Detection of dropped-behind user equipment in federated learning over air interface

Abstract

Detecting a fallen behind user device in federated learning over an air interface is provided. A method for detecting a fallen-behind user device in federated learning may include identifying one or more fallen-behind devices of a plurality of user devices; and suspending transmission of the aggregation model to the one or more dropped-behind devices for local model training. The method may also include resuming transmission of the aggregated model to at least one of the one or more dropped devices for local model training.

Description

Technical Field

[0001] Some example embodiments may generally relate to mobile or wireless telecommunications systems, such as Long Term Evolution (LTE) or Fifth Generation (5G) New Radio (NR) access technologies or access technologies beyond 5G or Sixth Generation (6G) or other communication systems. For example, certain example embodiments may relate to detecting straggler user equipment (UE) in Federated Learning (FL) via an air interface. Background Art

[0002] Examples of mobile or wireless telecommunications systems may include Universal Mobile Telecommunications System (UMTS) Terrestrial Radio Access Network (UTRAN), Long Term Evolution (LTE) Evolved UTRAN (E-UTRAN), Advanced LTE (LTE-A), MulteFire, LTE-A Pro, Fifth Generation (5G) radio access technology or New Radio (NR) access technology and / or Sixth Generation (6G) radio access technology. Fifth Generation (5G) and Sixth Generation (6G) wireless systems refer to the Next Generation (NG) of radio systems and network architectures. Most 5G and 6G network technologies are based on NR technology. However, 5G / 6G (or NG) networks can also be built on E-UTRAN radios. It is estimated that NR can provide a bit rate of about 10 - 20 Gbit / s or higher, and can support at least Enhanced Mobile Broadband (eMBB) and Ultra-Reliable Low-Latency Communication (URLLC) as well as Massive Machine Type Communication (mMTC). It is expected that NR will deliver ultra-wideband and ultra-robust, low-latency connectivity and large-scale networking to support the Internet of Things (IoT). Summary of the Invention

[0003] Various example embodiments may provide an apparatus including at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to at least: identify one or more straggler devices among a plurality of user devices. Also cause the apparatus to pause transmitting an aggregated model for local model training to one or more straggler devices, and resume transmitting the aggregated model for local model training to at least one of the one or more straggler devices.

[0004] Certain example embodiments may provide an apparatus including at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to at least: transmit a locally trained model generated by local model training to a network entity. Also cause the apparatus to receive an indication from the network entity that the apparatus is identified as a straggler device and the transmission of the aggregated model is paused, and resume receiving the aggregated model from the network entity based on determining that the latency of transmitting the locally trained model is reduced.

[0005] Some example embodiments may provide a method, including: identifying, by a device, one or more straggler devices among a plurality of user devices; suspending, by the device, transmitting an aggregated model to the one or more straggler devices for local model training; and resuming, by the device, transmitting the aggregated model to at least one of the one or more straggler devices for local model training.

[0006] Certain example embodiments may provide a method, including: transmitting, by a device, a locally trained model generated through local model training to a network entity; receiving, by the device, an indication that the device is identified as a straggler device and the transmission of the aggregated model is suspended from the network entity; and resuming, by the device, receiving the aggregated model from the network entity based on determining that the latency of transmitting the locally trained model is reduced.

[0007] Various example embodiments may provide a device, including: components for identifying one or more straggler devices among a plurality of user devices; components for suspending transmitting an aggregated model to the one or more straggler devices for local model training; and components for resuming transmitting the aggregated model to at least one of the one or more straggler devices for local model training.

[0008] Some example embodiments may provide a device, including: components for transmitting a locally trained model generated through local model training to a network entity; components for receiving, from the network entity, an indication that the device is identified as a straggler device and the transmission of the aggregated model is suspended; and components for resuming receiving the aggregated model from the network entity based on determining that the latency of transmitting the locally trained model is reduced.

[0009] Various example embodiments may provide a non-transitory computer-readable medium including program instructions that, when executed by a device, cause the device to at least identify one or more straggler devices among a plurality of user devices. The program instructions may also cause the device to suspend transmitting an aggregated model to the one or more straggler devices for local model training and resume transmitting the aggregated model to at least one of the one or more straggler devices for local model training.

[0010] Certain example embodiments may provide a non-transitory computer-readable medium including program instructions that, when executed by a device, cause the device to at least: transmit a locally trained model generated through local model training to a network entity. The program instructions may also cause the device to receive an indication that the device is identified as a straggler device and the transmission of the aggregated model is suspended from the network entity, and resume receiving the aggregated model from the network entity based on determining that the latency of transmitting the locally trained model is reduced.

[0011] Some example embodiments may provide a non - transitory computer - readable medium including program instructions that, when executed, cause a device to perform one of the methods described herein. Various example embodiments may provide a computer program that, when executed, causes a device to perform one of the methods described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To properly understand the example embodiments, reference should be made to the drawings as follows:

[0013] Figure 1 An example configuration of a federated learning (FL) model in a wireless network is shown;

[0014] Figure 2 An example graph showing the times at which different UEs provide model data to the network is shown;

[0015] Figure 3 An example of a signal diagram according to various example embodiments is shown;

[0016] Figure 4 is an example of a flowchart of a method according to some example embodiments;

[0017] Figure 5 is an example of a flowchart of another method according to certain example embodiments; and

[0018] Figure 6 A set of devices according to various example embodiments is shown. DETAILED DESCRIPTION

[0019] It will be readily understood that the components of certain example embodiments, as generally described and illustrated in the figures herein, can be arranged and designed in a wide variety of different configurations. The following is a detailed description of some example embodiments of a system, method, apparatus, and non - transitory computer program product for detecting straggler user equipment (UE) in federated learning (FL) over the air interface. Although the devices discussed below and shown in the figures refer to 5G, 6G, or next - generation node B (gNB) network entities and / or devices and user equipment (UE) devices, the present disclosure is not limited to only network entities / devices and UEs.

