Training framework

Abstract

Example embodiments of the present disclosure relate to a training framework. In an example method, a terminal device receives, from a network device, a training dataset for training an encoder at the terminal device and a distance threshold. The encoder is associated with a decoder at the network device, the distance threshold is used to evaluate a distance loss in the training of the encoder, and the distance loss is associated with a similarity performance related to the encoder and the decoder. Then, the terminal device trains the encoder based on the training dataset and the distance threshold. In this way, end-to-end (E2E) SGCS performance for CSI reconstruction can be ensured at the terminal device and the network device, and the overhead waste of the training dataset can be avoided or reduced.

Description

Technical Field

[0001] The exemplary embodiments of this disclosure generally relate to the field of communications, and more specifically to terminal devices, network devices, methods, apparatuses, and computer-readable media for training frameworks. Background Technology

[0002] A communication network can be viewed as a facility that enables communication between two or more communication devices, or provides access from communication devices to a data network. Mobile or wireless communication networks are an example of communication networks.

[0003] Such communication networks operate according to standards (such as those issued by 3GPP (3rd Generation Partnership Project) or ETSI (European Telecommunications Standards Institute). Examples of such standards include the so-called 5G (fifth generation) standard or other standards issued by 3GPP. Summary of the Invention

[0004] In general, the exemplary embodiments of this disclosure provide a solution for training frameworks, particularly for NW-priority split training frameworks.

[0005] In a first aspect, a terminal device is provided. The terminal device includes: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the terminal device to: receive from a network device a training dataset and a distance threshold for training an encoder at the terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and train the encoder based on the training dataset and the distance threshold.

[0006] In a second aspect, a network device is provided. The network device includes: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the terminal device to: determine a training dataset and a distance threshold for training an encoder at the terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and transmit the training dataset and the distance threshold to the terminal device.

[0007] In a third aspect, a method is provided. The method includes: receiving, at a terminal device and from a network device, a training dataset and a distance threshold for training an encoder at the terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate a distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and training the encoder based on the training dataset and the distance threshold.

[0008] In a fourth aspect, a method is provided. The method includes: determining, by a network device, a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold being used to evaluate distance loss during encoder training, and the distance loss being associated with similarity performance, which is related to both the encoder and the decoder; and transmitting the training dataset and the distance threshold at the network device and to the terminal device.

[0009] In a fifth aspect, an apparatus is provided. The apparatus includes: components for receiving a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at a network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and components for training the encoder based on the training dataset and the distance threshold.

[0010] In a sixth aspect, an apparatus is provided. The apparatus includes: components for determining a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at a network device, the distance threshold being used to evaluate distance loss during encoder training, and the distance loss being associated with similarity performance related to both the encoder and the decoder; and components for transmitting the training dataset and the distance threshold.

[0011] In a seventh aspect, a non-transitory computer-readable medium is provided. The non-transitory computer-readable medium has instructions stored thereon. When the instructions are executed on at least one processor, the at least one processor causes the processor to perform the method of the third or fourth aspect.

[0012] In an eighth aspect, a computer program is provided, including instructions, wherein when executed by a device, the instructions cause the device to at least: receive from a network device a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold being used to evaluate distance loss during the training of the encoder, and the distance loss being associated with similarity performance, which is related to both the encoder and the decoder; and train the encoder based on the training dataset and the distance threshold.

[0013] In a ninth aspect, a computer program is provided, including instructions, wherein when executed by a device, the instructions cause the device to at least: determine a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at a network device, the distance threshold being used to evaluate distance loss during the training of the encoder, and the distance loss being associated with similarity performance, which is related to both the encoder and the decoder; and send the training dataset and the distance threshold to the terminal device.

[0014] In a tenth aspect, a terminal device is provided. The terminal device includes: a receiving circuitry system configured to receive from a network device a training dataset and a distance threshold for training an encoder at the terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and a training circuitry system configured to train the encoder based on the training dataset and the distance threshold.

[0015] In an eleventh aspect, a network device is provided. The network device includes: a determining circuit system configured to determine a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and a transmitting circuit system configured to transmit the training dataset and the distance threshold to the terminal device.

[0016] It should be understood that the summary section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0017] Some exemplary embodiments will now be described with reference to the accompanying drawings, in which:

[0018] Figure 1A The illustration shows an example network environment in which some embodiments of this disclosure may be implemented;

[0019] Figure 1B The illustration shows a schematic diagram of the relationship between the performance of squared generalized cosine similarity (SGCS) and different training data sizes according to some embodiments of the present disclosure;

[0020] Figure 2 The diagram illustrates a signaling process according to some embodiments of the present disclosure;

[0021] Figure 3AThe illustration is a schematic diagram of the evaluation of the relationship between the mean square error (MSE) of the encoder output channel state information (CSI) codeword and the SGCS performance of the reconstructed CSI according to some embodiments of the present disclosure.

[0022] Figure 3B The illustration shows a schematic diagram of the convergence monotonic relationship between MSE and SGCS according to some embodiments of the present disclosure;

[0023] Figure 3C Another schematic diagram illustrating the convergence monotonic relationship between MSE and SGCS according to some embodiments of the present disclosure is shown.

[0024] Figure 4 The diagram illustrates a signaling process according to some exemplary embodiments of the present disclosure;

[0025] Figure 5 The illustration depicts a schematic workflow for deriving an example MSE-SGCS relationship according to some embodiments of the present disclosure;

[0026] Figure 6 The illustration shows a schematic diagram illustrating the relationship between MSE and different training data sizes according to some embodiments of the present disclosure;

[0027] Figure 7 The illustration shows a flowchart of another example method implemented at a terminal device according to some embodiments of the present disclosure;

[0028] Figure 8 The illustration shows a flowchart of yet another example method implemented at a network device according to some embodiments of the present disclosure;

[0029] Figure 9 The illustration shows a simplified block diagram of a device suitable for implementing some example embodiments of the present disclosure; and

[0030] Figure 10 A block diagram illustrating an example of a computer-readable medium according to some exemplary embodiments of the present disclosure is shown.

[0031] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. Detailed Implementation

[0032] The principles of this disclosure will now be described with reference to some exemplary embodiments. It should be understood that these embodiments are described for illustrative purposes and to assist those skilled in the art in understanding and implementing this disclosure, and do not imply any limitation on the scope of this disclosure. The disclosure described herein can be implemented in various ways other than those described below.

[0033] In the following description and claims, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0034] In this disclosure, references to "an embodiment," "embodiment," and "example embodiment," etc., indicate that the described embodiment may include a particular feature, structure, or characteristic, but not every embodiment must include that particular feature, structure, or characteristic. Furthermore, such phrases do not necessarily refer to the same embodiment. Moreover, when a particular feature, structure, or characteristic is described in connection with an embodiment, those skilled in the art will understand that, whether explicitly described or not, combining it with other embodiments to affect such a feature, structure, or characteristic is within the knowledge of those skilled in the art.

[0035] It should be understood that although the terms “first” and “second” may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. As used herein, the term “and / or” includes any and all combinations of one or more of the listed terms.

[0036] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. The singular forms “a,” “an,” and “the” used herein also include the plural forms unless the context clearly indicates otherwise. Further understanding, the terms “comprises,” “comprising,” “has,” “having,” “includes,” and / or “including”, when used herein, specify the presence of the stated features, elements, and / or components, but do not exclude the presence or addition of one or more other features, elements, components, and / or combinations thereof. As used herein, “at least one of the following: ” and “at least one of ” and similar wording (where a list of two or more elements is connected by “and” or “or”) means at least any one of these elements, or at least any two or more of these elements, or at least all of these elements.

