Distributed continual learning for ML life cycle

The distributed continual learning approach addresses the challenge of adapting ML models to changing environments by distributing target configurations across nodes for parallel evaluation, resulting in efficient adaptation and performance improvement with reduced overhead and maintaining previous knowledge.

WO2024230927A1PCT designated stage expired Publication Date: 2024-11-14NOKIA TECHNOLOGIES OY
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2023/062352
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-10
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

Current machine learning models face challenges in adapting to changing environments and new data distributions, leading to suboptimal performance over time, as they often require retraining with fixed architectures and hyperparameters, which can result in loss of previous knowledge and inefficiencies in updating model configurations.

Method used

The proposed solution involves a distributed continual learning approach that uses multiple candidate nodes to evaluate and adapt the configuration of an ML model's adaptive stage, distributing different target configurations across nodes for parallel processing, allowing for the identification of improved configurations based on performance indicators and capabilities, thereby ensuring rapid search and convergence to better model architectures.

Benefits of technology

This method enables efficient adaptation of ML models to new data distributions with reduced computational overhead and processing requirements, ensuring rapid performance improvement and minimizing the impact on overall network performance, while maintaining previous knowledge.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2023062352_14112024_PF_FP_ABST
    Figure EP2023062352_14112024_PF_FP_ABST
Patent Text Reader

Abstract

Method comprising deciding whether a trigger criterion is fulfilled; instructing each of plural nodes to train a ML model if the trigger criterion is fulfilled, wherein, for each of the nodes, an adaptive stage of the ML model has a respective target configuration; receiving, for each of the nodes, a respective first performance indication indicating a first value of a performance of the ML model in response to instructing the respective node to train the ML model; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of a top node among the nodes is instructed to train, and the first value of the performance for the top node is better than each of the first values of the performance for the other nodes; instructing at least one of the nodes to train and run inference from the ML model with the adaptive stage having the improved configuration.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Distributed continual learning for ML life cycle

[0002] Technical Field

[0003] The present disclosure relates to machine learning, in particular to continual learning.

[0004] Abbreviations

[0005] 3GPP 3rdGeneration Partnership Project

[0006] 5G / 6G / 7G 5th / 6th / 7thGeneration

[0007] ACK Acknowledgement

[0008] Al Artificial Intelligence

[0009] CL Continual Learning

[0010] CSI Channel State Information

[0011] CU Central Unit

[0012] DAG Direct Acyclic Graph

[0013] DCL Distributed Continual Learning

[0014] DU Distributed Unit

[0015] F1 Interface between CU and DU of gNB

[0016] F1AP F1 Application Part

[0017] F / A Freeze / Adaptive gNB next generation NodeB

[0018] HW Hardware

[0019] KPI Key Performance Indicators

[0020] ML Machine Learning

[0021] ML_MMF ML model management function

[0022] NACK No Acknowledgement

[0023] NAS Neural Architectural Search

[0024] NGAP Next Generation Application Part

[0025] NG-RAN Next Generation RAN

[0026] NWDAF Network Data Analytics Function

[0027] RAN Radio Access Network

[0028] TL Transfer Learning

[0029] UE User Equipment

[0030] ULPC Uplink Power Control Xn Interface between gNBs

[0031] XnAP Xn Application Part

[0032] Background

[0033] Continual Learning (CL) concept falls under the categories Stateful (re)training and online Training. In stateful (re)training, the ML model continues training on new data (fine-tuning) after the initial training of the ML model. Typically, it refers to data iteration; i.e., the model architecture and features are substantially maintained but trained with new data. It allows updating (fine-tuning) the model with less data. Stateful (re)training may use automated training procedures. In online training, the ML model being used for inference is (typically continuously) trained in (near) real-time with the arrival of new training samples. The notion of (near) real-time vs. non real-time is context-dependent and is relative to the inference timescale.

[0034] Continual training is typically performed because, due to continuous context changes, the model trained for a certain task at time t would not fit at time t + 5t. Thus, a continual adaptation of the model is recommendable to ensure the required performance accuracy for the desired functionality (for the certain task). An example of a task is CSI compression.

[0035] Summary

[0036] It is an object to improve the prior art.

[0037] According to a first aspect, there is provided an apparatus comprising: one or more processors, and memory storing instructions that, when executed by the one or more processors, cause the apparatus to perform: deciding whether a trigger criterion to adapt an original configuration of an adaptive stage of a machine learning model is fulfilled; instructing each of plural candidate nodes to train the machine learning model in response to deciding that the trigger criterion is fulfilled, wherein, for each of the plural candidate nodes, the adaptive stage of the machine learning model has a respective target configuration; receiving, for each of the plural candidate nodes, a respective first performance indication indicating a first value of a performance of the machine learning model or a first value of the performance of the adaptive stage of the machine learning model in response to instructing the respective candidate node to train the machine learning model with the adaptive stage of the machine learning model having the respective target configuration; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of a top candidate node among the plural candidate nodes is instructed to train, and the first value of the performance for the top candidate node is better than each of the first values of the performance for the plural candidate nodes different from the top candidate node; instructing at least one of the plural candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

[0038] The trigger criterion may comprise at least one of the following:

[0039] - an original value of the performance of the machine learning model is poorer than a predefined first threshold if the adaptive stage has the original configuration; or

[0040] - a predefined first time period has elapsed since it was previously decided whether or not the trigger criterion is fulfilled; or

[0041] - a predefined second time period has elapsed since at least one of the plural candidate nodes was instructed previously to train the machine learning model with the adaptive stage having the original configuration.

[0042] The instructions, when executed by the one or more processors, may further cause the apparatus to perform selecting each of the plural candidate nodes among plural nodes based on a capability of the respective node for training the adaptive stage.

[0043] The instructions, when executed by the one or more processors, may further cause the apparatus to perform requesting, from each of the plural nodes, an information on the capability of the respective node; receiving, from each of the plural nodes, the information on the capability of the respective node; the selecting each of the plural candidate nodes based on the capability according to the received information on the capability of the respective node. The instructions, when executed by the one or more processors, may further cause the apparatus to perform determining, for each of the plural candidate nodes, the respective target configuration of the adaptive stage based on the capability of the respective candidate node.

[0044] The instructions, when executed by the one or more processors, may further cause the apparatus to perform requesting each of the candidate nodes to provide a second performance indication indicating a second value of the performance of the machine learning model after the at least one of the plural candidate nodes is instructed to train and run inference from the machine learning model with the adaptive stage having the improved configuration; receiving, from each of the candidate nodes, the respective second performance indication in response to the request of the respective candidate node to provide the second performance indication; evaluating the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the second values of the performance; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to evaluating that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined second threshold.