[0020] It can be readily understood that the components of certain example embodiments, as generally described and illustrated in the figures herein, may be arranged and designed in a wide variety of different configurations. Different reference numerals from multiple figures may be used outside of the sequence in the specification to refer to the same elements to show their features or functions. If desired, the different functions or processes discussed herein may be performed in a different order and / or simultaneously with each other. Additionally, if desired, one or more of the described functions or processes may be optional or may be combined. Accordingly, the following description should be regarded as illustrative of the principles and teachings of certain example embodiments and not as limiting.

[0021] Modern mobile networks may require large amounts of data from multiple distributed data sources (e.g., nodes, UEs, etc.). Mobile networks employing one or more artificial intelligence / machine learning (ML) models may require even more data to train the models, such as a single general model. For example, the data may be generated by one or more distributed units on the mobile network, which may provide the generated data to one or more centralized units of the mobile network, and a common ML model may be created and trained. To minimize the data exchange between the distributed units and the centralized units, the concept of federated learning (FL) may be applied.

[0022] FL is a form of machine learning in which different versions of a model are trained at different distributed hosts rather than training the model at a single node. FL can be distinguished from distributed ML, in which a single ML model may be trained at distributed nodes to utilize the computing power of different nodes. For example, FL may be different from distributed learning because each distributed node in an FL scenario has its own local training data, which may not be from the same distribution as the data at other nodes. FL may also be different because each node computes the parameters for its local ML model, and the central host does not compute the version or parts of the model. Instead, the central host may combine the parameters of all the distributed models to generate a centralized or general model. This FL process may allow the training datasets to remain at the location where they are generated, and model training may be performed locally at each individual learner in the federation (e.g., a collection of networks or network elements).

[0023] Figure 1 An example configuration of a federated learning (FL) model in a wireless network is shown. Figure 1The wireless network can include multiple UEs, such as UE 101, UE 102, and UE 103, each having local storage (e.g., 104, 105, 106) to store the generated data. Each UE (e.g., 101, 102, 103) can transmit a local model to a network entity 107 (e.g., gNB), and the network entity 107 aggregates a single general model and feeds back the aggregated model to the UEs (e.g., 101, 102, 103).

[0024] After training the local models, each individual learner (e.g., UE 101, UE 102, and / or UE 103) transmits its local model data / parameters (instead of the original training dataset) to an aggregation unit, which can be located at the network entity 107. The aggregation unit of the network entity 107 can use the received local model parameters to update the global / general model, and the global / general model can be fed back to the UEs (e.g., 101, 102, 103) for further iterations until the global model converges. In this way, each UE (e.g., 101, 102, 103) can benefit from the datasets of other UEs (e.g., 101, 102, 103) through the global / general model shared by the aggregation unit, and there is no need to explicitly access the large amount of privacy-sensitive data available at each other UE. In each round of model collection from the UEs (e.g., 101, 102, 103), the amount of data transmitted from each UE (e.g., 101, 102, 103) to the network entity 107 consumes a large and heavy amount of network resources. Various example embodiments herein can provide solutions to the need for designing a communication-efficient model collection scheme that can reduce the load on the radio link.

[0025] The 3rd Generation Partnership Project (3GPP) specifications can enable various levels of cooperation between UEs and the network for performing machine learning. For example, 3GPP can provide level X without cooperation, level Y with signaling cooperation but without transmitting one or more models, and / or level Z with signaling cooperation and transmitting one or more models. Cooperation can also be applied to the cases of gNB-gNB, inside gNB, and UE-UE. In these examples, level X may not require cooperation between the UE and the network, while levels Y and Z may require coordination for cooperation between the UE and the network entity. Federated learning (FL) can be associated with level Z cooperation, which can provide signaling between the UE and the network and can allow the transmission of UE models between the UE and the network entity. When there are a large number of UEs (e.g., hundreds or thousands) and a large number of iterations (e.g., thousands), high communication and interaction between the UE and the network may occur. In this case, reducing communication over the air interface may be beneficial.

[0026] In synchronous mode FL model collection, the local training models from UEs may not be aggregated until all models are available at the network (e.g., gNB). A parameter aggregation algorithm (e.g., FedAvg) can be used to calculate the average of the model parameters when aggregating the parameters. It may be desirable for the local models from UEs to be available within a time limit or time period D max to avoid delays in unnecessary parameter aggregation. There may be situations where timely parameter aggregation cannot be performed within the time limit or time period D max because, for example, a UE may not have the computing power required to process the training in a timely manner, or the UE may have a larger amount of data compared to other UEs and may require additional time for processing and training. As another example, the link quality between the UE and the network may be relatively poor or blocked, preventing timely communication between the UE and the network.

[0027] Figure 2 shows an example graph depicting the times at which different UEs provide model data to the network. As Figure 2 shown, even when data from multiple UEs is available, one or more UEs may delay the model aggregation at the network. For example, the local training model from UE 6 never reaches the network, which may be due to signal blockage or limited coverage. The local model from UE 4 may be delayed by more than D max and may delay the model aggregation at the network. In this exemplary case, UE 6 and UE 4 can be identified as stragglers in FL.

[0028] Various example embodiments can provide technical advantages to support one or more processes for identifying persistent straggler UEs (which delay model aggregation) and temporarily removing these straggler UEs from the network configuration. By temporarily removing the communication between the straggler UEs and the network, the delay budget for model aggregation in a specific iteration can be achieved without significantly sacrificing performance in terms of FL model convergence. Some example embodiments can provide that the network (e.g., gNB) may not transmit the aggregated model to the straggler UEs and reduce the load on the radio link. Some example embodiments can provide the advantage that the aggregation unit of the network may not wait unnecessarily (e.g., be delayed) for the straggler UEs, which can enable faster iteration of the aggregated model.