[0037] As used in this application, the term "circuit system" may refer to one or more or all of the following: (a) Pure hardware circuit implementation (such as implementation using only analog and / or digital circuit systems), and (b) A combination of hardware circuitry and software, such as (if applicable): (i) A combination of (multiple) analog and / or digital hardware circuits and software / firmware, and (ii) Any part of a hardware processor(s) having software (including (multiple) digital signal processors), software, and (multiple) memories, which work together to enable a device (such as a mobile phone or server) to perform various functions, and (c) (Multiple) hardware circuits and / or (multiple) processors, such as (multiple) microprocessors or a portion thereof, which require software (e.g., firmware) to operate, but may be absent when operation does not require the software.

[0038] This definition of circuit system applies to all uses of the term in this application, including in any claim. As another example, as used in this application, the term circuit system also covers only hardware circuitry or a processor (or processors) or a portion of hardware circuitry or a processor and its accompanying software and / or firmware. For example, if applicable to a particular claim element, the term circuit system also covers baseband integrated circuits or processor integrated circuits for mobile devices, or similar integrated circuits in servers, cellular network devices, or other computing or network devices.

[0039] As used herein, the term "communication network" refers to a network that conforms to any suitable communication standard, such as Long Term Evolution (LTE), LTE-A Advanced (LTE-A), Wideband Code Division Multiple Access (WCDMA), High-Speed ​​Packet Access (HSPA), Narrowband Internet of Things (NB-IoT), etc. Furthermore, communication between terminal devices and network devices in a communication network can be performed according to any suitable generation of communication protocol, including but not limited to fourth-generation (4G), 4.5G, future fifth-generation (5G), the IEEE 802.11 communication protocol, and / or any other currently known or to be developed in the future. Embodiments of this disclosure can be applied to various communication systems. Given the rapid development of communications, there will naturally be communication technologies and systems of future types that can embody this disclosure. This should not be construed as limiting the scope of this disclosure to the systems described above.

[0040] As used herein, the term "network device" refers to a node in a communication network through which terminal devices access the network and receive services. Depending on the terminology and technology applied, a network device can refer to a base station (BS) or access point (AP), such as a Node B (NodeB or NB), an evolved Node B (eNodeB or eNB), an NR NB (also known as a gNB), a Remote Radio Unit (RRU), a Radio Header (RH), a Remote Radio Header (RRH), a WiFi device, a repeater, a low-power node (such as a femtosecond or picosecond), etc. In the following description, the terms "network device," "AP device," "AP," and "access point" are used interchangeably.

[0041] The term "terminal device" refers to any terminal device capable of wireless communication. As an example and not a limitation, a terminal device may also be referred to as a communication device, user equipment (UE), subscriber station (SS), portable subscriber station, mobile station (MS), station (STA), or station equipment, or access terminal (AT). Terminal devices may include, but are not limited to, mobile phones, cellular phones, smartphones, VoIP phones, wireless local loop phones, tablets, wearable terminal devices, personal digital assistants (PDAs), portable computers, desktop computers, image capture terminal devices (such as digital cameras), gaming terminal devices, music storage and playback devices, in-vehicle wireless terminal devices, wireless endpoints, mobile stations, laptop embedded devices (LEE), laptop in-vehicle devices (LME), USB dongles, smart devices, wireless customer premises equipment (CPE), Internet of Things (IoT) devices, watches or other wearable devices, VR (virtual reality) devices, XR (extended reality) devices, head-mounted displays (HMDs), vehicles, drones, medical devices and applications (e.g., remote surgery), industrial devices and applications (e.g., robots and / or other wireless devices operating in the context of industrial and / or automated processing chains), consumer electronics devices, devices operating on commercial and / or industrial wireless networks, etc. In the following description, the terms “station,” “station equipment,” “STA,” “terminal equipment,” “communication equipment,” “terminal,” “user equipment,” and “UE” are used interchangeably.

[0042] Data is communicated between terminal devices and network devices. An example of Channel State Information (CSI) data is presented. The proposed scheme is then introduced by describing examples of data that can be used for communication between terminal devices and network devices, using CSI as a primary data source. To process the data, artificial intelligence / machine learning (AI / ML) models can be trained and deployed at both the terminal and network devices.

[0043] Communication performance is crucial for communication. Similarly, Channel State Information (CSI) feedback is also critical for communication performance between terminal devices and network devices. CSI can be encoded (compressed) at the terminal device, for example, using an AI / ML model (also known as an encoder), and decoded (decompressed) at the network device, for example, using another AI / ML model corresponding to the one used at the terminal device (also known as a decoder). In most cases, in a communication network, a network device can serve multiple terminal devices, and these terminal devices may come from different vendors; that is, multiple terminal devices may have different encoders to encode CSI. Furthermore, multiple terminal devices may have different UE capabilities and be in different network conditions, which can affect the training of their encoders. Therefore, CSI feedback enhancement is being investigated to address (or at least mitigate) the problems of inter-vendor training collaboration and intra-UE network condition differences.

[0044] According to the specification in 3GPP TR 38.843 (Rel. 18), for two-sided model use cases like CSI compression, separate AI / ML model training methods are considered Type 3 for model training collaboration. In Type 3 training, the UE-side CSI generation and NW-side CSI reconstruction are trained separately and sequentially. The training process can be initiated by either the UE or the NW (i.e., gNB). In Type 3 AI / ML model training collaboration, separate AI / ML model training is performed on both the NW and UE sides, and the UE-side CSI generation and NW-side CSI reconstruction are trained accordingly by the UE and network sides. In other words, the UE-side CSI generation is trained on the UE side, while the NW-side CSI reconstruction is trained on the network side. 3GPP is considering standardizing aspects such as reference models (e.g., structure and / or parameters), reference model structures, model formats, datasets, and dataset formats to address / mitigate or mitigate issues related to AI / ML model training collaboration. When comparing different standardization options, factors such as complexity, performance, and feasibility need to be considered.

[0045] In Type 3 training, as described above, the NW-priority separate training process is divided into two phases: the NW-side model training phase and the UE-side model training phase. In the NW-side model training phase, similarity metrics (such as SGCS) can be used to evaluate the performance of the decoder at the network device. The decoder may employ a transformer (hereinafter referred to as "TF") structure. Furthermore, in the UE-side model training phase, distance metrics (such as MSE) can be used to evaluate the performance of the encoder at the terminal device. The encoder may also employ a transformer structure. Hereinafter, in some embodiments of this disclosure, it is assumed that both the decoder and encoder, as well as the encoder and the MSE-oriented assumed encoder, employ transformer structures.

[0046] However, potential problems may arise if the training process of the UE-side encoder is self-managed solely by the UE, and there is no explicit mutual alignment between the UE and the NW. The root cause of the UE's inability to guarantee the performance of its encoder training lies in the lack of an NW-side decoder during UE-side training, resulting in a lack of overall evaluation metrics (e.g., SGCS). Therefore, the UE cannot assess the quality of its UE-side encoder and its fit with the NW-side decoder. On the other hand, determining the amount of training data is challenging for the NW because the gNB also lacks details about the UE-side model (e.g., the number of trainable parameters). One option for the gNB to ensure the performance of the E2E SGCS reconstructed by CSI is to share a sufficiently large dataset for UE-side encoder training, considered a best-effort approach. However, this large dataset may be larger than a dataset with an appropriate data size, leading to wasted overhead.

