirelessly communicating with the base station 70 via a radio link. The UE 10 includes one or more processors 13, one or more memories 15, and other circuits 16. Another circuit 16 includes one or more receivers (Rx) 17 and one or more transmitters (Tx) 18. The program 12 is used to cause the UE 10 to perform the operations described herein. For the UE 10, the other circuits 16 may include circuits such as for user interface elements (not shown) such as a display.

[0137] As a network element of the cellular network 1, the base station 70 provides the UE 10 with access to the cellular network 1 and the data network 91 via the core network 90 (e.g., via the user plane function (UPF) of the core network 90). The base station 70 is shown as having one or more antennas 58. Generally, the base station 70 is referred to as a RAN node 70 in this document. An example of a RAN node 70 is a gNB. However, there are many other examples of RAN nodes, including eNBs (LTE base stations) or transmission reception points (TRPs). The base station 70 includes one or more processors 73, one or more memories 75, and other circuits 76. The other circuits 76 include one or more receivers (Rx) 77 and one or more transmitters (Tx) 78. The program 72 is used to enable the base station 70 to perform the operations described herein.

[0138] Note that base station 70 may alternatively be implemented via other wireless technologies, such as Wi-Fi (a wireless network protocol used by devices to communicate without a direct cable connection). In the case of Wi-Fi, link 11 may be characterized as a wireless link.

[0139] Two or more base stations 70 communicate using, for example, a link 79. Link 79 may be wired or wireless or both, and may implement, for example, an Xn interface for fifth generation (5G), an X2 interface for LTE, or other suitable interfaces for other standards.

[0140] The cellular network 1 may include a core network 90, as a third illustrated one or more elements, which may include core network functions and provide a connection to a data network 91, such as a telephone network and / or a data communications network (e.g., the Internet), via one or more links 81. The core network 90 includes one or more processors 93, one or more memories 95, and other circuits 96. The other circuits 96 include one or more receivers (Rx) 97 and one or more transmitters (Tx) 98. The program 92 is used to enable the core network 90 to perform the operations described herein.

[0141] The core network 90 may be a 5GC (5G core network). The core network 90 may implement or include a plurality of network functions (NFs) 99, and the program 92 may include one or more of the NFs 99. The 5G core network may use hardware such as memory and processors and a virtualization layer. It may be a single independent computing system, a distributed computing system, or a cloud computing system. The NF 99 as a network element of the core network may be a container or a virtual machine running on the hardware of the computing system constituting the core network 90.

[0142] The core network functions for 5G may include access and mobility management functions provided by network functions 99 such as access and mobility management functions (AMF), session management functions provided by network functions such as session management functions (SMF). The core network functions for access and mobility management in LTE networks may be provided by MME (mobility management entity) and / or SGW (serving gateway) functions, which route data to the data network. Many other functions are possible, such as Fig.13 As shown in the example in: AMF; SMF; MME; SGW; Gateway Mobile Location Center (GMLC); Location Management Function (LMF); Unified Data Management (UDM); Unified Data Repository (UDR); Network Repository Function (NRF); and / or Evolved Serving Mobile Location Center (E-SMLC). These are merely exemplary core network functions that may be provided by the core network 90, and note that both 5G and LTE core network functions may be provided by the core network 90. ​​A radio access network (RAN) node 70 is coupled to the core network 90 via a backhaul link 31. The RAN node 70 and the core network 90 may include an NG interface for 5G, or an S1 interface for LTE, or other suitable interfaces for other radio access technologies for communicating via the backhaul link 31.

[0143] In the data network 91, there is a computer-readable medium 94. The computer-readable medium 94 contains instructions, which, when downloaded and installed into the memory 15, memory 75 or memory 95 of the corresponding UE 10, base station 70 and / or core network element 90 and executed by the processor 13, processor 73 or processor 93, enable the corresponding device to perform the corresponding actions described herein. The computer-readable medium 94 can be implemented in other forms, such as via an optical disk or a memory stick.