[0029] Certain example embodiments can determine that a UE is a persistent straggler when it is a straggler in more than M consecutive iterations of FL model collection. When the delay D when the local model of the UE reaches the aggregation unit of the network entity is D > D maxWhen this occurs for at least M consecutive iterations, the UE can be declared as a persistent straggler, which can include local models, such as cases where the UE never reaches the network due to a complete lack of coverage. Additionally or alternatively, if the latency of the locally trained machine learning model of the UE exceeds a threshold time and persists for more than a threshold number of times at the network within a timer window of a specified number of iterations, the UE can be declared as a persistent straggler.

[0030] Various example embodiments can provide one or more processes for identifying persistent straggler UEs among multiple UEs, and pausing or stopping the transmission of the aggregated model for local model training to the identified persistent straggler UEs. The terms "straggler" and "persistent straggler" can be used interchangeably herein and both refer to persistent stragglers. Certain example embodiments can improve the link efficiency of the downlink transmission from the network entity to the UE, as transmissions with the aggregated model can be sent to those UEs that are not persistent stragglers, which reduces the waste of network resources and the latency of model aggregation. Persistent straggler UEs may not be permanently prohibited from participating in FL as they may have unique training data to contribute to FL. Some example embodiments can provide one or more processes for evaluating when the network condition and / or computing power of a straggler UE becomes good / high enough to resume its participation in the FL process.

[0031] Figure 3 An example of a signal diagram according to various example embodiments is shown. The signal diagram can represent one or more processes between an aggregator network entity 301 (e.g., gNB) and a UE 302 via, for example, an air interface. The training host can be located in the UE 302, and the aggregator can be located at the network entity 301. However, the embodiments are not limited to this scenario, and the training host and the aggregator can be located elsewhere on the network. For example, the training host can be located in the core network.

[0032] Various example embodiments can provide Figure 3 One or more of the processes can include: at 310, before starting FL, the straggler counter C can be reset at the FL management entity in the network entity 301 for all participating UEs (e.g., UE 302). i The straggler counter C i can be used to track straggler UEs in each iteration of FL. At 311, the network entity 301 can provide the aggregated FL model in the i-th iteration to all UEs (e.g., UE 302) for local model training using the FL model in the (i + 1)-th iteration. At 312, the UE 302 can generate (e.g., train) a locally trained model and provide the locally trained model to the aggregator of the network entity 301 for the (i + 1)-th iteration. At Figure 3In this example, it can be assumed that the locally trained model may be delayed by more than a threshold D set for receiving all locally trained models from the UE in each iteration max .

[0033] Some example embodiments may provide that, at 313, the network entity 301 may incrementally increase the straggler counter C for the delayed UE 302 i . The straggler counter C i can be used to track the model delay for each iteration. At 314, the network entity 301 may provide the aggregated FL model in the (i + 1)-th iteration to, for example, the UE 302 for local model training using the FL model in the (i + 2)-th iteration. At 315, the UE 302 may generate and train the locally trained model and provide the locally trained model to the network entity 301 for the (i + 2)-th iteration. At 316, when the locally trained model is delayed in the (i + 2)-th iteration, the network entity 301 may further incrementally increase the straggler counter C for the UE 302 i . For all UEs whose models all arrive at the network within the delay threshold D maxand , the straggler counter C is reset after each iteration i . At 317, when the straggler counter C i reaches the allowed maximum delay threshold M, the UE 302 may be declared as a persistent straggler UE. The threshold M may be reached after multiple consecutive or successive delays in the iteration

[0034] Some example embodiments may provide that, at 318, the network entity 301 may notify the UE 302 that the UE 302 has been declared as a persistent straggler UE. At 319, the UE 302 may use the information declaring the UE 302 as a persistent straggler UE to pause the training of the locally trained model. During the pause period, the UE 302 may not provide the locally trained model to the network entity 301 for a random or predetermined duration. The UE 302 may continue to collect data for the training dataset and wait for the training to be resumed or re-included. Some example embodiments may provide the process 310 - 319 to allow persistent straggler detection / determination

[0035] Various example embodiments may provide, at 320, after a random time within a predetermined time period, or after the predetermined time period, one or more processes for re - incorporating or resuming the UE 302's participation in the FL process begin. The network entity 301 may generate a query message to query the UE 302 to evaluate the network condition (e.g., link quality) and / or the computing ability of the UE 302. The query from the network entity 301 can be used to determine whether the conditions that caused the delay (previously causing the UE 302 to be declared as a persistent straggler UE) have improved sufficiently to resume the UE 302's participation in FL. At 321, the query message may be provided to the UE 302 to request context information from the UE 302, such as, for example, the availability of new data compared to the data used for the locally trained local model last time, the computing ability of the UE 302, the quality - of - service (QoS) metric information of the UE 302, and / or other information that can indicate whether the UE 302 can be re - incorporated into the FL process. At 322, the UE 302 may provide a response message to the network entity 301, which includes, for example, an acknowledgement (ACK) or non - acknowledgement (NACK) indication. The response message may also include context information. When the response message includes a NACK indication, it may also include a timer from the UE 302 if the UE 302 is not ready yet, which notifies the network entity 301 when to contact the UE 302.