[0047] Therefore, in view of the above, a method for enhancing CSI feedback is proposed in some embodiments of this disclosure. According to this method, on the network side, a network device determines a training dataset and a distance threshold for training an encoder at a terminal device. The encoder is associated with a decoder at the network device, and the distance threshold is used to evaluate the distance loss during encoder training, and the distance loss is associated with the similarity performance related to both the encoder and decoder. The network device also sends the training dataset and the distance threshold to the terminal device. At the terminal device, the terminal device receives the training dataset and the distance threshold for training the encoder from the network device. Then, using the received training dataset and distance threshold, the terminal device trains the encoder based on the training dataset and the distance threshold. In this way, end-to-end (E2E) SGCS performance for CSI reconstruction can be ensured at both the terminal device and the network device, and the overhead of wasting training datasets can be avoided or reduced.

[0048] Figure 1A An example communication system 100 in which some embodiments of this disclosure may be implemented is illustrated. Communication system 100 is part of a communication network and includes terminal devices (UEs) 110-1, UE 110-2, and network device 120. Terminal devices 110-1 and 110-2 may be, for example, Internet of Things (IoT) devices. Network device 120 may be, for example, a Random Access Network (RAN) device (such as an NG-RAN device, also known as a gNB), or a communication module thereof. Network device 120 is associated with cell 121 and provides communication services to terminal devices (such as UEs 110-1 and 110-2) in cell 121. UEs 110-1 and UE 110-2 may also be collectively referred to as UE 110. Figure 1A As shown, terminal devices 110-1 and 110-2 are both connected to network device 120.

[0049] In system 100, the link from network device 120 to terminal device 110 is referred to as the downlink (DL), and the link from terminal device 110 to network device 120 is referred to as the uplink (UL). In the downlink, network device 120 is a transmitting (TX) device (or transmitter), and terminal device 110 is a receiving (RX) device (or receiver). In the uplink, terminal device 110 is a transmitting TX device (or transmitter), and network device 120 is an RX device (or receiver).

[0050] The communication in communication system 100 can conform to any suitable standard, including but not limited to Long Term Evolution (LTE), LTE-A Advanced, Wideband Code Division Multiple Access (WCDMA), Code Division Multiple Access (CDMA), and Global System for Mobile Communications (GSM). Furthermore, communication can be performed according to any generation of communication protocols currently known or to be developed in the future. Examples of communication protocols include, but are not limited to, first-generation (1G), second-generation (2G), 2.5G, 2.75G, third-generation (3G), fourth-generation (4G), 4.5G, fifth-generation (5G), 5.5G, advanced 5G networks, or sixth-generation (6G) communication protocols.

[0051] It should be understood that Figure 1A The number, connection relationships, and types of devices (including terminal device 110 and network device 120) shown are for illustrative purposes only and do not imply any limitation. Communication system 100 may include any number of devices suitable for implementing embodiments of this disclosure.

[0052] Figure 1B The illustration 100B shows a schematic diagram of the relationship between SGCS performance and different training data sizes according to some embodiments of the present disclosure. Figure 1B The SGCS performance of different encoders (encoder 1 at UE1, encoder 2 at UE2, and encoder 3 at UE3) with different training data sizes is shown. These three encoders (i.e., encoder 1, encoder 2, and encoder 3) have different model complexities. In this example, if the required SGCS is 0.7245, UE3 might only require a training dataset of size 20K, while UE1 and UE2 might correspondingly require training datasets of size 40K to ensure the SGCS requirement is met. Sending the same 40K training data to UE3 via the gNB could be a waste of resources.

[0053] Figure 2 The illustration shows a signaling diagram illustrating an example communication process 200 according to some example embodiments of the present disclosure. For discussion purposes, reference will be made to... Figure 1A To describe communication process 200. Communication process 200 may involve terminal devices (e.g., such as...) Figure 1A As shown in the diagram, IoT devices (such as terminal device 110) and network devices (e.g., such as...) Figure 1A The network device 120 shown is described below. Reference will be made to the network device 120 shown below. Figure 1A The communication process 200 is described using the terminal device 110 and network device 120 shown. As described above, a decoder (e.g., ...) may be present in the network device 120. Figure 3A and Figure 5 The decoder marked "DEC" in the code, and a corresponding encoder may exist at the terminal device 110 (e.g., Figure 3A The encoder marked "ENC" in the middle.

[0054] like Figure 2 As shown, network device 120 determines (210) a training dataset and a distance threshold for training the encoder at terminal device 110. The encoder is associated with a decoder at network device 120. The distance threshold is used to evaluate the distance loss during encoder training, and the distance loss is associated with the E2E similarity performance related to both the encoder and decoder. The distance loss can be absolute direct distance, mean squared error (MSE), or cubic error. The similarity performance can be squared generalized cosine similarity (SGCS) performance. Hereinafter, it is described that MSE can be used as an example of distance loss, and the SGCS performance of the decoder can be used as an example of E2E similarity performance.

[0055] To determine the relationship between distance loss and similarity performance, network device 120 can train a decoder and a corresponding hypothesis encoder using a first loss function defined as similarity performance, and generate a training dataset between the hypothesis encoder and decoder. Then, until the number of training epochs reaches a predefined or predetermined maximum number of training epochs, in each training epoch, network device 120 can train a distance-oriented encoder based on the generated training dataset using a second loss function defined as the distance loss during training of the distance-oriented encoder. The average distance value (which can represent the distance loss) of the codewords output by the distance-oriented encoder for the training dataset is recorded. This codeword is then input into the frozen decoder to obtain reconstructed data, and an average similarity value (which can represent the similarity performance) is recorded based on the reconstructed data. After training for the maximum number of training epochs, network device 120 can determine the relationship between distance loss and similarity performance based on the recorded average distance value and average similarity value. This training can be performed using different training models and with different parameters and data sizes in the training dataset. Then, the relationship between distance loss and similarity performance can be plotted (e.g., as a curve) based on the recorded average distance and average similarity values.

[0056] For example, when distance loss is defined as MSE and similarity performance is defined as SGCS performance, in order to determine the training dataset and distance threshold, network device 120 can derive the MSE-SGCS relationship and the MSE threshold at the network device for the terminal device 110. This will refer to... Figure 5 To provide a more detailed description.

[0057] Network device 120 can proactively (i.e., without a request from terminal device 110 for the training dataset and distance threshold) determine the training dataset and distance threshold, or in response to a request from terminal device 110 for the training dataset and distance threshold (the latter case will be referred to in the following text). Figure 4 (Discussion) is needed to determine the training dataset and distance threshold.

[0058] In the former case, network device 120 can determine similarity performance values ​​(e.g., SGCS performance values) on its own in order to determine the training dataset and distance threshold at 210.

[0059] In the latter case, before network device 120 determines the training dataset and distance threshold, terminal device 110 may send a request for the training dataset and distance threshold to network device 120, and the request may include a similarity performance value (e.g., SGCS performance value), that is, instructing network device 120 to provide a training dataset and distance threshold that can satisfy the similarity performance value.