[0144] Programs 12, 72 and 92 contain instructions stored by corresponding one or more memories 15, 75 or 95. These instructions, when executed by corresponding one or more processors 13, 73 or 93, cause corresponding devices 10, 70 or 90 to perform the operations described herein. Computer-readable memory 15, 75 or 95 may be of any type suitable for the local technical environment and may be implemented using any suitable data storage technology, such as semiconductor-based memory devices, flash memory, firmware, magnetic memory devices and systems, optical memory devices and systems, fixed memory and removable memory. Computer-readable memory 15, 75 and 95 may be a device for performing storage functions. As non-limiting examples, processors 13, 73 and 93 may be of any type suitable for the local technical environment and may include one or more of a general-purpose computer, a special-purpose computer, a microprocessor, a digital signal processor (DSP) and a processor based on a multi-core processor architecture. Processors 13, 73 and 93 may be components for causing their respective devices to perform functions such as those described herein.

[0145] Receivers 17, 77 and 97 and transmitters 18, 78 and 98 may implement a wired or wireless interface. Receivers and transmitters may be grouped together as a transceiver.

[0146] Fig.14An example apparatus 1400 configured to implement the examples described herein that can be implemented in hardware. The apparatus 1400 includes at least one processor 1402 (e.g., an FPGA and / or a CPU), one or more memories 1404 including computer program code 1405, the computer program code 1405 having instructions for performing the methods described herein, wherein at least one memory 1404 and the computer program code 1405 are configured to, together with at least one processor 1402, cause the apparatus 1400 to implement circuits, processes, components, modules, or functions (implemented using a control module 1406) to implement the examples described herein, including NW-first individual sequential training with an original data set sharing scheme for AIML-enabled CSI compression. The memory 1404 can be a non-transitory memory, a temporary memory, a volatile memory (e.g., a RAM), or a non-volatile memory (e.g., a ROM). The training 1430 of the control module implements aspects described herein related to NW-first individual sequential training, wherein the original data set sharing scheme is used for AIML-enabled CSI compression.

[0147] The device 1400 includes a display and / or I / O interface 1408, which includes user interface (UI) circuits and elements that can be used to display aspects or states of the methods described herein (e.g., while one of the methods is being performed or at a subsequent time), or to receive input from a user, such as using a keypad, camera, touch screen, touch area, microphone, biometrics, one or more sensors, etc. The communication I / F 1410 can be wired and / or wireless and communicate via any communication technology (including via one or more links 1424 over the Internet / other network). Link 1424 can be from Fig.13 Links 11 and / or 79 and / or 31 and / or 81. Fig.13 The links 11 and / or 79 and / or 31 and / or 81 may also be implemented using the transceiver 1416 and the corresponding wireless link 1426. The communication I / F 1410 may include one or more transmitters or one or more receivers.

[0148] The transceiver 1416 includes one or more transmitters 1418 and one or more receivers 1420. The transceiver 1416 and / or the communication I / F 1410 may include standard well-known components such as amplifiers, filters, frequency converters, (de)modulators, and encoder / decoder circuits, and one or more antennas such as antenna 1414 for communicating over a wireless link 1426.

[0149] The control module 1406 of the device 1400 includes one or both of the parts 1406-1 and / or 1406-2, which can be implemented in a variety of ways. The control module 1406 can be implemented in hardware as the control module 1406-1, such as being implemented as a part of one or more processors 1402. The control module 1406-1 can also be implemented as an integrated circuit or by other hardware such as a programmable gate array. In another example, the control module 1406 can be implemented as a control module 1406-2, which is implemented as a computer program code (with corresponding instructions) 1405 and is executed by one or more processors 1402. For example, one or more memories 1404 store instructions that, when executed by one or more processors 1402, cause the device 1400 to perform one or more operations as described herein. In addition, one or more processors 1402, one or more memories 1404, and example algorithms (e.g., as flow charts and / or signaling diagrams) encoded as instructions, programs, or codes are components for causing the operations described herein to be performed.