[0045] The instructions, when executed by the one or more processors, may further cause the apparatus to perform requesting each of the candidate nodes to provide respective evaluation data for determining the performance of the machine learning model; receiving, from each of the candidate nodes, the respective evaluation data; determining the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the received evaluation data; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined third threshold.

[0046] The instructions, when executed by the one or more processors, may further cause the apparatus to perform instructing each of the plural candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

[0047] The machine learning model may comprise a frozen stage and the adaptive stage, and the instructions, when executed by the one or more processors, may cause the apparatus to perform the instructing each of plural candidate nodes to train the machine learning model with the adaptive stage of the machine learning model having a respective target configuration, wherein the configuration of the frozen stage is the same for all of the candidate nodes, and the parameters of the frozen stage are the same for all of the candidate nodes.

[0048] According to a second aspect, there is provided a combined system comprising the apparatus according to the first aspect; and one of the plural candidate nodes.

[0049] According to a third aspect, there is provided an apparatus comprising: one or more processors, and memory storing instructions that, when executed by the one or more processors, cause the apparatus to perform: receiving, by a first candidate node of plural candidate nodes, an instruction to train a machine learning model comprising an adaptive stage having a target configuration; determining a first value of a performance of the machine learning model with the adaptive stage having the target configuration being trained by the first candidate node; receiving, for each of the plural candidate nodes different from the first candidate node, a performance indication indicating the respective first value of the performance of the machine learning model; receiving, for each of the plural candidate nodes different from the first candidate node, a configuration information informing on a respective target configuration of the adaptive stage of the machine learning model trained by the respective candidate node; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of the machine learning model trained by a top candidate node among the plural candidate nodes, and the first value of the performance for the top candidate node is better than the first value of the performance for the one or more candidate nodes different from the top candidate node; instructing at least one of the candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

[0050] The instructions, when executed by the one or more processors, may further cause the apparatus to perform informing each of the plural candidate nodes different from the first candidate node on the first value of the performance of the machine learning model with the adaptive stage having the target configuration trained by the first candidate node and on the target configuration.

[0051] The instructions, when executed by the one or more processors, may further cause the apparatus to perform determining a second value of the performance of the machine learning model with the adaptive stage having the improved configuration trained by the first candidate node; inhibiting the first candidate node to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the second value of the performance of the machine learning model with the adaptive stage having the improved configuration is poorer than a predefined first threshold.

[0052] The machine learning model may comprises a frozen stage and the adaptive stage, and the configuration of the frozen stage is the same for all the candidate nodes and the parameters of the frozen stage are the same for all the candidate nodes.

[0053] The instructions, when executed by the one or more processors, may further cause the apparatus to perform instructing all the candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

[0054] According to a fourth aspect, there is provided combined system comprising the apparatus according to the third aspect; and the first candidate node.

[0055] According to a fifth aspect, there is provided a method comprising: deciding whether a trigger criterion to adapt an original configuration of an adaptive stage of a machine learning model is fulfilled; instructing each of plural candidate nodes to train the machine learning model in response to deciding that the trigger criterion is fulfilled, wherein, for each of the plural candidate nodes, the adaptive stage of the machine learning model has a respective target configuration; receiving, for each of the plural candidate nodes, a respective first performance indication indicating a first value of a performance of the machine learning model or a first value of the performance of the adaptive stage of the machine learning model in response to instructing the respective candidate node to train the machine learning model with the adaptive stage of the machine learning model having the respective target configuration; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of a top candidate node among the plural candidate nodes is instructed to train, and the first value of the performance for the top candidate node is better than each of the first values of the performance for the plural candidate nodes different from the top candidate node; instructing at least one of the plural candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

[0056] The trigger criterion may comprise at least one of the following:

[0057] - an original value of the performance of the machine learning model is poorer than a predefined first threshold if the adaptive stage has the original configuration; or

[0058] - a predefined first time period has elapsed since it was previously decided whether or not the trigger criterion is fulfilled; or

[0059] - a predefined second time period has elapsed since at least one of the plural candidate nodes was instructed previously to train the machine learning model with the adaptive stage having the original configuration.

[0060] The method may further comprise selecting each of the plural candidate nodes among plural nodes based on a capability of the respective node for training the adaptive stage.

[0061] The method may further comprise requesting, from each of the plural nodes, an information on the capability of the respective node; receiving, from each of the plural nodes, the information on the capability of the respective node; wherein each of the plural candidate nodes is selected based on the capability according to the received information on the capability of the respective node.

[0062] The method may further comprise determining, for each of the plural candidate nodes, the respective target configuration of the adaptive stage based on the capability of the respective candidate node.

[0063] The method may further comprise requesting each of the candidate nodes to provide a second performance indication indicating a second value of the performance of the machine learning model after the at least one of the plural candidate nodes is instructed to train and run inference from the machine learning model with the adaptive stage having the improved configuration; receiving, from each of the candidate nodes, the respective second performance indication in response to the request of the respective candidate node to provide the second performance indication; evaluating the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the second values of the performance; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to evaluating that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined second threshold.

[0064] The method may further comprise requesting each of the candidate nodes to provide respective evaluation data for determining the performance of the machine learning model; receiving, from each of the candidate nodes, the respective evaluation data; determining the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the received evaluation data; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined third threshold.

[0065] The method may further comprise instructing each of the plural candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

[0066] The machine learning model may comprise a frozen stage and the adaptive stage, and each of plural candidate nodes may be instructed to train the machine learning model with the adaptive stage of the machine learning model having a respective target configuration, wherein the configuration of the frozen stage is the same for all of the candidate nodes, and the parameters of the frozen stage are the same for all of the candidate nodes.

[0067] According to a sixth aspect, there is provided a method comprising: receiving, by a first candidate node of plural candidate nodes, an instruction to train a machine learning model comprising an adaptive stage having a target configuration; determining a first value of a performance of the machine learning model with the adaptive stage having the target configuration being trained by the first candidate node; receiving, for each of the plural candidate nodes different from the first candidate node, a performance indication indicating the respective first value of the performance of the machine learning model; receiving, for each of the plural candidate nodes different from the first candidate node, a configuration information informing on a respective target configuration of the adaptive stage of the machine learning model trained by the respective candidate node; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of the machine learning model trained by a top candidate node among the plural candidate nodes, and the first value of the performance for the top candidate node is better than the first value of the performance for the one or more candidate nodes different from the top candidate node; instructing at least one of the candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration. The method may further comprise informing each of the plural candidate nodes different from the first candidate node on the first value of the performance of the machine learning model with the adaptive stage having the target configuration trained by the first candidate node and on the target configuration.

