terminal

Abstract

We provide a terminal that enables optimal overhead reduction, channel estimation, and resource utilization. [Solution] In a next-generation mobile communication system, the user terminal has a receiving unit that receives student model information relating to student models in knowledge distillation, and a transmitting unit that performs inference processing using the student model based on the student model information and transmits information regarding the output of the inference processing to a base station via wireless communication. The student model information includes an index of the student model, and the information regarding the output of the inference processing includes an index of the student model that served as the basis for the student model to be retrained at the base station.

Description

[Technical Field]

[0001] This disclosure relates to terminals in next-generation mobile communication systems. [Background technology]

[0002] Long Term Evolution (LTE) was specified for Universal Mobile Telecommunications System (UMTS) networks with the aim of achieving even higher data rates and lower latency (Non-Patent Document 1). Furthermore, LTE-Advanced (3GPP Rel.10-14) was specified for the aim of further increasing capacity and sophistication of LTE (Third Generation Partnership Project (3GPP) Release (Rel.) 8, 9).

[0003] Successor systems to LTE (for example, 5th generation mobile communication system (5G), 5G+ (plus), 6th generation mobile communication system (6G), New Radio (NR), 3GPP Rel.15 and later, etc.) are also being considered. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] 3GPP TS 36.300 V8.12.0 “Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved Universal Terrestrial Radio Access Network (E-UTRAN); Overall description; Stage 2 (Release 8)”, April 2010 [Overview of the project] [Problems that the invention aims to solve]

[0005] Regarding future wireless communication technologies, the use of artificial intelligence (AI) technologies such as machine learning (ML) for network / device control and management is being considered.

[0006] It is desirable to be able to appropriately control AI model training / model inference on terminals (user terminals, User Equipment (UE)). However, the specific details of this control have not yet been thoroughly considered. If this control is not properly defined, appropriate overhead reduction / high-precision channel estimation / high-efficiency resource utilization may not be achieved, potentially hindering improvements in communication throughput / communication quality.

[0007] Therefore, one of the objectives of this disclosure is to provide a terminal that can achieve suitable overhead reduction, channel estimation, and resource utilization. [Means for solving the problem]

[0008] A terminal according to one aspect of the present disclosure includes a receiving unit that receives student model information relating to the student model in the knowledge distillation, and a transmitting unit that performs inference processing using the student model based on the student model information and transmits information regarding the output of the inference processing to a base station via wireless communication, wherein the student model information includes an index of the student model, and the information regarding the output of the inference processing includes an index of the student model that served as the basis for the student model to be retrained at the base station. [Effects of the Invention]

[0009] According to one aspect of this disclosure, suitable overhead reduction, channel estimation, and resource utilization can be achieved. [Brief explanation of the drawing]

[0010] [Figure 1]FIG. 1 is a diagram showing an example of a framework for managing an AI model. [Figure 2] FIG. 2 is a diagram showing an example of an AI model. [Figure 3] FIG. 3 is a diagram showing an example of specifying an AI model according to the first embodiment. [Figure 4] FIG. 4 is a diagram showing an example of updating an AI model according to the first embodiment. [Figure 5] FIG. 5 is a diagram showing an example of fine-tuning. [Figure 6] FIG. 6 is a diagram showing an example of transfer learning. [Figure 7] FIG. 7 is a diagram showing an example of updating an AI model according to the first embodiment. [Figure 8] FIG. 8 is a diagram showing an example of specifying an AI model according to the second embodiment. [Figure 9] FIG. 9 is a diagram showing an example of updating an AI model according to the second embodiment. [Figure 10] FIG. 10 is a diagram showing an example of updating an AI model according to the second embodiment. [Figure 11] FIG. 11 is a diagram showing an example of specifying an AI model according to the third embodiment. [Figure 12] FIG. 12 is a diagram showing an example of specifying an AI model according to the third embodiment. [Figure 13] FIGS. 13A-13B are diagrams showing an example of UE training AI model information according to the third embodiment. [Figure 14] FIGS. 14A-14B are diagrams showing an example of UE training AI model information according to the third embodiment. [Figure 15] FIG. 15 is a diagram showing an example of an exchange of AI model information regarding federated learning. [Figure 16] FIGS. 16A-16C are diagrams showing an example of updated AI model information. [Figure 17] FIGS. 17A and 17B are diagrams showing an example of learning using knowledge distillation. [Figure 18]Figure 18 shows an example of a schematic configuration of a wireless communication system according to one embodiment. [Figure 19] Figure 19 shows an example of the configuration of a base station according to one embodiment. [Figure 20] Figure 20 shows an example of the configuration of a user terminal according to one embodiment. [Figure 21] Figure 21 shows an example of the hardware configuration of a base station and a user terminal according to one embodiment. [Figure 22] Figure 22 shows an example of a vehicle according to one embodiment. [Modes for carrying out the invention]

[0011] (Application of Artificial Intelligence (AI) technology to wireless communication) Regarding future wireless communication technologies, the use of AI technologies such as machine learning (ML) for network / device control and management is being considered.

[0012] For example, AI technology is being considered for future wireless communication technologies to improve Channel State Information Reference Signal (CSI) feedback (e.g., overhead reduction, accuracy improvement, prediction), beam management (e.g., accuracy improvement, prediction in the spatiotemporal domain), and position measurement (e.g., improved position estimation / prediction).

[0013] Figure 1 shows an example of an AI model management framework. In this example, each stage related to the AI ​​model is shown as a block. This example can also be described as AI model lifecycle management.

[0014] The data collection stage is the phase in which data is collected for the generation / updating of an AI model. The data collection stage may also include data organization (e.g., deciding which data to transfer for model training / model inference) and data transfer (e.g., transferring data to entities that will be trained / inferred (e.g., UE, gNB)).

[0015] In the model training stage, the model is trained based on the data (training data) transferred from the collection stage. This stage may include data preparation (e.g., data preprocessing, cleaning, formatting, transformation, etc.), model training / validation, model testing (e.g., verifying whether the trained model meets performance thresholds), model exchange (e.g., transferring the model for distributed learning), and model deployment / update (deploying / updating the model to entities that perform model inference).

[0016] In the model inference stage, model inference is performed based on the data (inference data) transferred from the collection stage. This stage may include data preparation (e.g., data preprocessing, cleaning, formatting, and transformation), model inference, model performance feedback (feeding back model performance to the entities being trained), and output (providing the model output to the actors).

[0017] The actor stage may include action triggers (e.g., deciding whether or not to trigger an action on another entity), feedback (e.g., providing feedback on training data / inference data / information needed for performance feedback), etc.

[0018] Furthermore, training a model for mobility optimization, for example, may be performed in a network (NW) maintenance, administration, and maintenance (management) system (gNodeB / gNB). The former offers advantages in terms of interoperability, large-capacity storage, operator manageability, and model flexibility (feature engineering, etc.). The latter offers advantages in that it eliminates the need for latency in model updates and data exchange for model deployment. Inference for the above model may be performed in a gNB, for example.

[0019] Furthermore, the entities used for training / inference may differ depending on the use case.

[0020] For example, in AI-assisted beam management based on measurement reports, OAM / gNB may perform model training and gNB may perform model inference.

[0021] For AI-assisted UE-assisted positioning, a Location Management Function (LMF) may perform model training and model inference.

[0022] For CSI feedback / channel estimation using an autoencoder, the OAM / gNB / UE may perform model training, and the gNB / UE may perform model inference (jointly).

[0023] For AI-assisted beam management or AI-assisted UE-based positioning based on beam measurements, the OAM / gNB / UE may perform model training and the UE may perform model inference.

[0024] As described above, it is preferable to be able to train AI models in the UE. This is because the UE can directly acquire data that gNB / OAM cannot directly access. For example, training an autoencoder model for CSI feedback requires complete downlink CSI information (or channel information), and only the UE can directly acquire this information.

[0025] If an autoencoder model is trained in a gNB / OAM, the UE needs to transmit the complete downlink CSI information (or channel information), which incurs significant communication overhead. On the other hand, if an autoencoder model is trained in the UE, the UE needs to retain the complete downlink CSI information (or channel information), but the UE's storage / computation resources are less extensive compared to gNB / OAM.

[0026] Therefore, it is desirable to be able to appropriately control model training / model inference in the UE. However, the specific details of this control have not yet been thoroughly investigated. If this control is not properly defined, appropriate overhead reduction, high-precision channel estimation, and efficient resource utilization may not be achieved, potentially hindering improvements in communication throughput and communication quality.

[0027] Therefore, the inventors have conceived of a suitable control method for model training / model inference. Note that each embodiment of this disclosure may also be applicable when AI / prediction is not used.

[0028] In one embodiment of this disclosure, a terminal (User Equipment (UE)) / base station (BS) trains an ML model in training mode and runs the ML model in inference mode (also called inference mode, etc.). In inference mode, the accuracy of the trained ML model trained in training mode may be validated.

[0029] In this disclosure, the UE / BS may input channel status information, reference signal measurements, etc., to the ML model and output high-precision channel status information / measurements / beam selection / position, future channel status information / wireless link quality, etc.

[0030] In this disclosure, AI may be interpreted as an object (also called a subject, object, data, function, program, etc.) having (implementing) at least one of the following characteristics: • Estimation based on observed or collected information • Selection based on observed or collected information. • Predictions based on observed or collected information.

[0031] In this disclosure, an object may be, for example, a device such as a terminal or base station. Furthermore, in this disclosure, an object may refer to a program / model / entity operating on such device.

[0032] Furthermore, in this disclosure, the ML model may be replaced with an object having (implementing) at least one of the following features: • By providing information (feeding), estimates are generated. By providing information, predict the estimated value. By providing information, we can discover features. • By providing information, the user can select an action.

[0033] Furthermore, in this disclosure, ML models, models, AI models, predictive analytics, predictive analytics models, etc., may be interpreted interchangeably. Also, ML models may be derived using at least one of the following: regression analysis (e.g., linear regression analysis, multiple regression analysis, logistic regression analysis), support vector machines, random forests, neural networks, deep learning, etc. In this disclosure, models may be interpreted as at least one of the following: encoders, decoders, tools, etc.

[0034] An ML model outputs at least one piece of information based on the input information, such as an estimate, a prediction, a chosen action, or a classification.

[0035] ML models may include supervised learning, unsupervised learning, and reinforcement learning. Supervised learning may be used to learn general rules for mapping inputs to outputs. Unsupervised learning may be used to learn data features. Reinforcement learning may be used to learn actions to maximize an objective (goal).

[0036] In this disclosure, terms such as generation, calculation, and derivation may be interpreted interchangeably. In this disclosure, terms such as implementation, operation, function, and execution may be interpreted interchangeably. In this disclosure, terms such as training, learning, updating, and retraining may be interpreted interchangeably. In this disclosure, terms such as inference, after-training, production use, and actual use may be interpreted interchangeably. Signal may be interpreted interchangeably with signal / channel.

[0037] In this disclosure, the training mode may correspond to the mode in which the UE / BS transmits / receives signals for the ML model (in other words, the mode of operation during the training period). In this disclosure, the inference mode may correspond to the mode in which the UE / BS implements the ML model (e.g., implements the trained ML model to predict the output) (in other words, the mode of operation during the inference period).

[0038] In this disclosure, the training mode may mean a mode in which a particular signal transmitted in the inference mode is transmitted in a manner that has high overhead (e.g., high resource usage).

[0039] In this disclosure, the training mode may mean a mode that references a first configuration (e.g., a first DMRS configuration, a first CSI-RS configuration, a first CSI reporting configuration). In this disclosure, the inference mode may mean a mode that references a second configuration different from the first configuration (e.g., a second DMRS configuration, a second CSI-RS configuration, a second CSI reporting configuration). The first configuration may have at least one more time resource, frequency resource, code resource, or port (antenna port) related to measurement than the second configuration. For example, the CSI reporting configuration may include a configuration related to the autoencoder.

[0040] The embodiments of this disclosure will be described in detail below with reference to the drawings. Each wireless communication method according to the embodiments may be applied individually or in combination.