[0150] The device 1400 for implementing the function of controlling 1406 may be UE 10, base station 70 (e.g., gNB 70), or core network 90 including any network function in network function 99, which may be implemented by a network entity. Therefore, processor 1402 may correspond to processor 13, processor 73, and / or processor 93, memory 1404 may correspond to one or more memories 15, one or more memories 75, and / or one or more memories 95, computer program code 1405 may correspond to program 12, program 72, or program 92, communication I / F 1410 and / or transceiver 1416 may correspond to other circuits 16, other circuits 76, or other circuits 96, and antenna 1414 may correspond to antenna 28 or antenna 58.

[0151] Alternatively, apparatus 1400 and its elements may not correspond to any of UE 10, base station 70, or core network and their respective elements, as apparatus 1400 may be part of a Self-Organizing / Optimizing Network (SON) node or other node, such as a node in a cloud.

[0152] The apparatus 1400 may also be distributed throughout the network (eg, 91 ), including within and between the apparatus 1400 and any network elements (eg, core network 90 and / or base station 70 and / or UE 10 ).

[0153] Interface 1412 enables data communication and signaling between various items of apparatus 1400, such as Fig.14As shown. For example, interface 1412 can be one or more buses, such as an address, data, or control bus, and can include any interconnect mechanism, such as a series of lines on a motherboard or integrated circuit, optical fiber or other optical communication device, etc. The computer program code (e.g., instructions) 1405 including control 1406 may include object-oriented software that is configured to pass data or messages between objects within computer program code 1405. Device 1400 need not include each of the features mentioned, or may also include other features. The various components of device 1400 may reside at least partially in a common housing 1428, or a subset of the various components of device 1400 may be at least partially located in a different housing, which may include housing 1428.

[0154] Fig.15 Schematic representations of non-volatile memory media 1500a (e.g., a computer / compact disk (CD) or digital versatile disk (DVD)) and 1500b (e.g., a universal serial bus (USB) memory stick) and 1500c (e.g., cloud storage for downloading instructions and / or parameters 1502 or receiving instructions and / or parameters 1502 sent by email) storing instructions and / or parameters 1502 that, when executed by a processor, enable the processor to perform one or more steps of the methods described herein.

[0155] Fig.16 is an example method 1600 based on the example embodiments described herein. At 1610, the method includes training an artificial intelligence decoder using at least an input to a hypothetical artificial intelligence encoder. At 1620, the method includes determining a training data set, the training data set including the input to the hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback. At 1630, the method includes sending the training data set to a user equipment. At 1640, the method includes determining reconstructed channel state information using the trained artificial intelligence decoder. Method 1600 may be performed by a network node 70, a network entity of a core network 90 (e.g., performed by a network function 99), or an apparatus 1400.

[0156] Fig.17The method 1700 is based on the example embodiments described herein. At 1710, the method includes receiving a training data set from a network, the training data set including an input to a hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback. At 1720, the method includes training the artificial intelligence encoder using the training data set. At 1730, the method includes encoding the channel state information using the trained artificial intelligence encoder. The method 1700 may be performed by the user equipment 10 or the apparatus 1400.

[0157] The following examples are provided and described herein.

[0158] Example 1. An apparatus, comprising: at least one processor; and at least one memory storing instructions, wherein when executed by the at least one processor, the instructions cause the apparatus to at least: train an artificial intelligence decoder using at least an input of a hypothetical artificial intelligence encoder; determine a training data set, the training data set comprising the input of the hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; send the training data set to a user equipment; and determine reconstructed channel state information using the trained artificial intelligence decoder.

[0159] Example 2. The apparatus according to Example 1, wherein the input of the hypothetical artificial intelligence encoder includes at least one of the following: channel information, or a channel matrix, or a channel eigenvector, or a precoding matrix in the spatial frequency domain.

[0160] Example 3. The apparatus of any one of Examples 1 to 2, wherein the unquantized predicted channel state information feedback comprises a latent vector.