[0068] The method may further comprise determining a second value of the performance of the machine learning model with the adaptive stage having the improved configuration trained by the first candidate node; inhibiting the first candidate node to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the second value of the performance of the machine learning model with the adaptive stage having the improved configuration is poorer than a predefined first threshold.

[0069] The machine learning model may comprise a frozen stage and the adaptive stage, and the configuration of the frozen stage is the same for all the candidate nodes and the parameters of the frozen stage are the same for all the candidate nodes.

[0070] The method may further comprise instructing all the candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

[0071] Each of the methods of the fifth and sixth aspects may be a method of distributed continual learning

[0072] According to a seventh aspect, there is provided a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to carry out the method according to any of the fifth or sixth aspects. The computer program product may be embodied as a computer-readable medium or directly loadable into a computer.

[0073] According to some embodiments, at least one of the following advantages may be achieved:

[0074] • performance issues may be solved quickly;

[0075] • parallel processing reduces the processing requirements for each of the candidate nodes;

[0076] • a centralized and a distributed solution may be adopted; small overhead is required only;

[0077] Negative impact on overall performance on the network is reduced if previously trained model needs retraining and full retraining takes longer.

[0078] It is to be understood that any of the above modifications can be applied singly or in combination to the respective aspects to which they refer, unless they are explicitly stated as excluding alternatives.

[0079] Brief description of the drawings

[0080] Further details, features, objects, and advantages are apparent from the following detailed description of the preferred embodiments which is to be taken in conjunction with the appended drawings, wherein:

[0081] Fig. 1 depicts a high level view of the DCL process according to some example embodiments;

[0082] Fig. 2 illustrates a deployment according to some example embodiments;

[0083] Fig. 3 depicts a DCL process according to some example embodiments;

[0084] Fig. 4 illustrates a Neural Architecture Search method which may be used in some example embodiments;

[0085] Fig. 5 illustrates a message flow according to some example embodiments;

[0086] Fig. 6 illustrates a message flow according to some example embodiments;

[0087] Fig. 7 shows an apparatus according to an example embodiment;

[0088] Fig. 8 shows a method according to an example embodiment;

[0089] Fig. 9 shows an apparatus according to an example embodiment;

[0090] Fig. 10 shows a method according to an example embodiment; and

[0091] Fig. 11 shows an apparatus according to an example embodiment.

[0092] Detailed description of certain example embodiments

[0093] Herein below, certain example embodiments are described in detail with reference to the accompanying drawings, wherein the features of the embodiments can be freely combined with each other unless otherwise described. However, it is to be expressly understood that the description of certain embodiments is given by way of example only, and that it is by no way intended to be understood as limiting the disclosure to the disclosed details. Moreover, it is to be understood that the apparatus is configured to perform the corresponding method, although in some cases only the apparatus or only the method are described.

[0094] In ML learning (such as supervised learning and reinforcement learning), it is not guaranteed that updating the model will lead to a better knowledge space compared to past knowledge space. For new knowledge, one may freeze some layers of the current ML model (frozen stage of the ML model) and adapt other layers (adaptive stage) for retraining. However, the adapting (updating) of the adaptive layers may lead to losing current and / or previous knowledge represented in the adaptive stage.

[0095] The performance of the complete model (with adaptive + frozen stages) may depend on different factors, mainly:

[0096] • The selected ML architecture for the adaptive stage (such as number of hidden layers, number of hidden nodes)

[0097] • The additionally gathered data for the adaptive stage training

[0098] The (re)training is usually performed on a fixed ML configuration (also denoted ML architecture) since it takes time and overhead to optimize the adaptive stage’s hyperparameters (i.e. the configuration of the adaptive stage defined by hyperparameters such as number of layers and number of hidden nodes per layer). Though this simplifies learning process, it can be suboptimal if new data is substantially different from the one previously used for training the original ML model. Fixing ML model architecture and training to optimize parameter values may not be sufficient for all new training data distributions. In the context of the present application, the term “training” specifies the adaptation of the parameters based on training data, wherein hyperparameters (defining the architecture (configuration)) are fixed, unless otherwise specified or made clear from the context. The training data may be obtained by measurements or simulations.

[0099] Some example embodiments enable to improve the adaptive stage configuration (model architecture). The improvement may be performed with a small overhead and with small computational complexity. Convergence to an improved model architecture (or at least to a better model architecture) for a new (training) data distribution may be reached efficiently.

[0100] Some example embodiments provide a procedure and related signalling to enable distributed continual learning (DCL). Its goal is to ensure rapid search of an improved ML configuration (also denoted ML architecture) through the distribution of the search actions over different nodes (e.g. UEs, gNBs). The nodes may perform the search actions in parallel. The search actions comprise substantially at least running inference from an ML model with a certain configuration and determining the performance of this ML model. The term “improved ML configuration” does not necessarily mean to identify an absolute optimum (with respect to a certain KPI and use case) of all potential ML configurations but it includes to identify the best ML configuration (with respect to the certain KPI and use case) among a set of target ML configurations even though these ML configurations may not comprise the absolutely best ML configuration. The term “ML configuration” denotes the architecture of the ML model defined by hyperparameters, which may comprise the number of layers and hidden nodes of the ML model.

[0101] In some example embodiments, the ML model may be hosted by an ML model management function (ML_MMF). For example, ML_MMF may be stand-alone or may be embedded in gNB or any other entity in the NG-RAN or anywhere else in the core or radio network (or even in some other environment such as a power grid)). If it is decided to perform the continuous learning for the ML model (because a trigger condition is fulfilled), the ML model management function (ML_MMF) may identify the frozen part (frozen stage) valid for a group of nodes (e.g. for a group of UEs). Then, ML_MMF has the task to select the best configuration of the adaptive stage among a set of target configurations of the adaptive stage and to perform the training of the complete ML model comprising the frozen stage and the adaptive stage.

[0102] According to some example embodiments, to speed up the selection and training of the adaptive part of the ML model, ML_MMF may distribute / delegate different pre-selected DCL configurations (target configurations) over several candidate nodes (e.g. UEs). The nodes may be selected among the identified group for which the frozen layers are valid. In some example embodiments, if there are sufficient target configurations, ML_MMF may distribute the target configurations to all nodes of the group of nodes for which the frozen stage is valid.