[0060] In both cases, network device 120 can determine a distance threshold (e.g., an MSE threshold) based on the relationship between distance loss and similarity performance, and based on the similarity performance value. The relationship between distance loss and similarity performance can be represented by a table or formula. Therefore, using the similarity performance value, network device 120 can look up the table or substitute the similarity performance value into the formula to obtain the corresponding distance loss as the distance threshold. For details on how to determine the distance threshold based on this relationship and similarity performance value, please refer to [reference needed]. Figures 3A-3C and Figure 5 To illustrate further details, we will use the MSE threshold as an example of a distance threshold and the SGCS performance as an example of similarity performance.

[0061] After the network device 120 determines the training dataset and distance threshold, the network device sends (215) the training dataset and distance threshold 201 to the terminal device 110.

[0062] On the other side of the communication, terminal device 110 receives (217) a training dataset and a distance threshold 201 from network device 120. Using the training dataset and the distance threshold 201, terminal device 110 trains (220) an encoder based on the training dataset and the distance threshold 201. Based on the distance threshold determined and assigned from network device 120, terminal device 110 can monitor the average distance loss of the training dataset and manage its encoder training progress. More specifically, when the average distance loss of the training dataset falls below the distance threshold, terminal device 110 can terminate the training process. Alternatively, when terminal device 110 realizes that it cannot train the encoder to meet the distance threshold using the received training dataset within the maximum number of training epochs, it can request a supplementary training dataset from network device 120 for additional encoder training.

[0063] During training at point 220, for example, terminal device 110 may determine whether the trained encoder meets a distance threshold after a predetermined or predetermined maximum number of training epochs. For example, the predetermined or predetermined maximum number of training epochs may be 10,000. In other words, after 10,000 training epochs using the training dataset, terminal device 110 may determine whether the trained encoder meets a distance threshold (e.g., the distance loss expressed as the average MSE value over the training epochs is lower than a distance threshold expressed as an MSE threshold).

[0064] If the trained encoder meets a distance threshold (e.g., the distance loss is below the distance threshold), meaning the trained encoder is good enough to be used, the terminal device 110 can terminate the encoder training. If the trained encoder does not meet the distance threshold (e.g., the distance loss is greater than or equal to the distance threshold after a predetermined maximum number of training epochs), the terminal device 110 can send an additional request message to the network device 120 for an additional training dataset.

[0065] At network device 120, upon receiving an additional request message from terminal device 110 for a supplementary training dataset, network device 120 may send a supplementary training dataset to terminal device 110 in response to the additional request message. At terminal device 110, upon receiving the supplementary training dataset, terminal device 110 may train an encoder based on the supplementary training dataset and a distance threshold until the trained encoder meets the distance threshold. This example communication process 200 ensures E2E reconstruction SGCS performance between terminal device 110 and network device 120.

[0066] In the previous text, referring to the high-level signaling diagram Figure 2Some examples of this disclosure are generally described. Further examples of this disclosure are described below with reference to Figures 3-6. In the following description, the MSE threshold is used as an example of a distance threshold, and the SGCS performance is used as a similarity performance.

[0067] Figure 3A The illustration is a schematic diagram of the evaluation of the relationship between the MSE of the encoder output codeword and the SGCS performance of the reconstructed CSI according to some embodiments of the present disclosure. Figure 3A The upper part describes the NW-first separate training scheme, in which the relationship between SGCS and MSE is not established, and Figure 3A The lower half highlights the evaluation of the MSE-SGCS relationship. In this sense, some embodiments of this disclosure can be considered as extensions and modifications to this NW-priority separate training framework, and are intended to further enhance the framework, including facilitating performance monitoring during model training on the UE side.

[0068] like Figure 3A As shown, the NW-first separate training process consists of three phases. In the first phase, NW (e.g., as in...) Figure 1A and Figure 2 The network device 120 shown in the figure uses the E2E SGCS loss function to train the decoder using a hypothetical encoder, i.e., Loss1 = SGCS( In the second phase, the NW sends the training dataset to the UE (e.g., as in...). Figure 1A and Figure 2 The terminal device 110 shown in the figure. In the third stage, the UE uses the MSE loss function of the encoder's output codeword (i.e., Loss2 = MSE( To train the actual encoder (i.e., the encoder on the UE side).

[0069] When the actual encoder on the UE side is trained based on the trained (frozen) NW-side decoder, the average MSE of the encoder's output codewords in the training dataset (e.g., the MSE in Figure 3) The average SGCS of the real CSI and the reconstructed CSI (e.g., SGCS in Figure 3) The relationship between ) is evaluated in each round, for example, in the following Figure 3B and Figure 3C The model used to calculate the E2E SGCS is derived by cascading the UE-side encoder and the NW-side decoder.

[0070] The dataset parameters are given in Table 1. As shown in Table 1, the evaluation was conducted in a simulated scenario of UMa (urban macrocell), where the UE distribution was assumed to be 80% indoors and 20% outdoors, and the UE speeds for indoor and outdoor UEs were assumed to be 3 km / h and 30 km / h respectively. The carrier frequency used in the simulation was 4 GHz, the SCS was 30 kHz, and the bandwidth was 20 MHz. Seven sites were used in the simulation, with three sectors per site. The height of the BS antenna was 25 meters. The distance between macrocell base stations should be 200 meters. Table 1 Dataset Parameters

[0071] In the evaluation, both the NW-side decoder and its corresponding hypothetical encoder are converter backbones. Due to the mismatch between encoders and decoders that are not converter backbones, as observed in the Rel-18 study, leading to a greater performance degradation in separate training for CSI compression, only the converter backbone UE-side encoder was evaluated to examine the aforementioned MSE-SGCS relationship. All cases were evaluated using scalar quantization at 2 bits per element and CR (compression ratio) = 1 / 32. The hyperparameters involved in model training are shown in Table 2. Table 2 Hyperparameters used for model training

[0072] More specifically, in the example illustrated in Table 2, during model training, the training data samples for the NW-side decoder have a size of 80K, and the trainable parameters of the NW-side decoder are 1.1M. The training samples (training dataset) for the UE-side encoder can have a size of 20K, 40K, or 80K. The UE-side encoder can employ transformers TF1, TF2, TF3, and TF4, and the trainable parameters of TF1, TF2, TF3, and TF4 are 0.3M, 0.7M, 1.1M, and 4.5M respectively. The value of the parameter "Learning Rate" is selected as "5e-4". The maximum number of training epochs is set to 200. The batch size is set to 128, and the loss function is selected as SGCS.

[0073] It should be noted that the parameters shown in Tables 1 and 2 are for illustrative purposes only, and other parameters may be used for model training. For example, the data size of the training data samples used for the NW side decoder may also be 60K or 100K.

[0074] In each training epoch of the UE-side encoder, the average MSE value and corresponding average SGCS value of the training dataset were collected. Four different encoders (i.e., TF1, TF2, TF3, and TF4) and three training data sizes (i.e., 20K, 40K, and 80K) were experimented with. Some of the collected evaluation results are presented in... Figure 3B and Figure 3C It is shown in the middle.

[0075] Figure 3B The illustration 300B shows a schematic diagram of the convergence monotonic relationship between MSE and SGCS according to some embodiments of the present disclosure. Figure 3C Another schematic diagram 300C is illustrated for the convergence monotonic relationship between MSE and SGCS according to some embodiments of the present disclosure.