[0161] Example 4. The apparatus of any one of Examples 1 to 3, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: determine quantization and dequantization rules; and send the quantization and dequantization rules to the user equipment.

[0162] Example 5. An apparatus according to any one of Examples 1 to 4, wherein the instruction, when executed by the at least one processor, causes the apparatus to at least: send feedback of the dequantized predicted channel state information in the training data set to the user equipment when quantization and dequantization rules are not sent to the user equipment, or when the codebook and mapping scheme are not sent to the user equipment.

[0163] Example 6. The apparatus of any one of Examples 1 to 5, wherein the unquantized predicted channel state information feedback has a finer granularity than the dequantized predicted channel state information feedback.

[0164] Example 7. The apparatus according to any one of Examples 1 to 6, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: train the artificial intelligence decoder using a loss function, wherein input parameters of the loss function include the input of the assumed artificial intelligence encoder and the reconstructed channel state information.

[0165] Example 8. An apparatus according to any one of Examples 1 to 7, wherein the instructions, when executed by the at least one processor, cause the apparatus to receive at least one capability of the user device from the user device, wherein the at least one capability of the user device includes at least one of the following: a supported model identifier, or a capability for over-the-air dataset sharing, or a capability for general encoder training, or a configuration format for benchmark truth data sharing.

[0166] Example 9. An apparatus according to any one of Examples 1 to 8, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: receive a reference truth input of an artificial intelligence encoder from the user device; use the reference truth input of the artificial intelligence encoder to train the artificial intelligence decoder; and send the training data set to the user device, the training data set including at least one of the following: the unquantized predicted channel state information feedback, or the dequantized predicted channel state information feedback.

[0167] Example 10. The apparatus according to Example 9, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: send the trained quantization-dequantization rule and the training data set to the user equipment, the training data set comprising at least one of the following: the unquantized predicted channel state information feedback, or the dequantized predicted channel state information feedback.

[0168] Example 11. An apparatus according to any one of Examples 1 to 10, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: receive an instruction from the user device to complete training of the artificial intelligence encoder using the training data set.

[0169] Example 12. An apparatus comprising: at least one processor; and at least one memory storing instructions, wherein when executed by the at least one processor, the apparatus at least: receives a training data set from a network, the training data set comprising an input of a hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; trains the artificial intelligence encoder using the training data set; and encodes channel state information using the trained artificial intelligence encoder.

[0170] Example 13. The apparatus according to Example 12, wherein the input of the hypothetical artificial intelligence encoder includes at least one of the following: channel information, or a channel matrix, or a channel eigenvector, or a precoding matrix in the spatial frequency domain.

[0171] Example 14. The apparatus of any one of Examples 12 to 13, wherein the unquantized predicted channel state information feedback comprises a latent vector.

[0172] Example 15. The apparatus according to any one of Examples 12 to 14, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: receive quantization and dequantization rules from the network; and quantize the encoded channel state information and dequantize the quantized encoded channel state information based on the quantization and dequantization rules.

[0173] Example 16. An apparatus according to any one of Examples 12 to 15, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: receive the dequantized predicted channel state information feedback in the training data set from the network when quantization and dequantization rules are not received from the network, or when a codebook and mapping scheme are not received from the network; and learn a quantizer or generate a codebook based on the unquantized predicted channel state information feedback and the dequantized predicted channel state information feedback; wherein the learning of the quantizer or the generation of the codebook is performed when the unquantized predicted channel state information feedback and the dequantized predicted channel state information feedback are in the same data set or in different data sets.

[0174] Example 17. The apparatus of any one of Examples 12 to 16, wherein the unquantized predicted channel state information feedback has a finer granularity than the dequantized predicted channel state information feedback.

[0175] Example 18. The apparatus of any one of Examples 12 to 17, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: train the artificial intelligence encoder using a loss function, wherein input parameters of the loss function include the unquantized predicted channel state information feedback from the received training data set and the unquantized channel state information feedback obtained at the output of the artificial intelligence encoder.