[0036] At 323, when the network entity 301 receives a response message including an ACK indication, the network entity 301 may decide to re - incorporate the UE 302 into the FL process, and the straggler counter may be reset. At 324, if the response message received by the network entity 301 includes a NACK indication or no response message is received, the UE 302 may still be considered a persistent straggler. Since the UE 302 has been paused from the FL process, the straggler counter C i may continue to be ignored and will not be incremented anymore.

[0037] At 325, when the network entity 301 receives a response message with an ACK indication at process 323 to re - incorporate the UE 302 into the FL process, since the straggler counter C i has been reset, the network entity 301 may provide the FL model aggregated in the i - th iteration to the UE 302 for local model training in the (i + 1)-th iteration. At 326, the UE 302 may generate (e.g., train) a locally trained model and provide the locally trained model to the network entity 301 for the (i + 1)-th iteration. For example, the locally trained model may be provided to the network entity 301 within a time threshold D max .

[0038] Various example embodiments can advantageously provide an improvement in the selection of UEs participating in FL without, for example, network assistance. In the absence of network assistance for FL member selection in each iteration, since the network entity may not be aware of the radio conditions, coverage, etc. of the UEs, the latency differences during the collection of the trained model in each iteration may be reduced, and it may not be efficient to use the same process for handling persistently straggling UEs in each iteration. When persistently straggling UEs have been identified, some example embodiments allow for an improvement in the selection of such UEs and can avoid unnecessarily using network resources to collect the trained model from UEs with deep signal fading (e.g., minimum coverage).

[0039] Some example embodiments can provide an asynchronous process in which the network / network entity can perform the selection of UEs in the FL process. It can be assumed that the network can provide information regarding the radio conditions (e.g., uplink signal-to-interference and noise ratio (SINR) for each beam, which the UE can use to adjust its uplink transmission of its locally trained model) of the FL members (e.g., UEs) before applying the selection in each iteration. In this case, the network needs to provide information regarding the radio conditions of all FL members (e.g., UEs) in each iteration before the node selection is provided. The network is capable of detecting persistently straggling UEs to reduce unnecessary information flow and stop collecting information from the corresponding UEs.

[0040] Figure 4 An example flowchart of a method according to some example embodiments is shown. In an example embodiment, Figure 4 the method can be performed by a network element / entity or a group of multiple network entities in a 3GPP system, such as LTE or 5G-NR. For example, in an example embodiment, Figure 4 the method can be performed by a network node or network entity (e.g., gNB) similar to Figure 6 the apparatus 610 shown.

[0041] According to various example embodiments, Figure 4 the method can include: at 410, identifying one or more straggling devices among a plurality of user equipments, and at 420, pausing the transmission of the aggregated model for local model training to the one or more straggling devices. The method can further include: at 430, resuming the transmission of the aggregated model for local model training to at least one of the one or more straggling devices.

[0042] Some example embodiments may provide that, based on the delay of the device receiving the locally trained machine learning model exceeding a first threshold time period and lasting for more than a second threshold number of consecutive iterations, at least one user device among the plurality of user devices is identified as one or more straggler devices. The resumption of transmission may include: transmitting the aggregated model to at least one of the one or more straggler devices, which may result in resetting the straggler counter for the at least one straggler device. Some example embodiments may provide that the method further includes receiving the locally trained machine learning model from at least one user device. The delay in receiving the locally trained machine learning model exceeds the first threshold time period.

[0043] Various example embodiments may provide that the method further includes: sending an indication to the identified one or more straggler devices notifying the one or more straggler devices that they have been identified as stragglers. The method may further include: after pausing the transmission of the aggregated model to the one or more straggler devices, evaluating at least one of the following: the network condition under which the device and the plurality of user devices operate, or the computing power for the one or more straggler devices to support the local model training of the one or more straggler devices. Resuming the transmission of the aggregated model to at least one of the one or more straggler devices for local model training is based on an evaluation indicating that at least one of the network condition or the computing power to support local model training is higher than a third threshold. The evaluation indicating that at least one of the network condition or the computing power is higher than the third threshold is performed when the federated learning process is not completed at the time of pausing the transmission.

[0044] Some example embodiments may provide that the method further includes: sending a context query message to the one or more straggler devices, the context query message requesting information about at least one of the network condition or the computing power to support local model training; and receiving a response message from the one or more straggler devices. The response message includes an acknowledgement message or a non-acknowledgement message, and the acknowledgement message or the non-acknowledgement message indicates information for determining whether to resume the transmission of the aggregated model to at least one of the one or more straggler devices. When the response message includes a non-acknowledgement message, a timer is provided to the device from the one or more straggler devices or another network entity, and the timer enables the device to transmit the context query message again when the timer expires.

[0045] Certain example embodiments may provide that the method further includes: when receiving a response message including an acknowledgement message, resuming the transmission of the aggregated model to at least one of the one or more straggler devices for local model training. The one or more straggler devices are user devices, and the user devices cause the delay in the device receiving the locally trained machine learning model to exceed a fourth threshold number of times within a timer window of a specified number of iterations.

[0046] Figure 5Shows an example flowchart of a method according to some example embodiments. In an example embodiment, Figure 5 the method may be performed by a device or a user equipment within a network in a 3GPP system, such as LTE or 5G-NR. For example, in an example embodiment, Figure 5 the method may be performed by a UE, similar to Figure 6 the apparatus 620 shown in

[0047] According to various example embodiments, Figure 5 the method may include: at 510, transmitting a locally trained model generated through local model training to a network entity, and at 520, receiving an indication from the network entity that the device is identified as a straggler device and the transmission of the aggregated model is suspended. At 530, the method may further include: based on determining that the delay in transmitting the locally trained model is reduced, resuming receiving the aggregated model from the network entity.