[0076] like Figure 3B and Figure 3C As shown, in each training epoch, there is a convergent monotonic relationship between the average MSE of the training dataset targeting the actual encoder output codewords and the average SGCS performance of the reconstructed CSI. This convergence is particularly evident in the case of high SGCS, and can be seen as follows: Figure 3B and Figure 3C The figure shown is fitted as a line / curve.

[0077] The monotonic relationship between the MSE value and the corresponding SGCS value is independent of model complexity (TF in this case) and the size of the training dataset. In other words, the monotonic relationship between the MSE value and the corresponding SGCS value is independent of model complexity and the size of the training dataset. For example, ... Figure 3B As shown, with the same training dataset size (here, 80K), different models (here, TF1, TF2, TF3, and TF4) exhibit the same monotonic relationship between MSE values ​​and their corresponding SGCS values. Therefore, after the NW-side decoder is trained, regardless of the actual encoder's model complexity and training dataset size, the MSE-SGCS relationship always converges to the same relationship. Therefore, using a hypothesis encoder and an MSE-guided hypothesis encoder (referencing...) Figure 5 (As described in detail), the MSE-SGCS relationship can be obtained to guide the training of the UE-side encoder.

[0078] As mentioned above, since this MSE-SGCS relationship is unknown (and irrelevant) to the model complexity of the actual encoder and the size of the training dataset, the NW can pre-evaluate this MSE-SGCS relationship using any "hypothetical UE-side encoder" based on the training dataset. This pre-evaluated MSE-SGCS relationship can be stored as a numerical table at the NW side or fitted into a parameterized formula.

[0079] This MSE-SGCS relationship can be used to evaluate E2E performance at the UE side and provide training guidance to the UE side (e.g., to determine whether the encoder has been trained to achieve satisfactory performance, and to indicate whether the UE needs more training data to improve performance).

[0080] For the example above, observations can be made on the UE side or the NW side, and in each training epoch, there is a convergent monotonic relationship between the average MSE of the actual encoder output codewords and the average SGCS of the reconstructed CSI in the training dataset, regardless of the model used (e.g., TF1, TF2, TF3, or TF4) or the size of the training dataset (e.g., 20K, 40K, or 80K). This phenomenon is independent of the UE or NW. Therefore, if the SGCS performance value is determined (e.g., indicated by the UE, determined by the NW, or pre-determined or predefined), the NW can acquire (or determine) this relationship and determine the corresponding MSE threshold on the NW side.

[0081] Figure 4 The illustration shows a signaling diagram of an example communication process 400 according to some example embodiments of the present disclosure. Communication process 400 can be used for UE-side training guidance, as well as for additional on-demand supplementary training dataset transmission in an NW-priority split training scheme for CSI compression. For discussion purposes, communication process 400 will be referenced to... Figure 1A The communication process 400 may involve UE 410 (which may be as follows) Figure 1A Example of terminal device 110 shown in the figure), and gNB 420 (which can be as follows) Figure 1A (Example of network device 120 shown in the figure). As follows, communication process 400 will be described with reference to UE 410 and gNB 420.

[0082] like Figure 4 As illustrated in the diagram, at position 425, the gNB 420 maintains the normalized decoder. Simultaneously, the gNB 420 also maintains the MSE-SGCS relationship for the MSE-guided encoder during training.

[0083] Then, at 430, UE 410 sends a request to gNB 420 for the training dataset for the UE-side encoder and the MSE threshold for the target SGCS. In other words, in addition to requesting the transmission of the training dataset to gNB 420, UE 410 also requests the MSE threshold for its target SGCS from gNB 420.

[0084] The operation at 430 can be optional. In other words, gNB 420 can also determine the target SGCS by itself and send the corresponding training dataset and MSE threshold to UE 410.

[0085] At 435, after gNB 420 receives the target SGCS from UE 410 in the request sent from UE 410 at 430, or determines the target SGCS itself, gNB 420 checks a reference table or uses a pre-stored formula to obtain the corresponding MSE value as the MSE threshold, and then sends it back to UE 410. Here, the reference table is a table in which the relationship between MSE values ​​and SGCS performance values ​​is recorded (hereinafter referred to as "MSE-SGCS relationship" for simplicity), and the pre-stored formula is a formula representing the relationship between MSE values ​​and SGCS performance values. For example, if the SGCS value expected by UE 410 or determined by gNB 420 is 0.7245, gNB 420 can refer to the locally stored MSE-SGCS relationship (e.g., as in...). Figure 3B and Figure 3C (As shown in the diagram) and it is determined that when SGCS = 0.7245, 1 / log(MSE) = -0.2, therefore the corresponding MSE is approximately 0.0067. In this method, the MSE threshold corresponding to an SGCS value of 0.7245 is determined to be 0.0067.

[0086] At position 440, UE 410 uses the training dataset to train the actual encoder. The received MSE threshold serves as an indicator guiding the training of UE 410. Training can be terminated when the MSE loss falls below the received MSE threshold.

[0087] At 445, after a pre-determined or predefined maximum number of training epochs, UE 410 determines whether the trained encoder has reached the MSE threshold. If the encoder does not reach the MSE threshold, operations at 450, 455, and 460 are performed.

[0088] At position 450, UE 410 sends a request for a supplementary training dataset to gNB 420. On the other side of the communication, gNB 420 receives the request for the supplementary training dataset.

[0089] At 455, in response to the received request, gNB 420 sends the requested supplementary training dataset to UE 410. On the other side of the communication, UE 410 receives the supplementary training dataset from gNB 420. Using the supplementary training dataset received at 455, UE 410 further trains its encoder.

[0090] The operations at 460, 440, 445, 450, and 455 are repeated until the encoder training loss (here, MSE) at UE 410 is below the MSE threshold. For example, as referenced above. Figure 3B and Figure 3C The SGCS expected by UE 410 is assumed to be 0.7245, and the corresponding MSE threshold is 0.0067. If the relationship between the training dataset size and MSE is as described in... Figure 6 As shown in the diagram, if UE 410 uses TF2 as the encoder and trains it with a 20K training dataset, then UE 410 can only achieve an MSE of 0.0087, while the expected MSE is less than 0.0067. Therefore, this means that UE 410 needs more data for training. So UE 410 can request a 20K supplementary training dataset from gNB 420 at 450 to further train the encoder, and receive the requested 20K supplementary training dataset. Then, when UE 410 further trains the encoder using the requested 20K supplementary training dataset, its MSE will decrease to 0.0067, which meets the MSE requirement. On the other hand, under the same assumptions, if UE 410 uses TF4 as the encoder and trains it with a 20K training dataset, it can achieve the MSE requirement of 0.0067. Therefore, in this case, no supplementary training dataset is needed, which saves a lot of transmission overhead for UE 410 and gNB 420 compared to the best-effort approach.

[0091] Figure 5 The illustration depicts a schematic workflow 500 for deriving an example MSE-SGCS relationship according to some embodiments of the present disclosure. In workflow 500, NW (e.g., as...) Figure 1A and Figure 2 The network device 120 shown in the figure, or as... Figure 4 The gNB 420 shown in the figure uses a hypothetical UE-side encoder to derive the MSE-SGCS relationship (it should be noted that these "hypothetical UE-side encoders" actually exist at NW and are only "hypothetical" for deriving the MSE-SGCS relationship). This encoder is trained using the MSE loss function and can be regarded as an MSE-oriented encoder.