[0176] Example 19. The apparatus of any one of Examples 12 to 18, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least send at least one capability of the apparatus to the network, wherein the at least one capability of the apparatus comprises at least one of the following: supported model identifiers, or capabilities for over-the-air dataset sharing, or capabilities for general encoder training, or configuration formats for benchmark truth data sharing.

[0177] Example 20. The apparatus of any one of Examples 12 to 19, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: send a reference truth input of the artificial intelligence encoder to the network; and receive the training data set from the network, the training data set comprising at least one of: the unquantized predicted channel state information feedback or the dequantized predicted channel state information feedback.

[0178] Example 21. The apparatus according to Example 20, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: receive a trained quantization-dequantization rule and the training data set from the network, the training data set comprising at least one of the following: the unquantized predicted channel state information feedback or the dequantized predicted channel state information feedback.

[0179] Example 22. An apparatus according to any one of Examples 12 to 21, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: send an instruction to the network to complete the training of the artificial intelligence encoder using the training data set.

[0180] Example 23. A method comprising: training an artificial intelligence decoder using at least an input of a hypothetical artificial intelligence encoder; determining a training data set, the training data set comprising the input of the hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback, or dequantized predicted channel state information feedback; sending the training data set to a user equipment; and determining reconstructed channel state information using the trained artificial intelligence decoder.

[0181] Example 24. A method, comprising: receiving a training data set from a network, the training data set comprising an input of a hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; using the training data set to train an artificial intelligence encoder; and using the trained artificial intelligence encoder to encode channel state information.

[0182] Example 25. An apparatus comprising: means for training an artificial intelligence decoder using at least an input of a hypothetical artificial intelligence encoder; means for determining a training data set, the training data set comprising the input of the hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; means for sending the training data set to a user equipment; and means for determining reconstructed channel state information using the trained artificial intelligence decoder.

[0183] Example 26. An apparatus comprising: means for receiving a training data set from a network, the training data set comprising an input to a hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; means for training an artificial intelligence encoder using the training data set; and means for encoding channel state information using the trained artificial intelligence encoder.

[0184] Example 27. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine for performing operations, the operations comprising: training an artificial intelligence decoder using at least an input of a hypothetical artificial intelligence encoder; determining a training data set, the training data set comprising the input of the hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; sending the training data set to a user device; and determining reconstructed channel state information using the trained artificial intelligence decoder.

[0185] Example 28. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine for performing operations, the operations comprising: receiving a training data set from a network, the training data set comprising an input to a hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; training an artificial intelligence encoder using the training data set; and encoding channel state information using the trained artificial intelligence encoder.

[0186] References to "computers", "processors", etc. should be understood to cover not only computers with different architectures (such as single / multi-processor architectures and sequential or parallel architectures), but also special-purpose circuits (such as field programmable gate arrays (FPGAs), application-specific circuits (ASICs), signal processing devices and other processing circuits). References to computer programs, instructions, code, etc. should be understood to cover software for programmable processors or firmware, such as, for example, the programmable content of hardware devices, whether instructions for a processor or configuration settings for a fixed-function device, gate array or programmable logic device, etc.

[0187] The memory as described herein can be implemented using any suitable data storage technology, such as semiconductor-based memory devices, flash memory, magnetic memory devices and systems, optical memory devices and systems, non-transitory memory, temporary memory, fixed memory, and removable memory. The memory can include a database for storing data.

[0188] As used herein, the term "circuitry" may refer to the following: (a) a hardware circuit implementation, such as an implementation in analog and / or digital circuits, and (b) a combination of circuitry and software (and / or firmware), such as (where applicable): (i) a combination of processors or (ii) portions of processors / software, including digital signal processors, software, and memory that work together to enable the device to perform various functions, and (c) circuitry, such as a microprocessor or a portion of a microprocessor, that requires software or firmware to operate, even if the software or firmware is not physically present. As another example, as used herein, the term "circuitry" would also cover an implementation of only a processor (or multiple processors) or a portion of a processor and its (or its) accompanying software and / or firmware. The term "circuitry" would also cover, for example and if applicable to the particular element, a baseband integrated circuit or an application processor integrated circuit for a mobile phone or a similar integrated circuit in a server, cellular network device, or another network device.