[0048] Some example embodiments may provide that the indication that the device is identified as a straggler device is based on the delay in the network entity receiving the locally trained model exceeding a first threshold time period and continuing for more than a second threshold number of consecutive iterations. Certain example embodiments may provide that the method further includes: during the time period between receiving the indication that the transmission of the aggregated model is suspended and resuming receiving the aggregated model, collecting training data for the locally trained model. The method may further include: receiving a context query message from the network entity, the context query message requesting information about at least one of the following: network condition or the computing ability of the device to support local model training.

[0049] Various example embodiments may provide that the method further includes: transmitting a response message to the network entity. The response message includes information about at least one of the network condition or the computing ability of the device, where the information includes at least one of the following: a timer or a network condition indicator. The response message includes an acknowledgment message or a non-acknowledgment message, and the acknowledgment message or the non-acknowledgment message indicates the information for the network entity to determine whether to resume the transmission of the aggregated model. Certain example embodiments may provide that the method includes: when the response message includes a non-acknowledgment message, transmitting a timer together with the response message to the network entity; and monitoring when the timer expires to receive the context query message from the network entity again.

[0050] Some example embodiments may provide that the method further includes: after transmitting a response message including an acknowledgment message, resuming receiving the aggregated model from the network entity for local model training. When the delay in transmitting the locally trained model exceeds a fourth threshold number of times within a timer window of a specified number of iterations, the device is identified as a straggler device.

[0051] The following embodiments are provided and described herein.

[0052] Example 1. A device for wireless communication, comprising:

[0053] At least one processor; and

[0054] At least one memory storing instructions that, when executed by the at least one processor, cause the device to at least:

[0055] Identify one or more straggler devices among a plurality of user devices;

[0056] Pause transmitting an aggregated model for local model training to one or more straggler devices; and

[0057] Resume transmitting the aggregated model for local model training to at least one of the one or more straggler devices.

[0058] Example 2. The device according to Example 1, wherein at least one user device among the plurality of user devices is identified as one or more straggler devices based on that the delay of the device receiving a locally trained machine learning model exceeds a first threshold time period and lasts for more than a second threshold number of consecutive iterations.

[0059] Example 3. The device according to Example 2, wherein:

[0060] Resuming transmission includes: transmitting the aggregated model to at least one of the one or more straggler devices; and

[0061] The at least one memory stores instructions that, when executed by the at least one processor, cause the device to:

[0062] Receive a locally trained machine learning model from at least one user device, wherein the delay of receiving the locally trained machine learning model exceeds the first threshold time period.

[0063] Example 4. The device according to Example 1, wherein the at least one memory stores instructions that, when executed by the at least one processor, cause the device to:

[0064] Send an indication to the identified one or more straggler devices to notify the one or more straggler devices that they have been identified as stragglers.

[0065] Example 5. The device according to Example 1, wherein the at least one memory stores instructions that, when executed by the at least one processor, cause the device to:

[0066] After pausing transmitting the aggregated model to one or more straggler devices, evaluate at least one of the following: the network condition of the device and the plurality of user devices, or the computing power for one or more straggler devices to support local model training of the one or more straggler devices.

[0067] Example 6. The apparatus according to Example 5, wherein resuming transmission of the aggregated model to at least one straggler device among one or more straggler devices for local model training is based on an evaluation indicating that at least one of the network condition or the computing power to support local model training is higher than a third threshold.

[0068] Example 7. The apparatus according to Example 6, wherein the evaluation indicating that at least one of the network condition or the computing power is higher than a third threshold is performed when the federated learning process is not completed at the time of pausing transmission.

[0069] Example 8. The apparatus according to Example 5, wherein at least one memory stores instructions that, when executed by at least one processor, cause the apparatus to:

[0070] send a context query message to one or more straggler devices, the context query message requesting information about at least one of the network condition or the computing power to support local model training; and

[0071] receive a response message from one or more straggler devices.

[0072] Example 9. The apparatus according to Example 8, wherein the response message includes an acknowledgement message or a non-acknowledgement message, and the acknowledgement message or the non-acknowledgement message indicates information for determining whether to resume transmission of the aggregated model to at least one straggler device among one or more straggler devices.

[0073] Example 10. The apparatus according to Example 9, wherein when the response message includes a non-acknowledgement message, a timer is provided to the apparatus from one or more straggler devices or another network entity, and the timer enables the apparatus to transmit the context query message again when the timer expires.

[0074] Example 11. The apparatus according to Example 9, wherein at least one memory stores instructions that, when executed by at least one processor, cause the apparatus to:

[0075] when receiving a response message including an acknowledgement message, resume transmission of the aggregated model to at least one straggler device among one or more straggler devices for local model training.

[0076] Example 12. The apparatus according to any one of Examples 1 to 11, wherein one or more straggler devices are user devices, and the user devices cause the delay of the apparatus receiving the machine learning model trained locally within a timer window of a specified number of iterations to exceed a fourth threshold number of times.

[0077] Example 13. An apparatus for wireless communication, comprising:

[0078] at least one processor; and

[0079] At least one memory storing instructions which, when executed by at least one processor, cause the device to at least:

[0080] Transmit a locally trained model generated by local model training to a network entity;

[0081] Receive an indication from the network entity that the device is identified as a straggler and the transmission of the aggregated model is paused; and

[0082] Resume receiving the aggregated model from the network entity based on determining a reduction in the latency of transmitting the locally trained model.

[0083] Example 14. The device according to Example 13, wherein the indication that the device is identified as a straggler is based on the latency of the network entity receiving the locally trained model exceeding a first threshold time period and continuing for more than a second threshold number of consecutive iterations.

[0084] Example 15. The device according to Example 13, wherein the at least one memory stores instructions which, when executed by at least one processor, cause the device to:

[0085] Collect training data for the locally trained model during the time period between receiving the indication that the transmission of the aggregated model is paused and resuming receiving the aggregated model.