[0041] In the following embodiments, the relevant entities are the UE and BS to illustrate an ML model relating to communication between UEs and BSs, but the application of each embodiment of this disclosure is not limited thereto. For example, for communication between other entities (e.g., UE-UE communication), the UE and BS in the embodiments below may be replaced with a first UE and a second UE. In other words, any UE, BS, etc. in this disclosure may be replaced with any UE / BS.

[0042] In this disclosure, "A / B" and "at least one of A and B" may be interpreted as mutually exclusive. In this disclosure, "A / B / C" may mean "at least one of A, B, and C".

[0043] In this disclosure, terms such as activate, deactivate, indicate, select, configure, update, and determine may be interpreted interchangeably. In this disclosure, terms such as support, control, controllable, operate, and operable may be interpreted interchangeably.

[0044] In this disclosure, Radio Resource Control (RRC), RRC parameters, RRC messages, higher-layer parameters, fields, Information Elements (IE), settings, etc., may be interpreted interchangeably. In this disclosure, Medium Access Control elements (MAC Control Element (CE)), update commands, activation / deactivation commands, etc., may be interpreted interchangeably.

[0045] In this disclosure, the upper-layer signaling may be, for example, Radio Resource Control (RRC) signaling, Medium Access Control (MAC) signaling, broadcast information, or a combination thereof.

[0046] In this disclosure, MAC signaling may include, for example, MAC Control Elements (MAC CEs) and MAC Protocol Data Units (PDUs). Broadcast information may include, for example, Master Information Blocks (MIBs), System Information Blocks (SIBs), Remaining Minimum System Information (RMSIs), and Other System Information (OSIs).

[0047] In this disclosure, physical layer signaling may include, for example, Downlink Control Information (DCI) and Uplink Control Information (UCI).

[0048] In this disclosure, terms such as index, identifier (ID), indicator, and resource ID may be interpreted interchangeably. In this disclosure, terms such as sequence, list, set, group, cluster, and subset may be interpreted interchangeably.

[0049] In this disclosure, the terms used include: panel, UE panel, panel group, beam, beam group, precoder, Uplink (UL) transmit entity, Transmission / Reception Point (TRP), base station, Spatial Relation Information (SRI), spatial relationship, SRS Resource Indicator (SRI), Control Resource Set (CORESET), Physical Downlink Shared Channel (PDSCH), Codeword (CW), Transport Block (TB), Reference Signal (RS), antenna port (e.g., Demodulation Reference Signal (DMRS) port), antenna port group (e.g., DMRS port group), group (e.g., spatial relationship group, Code Division Multiplexing (CDM) group, reference signal group, CORESET group, Physical Uplink Control Channel (PUCCH) groups, PUCCH resource groups, resources (e.g., reference signal resources, SRS resources), resource sets (e.g., reference signal resource sets), CORESET pools, downlink Transmission Configuration Indication state (TCI state) (DL TCI state), uplink TCI state (UL TCI state), unified TCI state, common TCI state, quasi-co-location (QCL), QCL assumptions, etc., may be interpreted interchangeably.

[0050] In this disclosure, CSI-RS, Non Zero Power (NZP) CSI-RS, Zero Power (ZP) CSI-RS, and CSI Interference Measurement (CSI-IM) may be interpreted interchangeably. Furthermore, CSI-RS may include other reference signals.

[0051] In this disclosure, the measured / reported RS may mean the RS measured / reported for the CSI report.

[0052] In this disclosure, timing, time, duration, slot, sub-slot, symbol, subframe, etc., may be interpreted interchangeably.

[0053] In this disclosure, terms such as direction, axis, dimension, domain, polarization, and polarization component may be interpreted interchangeably.

[0054] In this disclosure, RS may be, for example, CSI-RS, SS / PBCH block (SS block (SSB)), etc. Also, RS index may be a CSI-RS resource indicator (CRI), SS / PBCH block resource indicator (SSBRI), etc.

[0055] In this disclosure, estimation, prediction, and inference may be interpreted interchangeably. Also, in this disclosure, estimate, predict, and infer may be interpreted interchangeably.

[0056] In this disclosure, autoencoders, encoders, decoders, etc., may be interpreted as at least one of a model, ML model, neural network model, AI model, AI algorithm, etc. Furthermore, autoencoders may be interpreted as any autoencoder, such as a stacked autoencoder or a convolutional autoencoder. The encoders / decoders in this disclosure may employ models such as Residual Network (ResNet), DenseNet, and RefineNet.

[0057] Furthermore, in this disclosure, terms such as encoder, encoding, encoding, and modification / change / control by an encoder may be interpreted interchangeably. Also, in this disclosure, terms such as decoder, decoding, decoding, and modification / change / control by a decoder may be interpreted interchangeably.

[0058] In this disclosure, channel measurement / estimation may be performed using, for example, at least one of the following: a Channel State Information Reference Signal (CSI-RS), a Synchronization Signal (SS), a Synchronization Signal / Physical Broadcast Channel (SS / PBCH) block, a DeModulation Reference Signal (DMRS), or a Sounding Reference Signal (SRS).

[0059] In this disclosure, CSI may include at least one of the following: Channel Quality Indicator (CQI), Precoding Matrix Indicator (PMI), CSI-RS Resource Indicator (CRI), SS / PBCH Block Resource Indicator (SSBRI), Layer Indicator (LI), Rank Indicator (RI), L1-RSRP (Layer 1 Reference Signal Received Power), L1-RSRQ (Reference Signal Received Quality), L1-SINR (Signal to Interference plus Noise Ratio), L1-SNR (Signal to Noise Ratio).

[0060] In this disclosure, UCI, CSI report, CSI feedback, feedback information, feedback bits, etc., may be interpreted interchangeably. Also, in this disclosure, bits, bit sequences, bit series, sequences, values, information, values ​​obtained from bits, information obtained from bits, etc., may be interpreted interchangeably.

[0061] In this disclosure, the term "layer" (referring to an encoder) may be interpreted interchangeably with the terms "input layer," "hidden layer," etc., used in an AI model. The layers in this disclosure may correspond to at least one of the following: an input layer, a hidden layer, an output layer, a batch normalization layer, a convolutional layer, an activation layer, a dense layer, a normalization layer, a pooling layer, an attention layer, a dropout layer, a fully connected layer, etc.

[0062] (Wireless communication method) In this disclosure, AI model information may mean information including at least one of the following: • Input / output information of the AI ​​model, • Pre-processing / post-processing information for AI model input / output, • Information on AI model parameters, • Training information for AI models • Inference information for AI models, • Performance information regarding the AI ​​model.

[0063] Here, the input / output information of the above AI model may include information about at least one of the following: • Contents of input / output data (e.g., RSRP, SINR, amplitude / phase information in the channel matrix (or precoding matrix), information on the angle of arrival (AoA), information on the angle of departure (AoD), position information), • Input / output data type (e.g., immutable value, floating-point number) • Quantization interval (quantization step size) of input / output data (e.g., 1 dBm for L1-RSRP), • The range of possible input / output data (e.g., [0, 1]).

[0064] In this disclosure, AoA information may include information on at least one of the azimuth angle of arrival and the zenith angle of arrival (ZoA). AoD information may include, for example, information on at least one of the azimuth angle of departure and the zenith angle of departure (ZoD).

[0065] In this disclosure, location information may be location information relating to the UE / NW. Location information may include at least one of the following: information obtained using a positioning system (e.g., satellite positioning system (Global Navigation Satellite System (GNSS), Global Positioning System (GPS), etc.)) (e.g., latitude, longitude, altitude); information about a base station adjacent to (or serving) the UE (e.g., base station / cell identifier (ID), distance between BS and UE, direction / angle of BS(UE) as seen from UE(BS), coordinates of BS(UE) as seen from UE(BS) (e.g., X / Y / Z axis coordinates), etc.); and a specific address of the UE (e.g., Internet Protocol (IP) address). The location information of the UE is not limited to information based on the location of the BS, but may also be information based on a specific point.

[0066] Location information may include information about its own implementation (for example, the location / position of the antenna, the location / position of the antenna panel, the number of antennas, the number of antenna panels, etc.).

[0067] Location information may include mobility information. Mobility information may include information indicating the mobility type, the movement speed of the UE, the acceleration of the UE, and the direction of movement of the UE, or at least one of these.

[0068] Here, the mobility type may fall under at least one of the following categories: fixed location UE, movable / moving UE, no mobility UE, low mobility UE, middle mobility UE, high mobility UE, cell-edge UE, not-cell-edge UE, etc.

[0069] The pre-processing / post-processing information for the input / output of the above AI model may include information about at least one of the following: Whether or not to apply normalization (e.g., Z-score normalization, min-max normalization), • Parameters for normalization (e.g., mean / variance for Z-score normalization, minimum / maximum value for minimum-maximum normalization), Whether or not to apply a specific numerical conversion method (e.g., one-hot encoding, label encoding, etc.) • Selection rules for whether or not to use the data for training.

[0070] For example, Z-score normalization (x) is performed as a preprocessing step for input information x. new Normalized input information x = (x-μ) / σ, where μ is the mean of x and σ is the standard deviation. new You can also input this into the AI ​​model, and the output y from the AI ​​model out The final output y may be obtained by applying post-processing to the result.

[0071] The parameter information for the above AI model may include information on at least one of the following: • Weight information in AI models (e.g., neuron coefficients (connection coefficients)), • Structure of the AI ​​model • Types of AI models as model components (e.g., Residual Network (ResNet), DenseNet, RefineNet, Transformer models, CRBlock, Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU)) • Functionality of the AI ​​model as a model component (e.g., decoder, encoder).

[0072] Furthermore, the weight information in the above AI model may include information about at least one of the following: • Bit width (size) of weight information, • Quantization interval of weight information, • The range of possible weight information, • Weight parameters in AI models • Information on the differences from the AI ​​model before the update (if updating), • Methods for weight initialization (e.g., zero initialization, random initialization (based on normal distribution / uniform distribution / truncated normal distribution), Xavier initialization (for sigmoid function), He initialization (for Rectified Linear Units (ReLU))).

[0073] Furthermore, the structure of the above AI model may include information about at least one of the following: • Number of layers, • Layer type (e.g., convolutional layer, activation layer, dense layer, normalization layer, pooling layer, attention layer), Layer information, • Time series-specific parameters (e.g., bidirectionality, time step), • Parameters for training (e.g., the type of function (L2 regularization, dropout function, etc.), and where to place this function (e.g., after which layer)).

[0074] The above layer information may include information about at least one of the following: • Number of neurons in each layer, • Kernel size, • Stride for pooling layer / convolutional layer, • Pooling methods (MaxPooling, AveragePooling, etc.) Command residual block information, • Number of heads, • Normalization methods (batch normalization, instance normalization, layer normalization, etc.) • Activation functions (sigmoid, tanh function, ReLU, leaky ReLU information, Maxout, Softmax).

[0075] Figure 2 shows an example of an AI model. This example shows an AI model that includes ResNet as model component #1, a transformer model as model component #2, a dense layer, and a normalization layer. Thus, one AI model may be included as a component of another AI model. Note that Figure 2 may also show an AI model where processing proceeds from left to right.

[0076] The training information for the above AI model may include information about at least one of the following: • Information for the optimization algorithm (e.g., type of optimization (Stochastic Gradient Descent (SGD)), AdaGrad, Adam, etc.), optimization parameters (learning rate, momentum information, etc.), • Information on the loss function (for example, information on the metrics of the loss function (Mean Absolute Error (MAE), Mean Squared Error (MSE), Cross-Entropy Loss, NLL Loss, KL Divergence, etc.)), • Parameters to be frozen for training (e.g., layers, weights) • Parameters to be updated (e.g., layer, weights) • Parameters that should be used as initial parameters for training (e.g., layers, weights), • How to train / update the AI ​​model (e.g., (recommended) number of epochs, batch size, and amount of data to use for training).

[0077] The inference information for the above AI model may include information regarding decision tree branch pruning, parameter quantization, and other related matters.

[0078] The performance information for the above AI model may include information regarding the expected value of the loss function defined for the AI ​​model.