[0103] That is, from a high-level point of view, a procedure according to some example embodiments may comprise one or more of the following actions:

[0104] - triggering the DCL procedure distribution of the ML models with different configurations of the adaptive stage

[0105] - selecting the target configurations of the adaptive stage. The target configurations are identified as the most relevant and these configurations should be evaluated. - selecting candidate nodes (e.g. UEs, potentially based on UE capabilities, for example)

[0106] - mapping of the selected target configurations to the candidate nodes

[0107] - determining the best configuration among the target configurations

[0108] - optionally validating the best configuration among the target configurations before the nodes are instructed to run inference from the best configuration among the target configurations.

[0109] Some example embodiments provide a signalling between the ML_MMF and the candidate nodes and / or between the candidate nodes to enable DCL procedure.

[0110] Fig. 1 depicts a high-level view of the process according to some example embodiments. When continual learning is triggered (S1), a first level CL process is run (S2). The trigger of S1 may be, for example, one or more of a lapse of a predefined time period after the last execution of the first level CL process, or a lapse of a predefined time period after the last change of the configuration of the adaptive stage of the ML model. In the first level CL process (S2), the configuration (architecture) of the ML model is fixed and only the parameters may be adapted by the training. More in detail, the configuration of the adaptive stage (defined by hyperparameters) stays unchanged and it is only re-trained to improve the parameters. Both the configuration and the parameters of the frozen stage remain unchanged.

[0111] The performance of the ML model is evaluated with regards to a predefined KPI (S3). If the performance is judged as being acceptable (S3 = yes), then the ML model with the unchanged architecture and the parameters derived from the training is good enough and the nodes may continue to run inference from the trained ML model (i.e. with the fixed ML architecture and the parameters derived from the training).

[0112] If the performance of the ML model after performing the first level CL process is not acceptable (S3 = no), then an architecture search may be performed. For the architecture search, the configuration (architecture) of the adaptive stage architecture is modified while the configuration (architecture) of the frozen stage remains unchanged. Different configurations may be feasible for the adaptive stage and searching the best configuration among the target configurations may require a lot of resources in terms of time and computation. Therefore, according to some example embodiments, the architecture search is performed in a distributed way. Thus, the architecture search is speed up and the impact on the overall performance on the network is reduced. That is, as outlined in Fig. 1 , if the performance after the continuous learning with the fixed ML architecture is not good enough (the performance of the ML model is not sufficient, i.e. lower than a threshold), several configurations of the adaptive stage are selected (S4). The ML model with the adaptive stage having different configurations is distributed to plural candidate nodes which run inference from the ML model and determine the performance (S5, “DCL search”). I.e., the configurations of the adaptive stage distributed to the different nodes are different from each other. Then, the performance of the assigned ML models is evaluated by each node (S6) and the best model performance is identified (with regards to a predefined threshold) to determine a best configuration of the adaptive stage among the configurations of the adaptive stage distributed to the candidate nodes (S6). At least one of the candidate nodes (typically: all the candidate nodes) may then run inference from the ML model with the adaptive stage having the determined best configuration among the target configurations (not shown in Fig. 1). In some example embodiments, the overall performance of the ML model comprising the adaptive stage with the best configuration among the target configurations may be verified prior to inference by the at least one candidate node.

[0113] In some example embodiments, the first level CL process (S2) may be omitted. In such example embodiments, the CL trigger (S1) may trigger directly the second level CL process (S4 to S6). As another option, the performance of the ML model may be measured continuously or periodically, independent from performing a first level CL process. If the performance is not acceptable (S3 = no), the second level CL process (S4 to S6) is performed. I.e., a poor performance of the ML model may be considered as a trigger to start the second level CL process (S4 to S6). In some example embodiments, each one of plural triggers such as the trigger of S1 and the trigger of (S3 = no) may trigger to start the second level CL process (S4 to S6). In some example embodiments, plural of these triggers have to be fulfilled simultaneously to start the second level CL process (S4 to S6).

[0114] Fig. 2 depicts an example network where the process according to some example embodiments is performed. In detail, in this example, gNB hosts the ML_MMF, and the candidate nodes performing the DCL search are UEs attached to the gNB. gNB sends to the UEs the configuration (config 1 , config 2,...) of the ML model (at least the configuration of the adaptive stage if the UEs know the configuration of the frozen stage). In response, the UEs provide the results of the performance measurement (res 1 , res 2,...) to the gNB. gNB may evaluate the results of the performance measurements (the values of the performance) to determine the best configuration among the configurations provided to the UEs.

[0115] The number n of the plural candidate nodes may be any number larger than 1 . Nodes (in Fig. 2: UEs) not being candidate nodes may be attached to the ML_MMF (in Fig. 2: gNB) in addition to the candidate nodes.

[0116] Once DCL is triggered (S3 = no in Fig. 1), ML_MMF may:

[0117] - Identify the candidate nodes (e.g. UEs) participating to the DCL search. For example, ML_MMF may account for the capabilities of the (candidate) nodes to select the candidate nodes.

[0118] - Identify the relevant configurations (target configurations) of the adaptive stage. For example, ML_MMF may perform a full search process (search through all possible configurations as the search space). As another example, ML_MMF may use a look up table storing configurations which performed well in the past as the search space. For example, the relevant configurations may be obtained from previous CL runs.

[0119] - Mapping the selected target configurations to the identified nodes. For this purpose, for example, ML_MMF may take the capabilities of the different nodes (e.g. UEs) into account. If the capabilities are not known to the ML_MMF anyway (e.g. from a previous DCL process), ML_MMF may explicitly request the nodes to indicate their capabilities.

[0120] - Analyse the obtained results (values of the performance) and select the best configuration among the target configurations.

[0121] - Optionally, evaluate the overall performance for the determined best configuration to ensure that the CL performance with the new configuration is satisfactory for all the nodes.

[0122] Some aspects of some example embodiments are now described at greater detail. Fig. 3 depicts a cycle of a DCL process according to some example embodiments.