[0086] Example 16. The device according to Example 13, wherein the at least one memory stores instructions which, when executed by at least one processor, cause the device to:

[0087] Receive a context query message from the network entity, the context query message requesting information about at least one of the following: network conditions or the computing capabilities of the device to support local model training.

[0088] Example 17. The device according to Example 16, wherein the at least one memory stores instructions which, when executed by at least one processor, cause the device to:

[0089] Transmit a response message to the network entity, the response message including information about at least one of the network conditions or the computing capabilities of the device, wherein the information includes at least one of the following: a timer or a network condition indicator.

[0090] Example 18. The device according to Example 16, wherein the response message includes an acknowledgement message or a non-acknowledgement message, the acknowledgement message or the non-acknowledgement message indicating information for the network entity to determine whether to resume the transmission of the aggregated model.

[0091] Example 19. The device according to Example 18, wherein the at least one memory stores instructions which, when executed by at least one processor, cause the device to:

[0092] When the response message includes a non-acknowledgment message, transmit a timer together with the response message to a network entity; and

[0093] Monitor when the timer expires to receive a context query message again from the network entity.

[0094] Example 20. The apparatus according to Example 18, wherein at least one memory stores instructions that, when executed by at least one processor, cause the apparatus to:

[0095] After transmitting a response message including an acknowledgment message, resume receiving an aggregated model from the network entity for local model training.

[0096] Example 21. The apparatus according to any one of Examples 13 to 20, wherein when the delay in transmitting the locally trained model exceeds a fourth threshold number of times within a timer window of a specified number of iterations, the apparatus is identified as a straggler device.

[0097] Example 22. A method for wireless communication, comprising:

[0098] Identifying, by a device, one or more straggler devices among a plurality of user devices;

[0099] Pausing, by the device, transmitting an aggregated model to one or more straggler devices for local model training; and

[0100] Resuming, by the device, transmitting an aggregated model to at least one of the one or more straggler devices for local model training.

[0101] Example 23. A method for wireless communication, comprising:

[0102] Transmitting, by the device, a locally trained model generated through local model training to a network entity;

[0103] Receiving, by the device, an indication from the network entity that the device is identified as a straggler device and the transmission of the aggregated model is suspended; and

[0104] Resuming, by the device, receiving an aggregated model from the network entity based on determining that the delay in transmitting the locally trained model has decreased.

[0105] Figure 6 Apparatuses 610 and 620 are shown in accordance with various example embodiments. In various example embodiments, apparatus 610 may be an element in or associated with a network, or a network entity, such as a gNB. gNB 301 may be an example of apparatus 610 according to various example embodiments described above. It should be noted that those of ordinary skill in the art will understand that apparatus 610 may include Figure 6Components or features not shown in [the figure]. Additionally, device 620 can be an element in a communication network or network entity, such as a UE, RedCap UE, SL UE, mobile device (ME), mobile station, mobile device, fixed device, IoT device, or another device. For example, UE 302 according to the various example embodiments described above can be an example of device 620. It should be noted that those of ordinary skill in the art will understand that device 620 can include Figure 6 components or features not shown in [the figure].

[0106] According to various example embodiments, device 610 and / or 620 can include one or more processors, one or more computer-readable storage media (e.g., memory, storage, etc.), one or more radio access components (e.g., modem, transceiver, etc.), and / or a user interface. In some example embodiments, device 610 and / or 620 can be configured to operate using one or more radio access technologies, such as GSM, LTE, LTE-A, NR, 5G, WLAN, WiFi, NB-IoT, Bluetooth, NFC, MulteFire, and / or any other radio access technology.

[0107] As Figure 6 shown in the example of [the figure], device 610 and / or 620 can respectively include or be coupled to processors 612 and 622 for processing information and executing instructions or operations. Processors 612 and 622 can be any type of general-purpose or special-purpose processor. In fact, by way of example, processors 612 and 622 can include one or more of a general computer, a special computer, a microprocessor, a digital signal processor (DSP), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), and a processor based on a multi-core processor architecture. Although Figure 6 a single processor 612 (and 622) is shown for each of devices 610 and / or 620 in [the figure], multiple processors can be utilized according to other example embodiments. For example, it should be understood that in certain example embodiments, device 610 and / or 620 can include two or more processors that can support multiprocessing and that can form a multiprocessor system (e.g., in this case, processors 612 and 622 can represent a multiprocessor). According to certain example embodiments, the multiprocessor system can be tightly coupled or loosely coupled to form, for example, a computer cluster.

[0108] Processors 612 and 622 can respectively perform functions associated with the operations of device 610 and / or 620, including, as some examples, precoding of antenna gain / phase parameters, encoding and decoding of the individual bits that form communication messages, formatting of information, and overall control of device 610 and / or 620, including Figures 3 to 5 the processes shown in [the figure].

[0109] The apparatus 610 and / or 620 may also respectively include or be coupled to memories 614 and / or 624 (internal or external), which may be respectively coupled to processors 612 and / or 622 for storing information and instructions that may be executed by processors 612 and 622. The memory 614 (and memory 624) may be one or more memories and may be of any type suitable for a local application environment and may be implemented using any suitable volatile or non-volatile data storage technology, such as semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory, and / or removable memory. For example, the memory 614 (and memory 624) may include any combination of random access memory (RAM), read-only memory (ROM), static storage such as a magnetic disk or optical disk, a hard disk drive (HDD), or any other type of non-transitory machine or computer-readable medium. The instructions stored in memories 614 and 624 may include program instructions or computer program code that, when executed by processors 612 and 622, cause the apparatus 610 and / or 620 to perform the tasks as described herein.