[0079] AI model information for a specific AI model may be predetermined in the standard, or it may be notified to the UE from the Network (NW). An AI model defined in the standard may be called a reference AI model. AI model information for a reference AI model may be called reference AI model information.

[0080] Furthermore, the AI ​​model information in this disclosure may include an index for identifying the AI ​​model (for example, which may be called an AI model index). The AI ​​model information in this disclosure may include, in addition to or instead of, the AI ​​model index, the input / output information of the AI ​​model described above. The association between the AI ​​model index and the AI ​​model information (for example, the input / output information of the AI ​​model) may be predetermined in the standard or notified from the network to the user entity.

[0081] <First Embodiment> The first embodiment relates to controlling the use of a reference AI model.

[0082] The UE may receive reference AI model information from the NW. Based on the received reference AI model information, the UE may deploy (make available, deploy) the above reference AI model.

[0083] Figure 3 shows an example of specifying an AI model according to the first embodiment. In this example, AI models #1-#3 are defined in the standard. A UE that receives AI model information indicating AI model #3 from the NW may deploy AI model #3.

[0084] The UE may use the reference AI model directly for inference, based on the received reference AI model information (for example, the reference AI model specified by the received reference AI model information).

[0085] Furthermore, the UE may train a reference AI model based on the received reference AI model information.

[0086] In this disclosure, the training / updating of the AI ​​model in the UE may be performed based on a dataset owned by the UE (e.g., data obtained from measurement results).

[0087] Figure 4 shows an example of updating an AI model according to the first embodiment. In this example, AI models #1 and #2 are defined in the standard. A UE that receives AI model information indicating AI model #2 from the NW may deploy AI model #2. Alternatively, the UE may update AI model #2.

[0088] Here, we will explain fine-tuning and transfer learning in relation to updating AI models. Figure 5 shows an example of fine-tuning. Figure 6 shows an example of transfer learning.

[0089] As shown in Figure 5, in fine tuning, the entity performing the training uses the parameters of the trained AI model in the source domain as the initial parameters (pre-training AI model) of the AI ​​model in the target domain to train and derive the trained AI model in the target domain. In other words, all layers of the trained AI model in the source domain may be transferred to all layers of the AI ​​model in the target domain, or some may be used as is without being updated (they may be fixed (frozen)).

[0090] As shown in Figure 6, in transfer learning (which may also be called network-based deep transfer learning), the entity performing the training reuses some of the trained layers in the source domain as layers in the target domain. These layers (the front (left) layers in the figure) can be used as feature extractors. In other words, only some of the layers of the trained AI model in the source domain may be transferred to the corresponding layers of the AI ​​model in the target domain, and these layers may be fixed (frozen) or fine-tuned. Furthermore, the remaining layers in the target domain may not have corresponding layers in the source domain and may be newly trained (or fine-tuned) in the target domain.

[0091] The UE does not need to update parameters that should be frozen for training (e.g., layers, weights) obtained from the reference AI model information during training. The UE may update parameters that should be updated (fine-tuned) obtained from the reference AI model information during training.

[0092] The UE may be trained using parameters that should be (should be used as) initial training parameters obtained from the reference AI model information.

[0093] Figure 7 shows an example of updating an AI model according to the first embodiment. In this example, the UE is instructed by the NW to update reference AI model #2. However, the instruction includes information indicating that the first three layers should be frozen and the last two layers should be fine-tuned. Based on the instruction, the UE may derive an updated AI model #2 in which the first three layers of reference AI model #2 are frozen and the last two layers are fine-tuned.

[0094] According to the first embodiment described above, the reference AI model can be appropriately utilized in the UE.

[0095] <Second Embodiment> The second embodiment relates to the control of the use of an AI model trained by a network (hereinafter also referred to as a network-trained AI model).

[0096] The network (NW) may train a reference AI model. The NW may send AI model information about the trained reference AI model (in other words, the NW-trained AI model) to the user interface (UE). This AI model information may be called the NW-trained AI model information.

[0097] The UE may receive reference AI model information from the NW, which will serve as the basis for the NW-trained AI model information. The UE may also receive NW-trained AI model information from the NW. The NW-trained AI model information may include, in place of or in addition to, the AI ​​model information already described, any or a combination of the following: • The reference AI model (and its index) that served as the basis for the NW training AI model. • Information on the difference between the reference AI model and the network-trained AI model.

[0098] Based on the received NW training AI model information, the UE may deploy (make available, or deploy) the NW training AI model.

[0099] Figure 8 shows an example of specifying an AI model according to the second embodiment. In this example, reference AI models #1 and #2 are defined in the standard. NW trains reference AI model #2 and derives NW-trained AI model #2.

[0100] The network notifies the user interface (UE) of AI model information indicating reference AI model #2, and also notifies it of information about the network-trained AI model #2 (e.g., updated differential information). Upon receiving this information, the UE can determine the network-trained AI model #2, taking into account the updated differential information for reference AI model #2.

[0101] The UE may use the NW-trained AI model, based on the received NW-trained AI model information, directly for inference.

[0102] Furthermore, the UE may train the NW training AI model based on the NW training AI model information it receives.

[0103] Figure 9 shows an example of updating the AI ​​model according to the second embodiment. In this example, as in Figure 8, it is assumed that the UE has determined the NW training AI model #2. The UE may further update the NW training AI model #2. For example, the UE may update the same layers (the first three layers shown) of the NW training AI model #2 as the original reference AI model #2, based on its own dataset.

[0104] The UE does not need to update parameters that should be frozen for training (e.g., layers, weights) obtained from the NW-trained AI model information during training. The UE may update parameters that should be updated obtained from the NW-trained AI model information during training.

[0105] The UE may perform training using parameters that should be (should be used as) initial training parameters obtained from the NW training AI model information.

[0106] Figure 10 shows an example of updating an AI model according to a second embodiment. In this example, the UE is instructed by the NW to update NW-trained AI model #2. However, the instruction includes information indicating that the first three layers should be frozen and the last two layers should be fine-tuned. Based on the instruction, the UE may derive an updated AI model #2 in which the first three layers of NW-trained AI model #2 are frozen and the last two layers are fine-tuned.

[0107] According to the second embodiment described above, the NW-trained AI model can be appropriately utilized in the UE.

[0108] <Third Embodiment> The third embodiment relates to the control of the use of an AI model trained by a UE (hereinafter also referred to as a UE-trained AI model).

[0109] The UE may train a reference AI model / Network training AI model to derive the UE training AI model.

[0110] The UE may transmit AI model information (hereinafter also referred to as UE training AI model information) about the UE training AI model to the NW using physical layer signaling (e.g., UCI), higher layer signaling (e.g., RRC signaling, MAC CE), specific signals / channels, or a combination thereof. The UE training AI model information may include, in place of or in addition to, the AI ​​model information already described, any or a combination thereof of the following: • The reference AI model / NW training AI model (and its index) that served as the basis for the UE training AI model. • Information on the difference between the reference AI model / network-trained AI model and the UE-trained AI model.

[0111] Figure 11 shows an example of specifying an AI model according to the third embodiment. In this example, the UE derives a UE training AI model. The UE transmits UE training AI model information (which may include AI model performance information) to the NW. The NW may apply the UE training AI model indicated by this information.

[0112] Figure 12 shows an example of specifying an AI model according to the third embodiment. In this example, the UE derives a UE-trained AI model from a reference AI model. The differences between these models are some of the layers highlighted in the figure. The UE transmits UE-trained AI model information to the NW. This information may include the reference AI model information, the difference information, and (AI model performance information). The NW can determine the UE-trained AI model for the reference AI model indicated by this information, taking into account the updated difference information.

[0113] The UE may report information about which weights / layers are being updated (for example, by including it in the UE training AI model information).

[0114] Figures 13A-13B and 14A-14B show examples of UE training AI model information according to the third embodiment. In this example, the UE trains an autoencoder (encoder / decoder) and, if necessary, reports UE training AI model information about the decoder to the NW.

[0115] Figure 13A represents the case where the UE does not train the AI ​​model, and since neither the encoder nor the decoder is updated, the UE-trained AI model information is not sent to the NW.

[0116] Figure 13B illustrates the case where only the encoder is updated. The UE does not need to include decoder information in the UE training AI model information, but it may include information about the estimated performance of the updated model, for example. The case in Figure 13B can be used when the change in the statistical properties of the input can be adequately addressed by modifying only the encoder.

[0117] Figure 14A illustrates a case where the encoder and part of the decoder are updated. The UE may include decoder information (information on the updated portion) in the UE training AI model information and transmit it, for example, information on the estimated performance of the updated model. The case in Figure 14A can be used when relatively small modifications to the decoder are sufficient to address changes in the statistical properties of the input.

[0118] Figure 14B illustrates the case where both the encoder and decoder are updated. The UE may include decoder information in the UE training AI model information, and may also transmit information such as the estimated performance of the updated model and a reference model report. The case in Figure 14B can be used when relatively large modifications to the decoder are required in response to changes in the statistical properties of the input.

[0119] From the standpoint of reducing communication overhead, it is preferable for the UE to apply the report in Figure 13B if the performance is sufficient in Figure 13B, the report in Figure 14A if the performance is sufficient in Figure 14A, and the report in Figure 14B if the performance is insufficient in Figure 14A. In this way, the UE may dynamically control the communication overhead associated with notifying the AI ​​model.

[0120] [Distributed Learning] The transmission of UE-trained AI model information in the third embodiment may be used for distributed learning performed using a server and multiple clients in a network. In this case, the UE-trained AI model information may be called client-trained AI model information.

[0121] Distributed learning may also be federated learning that includes at least one of the following steps: Each client reports its computing resources to the server. The server provides a global AI model to the selected client. · Each selected client trains a local AI model based on the global AI model and its own data. · Each selected client reports the parameters of the local AI model to the server. · The server combines the reported local AI models to update the global AI model and provides the updated global AI model to one or more clients (e.g., all clients).

[0122] At least some of the above steps may be repeated. FIG. 15 is a diagram showing an example of the exchange of AI model information for federated learning.

[0123] For example, in the i-th iteration (where i is an integer), the server (NW in this example) sends the AI model information of the global AI model G i to the client (UE in this example). The client updates the global AI model G i based on its own dataset to obtain the local AI model L i The client feeds back (reports) the update information i regarding the local AI model L i to the server.

[0124] The server updates the global AI model G i based on the update information i from one or more clients to obtain the global AI model G i+1 Federated learning may proceed in this way.

[0125] NW may send the AI model information regarding the global AI model to UE. The said AI model information may be called global AI model information. The global AI model information may include reference AI model information.

[0126] The global AI model information may include only information of any one or a combination of the following: • The reference AI model (and its index) that served as the basis for the global AI model. • Global AI models (index), • Information on the differences between the reference AI model and the global AI model.

[0127] Also, Global AI Model G i+1 The AI ​​model information is from Global AI Model G i This may include information about the differences from the AI ​​model (information).

[0128] Update information i may include the model ID for the global AI model / local AI model, or the global AI model G i This may include information about the difference from the AI ​​model (information), or it may correspond to the updated AI model information described later in the fourth embodiment.

[0129] UE, in update information i, or instead, local AI model L i Intermediate information (e.g., gradient, features, etc.) may also be reported.

[0130] The server may be a gNB / OAM / LMF / UE, and the client may be a UE / gNB. For example, if the UE is the client, there is the advantage that it does not need to send a large amount of training data to the gNB. For example, if the gNB is the client, there is the advantage that it does not need to send a large amount of training data to the OAM / LMF.

[0131] In this disclosure, server, NW, gNB, OAM, LMF, and UE may be interpreted interchangeably. In this disclosure, client, NW, UE, and gNB may be interpreted interchangeably. Furthermore, in this disclosure, gNB / UE may be interpreted interchangeably with other network / core network devices.

[0132] The client (e.g., UE) may report to the server (e.g., gNB) information about the resources for training the AI ​​model in the current iteration (hereinafter also referred to as resource information). The above resource information may include information about at least one of the following: • Capabilities related to training AI models (e.g., computing resources), • Information about available datasets (e.g., dataset structure / type / format / size), • Information about the client (e.g., UE status).