[0123] After the DCL process is triggered (S11), the ML_MMF (hosted by the gNB in the example of Fig. 2) may collect the capabilities of the nodes (e.g. UEs in the example of Fig. 2) sharing the same ML Model (having the same frozen stage) (S12). For each of the nodes, ML_MMF may decide, based on the capabilities, whether the respective node participates in the DCL search (i.e. whether the node becomes a candidate node), and, if the respective node becomes a candidate node, map one of the target configurations of the adaptive stage to the respective candidate node (S13). Table 1 indicates some example configurations with the respective number of parameters:

[0124] Table 1 : Example configurations of the adaptive stage and respective number of parameters (assuming 1000 trainable parameters per node)

[0125] Each of the configurations requires a different number of trainable parameters for the adaptive stage of the ML model. Thus, ML_MMF may map each configuration to one of the candidate nodes in accordance with the capabilities of the candidate node. For example, a UE (node) with low battery level or a slow processor may be assigned a model with a small number of trainable parameters.

[0126] In some example embodiments, ML_MMF may not take into account the capabilities of the nodes. For example, ML_MMF may map the target configurations to the nodes according to certain sequences of the candidate nodes and the target configurations or by a random procedure.

[0127] In some example embodiments, ML_MMF may take into account previous DCL iterations (if any) for the selection of the target configuration: If a configuration was already selected and tested with unsatisfactory performance this configuration may be omitted in the next DCL iteration.

[0128] Each of the candidate nodes runs inference from (infers) the assigned ML model (with the adaptive stage having the respective configuration)(S14). These runs may be performed fully or partly in parallel by the different candidate nodes. Each candidate node determines a value of the performance (e.g. KPI) for this run. For example, each node may determine a value of the performance of the entire ML model comprising the frozen stage and the adaptive stage. In some example embodiments, one or more of the nodes may determine the performance of the adaptive stage only but not of the frozen stage. In such example embodiments, the adaptive stage is using “latent space inputs" which are generated from the output of the frozen stage and then stored. Hence the frozen stage need is not run for the purpose of checking the performance of the adaptive stage because the latent space data and the output (labels) are available. Based on these performance values, a best configuration Cgamong the target configurations may be determined, namely the configuration showing the best performance among the target configurations.

[0129] In some example embodiments (not shown in Fig. 3), at least one of the nodes (typically each of the nodes) may run inference from the ML model comprising the adaptive stage having the best configuration among the target configurations, without an evaluation of the overall performance of the ML model comprising the adaptive stage having the best configuration among the target configurations being performed.

[0130] However, the performance determined at this level corresponds to performance values of each single node, based on its local data. For this reason, in some example embodiments, the performance of the ML model comprising the adaptive stage having the determined best configuration among the target configurations, is validated with respect to the rest of the nodes (evaluation of the overall performance).

[0131] In S15, it is decided how this validation is performed: o At a central node (e.g. ML_MMF, S16): This option assumes that validation data are available at the central node. The validation data should be well representative for all the nodes (e.g. the whole area over which the UEs representing the nodes are spread). The validation data may be used to estimate the performance with the ML model comprising the adaptive stage having the determined best configuration among the target configurations. For this purpose, the central node may receive validation data from each of the nodes (e.g. upon request from the central node, or stored at the central node in advance or from a previous request). o At node level: In this option, the central node informs each node on the determined best configuration Cgamong the target configurations (S17). Each node runs inference from the ML model with the determined best configuration among the target configurations and measures the performance (S18). After a predefined time, the nodes share their measured (estimated) performance value.

[0132] In some example embodiments, S15 may be omitted. In such example embodiments, the validation is performed always at the central node or always at node level, respectively.

[0133] Based on the estimated / measured performance values, an overall performance is estimated (S19). The value of the overall performance results from a combination of the single values of the performance determined at the nodes. Depending on the performance (KPI) which is measured, it may be, for example, a sum of the single values, an average of the single values, a product of the single values, a maximum of the single values, a minimum of the single values, a cumulative distribution function of the single values (such as a percentile), or some more complex combination.

[0134] If the overall performance is satisfactory (S19 = yes), the determined best configuration is considered to be validated and the nodes may use (train and / or run inference from) the determined best configuration among the target configurations. If the overall performance is not satisfactory (S19 = no), another cycle of a DCL process may be performed (potentially without collecting the UE capabilities because they are already known from the current cycle) and / or the nodes may return to the configuration of the adaptive stage of the ML model inferred prior to the current DCL process.

[0135] The selection of the most relevant adaptive stage configurations to be evaluated with the DCL process may be realized, for example, by an approach called Neural architecture search (NAS), which is an auto machine learning (or AutoML) technique. It aims to automatically identify deep neural models and adaptively learn appropriate deep neural structures. Fig. 4 shows the overall components of NAS [8] to find a best configuration among the target configurations of the adaptive stage which each UE can use. As shown in Fig. 4, a search strategy 21 selects an architecture A from a predefined search space 21. The selected architecture A is passed to a performance estimation strategy 23 which returns the estimated performance of the architecture A to the search strategy 22. The search strategy 22 may select the architecture (configuration) A having the best performance among the target configurations. Fig. 5 shows a signalling according to some example embodiments for the case where ML_MMF is hosted at a serving gNB and DCL is performed over its attached UEs running inference from the same ML model having a frozen stage and an adaptive stage. In Fig. 5, the number of UEs is 3, but any number of plural UEs may be involved.

[0136] • 1 : Use of the current ML model (showing poor performance) is suspended

[0137] • 2: DCL is triggered (corresponds to S11 in Fig. 3).

[0138] • 3 to 5: ML_MMF requests the capabilities from the attached UEs and optionally requests local labelled data set (validation data; for the case where the determined best configuration among the target configurations should be tested and validated at ML_MMF side). Depending on implementation, the request for validation data, if any, may be performed at a later point in time, separate from the request for capabilities.

[0139] • 6 to 8: The UEs provide information on their capabilities and optionally validation data.

[0140] • 9: ML_MMF identifies the appropriate configurations of the adaptive stage and maps them to the different UEs.

[0141] • 10 to 12: ML_MMF informs each of the UEs on the respective ML model comprising the frozen stage and the adaptive stage having one of the selected configurations to be trained and evaluated (Steps 10 to 12).

[0142] • 13 to 15: Each of the UEs trains the respective ML model with its locally available data in parallel in the same time and measures the performance.

[0143] • 16 to 18: Each UE provides feedback on the measured performance to ML_MMF.

[0144] • 19: ML_MMF identifies the best configuration among the target configurations (e.g. config 1) which corresponds to the one achieving the best performance in 13 to 15.

[0145] • 20 and 21 (optional validation, distributed case): ML_MMF requests the other UEs to check the performance of the identified best configuration among the target configurations.

[0146] • If the optional validation is centralized (not shown in Fig. 5), then the ML_MMF may use the data shared by the UEs in 3 to 5 to check the performance of the ML model where the adaptive stage has the best configuration among the target configurations identified in 19.