[0110] In certain example embodiments, the apparatus 610 and / or 620 may also include a drive or port or be coupled to a drive or port (internal or external) configured to receive and read an external computer-readable storage medium, such as an optical disc, a USB drive, a flash drive, or any other storage medium. For example, the external computer-readable storage medium may store a computer program or software for execution by processors 612 and 622 and / or the apparatus 610 and / or 620 to perform Figures 3 to 5 any of the methods shown herein.

[0111] In some example embodiments, apparatus 610 and / or 620 may further include or be coupled to one or more antennas 615 and 625 respectively for receiving downlink signals and for transmitting via an uplink from apparatus 610 and / or 620. Apparatus 610 and / or 620 may also include transceivers 616 and 626 configured to transmit and receive information. Transceivers 616 and 626 may also include radio interfaces (e.g., modems) coupled to antennas 615 and 625 respectively. The radio interfaces may correspond to a variety of radio access technologies, including one or more of GSM, LTE, LTE-A, 5G, NR, WLAN, NB-IoT, Bluetooth, BT-LE, NFC, RFID, UWB, etc. The radio interfaces may include other components, such as filters, converters (e.g., digital-to-analog converters, etc.), symbol demappers, signal shaping components, inverse fast Fourier transform (IFFT) modules, etc., to process symbols carried by the downlink or uplink, such as OFDMA symbols.

[0112] For example, transceivers 616 and 626 may be configured to modulate information about a carrier waveform for transmission by antennas 615 and 625 respectively, and to demodulate information received via antennas 615 and 625 for further processing by other elements of apparatus 610 and / or 620. In other example embodiments, transceivers 616 and 626 may be capable of directly transmitting and receiving signals or data. Additionally or alternatively, in some example embodiments, apparatus 610 and / or 620 may include input and / or output devices (I / O devices). In certain example embodiments, apparatus 610 and / or 620 may also include a user interface, such as a graphical user interface or a touch screen.

[0113] In certain example embodiments, memories 614 and 624 may store software modules that provide functionality when executed by processors 612 and 622 respectively. The modules may include, for example, an operating system that provides operating system functionality for apparatus 610 and / or 620. The memories may also store one or more functional modules, such as applications or programs, to provide additional functionality for apparatus 610 and / or 620. The components of apparatus 610 and / or 620 may be implemented in hardware or as any suitable combination of hardware and software. According to certain example embodiments, apparatus 610 may optionally be configured to communicate with apparatus 620 via a wireless or wired communication link 630 according to any radio access technology, such as NR.

[0114] According to certain example embodiments, processors 612 and 622 and memories 614 and 624 may be included in or may form part of a processing circuit or a control circuit. Additionally, in some example embodiments, transceivers 616 and 626 may be included in or may form part of a transceiver circuit.

[0115] For example, in some example embodiments, the apparatus 610 may be controlled by the memory 614 and the processor 612 to identify one or more straggler devices among a plurality of user devices, and to suspend the transmission of the aggregated model to the one or more straggler devices for local model training. The apparatus 610 may also be controlled to resume the transmission of the aggregated model to at least one of the one or more straggler devices for local model training.

[0116] Some example embodiments may provide that the apparatus 620 may be controlled by the memory 624 and the processor 622 to send a locally trained model generated from local model training to a network entity, and to receive an indication that the apparatus 620 is identified as a straggler device and the transmission of the aggregated model is suspended from the network entity. The apparatus 620 may also be controlled to resume receiving the aggregated model from the network entity based on a determination of a reduction in the latency of sending the locally trained model.

[0117] In some example embodiments, an apparatus (e.g., apparatus 610 and / or apparatus 620) may include units for performing the methods, processes, or any variations discussed herein. Examples of the apparatus may include one or more processors, memories, controllers, transmitters, receivers, and / or computer program code for causing the execution of operations.

[0118] Various example embodiments may be directed to an apparatus, such as apparatus 610, that includes: a module for identifying one or more straggler devices among a plurality of user devices; and a module for suspending the transmission of the aggregated model to the one or more straggler devices for local model training. The apparatus 610 may also include a unit for resuming the transmission of the aggregated model to at least one of the one or more straggler devices for local model training.

[0119] Certain example embodiments may be directed to an apparatus, such as apparatus 620, that includes: a unit for sending a locally trained model generated from local model training to a network entity; and a unit for receiving an indication that the apparatus 620 is identified as a straggler device and the transmission of the aggregated model is suspended from the network entity. The apparatus 620 may also include: a unit for resuming the aggregated model from the network entity based on a determination of a reduction in the latency of sending the locally trained model.

[0120] As used herein, the term "circuitry" can refer to a hardware circuit implementation (e.g., analog and / or digital circuitry), a combination of hardware circuitry and software, a combination of analog and / or digital hardware circuitry and software / firmware, any portion of a (one or more) hardware processor with software (including a digital signal processor), which work together to cause a device (e.g., device 610 and / or 620) to perform various functions, and / or hardware circuitry and / or a (one or more) processor, or portions thereof, which operate using software but may not have software present when not needed for operation. As another example, as used herein, the term "circuitry" can also encompass an implementation of only hardware circuitry or a processor or multiple processors or portions of a hardware circuitry or processor along with accompanying software and / or firmware. The term circuitry can also cover, for example, a baseband integrated circuit in a server, a cellular network node or device, or other computing or network device.

[0121] A computer program product can include one or more computer-executable components that are configured to perform some example embodiments when the program is run. The one or more computer-executable components can be at least one software code or portions thereof. Modifications and configurations required to implement the functions of certain example embodiments can be performed as routines, which can be implemented as added or updated software routines. The software routines can be downloaded to a device.