[0133] The computing resources mentioned above may include, for example, at least one of the following: available floating-point operations per second (FLOPS), Central Processing Unit (CPU) / Graphics Processing Unit (GPU) performance, and available memory.

[0134] The above information about the client may include at least one of the following: • Information about the client (device) type (e.g., device model name / model number, communication chip name / model number, etc.) • Estimated results / quality (e.g., Block Error Rate (BLER) / CQI / RSRP / RSRQ / SINR), • The client's willingness to handle the training (for example, whether one or more clients are willing to handle the training of an AI model). • The amount of data that the client can exchange with the server (for example, the amount of data the client can exchange), • Battery power level.

[0135] According to the third embodiment described above, the UE training AI model can be used appropriately.

[0136] <Fourth Embodiment> The fourth embodiment relates to information about the difference between one AI model and another AI model (information about the difference between one AI model and another AI model relative to one AI model). This difference information may be called updated AI model information.

[0137] The updated AI model information may be used, for example, as at least one of the following pieces of information already mentioned: • Information on the differences between the AI ​​model before the update and the AI ​​model after the update. • Information on the difference between the reference AI model and the NW-trained AI model. • Information on the difference between the reference AI model / network-trained AI model and the UE-trained AI model.

[0138] The updated AI model information may include weight information expressed as absolute values, or it may include weight information expressed as the difference between the AI ​​model before the update and the AI ​​model after the update.

[0139] Figures 16A-16C show an example of updated AI model information. In this example, the updated AI model information is shown when the NW-trained AI model is trained from the reference model shown in Figure 16A.

[0140] Figure 16B shows the weight information, expressed as absolute values, included in the updated AI model information. In this example, the last layer has not been updated and therefore does not need to be reported. The UE does not need to expect to report / receive AI model information that has not been updated.

[0141] Figure 16C shows the weight information, represented by difference values, included in the updated AI model information. Since difference values ​​are generally expected to be smaller than absolute values, a reduction in the amount of information can be expected. Furthermore, if the amount of information in Figure 16C is the same as in Figure 16B, the granularity of each weight information can be reduced, enabling a more detailed representation of the model.

[0142] Furthermore, if the reference model information includes information about the structure of the AI ​​model, the updated AI model information does not need to include information about the structure of the AI ​​model. This is because the UE and the base station can determine the AI ​​model based on the same AI model structure by exchanging only weight information.

[0143] According to the fourth embodiment described above, the signaling overhead for AI model deployment can be appropriately reduced.

[0144] <Fifth Embodiment> The fifth embodiment relates to knowledge distillation.

[0145] In knowledge distillation, a single large model or an ensemble of models with good predictive accuracy is prepared as a teacher model, and the knowledge contained in the teacher model is used to train a lightweight and easily deployable student model. The student model obtained through distillation is expected to be a model with similar accuracy to the teacher model but lighter in weight.

[0146] Figures 17A and 17B illustrate an example of learning using knowledge distillation. In typical learning, an AI model is trained so that the output obtained from a given dataset (training data) (output from layer m) becomes as close to the correct answer as possible.

[0147] Furthermore, the variables / values ​​of the output from layer m before one or more functions (e.g., sigmoid function / softmax function) are applied may be called logits, and the variables / values ​​after the application of those one or more functions may be called probs. The sum of the probs for all possible classes of the correct answer is 1.

[0148] In the typical learning process shown in Figure 17A, the correct labels of the training data are used as hard targets, and learning is performed in a way that minimizes the loss between the problems and the hard targets (also called hard target loss, inference reference loss, etc.).

[0149] On the other hand, in learning using knowledge distillation as shown in Figure 17B, the teacher's output (probs) is used as a soft target, and learning is performed in a way that minimizes the loss between the student's probs and the soft target (also called soft target loss, distillation loss, etc.). In knowledge distillation, the one or more functions mentioned above may be, for example, a temperature-controlled softmax function. The soft target may correspond to information about the probability that the input belongs to each class.

[0150] If we denote the teacher model's logitz as v and its prob as p, and the student model's logitz as z and its prob as q, then the prob p for class i of the teacher model when the temperature parameter is T is given by i is, p i =exp(v i / T) / Σ j exp(v j It may also be obtained by ( / T), and the problem q for class i of the teacher model when the temperature parameter is T i is, q i =exp(z i / T) / Σ j exp(z j It may also be determined by / T).

[0151] Soft target loss is p of the teacher model. i and student model p i Based on this, the cross-entropy can also be taken to determine the soft target loss, for example, soft target loss = -Σ i p i log(q i ) may also be requested.

[0152] Note that the hard target loss is q for T=1. i It may be determined based on the following criteria.

[0153] In the fifth embodiment, the UE may receive AI model information (hereinafter also referred to as student model information) relating to a student model. The UE may obtain an output using the student model specified based on the received student model information (train the student model) and report information about the output to the NW. This step may be called retraining the student model.

[0154] The student model information may include, in place of or in addition to, the AI ​​model information already described, any or a combination of the following: • Student model (index), • Teacher model (index), • Information regarding the dataset used to calculate the teacher model / student model. • Soft target values ​​for teacher model / student model, • Output values ​​from a layer of the teacher / student model, • Values ​​of the loss function for the teacher model / student model, • Soft target values ​​for teacher / student models, • Output values ​​from a certain layer for the teacher model / student model, • Values ​​of the loss function for the teacher model / student model.

[0155] The information regarding the dataset may include, for example, at least one of the following: a batch of data, an index indicating the dataset (or data), or the dataset (or data) itself. The information regarding the value of the loss function may include information on either or both of the soft target loss and the hard target loss.

[0156] The UE may transmit information about the student model to be retrained (hereinafter also referred to as retrained student model information) to the NW using physical layer signaling (e.g., UCI), higher layer signaling (e.g., RRC signaling, MAC CE), specific signals / channels, or a combination thereof. The retrained student model information may include any or a combination thereof of the following information instead of or in addition to the AI ​​model information already described: • The student model (and its index) that served as the basis for the retraining student model. • Information about the knowledge of the retrained student model (e.g., soft targets, output from a certain layer), • Information about the dataset used for retraining. • Intermediate information of the retrained student model (e.g., gradients, features, etc.) • Information regarding hard target loss in the retraining student model.

[0157] The information regarding the above dataset may include information regarding the content of the data contained in that dataset.

[0158] The network (NW) may further update the student model based on soft targets identified by information about the student model being retrained and soft targets in the training data. This update of the NW may also be referred to as retraining.

[0159] Furthermore, if the UE receives information about the teacher model from the NW (hereinafter also referred to as teacher model information), it may use such teacher model information to retrain the student model mentioned above.

[0160] According to the fifth embodiment, even if the UE uses a lightweight student model considering computational power, by reporting retrained student model information to the NW, the NW can perform knowledge distillation again to generate a more accurate / lighter student model.

[0161] <Supplement> Any notification from the NW to the UE in the above-described embodiments may be made using physical layer signaling (e.g., DCI), higher layer signaling (e.g., RRC signaling, MAC CE), specific signals / channels (e.g., PDCCH, PDSCH, reference signal), or a combination thereof.

[0162] If the above notification is made by a MAC CE, the MAC CE may be identified by the inclusion of a new Logical Channel ID (LCID) in the MAC subheader.

[0163] If the above notification is made by a DCI, the notification may be made by a specific field of the DCI, a Radio Network Temporary Identifier (RNTI) used to scramble the Cyclic Redundancy Check (CRC) bits assigned to the DCI, or the format of the DCI.

[0164] Furthermore, any notifications from the NW to the UE in the above-described embodiments may be made periodically, semi-persistently, or aperiodically.

[0165] In the embodiments described above, the encoder / decoder may be interpreted as the AI ​​model deployed in the UE / base station. In other words, this disclosure is not limited to the use of an autoencoder, but may also apply to inference using any model. Furthermore, the information that the UE / base station compresses using the encoder in this disclosure is not limited to the CSI (or channel / precoding matrix), but may be any information.

[0166] At least one of the embodiments described above may apply only to a UE that has reported or supports a particular UE capability.

[0167] The specific UE capability may represent at least one of the following: • To support specific processing / operation / control / information for at least one of the above embodiments, • Support for acquiring / reporting channel characteristic information (e.g., LOS, NLOS, location information). • The maximum FLOPs of AI models that UE can deploy, • Maximum number of parameters for an AI model that UE can deploy, • Reference models supported by UE, • Layers / algorithms / functions supported by UE, ·Numeracy ability, • Capability to collect data.

[0168] Furthermore, the specific UE capabilities described above may be capabilities that apply across all frequencies (commonly regardless of frequency), capabilities per frequency (e.g., cell, band, BWP), capabilities per frequency range (e.g., Frequency Range 1 (FR1), FR2, FR3, FR4, FR5, FR2-1, FR2-2), or capabilities per subcarrier spacing (SCS).

[0169] Furthermore, the specific UE capabilities described above may be capabilities that apply across all duplexing schemes (common to all duplexing schemes), or they may be capabilities specific to each duplexing scheme (e.g., Time Division Duplex (TDD), Frequency Division Duplex (FDD)).

[0170] Furthermore, at least one of the embodiments described above may be applied when the UE is configured with specific information related to the embodiments described above by upper-layer signaling. For example, such specific information may be information indicating the activation of the AI ​​model, or arbitrary RRC parameters for a particular release (e.g., Rel.18).

[0171] If the UE does not support at least one of the above-mentioned specific UE capabilities or does not have the above-mentioned specific information configured, the behavior of, for example, Rel.15 / 16 may be applied.

[0172] Furthermore, at least one of the embodiments described above may be used for the transmission (or compression) of information between the UE and the base station other than CSI feedback. For example, the UE may report information regarding its location (or positioning) / location estimation in the Location Management Function (LMF) to the network according to at least one of the embodiments described above (e.g., generated using an encoder). This information may be channel impulse response (CIR) information for each subband / antenna port. By reporting this, the base station can estimate the UE's location without reporting the angle / time difference of the received signal, etc.

[0173] (Note) The following invention is added with respect to one embodiment of this disclosure. [Note 1] A receiving unit that receives NW training model information regarding a model trained by a network (NW) and reference model information regarding a reference model that serves as the basis for the NW training model. A terminal having a control unit that determines the NW training model based on the NW training model information and the reference model information. [Note 2] The NW training model information indicates the difference between the NW training model and the reference model, and the reference model information indicates the terminal described in Appendix 1, which shows an index related to the reference model. [Note 3] The control unit further trains the NW training model. The terminals listed in Appendix 1 or Appendix 2. [Note 4] The aforementioned NW training model information includes information on parameters that should be frozen for training. The control unit further trains the NW training model, where the parameters are not updated, as described in any of the terminals listed in Appendix 1 to Appendix 3.

[0174] (Note) The following invention is added with respect to one embodiment of this disclosure. [Note 1] A receiving unit that receives NW training model information regarding a model trained by a network (NW training model) or reference model information regarding a reference model that serves as the basis for the NW training model, A control unit that further trains a model determined based on the aforementioned NW training model information or the aforementioned reference model information, A terminal having a transmitting unit that transmits information about the trained model. [Note 2] The information regarding the trained model is provided in the terminal described in Appendix 1, which includes information on which weights or layers are updated. [Note 3] The information regarding the aforementioned trained model is provided by the terminal described in Appendix 1 or Appendix 2, which includes updates to the global model in federative learning. [Note 4] The transmitting unit is a terminal according to any one of the appendices 1 to 3 that transmits information about resources for model training in the current iteration in associative learning.

[0175] (Note) The following invention is added with respect to one embodiment of this disclosure. [Note 1] A receiving unit that receives student model information regarding the student model in knowledge distillation, A terminal having a transmitting unit that transmits information relating to output based on the aforementioned student model information. [Note 2] The information relating to the output is the terminal described in Appendix 1, which includes information relating to the soft target of the student model. [Note 3] The aforementioned student model information is the terminal specified in Appendix 1 or Appendix 2, which includes the value of the loss function. [Note 4] The aforementioned student model information is a terminal described in any of Appendix 1 to Appendix 3, which includes information on either or both soft target loss and hard target loss.