[0092] like Figure 5 As shown, at step 1, NW utilizes the E2E loss function (e.g., Loss1 = SGCS ( The training decoder (labeled "DEC") and the corresponding hypothetical encoder (labeled "Hypo ENC") for CSI compression are shown, where X is the original CSI (i.e., the input to Hypo ENC), Y is the codeword between the hypothetical encoder and decoder, and... The CSI (from the DEC output) is reconstructed from its input Y, which is also the output of the hypothetical encoder. In other words, in the reconstructed CSI (i.e., Once obtained, the E2E loss function (e.g., Loss1 = SGCS) is used. The E2E similarity performance between the Hypo ENC (hypothetically represented as the encoder at the UE) and the DEC (represented as the decoder at the NW / gNB) can be evaluated. In step 2, the NW generates the training dataset. , where Y is the codeword between the hypothetical encoder and decoder, as described above.

[0093] In step 3, NW bases its data on the generated dataset. Using MSE ( ) as the loss function (i.e., Loss2 = MSE( To train an MSE-guided hypothesis encoder, where Y is the codeword between the MSE-guided hypothesis encoder and decoder as described above, and This is the output from the MSE-guided hypothesis encoder. Note that the "MSE-guided hypothesis encoder" here differs from the "hypothesis encoder" in step 1, where the "hypothesis encoder" is used to evaluate the E2E similarity loss, i.e., SGCS, while the "MSE-guided hypothesis encoder" is used to evaluate the distance loss, i.e., the MSE loss here. To obtain the MSE loss, several different MSE-guided hypothesis encoders can be used, and these MSE-guided hypothesis encoders can be trained using training datasets with different data sizes. For example, as in... Figure 4 B and Figure 4 TF1, TF2, TF3, and TF4, illustrated in C, can be examples of MSE-guided hypothetical encoders, and 80K, 40K, and 20K can be examples of different data sizes for the training dataset. NW records the average MSE value of the codewords in the training dataset in each training epoch.

[0094] In step 4, during the same training process, the output codewords of the MSE-guided encoder are fed into the frozen decoder to derive the reconstructed CSI by freezing the decoder. NW records the average SGCS value of the CSI in the training dataset in each round.

[0095] In step 5, data from training datasets of the same or different sizes are used (see also...). Figure 3B and Figure 3CThe NW (Near-Wide Wavelength Controller) calculates the average MSE and average SGCS values ​​for each training epoch of the different MSE-oriented hypothetical encoders. The NW obtains the MSE-SGCS relationship, which it can use to assign MSE thresholds to the UE, for example, in response to the UE's required SGCS. This relationship can be stored as a numerical table or fitted into a parameterized formula. This MSE-SGCS relationship can be used to evaluate the UE-side E2E performance and provide training guidance to the UE-side training (e.g., determining whether the encoder has been trained to satisfactory performance, or indicating whether the UE needs more training data to improve performance).

[0096] In this way, based on the MSE threshold determined and allocated from the NW, the UE can monitor the average MSE of the training dataset and manage its encoder training progress. More specifically, when the average MSE of the training dataset falls below the MSE threshold at the UE's side, the UE can terminate the training process. Alternatively, when the UE realizes that it cannot train the encoder to meet the MSE threshold within the maximum training epochs, it will request the NW for supplementary training datasets for additional training.

[0097] Figure 6 The illustration shows a schematic diagram of the relationship between MSE and different training data sizes according to some embodiments of the present disclosure. Figure 6 The MSE of different encoders (TF1, TF2, TF3, and TF4) with different training data sizes are shown for different UEs. These four encoders (i.e., TF1, TF2, TF3, and TF4) represent different model complexities. In this example, if the required SGCS is 0.7245 and the corresponding MSE threshold is 0.0067, a UE with the TF4 encoder may only need 20K data for training, while a UE with the TF2 encoder and a UE with the TF3 encoder may require 40K training data to ensure the SGCS requirement. Therefore, sending the same 40K training data to a UE with the TF4 encoder for a gNB may be a waste of resources.

[0098] Figure 7 The illustrations depict, at some other embodiments of the present disclosure, at the terminal device (e.g., as in...). Figure 1A and Figure 2 The flowchart illustrates an example method 700 implemented by the terminal device 110 shown in the figure. For discussion purposes, method 700 will be referenced from the perspective of the terminal device 110. Figure 1A and Figure 2 Described.

[0099] As in Figure 7 As illustrated in the diagram, at box 710, terminal device 110 receives data from a network device (e.g., as shown in...). Figure 1A and Figure 2 The network device 120 shown in the figure receives a training dataset and a distance threshold (e.g., as shown in the figure) for training the encoder at the terminal device 110. Figure 2 The training dataset and distance threshold are illustrated in Figure 201. The encoder is associated with the decoder at the network device, and the distance threshold is used to evaluate the distance loss during encoder training. The distance loss is correlated with similarity performance, which is related to both the encoder and decoder. In box 720, the terminal device 110 trains the encoder based on the training dataset and the distance threshold.

[0100] In some example embodiments, in order to receive the training dataset and distance threshold, the terminal device 110 may also send a request for the training dataset and distance threshold to the network device; and receive the training dataset and distance threshold as a response to the request from the network device.

[0101] In some example embodiments, the request may include a similarity performance value.

[0102] In some example embodiments, in order to train the encoder based on the training dataset and the distance threshold, the terminal device 110 can determine whether the trained encoder meets the distance threshold.

[0103] In some example embodiments, if the terminal device 110 determines that the trained encoder meets a distance threshold, the terminal device 110 may terminate the training of the encoder.

[0104] In some example embodiments, if the terminal device 110 determines that the trained encoder does not meet the distance threshold after a predetermined maximum number of training epochs, the terminal device 110 may also send an additional request message to the network device for an additional training dataset; and receive an additional response message from the network device including the additional training dataset; then the terminal device 110 may also train the encoder based on the additional training dataset and the distance threshold.

[0105] In some example embodiments, in order to determine whether the trained encoder meets the distance threshold, the terminal device 110 may determine whether the distance loss is below the distance threshold.

[0106] In some example embodiments, the distance loss may include one of absolute direct distance, mean squared error (MSE), or cubic error, and / or the similarity performance may include squared generalized cosine similarity (SGCS) performance.

[0107] Figure 8 The illustrations depict, at some other embodiments of the present disclosure, at a network device (e.g., as in...). Figure 1A and Figure 2The flowchart illustrates an example method 800 implemented by network device 120. For discussion purposes, method 800 will be referenced from the perspective of network device 120. Figure 1A and Figure 2 Described.

[0108] As in Figure 8 As illustrated in the diagram, at box 810, network device 120 determines a training dataset and a distance threshold for training the encoder at the terminal device. The encoder is associated with a decoder at the network device, and the distance threshold is used to evaluate the distance loss during encoder training. This distance loss is correlated with similarity performance, which is related to both the encoder and decoder. At box 820, network device 120 sends the training dataset and distance threshold to the terminal device (e.g., as shown in...). Figure 2 The training dataset and distance threshold 201 are illustrated in the figure.

[0109] In some example embodiments, before sending the training dataset and distance threshold, network device 120 may receive a request for the training dataset and distance threshold from the terminal device; and send the training dataset and distance threshold to the terminal device as a response to the request.

[0110] In some example embodiments, the request may include a similarity performance value.