[0189] It should be understood that the foregoing description is illustrative only. Various alternatives and modifications may be devised by those skilled in the art. For example, the features described in the various dependent claims may be combined with each other in any suitable combination. In addition, features from the different example embodiments described above may be selectively combined into new example embodiments. Therefore, this specification is intended to cover all such alternatives, modifications and variations that fall within the scope of the appended claims.

[0190] The following abbreviations and abbreviations that may be found in the specification and / or the drawings are given below (the abbreviations and abbreviations may be appended to each other or to other characters using, for example, dashes, hyphens, slashes or numbers, and may be non-case sensitive):

[0191] 3GPP Third Generation Partnership Project

[0192] 4G Fourth Generation

[0193] 5G Fifth Generation

[0194] 5GC 5G Core Network

[0195] AI

[0196] AIML Artificial Intelligence Machine Learning

[0197] AMF Access and Mobility Management Function

[0198] Ant

[0199] API Application Programming Interface

[0200] ASIC Application-Specific Integrated Circuit

[0201] BLER Block Error Rate

[0202] BS Base Station

[0203] CD Compact / Computer Disk

[0204] CE control components

[0205] CH Channel

[0206] CPU Central Processing Unit

[0207] CR Compression ratio

[0208] CSI Channel State Information

[0209] CSI-RS Channel State Information Reference Signal

[0210] DCI Downlink Control Information

[0211] DEC decoder

[0212] DFT Discrete Fourier Transform

[0213] dim dimension

[0214] DL Downlink

[0215] DS dataset

[0216] DSP Digital Signal Processor

[0217] dQ Dequantization

[0218] DVD Digital Versatile Disc

[0219] E2E End to End

[0220] eNB Evolved Node B (e.g., LTE base station)

[0221] ENC Encoder

[0222] EPC Evolved Packet Core

[0223] E-SMLC Evolved Serving Mobile Location Center

[0224] FB Feedback

[0225] FPGA Field Programmable Gate Array

[0226] GMLC Gateway Mobile Location Center

[0227] gNB Next generation Node B, 5G / NR base station, that is, the node that provides NR user plane and control plane protocol terminals to UE, connected to 5GC through the NG interface.

[0228] ID

[0229] I / F Interface

[0230] I / O Input / Output

[0231] IODT Interoperability Development Test

[0232] KPI performance indicator

[0233] LCM Lifecycle Management

[0234] LMF Location Management Function

[0235] LTE Long Term Evolution (4G)