[0147] The DCL process according to some example embodiments may be applied not only to the case of gNB and UEs (as shown in Fig. 5) but also to intra-RAN and / or inter-RAN cases, for example. Some example cases are the following:

[0148] Case 1 : gNB shares the DCL configuration with the neighboring gNBs via XnAP Case 2: gNB-Cll shares the DCL configuration with the gNB-DUs via F1AP

[0149] • Case 3: 5GC (e.g., NWDAF) shares the DCL configuration with the gNBs via NGAP

[0150] • other cases

[0151] Fig. 6 illustrates an example of Case 1. A use case may be e.g., LILPC.

[0152] • 0: Source gNB (hosting ML_MMF) decides the target gNBs for coordination. The selection criteria may be similarities between various radio or traffic conditions, mean cell throughput variation, etc. The coordination configuration may also include the constraints when to release the DCL operation between correlated cells. This means the similarity does not hold true any longer.

[0153] The gNBs suspend using the current ML model (showing poor performance).

[0154] • 1 : The DCL process is triggered at source gNB.

[0155] • 2 and 3: Optionally, source gNB requests validation data (local labelled data set) from the target gNBs (for the case where the final selected configuration should be tested and validated at ML_MMF side). Depending on implementation, this request, if any, may be performed later in the process.

[0156] • 4: Source gNB selects configurations of the adaptive stage of the ML model inferred by all the involved gNBs and maps them to the different gNBs

[0157] • 5: Source gNB provides a respective configuration (“DCL configuration”) of the adaptive stage of the ML model to each of the target gNBs and to itself

[0158] • 6: Source gNB and target gNBs perform the training and the run inference for the ML related functionalities and measure the performance.

[0159] Fig. 6 includes two options for the evaluation of the measured performance values to identify the best configuration among the target configurations of the adaptive stage. Typically, only one of these options is performed.

[0160] • 7 - Option 1: The evaluation may be executed jointly with all the target gNBs or per gNB. This option is preferred to verify the global improvement target. For instance, when two coordinated cells (source and target) are sharing the common DCL configuration to determine the best power control parameters, the aggregate cell throughput should be maximized but the inter-cell interference should be minimized. 8 and 9 - Option 2: The gNBs evaluate the performance values in parallel and feedback the obtained performance to ML_MMF. This option is preferred for local improvement objective.

[0161] • 10. The outcome of the evaluation may be buffered by source gNB. Based on the outcome of the evaluation, source gNB may reconfigure the ML model on the involved gNBs (not shown in Fig. 6).

[0162] • In Step 11 , source gNB (hosting ML_MMF) selects the best configuration among the target configurations which corresponds to the one achieving the best performance via the outcome assessment. If the overall performance is satisfactory (ACK case), the procedure is ended (“EXIT”). If the overall performance is not satisfactory (NACK case), another iteration of the DCL process may be performed (“Return to 1”).

[0163] Fig. 7 shows an apparatus according to an example embodiment. The apparatus may be a management function (such as ML_MMF) or an element thereof. Fig. 8 shows a method according to an example embodiment. The apparatus according to Fig. 7 may perform the method of Fig. 8 but is not limited to this method. The method of Fig. 8 may be performed by the apparatus of Fig. 7 but is not limited to being performed by this apparatus.

[0164] The apparatus comprises means for deciding 110, first means for instructing 120, means for receiving 130, means for identifying 140, and second means for instructing 150. The means for deciding 110, first means for instructing 120, means for receiving 130, means for identifying 140, and second means for instructing 150 may be a deciding means, first instructing means, receiving means, identifying means, and second instructing means, respectively. The means for deciding 110, first means for instructing 120, means for receiving 130, means for identifying 140, and second means for instructing 150 may be a decider, first instructor, identifier, receiver, identifier, and second instructor, respectively. The means for deciding 110, first means for instructing 120, means for receiving 130, means for identifying 140, and second means for instructing 150 may be a deciding processor, first instructing processor, receiving processor, identifying processor, and second instructing processor, respectively.

[0165] The means for deciding 110 decides whether a trigger criterion is fulfilled (S110). The trigger criterion triggers to adapt an original configuration of an adaptive stage of a machine learning model. I.e., the machine learning model comprises the adaptive stage and may additionally comprise a frozen stage. In response to deciding that the trigger criterion is fulfilled (S110 = yes), the first means for instructing 120 instructs each of plural candidate nodes to train the machine learning model, wherein the adaptive stage of the machine learning model has a respective target configuration (S120). For example, target configurations may be mapped to the nodes according to the capabilities of the nodes.

[0166] In response to instructing the respective candidate node to infer the machine learning model with the adaptive stage of the machine learning model having the respective target configuration (i.e., in response to the instructing of S120), the means for receiving 130, receives for each of the plural candidate nodes, a respective first performance indication (S130). Each of the first performance indications indicates a first value of a performance of the machine learning model or a first value of the performance of the adaptive stage.

[0167] The means for identifying 140 identifies an improved configuration of the adaptive stage as follows (S140). The improved configuration is the target configuration of the adaptive stage a top candidate node among the plural candidate nodes is instructed to train, and the first value of the performance for the top candidate node is better than each of the first values of the performance for the plural candidate nodes different from the top candidate node. In other words, the performance value indicated by the top candidate node is the best among the indicated performance values. Whether the best performance value is a maximum value or a minimum value depends on the measured performance (KPI) and use case.

[0168] The second means for instructing 150 instructs at least one of the plural candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration (S150). Typically, the second means for instructing 150 instructs plural or even all of the plural candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

[0169] Fig. 9 shows an apparatus according to an example embodiment. The apparatus may be a node in DCL or an element thereof. Fig. 10 shows a method according to an example embodiment. The apparatus according to Fig. 9 may perform the method of Fig. 10 but is not limited to this method. The method of Fig. 10 may be performed by the apparatus of Fig. 9 but is not limited to being performed by this apparatus. The apparatus comprises first means for receiving 210, means for determining 220, second means for receiving 230, third means for receiving 240, means for identifying 250, and means for instructing 260. The comprises first means for receiving 210, means for determining 220, second means for receiving 230, third means for receiving 240, means for identifying 250, and means for instructing 260 may be a first receiving means, determining means, second receiving means, third receiving means, identifying means, and instructing means, respectively. The comprises first means for receiving 210, means for determining 220, second means for receiving 230, third means for receiving 240, means for identifying 250, and means for instructing 260 may be a first receiver, determiner, second receiver, third receiver, identifier, and instructor, respectively. The comprises first means for receiving 210, means for determining 220, second means for receiving 230, third means for receiving 240, means for identifying 250, and means for instructing 260 may be a first receiving processor, determining processor, second receiving processor, third receiving processor, identifying processor, and instructing processor, respectively.