[0111] In some example embodiments, network device 120 may also determine similarity performance values ​​before determining the training dataset and distance threshold.

[0112] In some example embodiments, network device 120 may also receive additional request messages from terminal device for additional training datasets; and send additional training datasets to terminal device as a response to the additional request messages.

[0113] In some example embodiments, network device 120 may determine the distance threshold based on the relationship between distance loss and similarity performance, as well as the similarity performance value.

[0114] In some example implementations, the relationship between distance loss and similarity performance can be represented by a table or formula.

[0115] In some example embodiments, to determine the relationship between distance loss and similarity performance, network device 120 may train a decoder and a corresponding hypothesis encoder using a first loss function defined as similarity performance; and generate a training dataset between the hypothesis encoder and the decoder. Furthermore, until the number of training epochs reaches a predetermined maximum, in each training epoch at network device 120, network device 120 may also train a distance-oriented encoder based on the generated training dataset using a second loss function defined as the distance loss during the training of the distance-oriented encoder, recording the average distance value (which can represent the distance loss) of the codewords output by the distance-oriented encoder to the training dataset, inputting codewords to the frozen decoder to obtain reconstructed data, and recording an average similarity value (which represents similarity performance) based on the reconstructed data. Then, using the recorded average distance value and average similarity value, network device 120 may determine the relationship between distance loss and similarity performance.

[0116] In some example embodiments, the distance loss may include one of absolute direct distance, mean squared error (MSE), or cubic error, and / or the similarity performance may include squared generalized cosine similarity (SGCS) performance.

[0117] In some embodiments, the apparatus capable of performing method 700 may include components for performing corresponding steps of method 700. These components may be implemented in any suitable form. For example, the components may be implemented in a circuit system or a software module.

[0118] In some example embodiments, the apparatus includes: components for receiving a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at a network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and components for training the encoder based on the training dataset and the distance threshold.

[0119] In some example embodiments, the component for receiving the training dataset and the distance threshold may include: a component for sending a request for the training dataset and the distance threshold to the network device; and a component for receiving the training dataset and the distance threshold from the network device as a response to the request.

[0120] In some example embodiments, the request includes a similarity performance value.

[0121] In some example embodiments, the component for training the encoder based on the training dataset and the distance threshold may include a component for determining whether the trained encoder meets the distance threshold.

[0122] In some example embodiments, the device may also include a component for terminating encoder training if the determining component determines that the trained encoder meets a distance threshold.

[0123] In some example embodiments, if the determining component determines, after a predetermined maximum number of training epochs, that the trained encoder does not meet a distance threshold, the apparatus may further include: components for sending an additional request message to a network device for an additional training dataset; and components for receiving an additional response message from the network device including the additional training dataset. The apparatus may also include components for training the encoder based on the additional training dataset and the distance threshold.

[0124] In some example embodiments, the components for determining whether the trained encoder meets a distance threshold may include components for determining whether the distance loss is below the distance threshold.

[0125] In some example embodiments, the distance loss may include one of absolute direct distance, mean squared error (MSE), or cubic error, and / or the similarity performance may include squared generalized cosine similarity (SGCS) performance.

[0126] In some embodiments, the apparatus further includes components for performing additional steps in some embodiments of method 700. In some embodiments, the components include at least one processor; and at least one memory including computer program code, the at least one memory and the computer program code being configured together with the at least one processor to cause the apparatus to execute.

[0127] In some embodiments, the apparatus capable of performing method 800 may include components for performing corresponding steps of method 800. These components may be implemented in any suitable form. For example, the components may be implemented in a circuit system or a software module.

[0128] In some example embodiments, the apparatus includes: components for determining a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at a network device, the distance threshold is used to evaluate distance loss during encoder training, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; and components for transmitting the training dataset and the distance threshold.

[0129] In some example embodiments, before sending the training dataset and distance threshold, the apparatus may include: a component for receiving a request for the training dataset and distance threshold from a terminal device; and a component for sending the training dataset and distance threshold to the terminal device as a response to the request.

[0130] In some example embodiments, the request includes a similarity performance value.

[0131] In some example embodiments, the apparatus may include components for determining similarity performance values ​​before determining the training dataset and distance threshold.

[0132] In some example embodiments, the apparatus may further include: a component for receiving a further request message from a terminal device for a further training dataset; and a component for sending a further training dataset to the terminal device as a response to the further request message.

[0133] In some example embodiments, the apparatus may include a component for determining a distance threshold based on the relationship between distance loss and similarity performance, and the similarity performance value.

[0134] In some example implementations, the relationship between distance loss and similarity performance is represented by a table or formula.

[0135] In some example embodiments, the components for determining the relationship between distance loss and similarity performance may include: components for training a decoder and a corresponding hypothesis encoder using a first loss function defined as similarity performance; and components for generating a training dataset between the hypothesis encoder and the decoder. Furthermore, until the number of training epochs reaches a predetermined maximum number of training epochs, the components for determining the relationship may further include: components for training a distance-oriented encoder based on the generated training dataset using a second loss function defined as the distance loss during the training of the distance-oriented encoder; components for recording the average distance value (representing distance loss) of the codewords output by the distance-oriented encoder for the training dataset; components for inputting codewords into a frozen decoder to derive reconstructed data; and components for recording an average similarity value (representing similarity performance) based on the reconstructed data. Then, using the recorded average distance value and average similarity value, the apparatus may further include: components for determining the relationship between distance loss and similarity performance based on the recorded average distance value and average similarity value.

[0136] In some example embodiments, distance loss may include one of absolute direct distance, mean squared error (MSE), or cubic error; and / or similarity performance may include squared generalized cosine similarity (SGCS) performance.

[0137] In some embodiments, the apparatus further includes components for performing additional steps in some embodiments of method 800. In some embodiments, the components include at least one processor; and at least one memory including computer program code, the at least one memory and the computer program code being configured together with the at least one processor to cause the apparatus to execute.

[0138] Figure 9 A simplified block diagram of a device 900 suitable for implementing some example embodiments of the present disclosure is illustrated. Device 900 can be provided to implement a communication device, such as in… Figure 1A and Figure 2 The terminal device 110 or network device 120 shown is illustrated. As shown, device 900 includes one or more processors 910, one or more memories 920 coupled to processors 910, and one or more communication modules 940 coupled to processors 910.

[0139] The communication module 940 is used for bidirectional communication. The communication module 940 has at least one antenna to facilitate communication. The communication interface can represent any interface necessary for communication with other network elements.

[0140] Processor 910 can be any type suitable for a local network, and as a non-limiting example, can include one or more of the following: general-purpose computer, special-purpose computer, microprocessor, digital signal processor (DSP), and processor based on a multi-core processor architecture. Device 900 can have multiple processors, such as application-specific integrated circuit chips that are time-dependent on a clock synchronized with the main processor.

[0141] Memory 920 may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, read-only memory (ROM) 924, electrically programmable read-only memory (EPROM), flash memory, hard disk, optical disc (CD), digital video disc (DVD), and other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, random access memory (RAM) 922 and other volatile memories that do not persist during power outages.

[0142] Computer program 930 includes computer-executable instructions that are executed by the associated processor 910. Program 930 may be stored in ROM 924. Processor 910 may perform any suitable actions and processes by loading program 930 into RAM 922.