[0176] (Wireless communication system) The configuration of a wireless communication system according to one embodiment of this disclosure will be described below. In this wireless communication system, communication is performed using any or a combination thereof of the wireless communication methods according to the above embodiments of this disclosure.

[0177] Figure 18 shows an example of a schematic configuration of a wireless communication system according to one embodiment. The wireless communication system 1 may be a system that realizes communication using Long Term Evolution (LTE), 5th generation mobile communication system New Radio (5G NR), etc., as specified by the Third Generation Partnership Project (3GPP).

[0178] Furthermore, the wireless communication system 1 may support dual connectivity between multiple Radio Access Technologies (RATs) (Multi-RAT Dual Connectivity (MR-DC)). MR-DC may include dual connectivity between LTE (Evolved Universal Terrestrial Radio Access (E-UTRA)) and NR (E-UTRA-NR Dual Connectivity (EN-DC)), dual connectivity between NR and LTE (NR-E-UTRA Dual Connectivity (NE-DC)), and so on.

[0179] In EN-DC, the LTE (E-UTRA) base station (eNB) is the Master Node (MN), and the NR base station (gNB) is the Secondary Node (SN). In NE-DC, the NR base station (gNB) is the MN, and the LTE (E-UTRA) base station (eNB) is the SN.

[0180] The wireless communication system 1 may support dual connectivity between multiple base stations within the same RAT (for example, dual connectivity where both MN and SN are NR base stations (gNB) (NR-NR Dual Connectivity (NN-DC))).

[0181] The wireless communication system 1 may include a base station 11 that forms a macrocell C1 with relatively wide coverage, and base stations 12 (12a-12c) located within the macrocell C1 that form a small cell C2 that is narrower than the macrocell C1. User terminals 20 may be located within at least one cell. The arrangement and number of each cell and user terminal 20 are not limited to the configuration shown in the figure. Hereinafter, when base stations 11 and 12 are not distinguished, they will be collectively referred to as base station 10.

[0182] The user terminal 20 may be connected to at least one of the multiple base stations 10. The user terminal 20 may utilize at least one of Carrier Aggregation (CA) using multiple Component Carriers (CC) and Dual Connectivity (DC).

[0183] Each CC may be included in at least one of the first frequency band (Frequency Range 1 (FR1)) and the second frequency band (Frequency Range 2 (FR2)). A macrocell C1 may be included in FR1, and a small cell C2 may be included in FR2. For example, FR1 may be a frequency band of 6 GHz or less (sub-6 GHz), and FR2 may be a frequency band above 24 GHz (above-24 GHz). Note that the frequency bands and definitions of FR1 and FR2 are not limited to these, and for example, FR1 may fall in a frequency band higher than FR2.

[0184] Furthermore, the user terminal 20 may communicate using at least one of the following methods at each CC: Time Division Duplex (TDD) and Frequency Division Duplex (FDD).

[0185] Multiple base stations 10 may be connected by wire (e.g., optical fiber compliant with Common Public Radio Interface (CPRI), X2 interface, etc.) or wireless (e.g., NR communication). For example, if NR communication is used as a backhaul between base stations 11 and 12, base station 11, which is the upstream station, may be called an Integrated Access Backhaul (IAB) donor, and base station 12, which is the relay station, may be called an IAB node.

[0186] Base station 10 may be connected to the core network 30 via other base stations 10 or directly. The core network 30 may include at least one of the following: Evolved Packet Core (EPC), 5G Core Network (5GCN), Next Generation Core (NGC), etc.

[0187] The user terminal 20 may be a terminal that supports at least one of the following communication methods: LTE, LTE-A, 5G, etc.

[0188] In the wireless communication system 1, an orthogonal frequency division multiplexing (OFDM)-based wireless access scheme may be used. For example, Cyclic Prefix OFDM (CP-OFDM), Discrete Fourier Transform Spread OFDM (DFT-s-OFDM), Orthogonal Frequency Division Multiple Access (OFDMA), Single Carrier Frequency Division Multiple Access (SC-FDMA), etc., may be used in at least one of the downlink (DL) and uplink (UL).

[0189] The wireless access method may also be called a waveform. In wireless communication system 1, other wireless access methods (for example, other single-carrier transmission methods, other multi-carrier transmission methods) may be used for the UL and DL wireless access methods.

[0190] In the wireless communication system 1, a Physical Downlink Shared Channel (PDSCH), a Broadcast Channel (PBCH), or a Physical Downlink Control Channel (PDCCH) may be used as the downlink channel, shared by each user terminal 20.

[0191] Furthermore, in the wireless communication system 1, the uplink channel may include a Physical Uplink Shared Channel (PUSCH), a Physical Uplink Control Channel (PUCCH), a Physical Random Access Channel (PRACH), or the like, all of which are shared by each user terminal 20.

[0192] User data, higher-layer control information, and System Information Blocks (SIBs) are transmitted via PDSCH. User data and higher-layer control information may also be transmitted via PUSCH. Furthermore, Master Information Blocks (MIBs) may be transmitted via PBCH.

[0193] Lower-layer control information may be transmitted by PDCCH. The lower-layer control information may include, for example, Downlink Control Information (DCI) which includes scheduling information for at least one of PDSCH and PUSCH.

[0194] Furthermore, the DCI that schedules PDSCH may be called a DL assignment or DL ​​DCI, and the DCI that schedules PUSCH may be called a UL grant or UL DCI. Furthermore, PDSCH may be interpreted as DL data, and PUSCH may be interpreted as UL data.

[0195] PDCCH detection may utilize a Control Resource Set (CORESET) and a search space. A CORESET corresponds to the resources used to search for DCIs. A search space corresponds to the search area and search method for PDCCH candidates. A single CORESET may be associated with one or more search spaces. The UE may monitor CORESETs associated with a particular search space based on the search space configuration.

[0196] A single search space may correspond to one or more PDCCH candidates corresponding to aggregation levels. One or more search spaces may be referred to as a search space set. In this disclosure, "search space," "search space set," "search space configuration," "search space set configuration," "CORESET," and "CORESET configuration" may be interpreted interchangeably.

[0197] PUCCH may transmit uplink control information (UCI) which includes at least one of the following: channel state information (CSI), delivery acknowledgment (e.g., Hybrid Automatic Repeat reQuest ACKnowledgement (HARQ-ACK), ACK / NACK, etc.), and scheduling request (SR). PRACH may transmit a random access preamble for establishing a connection with the cell.

[0198] In this disclosure, downlinks, uplinks, etc., may be expressed without the prefix "link." Also, the prefix "physical" may be omitted when describing various channels.

[0199] In the wireless communication system 1, a synchronization signal (SS), a downlink reference signal (DL-RS), etc., may be transmitted. In the wireless communication system 1, as DL-RS, a cell-specific reference signal (CRS), a channel state information reference signal (CSI-RS), a demodulation reference signal (DMRS), a positioning reference signal (PRS), a phase tracking reference signal (PTRS), etc., may be transmitted.

[0200] The synchronization signal may be, for example, at least one of a Primary Synchronization Signal (PSS) and a Secondary Synchronization Signal (SSS). A signal block including SS (PSS, SSS) and PBCH (and DMRS for PBCH) may be called an SS / PBCH block, SS Block (SSB), etc. SS, SSB, etc., may also be called reference signals.

[0201] Furthermore, in the wireless communication system 1, the Uplink Reference Signal (UL-RS) may transmit the Sounding Reference Signal (SRS), Demodulation Reference Signal (DMRS), etc. The DMRS may also be called the User-Specific Reference Signal (UE-specific Reference Signal).

[0202] (base station) Figure 19 shows an example of the configuration of a base station according to one embodiment. The base station 10 includes a control unit 110, a transceiver unit 120, a transceiver antenna 130, and a transmission line interface 140. Note that one or more of the control unit 110, transceiver unit 120, transceiver antenna 130, and transmission line interface 140 may be provided.

[0203] In this example, the functional blocks of the characteristic parts of this embodiment are mainly shown, and it may be assumed that the base station 10 also has other functional blocks necessary for wireless communication. Some of the processing of each part described below may be omitted.

[0204] The control unit 110 controls the entire base station 10. The control unit 110 can consist of a controller, control circuit, etc., as described based on common understanding in the art relating to this disclosure.

[0205] The control unit 110 may control signal generation, scheduling (e.g., resource allocation, mapping), etc. The control unit 110 may also control transmission and reception, measurement, etc., using the transceiver unit 120, the transceiver antenna 130, and the transmission path interface 140. The control unit 110 may generate data to be transmitted as signals, control information, sequences, etc., and transfer them to the transceiver unit 120. The control unit 110 may also perform call processing of communication channels (setting, releasing, etc.), status management of the base station 10, management of radio resources, etc.

[0206] The transmitting / receiving unit 120 may include a baseband unit 121, a radio frequency (RF) unit 122, and a measurement unit 123. The baseband unit 121 may include a transmission processing unit 1211 and a reception processing unit 1212. The transmitting / receiving unit 120 can be composed of a transmitter / receiver, RF circuit, baseband circuit, filter, phase shifter, measurement circuit, transmitting / receiving circuit, etc., as described based on common understanding in the art relating to this disclosure.

[0207] The transmitting / receiving unit 120 may be configured as an integrated transmitting / receiving unit, or it may be composed of a transmitting unit and a receiving unit. The transmitting unit may consist of a transmitting processing unit 1211 and an RF unit 122. The receiving unit may consist of a receiving processing unit 1212, an RF unit 122 and a measuring unit 123.

[0208] The transmitting and receiving antenna 130 can be composed of an antenna described based on common understanding in the art relating to this disclosure, such as an array antenna.

[0209] The transmitting / receiving unit 120 may transmit the downlink channel, synchronization signal, downlink reference signal, etc. The transmitting / receiving unit 120 may also receive the uplink channel, uplink reference signal, etc.

[0210] The transmitting / receiving unit 120 may form at least one of the transmitting beam and the receiving beam using digital beamforming (e.g., precoding), analog beamforming (e.g., phase rotation), or the like.

[0211] The transmitting / receiving unit 120 (transmission processing unit 1211) may perform processing on data and control information acquired from the control unit 110, for example, at the Packet Data Convergence Protocol (PDCP) layer, the Radio Link Control (RLC) layer (e.g., RLC retransmission control), the Medium Access Control (MAC) layer (e.g., HARQ retransmission control), etc., to generate a bit sequence to be transmitted.

[0212] The transmitting / receiving unit 120 (transmission processing unit 1211) may perform transmission processing on the bit sequence to be transmitted, such as channel coding (which may include error correction coding), modulation, mapping, filtering, discrete Fourier transform (DFT) processing (if necessary), inverse fast Fourier transform (IFFT) processing, precoding, and digital-to-analog conversion, and output a baseband signal.

[0213] The transmitting / receiving unit 120 (RF unit 122) may perform modulation, filtering, amplification, etc., of the baseband signal to the radio frequency band and transmit the signal in the radio frequency band via the transmitting / receiving antenna 130.

[0214] On the other hand, the transmitting / receiving unit 120 (RF unit 122) may perform amplification, filtering, demodulation to a baseband signal, etc., on the radio frequency band signal received by the transmitting / receiving antenna 130.

[0215] The transmitting / receiving unit 120 (receiving processing unit 1212) may apply reception processing to the acquired baseband signal, such as analog-to-digital conversion, Fast Fourier Transform (FFT) processing, Inverse Discrete Fourier Transform (IDFT) processing (if necessary), filtering, demapping, demodulation, decoding (may include error correction decoding), MAC layer processing, RLC layer processing, and PDCP layer processing, to acquire user data, etc.

[0216] The transmitting / receiving unit 120 (measurement unit 123) may perform measurements related to the received signal. For example, the measurement unit 123 may perform Radio Resource Management (RRM) measurements, Channel State Information (CSI) measurements, etc., based on the received signal. The measurement unit 123 may also measure received power (e.g., Reference Signal Received Power (RSRP)), reception quality (e.g., Reference Signal Received Quality (RSRQ), Signal to Interference plus Noise Ratio (SINR), Signal to Noise Ratio (SNR)), signal strength (e.g., Received Signal Strength Indicator (RSSI)), propagation path information (e.g., CSI), etc. The measurement results may be output to the control unit 110.