[0236] MAC Media Access Control

[0237] MME Mobility Management Entity

[0238] MSE Mean Squared Error

[0239] NF Network Function

[0240] NG New Generation

[0241] NG-RAN Next Generation Radio Access Network

[0242] NN Neural Network

[0243] NR Next Generation Wireless

[0244] NRF Network Storage Function

[0245] NW Network

[0246] N / W Network

[0247] Opt. Optional

[0248] PDSCH Physical Downlink Shared Channel

[0249] PMI Precoding matrix indicator

[0250] prj.forecast

[0251] PUSCH Physical Uplink Shared Channel

[0252] Q Quantization

[0253] Q-aware

[0254] Q-dQ Quantization and Dequantization

[0255] RAM Random Access Memory

[0256] RAN Radio Access Network

[0257] RAN1 Radio Layer 1

[0258] RAN1#RAN1 Meeting

[0259] Rel Release

[0260] Rev

[0261] RF

[0262] RFIC Radio Frequency Integrated Circuit

[0263] ROM Read Only Memory

[0264] RRC Radio Resource Control

[0265] RS reference signal

[0266] Rx Receiver or Receiver

[0267] S1 connects the eNB to the S1 interface of the EPC

[0268] sb subband or subband

[0269] SCS Subcarrier Spacing

[0270] sep segmentation

[0271] sepNwFirstwFnT NW first single training with freezing and training

[0272] sepNwFirstwRawDS NW first separate training with original dataset sharing

[0273] CSGS Squared Generalized Cosine Similarity

[0274] SGW Service Gateway

[0275] SLS System Level Simulation

[0276] SMF session management functions

[0277] SON self-organizing / optimizing network

[0278] SQ Scalar Quantization

[0279] SVD Singular Value Decomposition

[0280] TDOC Technical Documentation

[0281] TF Transformer

[0282] TRP Transmission Reception Point

[0283] TS Technical Specifications

[0284] Tx Transmitter or Transmission

[0285] UDM Unified Data Management

[0286] UDR Unified Data Repository

[0287] UE User Equipment (e.g., wireless, usually mobile)

[0288] UI User Interface

[0289] UMA City Macro

[0290] UPF User Plane Function

[0291] USB Universal Serial Bus

[0292] VQ Vector Quantization

[0293] Wi-Fi wireless networking protocol used by Wi-Fi devices to communicate without a direct cable connection

[0294] X2 network interface between RAN nodes and between RAN and core network

[0295] Xn Network interface between NG-RAN nodes

Claims

1. A device for communication, comprising: at least one processor; as well as at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to at least: receiving a training data set from a network, the training data set comprising an input of a hypothetical artificial intelligence encoder and at least one of: unquantized predicted channel state information feedback or dequantized predicted channel state information feedback; Using the training data set to train an artificial intelligence encoder; as well as The channel state information is encoded using a trained artificial intelligence encoder.

2. The apparatus of claim 1, wherein the input to the hypothetical artificial intelligence encoder comprises at least one of: Channel information, or channel matrix, or The channel eigenvector, or Precoding matrix in the spatial frequency domain.

3. The apparatus according to any one of claims 1 to 2, wherein the unquantized predicted channel state information feedback comprises a latent vector.

4. The apparatus according to any one of claims 1 to 2, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: receiving quantization and dequantization rules from the network; Based on the quantization and dequantization rules, the encoded channel state information is quantized and the quantized encoded channel state information is dequantized.

5. The apparatus according to any one of claims 1 to 2, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: When no quantization and dequantization rules are received from the network, or when no codebook and mapping scheme are received from the network, receiving the dequantized predicted channel state information feedback in the training data set from the network; and learning a quantizer or generating a codebook based on the unquantized predicted channel state information feedback and the dequantized predicted channel state information feedback; Wherein the learning of the quantizer or the generation of the codebook is performed when the unquantized predicted channel state information feedback and the dequantized predicted channel state information feedback are in the same data set or in different data sets.

6. The apparatus according to any one of claims 1 to 2, wherein the unquantized predicted channel state information feedback has a finer granularity than the dequantized predicted channel state information feedback.

7. The apparatus according to any one of claims 1 to 2, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: The artificial intelligence encoder is trained using a loss function, wherein input parameters of the loss function include the unquantized predicted channel state information feedback from the received training data set and the unquantized channel state information feedback obtained at the output of the artificial intelligence encoder.

8. The apparatus according to any one of claims 1 to 2, wherein when the instructions are executed by the at least one processor, the apparatus at least sends at least one capability of the apparatus to the network, wherein the at least one capability of the apparatus comprises at least one of the following: A supported model identifier, or Capabilities for aerial dataset sharing, or Capabilities for general encoder training, or Configuration format for benchmark truth data sharing.

9. The apparatus according to any one of claims 1 to 2, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: sending a ground truth input of the artificial intelligence encoder to the network; and The training data set is received from the network, wherein the training data set includes at least one of the following: the unquantized predicted channel state information feedback or the dequantized predicted channel state information feedback.

10. The apparatus of claim 9, wherein the instructions, when executed by the at least one processor, cause the apparatus to at least: A trained quantization-dequantization rule and the training data set are received from the network, the training data set comprising at least one of: the unquantized predicted channel state information feedback or the dequantized predicted channel state information feedback.

Images