[0143] The embodiments of this disclosure can be implemented via program 930, enabling device 900 to execute as described in the reference. Figure 2 , Figure 4 and Figure 7-8 Any process discussed in this disclosure. Embodiments of this disclosure may also be implemented by hardware or by a combination of software and hardware.

[0144] In some example embodiments, program 930 may be tangibly contained in a computer-readable medium that may be included in device 900 (such as in memory 920) or in other storage accessible by device 900. Device 900 may load program 930 from the computer-readable medium into RAM 922 for execution. The computer-readable medium may include any type of tangible non-volatile storage, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0145] Figure 10 A block diagram of an example of a computer-readable medium 1000 according to some exemplary embodiments of the present disclosure is shown. A program 930 is stored on the computer-readable medium 1000. It should be noted that, although in Figure 10 The computer-readable medium 1000 is described as being in the form of a CD or DVD, but the computer-readable medium 1000 may be any other form suitable for carrying or storing the program 930.

[0146] Generally, the various embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software executable by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are illustrated and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0147] This disclosure also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which are executed on a target real or virtual processor on a device to perform the functions described above. Figure 2 , Figure 4 and Figure 7-8The method described is any one of 200, 400, and 700-800. Typically, a program module includes routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. The functionality of a program module can be combined or split among program modules as needed in various embodiments. The machine-executable instructions used for a program module can execute on a local or distributed device. In a distributed device, a program module can reside on both local and remote storage media.

[0148] Program code used to perform the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that, when executed by the processor or controller, the program code enables the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine as a stand-alone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0149] In the context of this disclosure, computer program code or related data may be carried by any suitable carrier wave to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carrier waves include signals, computer-readable media, etc.

[0150] Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable media can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing. More specific examples of computer-readable storage media will include electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable optical disc read-only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing. The term "non-transient" as used herein is a limitation on the medium itself (i.e., tangible, non-signal), not a limitation on the persistence of data storage (e.g., RAM versus ROM).

[0151] Furthermore, although the operations are depicted in a specific order, this should not be construed as requiring such operations to be performed in the specific order shown or in sequential order, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this disclosure, but rather as descriptions of features that may be specific to particular embodiments. Certain features described in the context of a single embodiment may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0152] Although this disclosure has been described in language specific to structural features and / or methodological actions, it should be understood that the disclosure as defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are disclosed as exemplary forms of implementing the claims.

Claims

1. A terminal device, 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 terminal device to: The network device receives a training dataset and a distance threshold for training an encoder at the terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate the distance loss in the training of the encoder, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder. as well as The encoder is trained based on the training dataset and the distance threshold.

2. The terminal device of claim 1, wherein the terminal device is configured to receive the training dataset and the distance threshold in the following manner: Send a request to the network device for the training dataset and the distance threshold; and Receive the training dataset and the distance threshold from the network device as a response to the request.

3. The terminal device according to claim 2, wherein the request includes the similarity performance value of the similarity performance.

4. The terminal device according to any one of claims 1 to 3, wherein the terminal device is configured to train the encoder based on the training dataset and the distance threshold in the following manner: Determine whether the trained encoder satisfies the distance threshold.

5. The terminal device according to claim 4, wherein the terminal device is further configured to: The training of the encoder is terminated based on the determination that the trained encoder meets the distance threshold.

6. The terminal device according to claim 4, wherein the terminal device is further configured to: After a predetermined maximum number of training epochs, based on the determination that the trained encoder does not meet the distance threshold, Send a separate request message to the network device for a different training dataset; Receive additional response messages from the network device, including the additional training dataset; as well as The encoder is trained based on the additional training dataset and the distance threshold.

7. The terminal device according to any one of claims 4 to 6, wherein the terminal device is configured to determine whether the trained encoder satisfies the distance threshold by: Determine whether the distance loss is lower than the distance threshold.

8. The terminal device according to any one of claims 1 to 7, wherein at least one of the following: The distance loss includes one of absolute direct distance, mean square error (MSE), or cubic error; or The similarity performance includes the squared generalized cosine similarity (SGCS) performance.

9. A network device, 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 network device to: A training dataset and a distance threshold are determined for training an encoder at a terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate the distance loss in the training of the encoder, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; as well as The training dataset and the distance threshold are sent to the terminal device.

10. The network device of claim 9, wherein the network device is configured to transmit the training dataset and the distance threshold in the following manner: Receive a request from the terminal device for the training dataset and the distance threshold; and The training dataset and the distance threshold are sent to the terminal device as a response to the request.

11. The network device of claim 10, wherein the request includes a similarity performance value of the similarity performance.

12. The network device of claim 9, wherein before determining the training dataset and the distance threshold, the network device is further configured to: Determine the similarity performance value.

13. The network device according to any one of claims 9 to 12, wherein the network device is further configured to: Receive additional request messages from the terminal device for a different training dataset; and The additional training dataset is sent to the terminal device as a response to the additional request message.

14. The network device according to any one of claims 11 to 13, wherein the network device is configured to determine the distance threshold based on: The relationship between the distance loss and the similarity performance; and The similarity performance value.

15. The network device of claim 14, wherein the relationship between the distance loss and the similarity performance is represented by a table or formula.

16. The network device of claim 14 or 15, wherein the network device is configured to determine the relationship between the distance loss and the similarity performance in the following manner: At the network device, the decoder and the corresponding hypothesis encoder are trained using a first loss function defined as the similarity performance; A training dataset is generated between the hypothetical encoder and the decoder; The training continues until the number of training rounds reaches a predetermined maximum, during which time the training rounds occur at the network device. Based on the generated training dataset, a distance-oriented encoder is trained using a second loss function, which is defined as the distance loss during the training of the distance-oriented encoder. The average distance value of the codewords output by the distance-oriented encoder for the training dataset is recorded, where the average distance value represents the distance loss. The codeword is input into the frozen decoder to obtain the reconstructed data, and An average similarity value is recorded based on the reconstructed data, and the average similarity value represents the similarity performance; as well as Based on the recorded average distance value and average similarity value, the relationship between the distance loss and the similarity performance is determined.

17. The network device according to any one of claims 9 to 16, wherein at least one of the following: The distance loss includes one of absolute direct distance, mean square error (MSE), or cubic error; or The similarity performance includes the squared generalized cosine similarity (SGCS) performance.

18. A method comprising: At the terminal device, a training dataset and a distance threshold for training an encoder at the terminal device are received from a network device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate the distance loss in the training of the encoder, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder. as well as The encoder is trained based on the training dataset and the distance threshold.

19. A method comprising: A network device determines a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate the distance loss in the training of the encoder, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; as well as At the network device, the training dataset and the distance threshold are sent to the terminal device.

20. An apparatus comprising: A component for receiving a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at a network device, the distance threshold is used to evaluate the distance loss in the training of the encoder, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; as well as A component used to train the encoder based on the training dataset and the distance threshold.

21. An apparatus comprising: Components for determining a training dataset and a distance threshold for training an encoder at a terminal device, wherein the encoder is associated with a decoder at the network device, the distance threshold is used to evaluate the distance loss in the training of the encoder, and the distance loss is associated with similarity performance, which is related to both the encoder and the decoder; as well as A component used to send the training dataset and the distance threshold.

22. A non-transitory computer-readable medium comprising program instructions stored thereon, which, when executed on at least one processor, cause the at least one processor to perform the method according to claim 18 or 19.

Images