[0217] The transmission path interface 140 may send and receive signals (backhaul signaling) with devices included in the core network 30, other base stations 10, etc., and may acquire and transmit user data (user plane data), control plane data, etc. for the user terminal 20.

[0218] In this disclosure, the transmitting and receiving units of the base station 10 may consist of at least one of a transmitting / receiving unit 120, a transmitting / receiving antenna 130, and a transmission path interface 140.

[0219] The control unit 110 may also train a reference model. The transmitting / receiving unit 120 may transmit information about the trained model and reference model information about the reference model to the user terminal 20.

[0220] Furthermore, the transmitting / receiving unit 120 may transmit to the user terminal 20 reference model information relating to the reference model or information relating to a model trained based on the reference model. The transmitting / receiving unit 120 may also receive information relating to the reference model or a model further trained based on the model.

[0221] In addition, the transmission / reception unit 120 may transmit pupil model information regarding the pupil model in knowledge distillation to the user terminal 20. The transmission / reception unit 120 may receive information regarding an output based on the pupil model information.

[0222] (User terminal) FIG. 20 is a diagram showing an example of the configuration of a user terminal according to an embodiment. The user terminal 20 includes a control unit 210, a transmission / reception unit 220, and a transmission / reception antenna 230. Note that one or more of the control unit 210, the transmission / reception unit 220, and the transmission / reception antenna 230 may be provided.

[0223] Note that in this example, the functional blocks of the characteristic parts in the present embodiment are mainly shown, and the user terminal 20 may be assumed to have other functional blocks necessary for wireless communication. A part of the processing of each unit described below may be omitted.

[0224] The control unit 210 controls the entire user terminal 20. The control unit 210 can be composed of a controller, a control circuit, etc., which are described based on the common knowledge in the technical field related to the present disclosure.

[0225] The control unit 210 may control signal generation, mapping, etc. The control unit 210 may control transmission, reception, measurement, etc. using the transmission / reception unit 220 and the transmission / reception antenna 230. The control unit 210 may generate data, control information, sequences, etc. to be transmitted as signals and transfer them to the transmission / reception unit 220.

[0226] The transmission / reception unit 220 may include a baseband unit 221, an RF unit 222, and a measurement unit 223. The baseband unit 221 may include a transmission processing unit 2211 and a reception processing unit 2212. The transmission / reception unit 220 can be composed of a transmitter / receiver, an RF circuit, a baseband circuit, a filter, a phase shifter, a measurement circuit, a transmission / reception circuit, etc., which are described based on the common knowledge in the technical field related to the present disclosure.

[0227] The transmitting / receiving unit 220 may be configured as an integrated transmitting / receiving unit, or it may be composed of a transmitting unit and a receiving unit. The transmitting unit may consist of a transmitting processing unit 2211 and an RF unit 222. The receiving unit may consist of a receiving processing unit 2212, an RF unit 222 and a measuring unit 223.

[0228] The transmitting and receiving antenna 230 can be composed of an antenna described based on common understanding in the art relating to this disclosure, such as an array antenna.

[0229] The transmitting / receiving unit 220 may receive the downlink channel, synchronization signal, downlink reference signal, etc. The transmitting / receiving unit 220 may also transmit the uplink channel, uplink reference signal, etc.

[0230] The transmitting / receiving unit 220 may form at least one of the transmitting beam and the receiving beam using digital beamforming (e.g., precoding), analog beamforming (e.g., phase rotation), or the like.

[0231] The transmitting / receiving unit 220 (transmission processing unit 2211) may perform PDCP layer processing, RLC layer processing (e.g., RLC retransmission control), MAC layer processing (e.g., HARQ retransmission control), etc., on data and control information acquired from the control unit 210, etc., to generate a bit sequence to be transmitted.

[0232] The transmitting / receiving unit 220 (transmission processing unit 2211) may perform transmission processing on the bit sequence to be transmitted, such as channel coding (which may include error correction coding), modulation, mapping, filtering, DFT processing (if necessary), IFFT processing, precoding, and digital-to-analog conversion, and output a baseband signal.

[0233] Whether or not to apply DFT processing may be based on the transform precoding settings. The transmitting / receiving unit 220 (transmission processing unit 2211) may perform DFT processing as part of the transmission process to transmit a channel (for example, PUSCH) using a DFT-s-OFDM waveform if transform precoding is enabled for that channel, or it may not perform DFT processing as part of the transmission process if transform precoding is not enabled for that channel.

[0234] The transmitting / receiving unit 220 (RF unit 222) may perform modulation, filtering, amplification, etc., of the baseband signal to the radio frequency band and transmit the signal in the radio frequency band via the transmitting / receiving antenna 230.

[0235] On the other hand, the transmitting / receiving unit 220 (RF unit 222) may perform amplification, filtering, demodulation to a baseband signal, etc., on the radio frequency band signal received by the transmitting / receiving antenna 230.

[0236] The transmitting / receiving unit 220 (receiving processing unit 2212) may apply reception processing such as analog-to-digital conversion, FFT processing, IDFT processing (if necessary), filtering, demapping, demodulation, decoding (may include error correction decoding), MAC layer processing, RLC layer processing, and PDCP layer processing to the acquired baseband signal to acquire user data, etc.

[0237] The transmitting / receiving unit 220 (measuring unit 223) may perform measurements related to the received signal. For example, the measuring unit 223 may perform RRM measurement, CSI measurement, etc., based on the received signal. The measuring unit 223 may also measure received power (e.g., RSRP), received quality (e.g., RSRQ, SINR, SNR), signal strength (e.g., RSSI), propagation path information (e.g., CSI), etc. The measurement results may be output to the control unit 210.

[0238] Note that the transmission unit and reception unit of the user terminal 20 in the present disclosure may be configured by at least one of a transmission / reception unit 220 and a transmission / reception antenna 230.

[0239] Note that the transmission / reception unit 220 may receive NW training model information regarding a model (NW training model) trained by a network (Network (NW)) and reference model information regarding a reference model that is the basis of the NW training model. The control unit 210 may determine the NW training model based on the NW training model information and the reference model information.

[0240] The NW training model information may indicate a difference between the NW training model and the reference model, and the reference model information may indicate an index regarding the reference model.

[0241] The control unit 210 may further train the NW training model.

[0242] The NW training model information includes information on parameters to be frozen for training, and the control unit 210 further trains the NW training model, where the parameters may not be updated.

[0243] Also, the transmission / reception unit 220 may receive NW training model information regarding a model (NW training model) trained by a network (Network (NW)) or reference model information regarding a reference model that is the basis of the NW training model. The control unit 210 may further train a model determined based on the NW training model information or the reference model information. The transmission / reception unit 220 may transmit information regarding the trained model.

[0244] The information regarding the trained model may include information on which weights or layers are updated.

[0245] The information regarding the trained model may include update information of the global model in federated learning.

[0246] The transmitter 220 may transmit information about the resources for model training in the current iteration in associative learning.

[0247] The transmitting / receiving unit 220 may also receive student model information relating to the student model in knowledge distillation. The transmitting / receiving unit 220 may also transmit information relating to the output based on the student model information.

[0248] The information relating to the output may include information relating to the soft target of the student model.

[0249] The aforementioned student model information may include the value of the loss function.

[0250] The aforementioned student model information may include information on either or both of the soft target loss and hard target loss.

[0251] (Hardware configuration) The block diagrams used in the description of the above embodiments show functional units. These functional blocks (components) are realized by any combination of at least one of hardware and software. Furthermore, the method of realizing each functional block is not particularly limited. That is, each functional block may be realized using one device that is physically or logically coupled, or it may be realized using two or more physically or logically separated devices that are directly or indirectly connected (for example, using wired or wireless connections). A functional block may also be realized by combining the above one device or the above multiple devices with software.

[0252] Here, functions include, but are not limited to, judgment, decision, determination, calculation, calculation, processing, derivation, investigation, search, confirmation, reception, transmission, output, access, resolution, selection, selection, establishment, comparison, assumption, expectation, consideration, broadcasting, notifying, communicating, forwarding, configuring, reconfiguring, allocating (mapping), and assigning. For example, a functional block (configuration part) that enables transmission may be called a transmitting unit or transmitter. In all cases, as mentioned above, the method of implementation is not particularly limited.

[0253] For example, a base station, user terminal, etc. in one embodiment of the present disclosure may function as a computer that processes the wireless communication method of the present disclosure. Figure 21 is a diagram showing an example of the hardware configuration of a base station and user terminal according to one embodiment. The base station 10 and user terminal 20 described above may be physically configured as a computer device including a processor 1001, memory 1002, storage 1003, communication device 1004, input device 1005, output device 1006, bus 1007, etc.

[0254] In this disclosure, terms such as apparatus, circuit, device, section, and unit are interchangeable. The hardware configuration of the base station 10 and the user terminal 20 may include one or more of the devices shown in the figure, or it may be configured to omit some of the devices.

[0255] For example, although only one processor 1001 is shown in the diagram, there may be multiple processors. Furthermore, processing may be performed by one processor, or by two or more processors simultaneously, sequentially, or by other means. Note that processor 1001 may be implemented using one or more chips.

[0256] Each function in the base station 10 and the user terminal 20 is realized, for example, by loading predetermined software (programs) onto hardware such as the processor 1001 and memory 1002, which allows the processor 1001 to perform calculations and control communication via the communication device 1004, or to control at least one of the reading and writing of data in the memory 1002 and storage 1003.

[0257] The processor 1001 controls the entire computer, for example, by running an operating system. The processor 1001 may be composed of a central processing unit (CPU) that includes interfaces with peripheral devices, control units, arithmetic units, registers, etc. For example, at least a part of the control unit 110 (210) and the transmitting / receiving unit 120 (220) described above may be implemented by the processor 1001.

[0258] Furthermore, the processor 1001 reads programs (program code), software modules, data, etc., from at least one of the storage 1003 and the communication device 1004 into the memory 1002 and executes various processes accordingly. The program used is one that causes the computer to execute at least a part of the operations described in the above embodiment. For example, the control unit 110 (210) may be implemented by a control program stored in the memory 1002 and running on the processor 1001, and other functional blocks may be implemented similarly.

[0259] Memory 1002 is a computer-readable recording medium and may consist of at least one of the following: Read Only Memory (ROM), Erasable Programmable ROM (EPROM), Electrically EPROM (EEPROM), Random Access Memory (RAM), or other suitable storage medium. Memory 1002 may also be called a register, cache, or main memory. Memory 1002 can store executable programs (program code), software modules, etc., for carrying out a wireless communication method according to one embodiment of this disclosure.

[0260] Storage 1003 is a computer-readable recording medium and may consist of at least one of the following: a flexible disk, a floppy disk, a magneto-optical disk (e.g., a compact disk (Compact Disc ROM (CD-ROM)), a digital multipurpose disk, a Blu-ray disk), a removable disk, a hard disk drive, a smart card, a flash memory device (e.g., a card, stick, key drive), a magnetic stripe, a database, a server, or other suitable storage medium. Storage 1003 may also be called an auxiliary storage device.

[0261] The communication device 1004 is hardware (transmitting / receiving device) for communicating between computers via at least one of a wired network and a wireless network, and is also referred to as a network device, network controller, network card, communication module, etc. The communication device 1004 may be configured to include, for example, a high-frequency switch, duplexer, filter, frequency synthesizer, etc., in order to implement at least one of frequency division duplex (FDD) and time division duplex (TDD). For example, the above-mentioned transmitting / receiving unit 120 (220), transmitting / receiving antenna 130 (230), etc., may be implemented by the communication device 1004. The transmitting / receiving unit 120 (220) may be implemented with physically or logically separated implementations of a transmitting unit 120a (220a) and a receiving unit 120b (220b).

[0262] The input device 1005 is an input device that accepts input from an external source (e.g., a keyboard, mouse, microphone, switch, button, sensor, etc.). The output device 1006 is an output device that outputs to an external source (e.g., a display, speaker, light-emitting diode (LED) lamp, etc.). The input device 1005 and the output device 1006 may be configured as an integrated unit (e.g., a touch panel).