[0170] The first means for receiving 210 receives, by a first candidate node of plural candidate nodes, an instruction to train a machine learning model (S210). The machine learning model comprises an adaptive stage and may comprise a frozen stage. According to the received instruction, the adaptive stage has a target configuration.

[0171] The means for determining 220 determines (estimates) a first value of a performance of the machine learning model with the adaptive stage having the target configuration (S220). For example, the first value of the performance may be determined based on labelled data.

[0172] The means for receiving 230 receives, for each of the plural candidate nodes different from the first candidate node, a performance indication (S230). For each of the plural candidate nodes different from the first candidate node, the performance indication indicates the respective first value of the performance of the machine learning model.

[0173] The second means for receiving 240 receives, for each of the plural candidate nodes different from the first candidate node, a configuration information (S240). For each of the plural candidate nodes different from the first candidate node, the configuration information informs on a respective target configuration of the adaptive stage of the machine learning model trained by the respective candidate node. The means for identifying 250 identifies an improved configuration of the adaptive stage as follows (S250): the improved configuration is the target configuration of the adaptive stage of the machine learning model trained by a top candidate node among the plural candidate nodes (including the first candidate node). The first value of the performance for the top candidate node is better than the first value of the performance for the one or more candidate nodes different from the top candidate node. In other words, the performance value indicated by the top candidate node is the best among the indicated performance values. Whether the best performance value is a maximum value or a minimum value depends on the measured performance (KPI) and use case.

[0174] The means for instructing 260 instructs at least one of the candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration (S260). Typically, the means for instructing 260 may instruct plural or all of the candidate nodes to train and run inference from the machine learning model with the adaptive stage having the best configuration.

[0175] Fig. 11 shows an apparatus according to an example embodiment. The apparatus comprises at least one processor 810, at least one memory 820 storing instructions that, when executed by the at least one processor 810, cause the apparatus at least to perform the method according to at least one of the following figures and related description: Fig. 8, or Fig. 10.

[0176] Some example embodiments are described where the ML model comprises a frozen stage and an adaptive stage, and both the architecture and the parameters of the frozen stage are the same for all candidate nodes. Some example embodiments may not comprise a frozen stage. I.e., in these example embodiments, in the DCL search step, potentially the entire architecture and all parameters of the ML model may be modified. In other terms, in these example embodiments, the ML model consists of the adaptive stage.

[0177] Some example embodiments are described where a poor performance of the ML model triggers the process of a potential reconfiguration of the adaptive stage of the ML model (continual learning). However, in some example embodiments, other conditions may trigger the potential reconfiguration in addition to or instead of the poor performance of the ML model. For example, the process of continual learning may be triggered periodically, or after a predefined time period has elapsed after the last actual reconfiguration of the adaptive stage.

[0178] In some example embodiments, the identified best configuration of the adaptive stage is configured in only one of the candidate nodes for training and inference. In other example embodiments, the identified best configuration of the adaptive stage is configured in plural candidate nodes or even all of the candidate nodes for training and inference.

[0179] One piece of information may be transmitted in one or plural messages from one entity to another entity. Each of these messages may comprise further (different) pieces of information.

[0180] Names of network elements, network functions, protocols, and methods are based on current standards. In other versions or other technologies, the names of these network elements and / or network functions and / or protocols and / or methods may be different, as long as they provide a corresponding functionality. The same applies correspondingly to the terminal.

[0181] If not otherwise stated or otherwise made clear from the context, the statement that two entities are different means that they perform different functions. It does not necessarily mean that they are based on different hardware. That is, each of the entities described in the present description may be based on a different hardware, or some or all of the entities may be based on the same hardware. It does not necessarily mean that they are based on different software. That is, each of the entities described in the present description may be based on different software, or some or all of the entities may be based on the same software. Each of the entities described in the present description may be deployed in the cloud.

[0182] According to the above description, it should thus be apparent that example embodiments provide, for example, a central node (such as ML_MMF) or a component thereof, an apparatus embodying the same, a method for controlling and / or operating the same, and computer program(s) controlling and / or operating the same as well as mediums carrying such computer program(s) and forming computer program product(s). The central node may be implemented in a gNB or a AMF, for example. According to the above description, it should thus be apparent that example embodiments provide, for example, a distributed node or a component thereof, an apparatus embodying the same, a method for controlling and / or operating the same, and computer program(s) controlling and / or operating the same as well as mediums carrying such computer program(s) and forming computer program product(s). The distributed node may be implemented in a UE or a gNB, for example.

[0183] Implementations of any of the above described blocks, apparatuses, systems, techniques or methods include, as non-limiting examples, implementations as hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof. Each of the entities described in the present description may be embodied in the cloud.

[0184] It is to be understood that what is described above is what is presently considered the preferred example embodiments. However, it should be noted that the description of the preferred example embodiments is given by way of example only and that various modifications may be made without departing from the scope of the disclosure as defined by the appended claims.

[0185] The terms “first X” and “second X” include the options that “first X” is the same as “second X” and that “first X” is different from “second X”, unless otherwise specified. As used herein, “at least one of the following: ” and “at least one of ” and similar wording, where the list of two or more elements are joined by “and” or “or”, mean at least any one of the elements, or at least any two or more of the elements, or at least all the elements.

Claims

Claims:

1. Apparatus comprising: one or more processors, and memory storing instructions that, when executed by the one or more processors, cause the apparatus to perform: deciding whether a trigger criterion to adapt an original configuration of an adaptive stage of a machine learning model is fulfilled; instructing each of plural candidate nodes to train the machine learning model in response to deciding that the trigger criterion is fulfilled, wherein, for each of the plural candidate nodes, the adaptive stage of the machine learning model has a respective target configuration; receiving, for each of the plural candidate nodes, a respective first performance indication indicating a first value of a performance of the machine learning model or a first value of the performance of the adaptive stage of the machine learning model in response to instructing the respective candidate node to train the machine learning model with the adaptive stage of the machine learning model having the respective target configuration; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of a top candidate node among the plural candidate nodes is instructed to train, and the first value of the performance for the top candidate node is better than each of the first values of the performance for the plural candidate nodes different from the top candidate node; instructing at least one of the plural candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

2. The apparatus according to claim 1 , wherein the trigger criterion comprises at least one of the following:- an original value of the performance of the machine learning model is poorer than a predefined first threshold if the adaptive stage has the original configuration; or- a predefined first time period has elapsed since it was previously decided whether or not the trigger criterion is fulfilled; or- a predefined second time period has elapsed since at least one of the plural candidate nodes was instructed previously to train the machine learning model with the adaptive stage having the original configuration.