[0122] As an example, the software or computer program code or portions thereof can be in source code form, object code form, or in some intermediate form, and it can be stored in some carrier, distribution medium, or computer-readable medium, which can be any entity or device capable of carrying the program. For example, such a carrier can include a recording medium, a computer memory, a read-only memory, an optical and / or electrical carrier signal, a telecommunication signal, and a software distribution package. Depending on the required processing power, the computer program can be executed in a single electronic digital computer, or it can be distributed among multiple computers. The computer-readable medium or computer-readable storage medium can be a non-transitory medium.

[0123] In other example embodiments, the functions can be performed by hardware or circuitry included in a device (e.g., device 610 and / or 620), such as by using an application specific integrated circuit (ASIC), a programmable gate array (PGA), a field programmable gate array (FPGA), or any other combination of hardware and software. In yet another example embodiment, the functions can be implemented as a signal, a non-tangible device that can be carried by an electromagnetic signal downloaded from the Internet or other network.

[0124] According to certain example embodiments, an apparatus such as a node, device, or corresponding component may be configured as a circuit, computer, or microprocessor, such as a single-chip computer element or chipset, including at least a memory for providing storage capacity for arithmetic operations and an arithmetic processor for performing arithmetic operations.

[0125] The features, structures, or characteristics of the example embodiments described throughout this specification may be combined in any suitable manner in one or more example embodiments. For example, the use of phrases such as "certain embodiments", "example embodiments", "some embodiments", or other similar language throughout this specification refers to the fact that the specific features, structures, or characteristics described in connection with the embodiments may be included in at least one embodiment. Thus, the appearances of phrases such as "in certain embodiments", "example embodiments", "in some embodiments", "in other embodiments", or other similar language throughout this specification do not necessarily refer to the same set of embodiments, and the described features, structures, or characteristics may be combined in any suitable manner in one or more example embodiments. Additionally, the terms "cell", "node", "gNB", or other similar language throughout this specification may be used interchangeably.

[0126] As used herein, "at least one of the following: <list of two or more elements>" and "<list of two or more elements>" and similar phrasings, where the list of two or more elements is joined by "and" or "or", means at least any one of the elements, or at least any two or more of the elements, or at least all of the elements.

[0127] Those of ordinary skill in the art will readily understand that the disclosures described above may be practiced in a different order and / or with hardware elements configured differently from those disclosed. Thus, although this disclosure has been described based on these example embodiments, it will be apparent to those skilled in the art that certain modifications, variations, and alternative configurations will be apparent while remaining within the spirit and scope of the example embodiments. Although the above embodiments relate to 5G NR and LTE technologies, the above embodiments may also be applied to any other current or future 3GPP technologies, such as LTE Advanced technology and / or Fourth Generation (4G) and / or Sixth Generation (6G) technologies.

[0128] Partial Glossary: 3GPP Third Generation Partnership Project 5G Fifth Generation 6G Sixth Generation EMBB Enhanced Mobile Broadband FL Federated Learning gNB 5G or Next Generation NodeB LTE Long Term Evolution ML Machine Learning NR New Radio UE User Equipment

Claims

1. A device for wireless communication, comprising: at least one processor; as well as at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus to at least: identifying one or more stragglers among a plurality of user devices; suspending transmission of the aggregated model to the one or more stragglers for local model training; as well as Resume transmitting the aggregated model to at least one of the one or more stragglers for the local model training.

2. The device according to claim 1 identifies at least one user device among the multiple user devices as the one or more stragglers based on the delay of the device receiving the locally trained machine learning model exceeding a first threshold time period and lasting for more than a second threshold number of consecutive iterations.

3. The device according to claim 2, wherein: The resuming the transmission comprises: transmitting the aggregate model to at least one straggler device among the one or more stragglers; and The at least one memory stores instructions that, when executed by the at least one processor, cause the apparatus to: Receiving the locally trained machine learning model from the at least one user device, wherein the delay in receiving the locally trained machine learning model exceeds the first threshold time period.

4. The apparatus of claim 1 , wherein the at least one memory stores instructions that, when executed by the at least one processor, cause the apparatus to: An indication is sent to the identified one or more stragglers, notifying the one or more stragglers that they have been identified as stragglers.

5. The apparatus of claim 1 , wherein the at least one memory stores instructions that, when executed by the at least one processor, cause the apparatus to: After pausing transmission of the aggregate model to the one or more stragglers, evaluating at least one of: a network condition in which the apparatus and the plurality of user devices operate, or computing power available to the one or more stragglers to support training of the local model for the one or more stragglers.

6. An apparatus according to claim 5, wherein the resuming transmission of the aggregated model to the at least one stragglers among the one or more stragglers for the local model training is based on the assessment indicating that at least one of the network condition or the computing power to support the local model training is above a third threshold. 7 . The apparatus of claim 6 , wherein the evaluation indicating that the at least one of the network condition or the computing capacity is above the third threshold is performed when the transmission is suspended while a joint learning process is not completed.

8. The apparatus of claim 5, wherein the at least one memory stores instructions that, when executed by the at least one processor, cause the apparatus to: sending a context query message to the one or more stragglers, the context query message requesting information about at least one of the network condition or the computing capability to support local model training; and A response message is received from the one or more stragglers.

9. The apparatus according to claim 8, wherein the response message comprises a confirmation message or a non-confirmation message, wherein the confirmation message or the non-confirmation message indicates information used to determine whether to resume transmitting the aggregation model to the at least one stragglers among the one or more stragglers.

10. The apparatus of claim 9, wherein when the response message includes the non-confirmation message, a timer is provided to the apparatus from the one or more stragglers or another network entity, the timer enabling the apparatus to transmit the context query message again when the timer expires.

Images