[0263] Furthermore, each device, such as the processor 1001 and memory 1002, is connected by a bus 1007 for communicating information. The bus 1007 may be configured using a single bus, or different buses may be configured for each device.

[0264] Furthermore, the base station 10 and the user terminal 20 may be configured to include hardware such as a microprocessor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a programmable logic device (PLD), and a field programmable gate array (FPGA), and some or all of each functional block may be implemented using such hardware. For example, the processor 1001 may be implemented using at least one of these hardware components.

[0265] (modified version) In addition, terms used in this disclosure and terms necessary for understanding this disclosure may be replaced with terms having the same or similar meanings. For example, channel, symbol, and signal (signal or signaling) may be used interchangeably. Also, a signal may be a message. A reference signal may be abbreviated as RS and may be called a pilot, pilot signal, etc., depending on the applicable standard. Also, a component carrier (CC) may be called a cell, frequency carrier, carrier frequency, etc.

[0266] A wireless frame may consist of one or more periods (frames) in the time domain. Each of these periods (frames) constituting a wireless frame may be called a subframe. Furthermore, a subframe may consist of one or more slots in the time domain. A subframe may have a fixed time length (e.g., 1 ms) that is independent of numerology.

[0267] Here, the neuralelogy may be communication parameters applied to at least one of the transmission and reception of a signal or channel. The neuralelogy may be, for example, at least one of the following: subcarrier spacing (SCS), bandwidth, symbol length, cyclic prefix length, transmission time interval (TTI), number of symbols per TTI, radio frame configuration, specific filtering processes performed by the transceiver in the frequency domain, or specific windowing processes performed by the transceiver in the time domain.

[0268] A slot may consist of one or more symbols in the time domain (such as Orthogonal Frequency Division Multiplexing (OFDM) symbols or Single Carrier Frequency Division Multiple Access (SC-FDMA) symbols). Alternatively, a slot may be a time unit based on neurology.

[0269] A slot may include multiple mini-slots. Each mini-slot may consist of one or more symbols in the time domain. Mini-slots may also be called sub-slots. Mini-slots may consist of fewer symbols than a slot. A PDSCH (or PUSCH) transmitted in a time unit larger than a mini-slot may be called a PDSCH (PUSCH) mapping type A. A PDSCH (or PUSCH) transmitted using a mini-slot may be called a PDSCH (PUSCH) mapping type B.

[0270] Wireless frames, subframes, slots, minislots, and symbols all represent units of time when transmitting a signal. Wireless frames, subframes, slots, minislots, and symbols may each be referred to by different names. Furthermore, the units of time such as frames, subframes, slots, minislots, and symbols in this disclosure may be interpreted as interchangeable.

[0271] For example, one subframe may be called TTI, multiple consecutive subframes may be called TTI, or one slot or one mini-slot may be called TTI. In other words, at least one of the subframe and TTI may be a subframe (1ms) in existing LTE, a period shorter than 1ms (e.g., 1-13 symbols), or a period longer than 1ms. Note that the unit representing TTI may be called a slot, mini-slot, etc., instead of a subframe.

[0272] Here, TTI refers to, for example, the smallest unit of time for scheduling in wireless communication. For example, in an LTE system, the base station schedules each user terminal to allocate wireless resources (such as the frequency bandwidth and transmission power available to each user terminal) in TTI units. However, the definition of TTI is not limited to this.

[0273] TTI may be a transmission time unit for channel-encoded data packets (transport blocks), code blocks, code words, etc., or it may be a processing unit for scheduling, link adaptation, etc. Given a TTI, the actual time interval (e.g., number of symbols) to which the transport block, code block, code word, etc. are mapped may be shorter than the given TTI.

[0274] Furthermore, if one slot or one mini-slot is referred to as TTI, then one or more TTIs (i.e., one or more slots or one or more mini-slots) may constitute the minimum time unit of scheduling. In addition, the number of slots (number of mini-slots) that constitute the minimum time unit of scheduling may be controlled.

[0275] A TTI with a time length of 1 ms may also be called a normal TTI (TTI in 3GPP Rel.8-12), a long TTI, a normal subframe, a long subframe, or a slot. A TTI shorter than a normal TTI may also be called a shortened TTI, a short TTI, a partial or fractional TTI, a shortened subframe, a short subframe, a mini slot, a sub slot, or a slot.

[0276] Furthermore, long TTIs (e.g., normal TTIs, subframes, etc.) may be interpreted as TTIs with a time length exceeding 1 ms, and short TTIs (e.g., shortened TTIs, etc.) may be interpreted as TTIs with a TTI length less than that of a long TTI but 1 ms or more.

[0277] A Resource Block (RB) is a resource allocation unit in the time domain and frequency domain, and in the frequency domain, it may contain one or more consecutive subcarriers. The number of subcarriers in an RB may be the same regardless of the neurology, for example, 12. The number of subcarriers in an RB may be determined based on the neurology.

[0278] Furthermore, an RB may contain one or more symbols in the time domain and may have the length of one slot, one minislot, one subframe, or one TTI. Each TTI, subframe, etc., may consist of one or more resource blocks.

[0279] One or more RBs may also be called Physical RBs (PRBs), Sub-Carrier Groups (SCGs), Resource Element Groups (REGs), PRB pairs, RB pairs, etc.

[0280] Furthermore, a resource block may consist of one or more resource elements (REs). For example, one RE may be a radio resource area comprising one subcarrier and one symbol.

[0281] A Bandwidth Part (BWP) (also called a partial bandwidth) may represent a subset of consecutive common resource blocks (RBs) for a given neurology in a given carrier. Here, the common RBs may be identified by an index of the RBs relative to the carrier's common reference point. PRBs may be defined and numbered within a BWP.

[0282] A BWP may include UL BWPs (BWPs for UL) and DL BWPs (BWPs for DL). One or more BWPs may be configured within a single carrier for a UE.

[0283] At least one of the configured BWPs may be active, and the UE does not need to assume that it will send or receive a given signal / channel outside of the active BWP. In this disclosure, terms such as "cell" and "carrier" may be read as "BWP".

[0284] The structures described above, such as wireless frames, subframes, slots, minislots, and symbols, are merely illustrative examples. For instance, the number of subframes included in a wireless frame, the number of slots per subframe or wireless frame, the number of minislots within a slot, the number of symbols and RBs included in a slot or minislot, the number of subcarriers included in an RB, and the number of symbols, symbol length, and cyclic prefix (CP) length within a TTI can be varied in various ways.

[0285] Furthermore, the information, parameters, etc., described in this disclosure may be expressed using absolute values, relative values ​​from a predetermined value, or corresponding other information. For example, wireless resources may be indicated by a predetermined index.

[0286] The names used for parameters and other elements in this disclosure are not restrictive in any way. Furthermore, mathematical formulas and other elements that use these parameters may differ from those expressly disclosed in this disclosure. Various channels (PUCCH, PDCCH, etc.) and information elements can be identified by any suitable name, and therefore, the various names assigned to these various channels and information elements are not restrictive in any way.

[0287] The information, signals, etc. described in this disclosure may be represented using any of the various different techniques. For example, the data, instructions, commands, information, signals, bits, symbols, chips, etc. that may be referred to throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or photons, or any combination thereof.

[0288] Furthermore, information, signals, etc., can be output from upper layers to lower layers and from lower layers to upper layers, or to at least one of the two. Information, signals, etc., may also be input and output via multiple network nodes.

[0289] Input and output information and signals may be stored in a specific location (e.g., memory) or managed using a management table. Input and output information and signals may be overwritten, updated, or appended to. Output information and signals may be deleted. Input information and signals may be transmitted to other devices.

[0290] Information notification is not limited to the embodiments described herein and may be carried out by other means. For example, information notification in this disclosure may be carried out by physical layer signaling (e.g., Downlink Control Information (DCI), Uplink Control Information (UCI)), higher layer signaling (e.g., Radio Resource Control (RRC) signaling, broadcast information (Master Information Block (MIB), System Information Block (SIB)), Medium Access Control (MAC) signaling), other signals, or a combination thereof).

[0291] Physical layer signaling may also be called Layer 1 / Layer 2 (L1 / L2) control information (L1 / L2 control signals), L1 control information (L1 control signals), etc. RRC signaling may also be called RRC messages, for example, RRC Connection Setup messages, RRC Connection Reconfiguration messages, etc. MAC signaling may also be communicated using, for example, MAC Control Element (CE).

[0292] Furthermore, notification of the specified information (for example, notification that "X is the case") is not limited to explicit notification, but may also be made implicitly (for example, by not notifying the specified information or by notifying other information).

[0293] The determination may be made by a value represented by 1 bit (0 or 1), by a boolean value represented as true or false, or by a numerical comparison (for example, a comparison with a predetermined value).

[0294] Software should be broadly interpreted to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, functions, and so on, whether they are called software, firmware, middleware, microcode, hardware description languages, or by any other name.

[0295] Furthermore, software, instructions, information, etc., may be transmitted and received via a transmission medium. For example, if software is transmitted from a website, server, or other remote source using at least one of wired technology (such as coaxial cable, fiber optic cable, twisted pair, or Digital Subscriber Line (DSL)) and wireless technology (such as infrared or microwave), then at least one of these wired and wireless technologies is included in the definition of a transmission medium.

[0296] The terms “system” and “network” as used in this disclosure may be used interchangeably. “Network” may also mean the equipment included in the network (e.g., base stations).

[0297] In this disclosure, terms such as "precoding," "precoder," "weight (precoding weight)," "quasi-co-location (QCL)," "transmission configuration indication state (TCI state)," "spatial relation," "spatial domain filter," "transmit power," "phase rotation," "antenna port," "antenna port group," "layer," "number of layers," "rank," "resource," "resource set," "resource group," "beam," "beam width," "beam angle," "antenna," "antenna element," and "panel" may be used interchangeably.

[0298] In this disclosure, terms such as "Base Station (BS)", "wireless base station", "fixed station", "NodeB", "eNB (eNodeB)", "gNB (gNodeB)", "access point", "Transmission Point (TP)", "Reception Point (RP)", "Transmission / Reception Point (TRP)", "panel", "cell", "sector", "cell group", "carrier", and "component carrier" may be used interchangeably. Base stations may also be referred to by terms such as macrocell, small cell, femtocell, and picocell.

[0299] A base station can house one or more (e.g., three) cells. If a base station houses multiple cells, the entire coverage area of ​​the base station can be divided into several smaller areas, each of which may also be provided with communication services by a base station subsystem (e.g., a small indoor base station (Remote Radio Head (RRH))). The terms “cell” or “sector” refer to part or all of the coverage area of ​​at least one of the base station and / or base station subsystems that provide communication services in that coverage.

[0300] In this disclosure, the transmission of information by a base station to a terminal may be interpreted as the base station instructing the terminal to perform a control / operation based on said information.

[0301] In this disclosure, terms such as "Mobile Station (MS)," "user terminal," "User Equipment (UE)," and "terminal" may be used interchangeably.

[0302] A mobile station may also be called a subscriber station, mobile unit, subscriber unit, wireless unit, remote unit, mobile device, wireless device, wireless communication device, remote device, mobile subscriber station, access terminal, mobile terminal, wireless terminal, remote terminal, handset, user agent, mobile client, client, or some other appropriate term.

[0303] At least one of the base station and the mobile station may be called a transmitting device, a receiving device, a wireless communication device, etc. At least one of the base station and the mobile station may also be a device mounted on a moving object, the moving object itself, etc.

[0304] The term "mobile object" refers to any movable object, regardless of its speed, and naturally includes cases where the mobile object is stationary. Examples of such mobile objects include, but are not limited to, vehicles, transport vehicles, automobiles, motorcycles, bicycles, connected cars, excavators, bulldozers, wheel loaders, dump trucks, forklifts, trains, buses, handcarts, rickshaws, ships and other watercraft, airplanes, rockets, satellites, drones, multicopters, quadcopters, balloons, and items carried on them. Furthermore, such mobile objects may be autonomously driven objects operating based on operational commands.

[0305] The mobile entity may be a vehicle (e.g., a car, an airplane), an unmanned mobile entity (e.g., a drone, an autonomous vehicle), or a robot (manned or unmanned). At least one of the base station and the mobile station may be a device that does not necessarily move during communication operations. For example, at least one of the base station and the mobile station may be an Internet of Things (IoT) device such as a sensor.

[0306] Figure 22 shows an example of a vehicle according to one embodiment. The vehicle 40 includes a drive unit 41, a steering unit 42, an accelerator pedal 43, a brake pedal 44, a shift lever 45, left and right front wheels 46, left and right rear wheels 47, an axle 48, an electronic control unit 49, various sensors (including a current sensor 50, a rotation speed sensor 51, a pneumatic pressure sensor 52, a vehicle speed sensor 53, an acceleration sensor 54, an accelerator pedal sensor 55, a brake pedal sensor 56, a shift lever sensor 57, and an object detection sensor 58), an information service unit 59, and a communication module 60.

[0307] The drive unit 41 consists of, for example, at least one of an engine, a motor, or an engine-motor hybrid. The steering unit 42 includes at least a steering wheel (also called a handle) and is configured to steer at least one of the front wheels 46 and the rear wheels 47 based on the operation of the steering wheel operated by the user.

[0308] The electronic control unit 49 consists of a microprocessor 61, memory (ROM, RAM) 62, and communication ports (e.g., input / output (IO) ports) 63. Signals from various sensors 50-58 installed in the vehicle are input to the electronic control unit 49. The electronic control unit 49 may also be called an Electronic Control Unit (ECU).

[0309] Signals from various sensors 50-58 include current signals from current sensor 50 for sensing motor current, rotational speed signals of front wheels 46 / rear wheels 47 acquired by rotational speed sensor 51, air pressure signals of front wheels 46 / rear wheels 47 acquired by air pressure sensor 52, vehicle speed signals acquired by vehicle speed sensor 53, acceleration signals acquired by acceleration sensor 54, accelerator pedal depression signal of accelerator pedal 43 acquired by accelerator pedal sensor 55, brake pedal depression signal of brake pedal 44 acquired by brake pedal sensor 56, operation signals of shift lever 45 acquired by shift lever sensor 57, and detection signals for detecting obstacles, vehicles, pedestrians, etc., acquired by object detection sensor 58.

[0310] The information service unit 59 consists of various devices for providing (outputting) various types of information such as driving information, traffic information, and entertainment information, including a car navigation system, audio system, speakers, displays, television, and radio, and one or more ECUs that control these devices. The information service unit 59 uses information acquired from external devices via a communication module 60 or the like to provide various types of information / services (e.g., multimedia information / multimedia services) to the occupants of the vehicle 40.

[0311] The information service unit 59 may include input devices that accept input from the outside (e.g., keyboard, mouse, microphone, switch, button, sensor, touch panel, etc.) and output devices that perform output to the outside (e.g., display, speaker, LED lamp, touch panel, etc.).

[0312] The driver assistance system unit 64 consists of various devices that provide functions to prevent accidents or reduce the driver's workload, such as millimeter-wave radar, Light Detection and Ranging (LiDAR), cameras, positioning locators (e.g., Global Navigation Satellite System (GNSS)), map information (e.g., High Definition (HD) maps, Autonomous Vehicle (AV) maps), gyro systems (e.g., Inertial Measurement Unit (IMU), Inertial Navigation System (INS)), artificial intelligence (AI) chips, and AI processors, as well as one or more ECUs that control these devices. The driver assistance system unit 64 also transmits and receives various information via the communication module 60 to realize driver assistance functions or autonomous driving functions.

[0313] The communication module 60 can communicate with the microprocessor 61 and components of the vehicle 40 via the communication port 63. For example, the communication module 60 sends and receives data (information) via the communication port 63 to the drive unit 41, steering unit 42, accelerator pedal 43, brake pedal 44, shift lever 45, left and right front wheels 46, left and right rear wheels 47, axle 48, the microprocessor 61 and memory (ROM, RAM) 62 in the electronic control unit 49, and various sensors 50-58 provided in the vehicle 40.

[0314] The communication module 60 is a communication device that can be controlled by the microprocessor 61 of the electronic control unit 49 and can communicate with external devices. For example, it can send and receive various types of information to and from external devices via wireless communication. The communication module 60 may be located either inside or outside the electronic control unit 49. The external device may be, for example, the base station 10 or the user terminal 20 described above. Alternatively, the communication module 60 may be, for example, at least one of the base station 10 and the user terminal 20 (it may function as at least one of the base station 10 and the user terminal 20).

[0315] The communication module 60 may transmit at least one of the following to an external device via wireless communication: signals from the various sensors 50-58 input to the electronic control unit 49, information obtained based on said signals, and information based on input from an external source (user) obtained via the information service unit 59. The electronic control unit 49, the various sensors 50-58, the information service unit 59, etc., may also be called input units that accept input. For example, the PUSCH transmitted by the communication module 60 may include information based on the above input.

[0316] The communication module 60 receives various information (traffic information, signal information, inter-vehicle information, etc.) transmitted from an external device and displays it on the information service unit 59 installed in the vehicle. The information service unit 59 may also be called an output unit, which outputs information (for example, it outputs information to devices such as displays and speakers based on the PDSCH (or data / information decoded from the PDSCH) received by the communication module 60).

[0317] Furthermore, the communication module 60 stores various information received from external devices in a memory 62 that can be used by the microprocessor 61. Based on the information stored in the memory 62, the microprocessor 61 may control the drive unit 41, steering unit 42, accelerator pedal 43, brake pedal 44, shift lever 45, left and right front wheels 46, left and right rear wheels 47, axle 48, various sensors 50-58, etc., which are provided in the vehicle 40.

[0318] Furthermore, the term "base station" in this disclosure may be interpreted as "user terminal." For example, the various aspects / embodiments of this disclosure may be applied to a configuration in which communication between a base station and a user terminal is replaced with communication between multiple user terminals (which may be called, for example, Device-to-Device (D2D), Vehicle-to-Everything (V2X)). In this case, the user terminal 20 may have the functions that the base station 10 has. Also, terms such as "uplink" and "downlink" may be interpreted as terms corresponding to terminal-to-terminal communication (for example, "sidelink"). For example, uplink channel and downlink channel may be interpreted as sidelink channel.

[0319] Similarly, the term "user terminal" in this disclosure may be replaced with "base station." In this case, the base station 10 may be configured to have the same functions as the user terminal 20 described above.

[0320] In this disclosure, operations performed by a base station may, in some cases, be performed by its upper node. In a network including one or more network nodes with base stations, it is clear that various operations performed for communication with terminals may be performed by the base station, one or more network nodes other than the base station (for example, a Mobility Management Entity (MME), a Serving Gateway (S-GW), etc., but not limited to these), or a combination thereof.

[0321] Each aspect / embodiment described in this disclosure may be used individually, in combination, or switched between during execution. Furthermore, the processing procedures, sequences, flowcharts, etc., of each aspect / embodiment described in this disclosure may be rearranged in order, provided they are consistent. For example, the methods described in this disclosure present various step elements in an exemplary order and are not limited to that specific order.

[0322] Each aspect / embodiment described in this disclosure includes Long Term Evolution (LTE), LTE-Advanced (LTE-A), LTE-Beyond (LTE-B), SUPER 3G, IMT-Advanced, 4th generation mobile communication system (4G), 5th generation mobile communication system (5G), 6th generation mobile communication system (6G), xth generation mobile communication system (xG (where x is, for example, an integer or decimal)), Future Radio Access (FRA), New-Radio Access Technology (RAT), New Radio (NR), New radio access (NX), Future generation radio access (FX), Global System for Mobile communications (GSM®), CDMA2000, Ultra Mobile Broadband (UMB), IEEE 802.11 (Wi-Fi®), IEEE 802.16 (WiMAX®), and IEEE This may apply to systems utilizing 802.20, Ultra-WideBand (UWB), Bluetooth®, or other appropriate wireless communication methods, as well as next-generation systems that are extended, modified, created, or defined based on these. It may also apply to combinations of multiple systems (e.g., a combination of LTE or LTE-A and 5G).

[0323] In this disclosure, the phrase "based on" does not mean "based solely on" unless otherwise specified. In other words, the phrase "based on" means both "based solely on" and "based at least on."

[0324] Any reference to elements using the designations “first,” “second,” etc., as used in this disclosure does not generally limit the quantity or order of those elements. These designations may be used in this disclosure as a convenient way to distinguish between two or more elements. Accordingly, the references to the first and second elements do not imply that only two elements may be employed or that the first element must precede the second element in any way.

[0325] The term “determining” as used in this disclosure may encompass a wide variety of actions. For example, “determining” may be considered to include judging, calculating, computing, processing, deriving, investigating, looking up, searching, inquiry (e.g., searching in tables, databases, or other data structures), ascertaining, etc.

[0326] Furthermore, "judgment (decision)" may be considered as "judging (deciding)" things like receiving (e.g., receiving information), transmitting (e.g., sending information), input, output, accessing (e.g., accessing data in memory).

[0327] Furthermore, "judgment (decision)" can be considered as "judging (deciding)" something like resolving, selecting, choosing, establishing, comparing, etc. In other words, "judgment (decision)" can be considered as "judging (deciding)" something about an action.

[0328] Furthermore, "judgment (decision)" can be replaced with "assuming," "expecting," or "considering."

[0329] As used in this disclosure, the terms “connected,” “coupled,” and any variations thereof mean any direct or indirect connection or coupling between two or more elements, and may include one or more intermediate elements between two elements that are “connected” or “coupled” with each other. The coupling or connection between elements may be physical, logical, or a combination thereof. For example, “connection” may be replaced with “access.”

[0330] In this disclosure, when two elements are connected, they can be considered to be “connected” or “coupled” to each other using one or more wires, cables, printed electrical connections, etc., and, in some non-exclusive and non-exclusive examples, electromagnetic energy having wavelengths in the radio frequency domain, microwave domain, or optical domain (both visible and invisible).

[0331] In this disclosure, the term "A and B are different" may mean "A and B are different from each other." The term may also mean "A and B are each different from C." Terms such as "separate" and "combine" may be interpreted similarly to "different."

[0332] Where the terms “include,” “including,” and variations thereof are used in this disclosure, these terms are intended to be inclusive, as is the term “comprising.” Furthermore, the term “or” as used in this disclosure is not intended to mean exclusive OR.

[0333] In this disclosure, if articles are added by translation, such as a, an, and the in English, this disclosure may include the fact that the noun following these articles is plural.

[0334] In this disclosure, "less than or equal to," "less than," "greater than or equal to," "more than," and "equal to" may be interpreted interchangeably. Also, in this disclosure, words meaning "good," "bad," "big," "small," "high," "low," "early," and "slow" may be interpreted interchangeably (not limited to the positive, comparative, and superlative degrees). Also, in this disclosure, words meaning "good," "bad," "big," "small," "high," "low," "early," and "slow" may be interpreted interchangeably as expressions with "i-th" added to them (not limited to the positive, comparative, and superlative degrees) (for example, "highest" may be interpreted interchangeably as "i-th highest").

[0335] In this disclosure, "of," "for," "regarding," "related to," and "associated with" may be interpreted as being interchangeable.

[0336] Although the invention described herein has been explained in detail above, it will be clear to those skilled in the art that the invention described herein is not limited to the embodiments described herein. The invention described herein can be implemented in modified and altered forms without departing from the spirit and scope of the invention as defined in the claims. Therefore, the descriptions herein are for illustrative purposes only and do not imply any limitation on the invention described herein.

Claims

[Claim 1] A receiving unit that receives student model information relating to the student model in the aforementioned knowledge distillation, The system includes a transmitting unit that performs inference processing using the student model based on the student model information and transmits information regarding the output of the inference processing to a base station via wireless communication. The terminal wherein the student model information includes the index of the student model, and the information relating to the output of the inference process includes the index of the student model on which the student model to be retrained at the base station is based.

Images