3. The apparatus according to any of claims 1 and 2, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform selecting each of the plural candidate nodes among plural nodes based on a capability of the respective node for training the adaptive stage.

4. The apparatus according to claim 3, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform requesting, from each of the plural nodes, an information on the capability of the respective node; receiving, from each of the plural nodes, the information on the capability of the respective node; the selecting each of the plural candidate nodes based on the capability according to the received information on the capability of the respective node.

5. The apparatus according to any of claims 3 and 4, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform determining, for each of the plural candidate nodes, the respective target configuration of the adaptive stage based on the capability of the respective candidate node.

6. The apparatus according to any of claims 1 to 5, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform requesting each of the candidate nodes to provide a second performance indication indicating a second value of the performance of the machine learning model after the at least one of the plural candidate nodes is instructed to train and run inference from the machine learning model with the adaptive stage having the improved configuration; receiving, from each of the candidate nodes, the respective second performance indication in response to the request of the respective candidate node to provide the second performance indication; evaluating the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the second values of the performance; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to evaluating that the performance of the machine learning model with the adaptive stage ofthe at least one candidate node having the improved configuration is poorer than a predefined second threshold.

7. The apparatus according to any of claims 1 to 5, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform requesting each of the candidate nodes to provide respective evaluation data for determining the performance of the machine learning model; receiving, from each of the candidate nodes, the respective evaluation data; determining the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the received evaluation data; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined third threshold.

8. The apparatus according to any of claims 1 to 7, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform instructing each of the plural candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

9. The apparatus according to any of claims 1 to 8, wherein the machine learning model comprises a frozen stage and the adaptive stage, and the instructions, when executed by the one or more processors, cause the apparatus to perform the instructing each of plural candidate nodes to train the machine learning model with the adaptive stage of the machine learning model having a respective target configuration, wherein the configuration of the frozen stage is the same for all of the candidate nodes, and the parameters of the frozen stage are the same for all of the candidate nodes.

10. A combined system comprising the apparatus according to any of claims 1 to 9; and one of the plural candidate nodes.

11. Apparatus comprising:one or more processors, and memory storing instructions that, when executed by the one or more processors, cause the apparatus to perform: receiving, by a first candidate node of plural candidate nodes, an instruction to train a machine learning model comprising an adaptive stage having a target configuration; determining a first value of a performance of the machine learning model with the adaptive stage having the target configuration being trained by the first candidate node; receiving, for each of the plural candidate nodes different from the first candidate node, a performance indication indicating the respective first value of the performance of the machine learning model; receiving, for each of the plural candidate nodes different from the first candidate node, a configuration information informing on a respective target configuration of the adaptive stage of the machine learning model trained by the respective candidate node; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of the machine learning model trained by a top candidate node among the plural candidate nodes, and the first value of the performance for the top candidate node is better than the first value of the performance for the one or more candidate nodes different from the top candidate node; instructing at least one of the candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

12. The apparatus according to claim 11 , wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform informing each of the plural candidate nodes different from the first candidate node on the first value of the performance of the machine learning model with the adaptive stage having the target configuration trained by the first candidate node and on the target configuration.

13. The apparatus according to any of claims 11 to 12, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform determining a second value of the performance of the machine learning model with the adaptive stage having the improved configuration trained by the first candidate node; inhibiting the first candidate node to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the second value of the performance of the machine learning model with the adaptive stage having the improved configuration is poorer than a predefined first threshold.

14. The apparatus according to any of claims 11 to 13, wherein the machine learning model comprises a frozen stage and the adaptive stage, and the configuration of the frozen stage is the same for all the candidate nodes and the parameters of the frozen stage are the same for all the candidate nodes.

15. The apparatus according to any of claims 11 to 14, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform instructing all the candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

16. A combined system comprising the apparatus according to any of claims 11 to 15; and the first candidate node.

17. Method comprising: deciding whether a trigger criterion to adapt an original configuration of an adaptive stage of a machine learning model is fulfilled; instructing each of plural candidate nodes to train the machine learning model in response to deciding that the trigger criterion is fulfilled, wherein, for each of the plural candidate nodes, the adaptive stage of the machine learning model has a respective target configuration; receiving, for each of the plural candidate nodes, a respective first performance indication indicating a first value of a performance of the machine learning model or a first value of the performance of the adaptive stage of the machine learning model in response to instructing the respective candidate node to train the machine learning model with the adaptive stage of the machine learning model having the respective target configuration; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of a top candidate node among the plural candidate nodes is instructed to train, and the first value of the performance for the top candidate node is better than each of the first values of the performance for the plural candidate nodes different from the top candidate node; instructing at least one of the plural candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

18. The method according to claim 17, wherein the trigger criterion comprises at least one of the following:- an original value of the performance of the machine learning model is poorer than a predefined first threshold if the adaptive stage has the original configuration; or- a predefined first time period has elapsed since it was previously decided whether or not the trigger criterion is fulfilled; or- a predefined second time period has elapsed since at least one of the plural candidate nodes was instructed previously to train the machine learning model with the adaptive stage having the original configuration.

19. The method according to any of claims 17 and 18, further comprising selecting each of the plural candidate nodes among plural nodes based on a capability of the respective node for training the adaptive stage.

20. The method according to claim 19, further comprising requesting, from each of the plural nodes, an information on the capability of the respective node; receiving, from each of the plural nodes, the information on the capability of the respective node; wherein each of the plural candidate nodes is selected based on the capability according to the received information on the capability of the respective node.

21. The method according to any of claims 19 and 20, further comprising determining, for each of the plural candidate nodes, the respective target configuration of the adaptive stage based on the capability of the respective candidate node.

22. The method according to any of claims 17 to 21 , further comprising requesting each of the candidate nodes to provide a second performance indication indicating a second value of the performance of the machine learning model after the at least one of the plural candidate nodes is instructed to train and run inference from the machine learning model with the adaptive stage having the improved configuration; receiving, from each of the candidate nodes, the respective second performance indication in response to the request of the respective candidate node to provide the second performance indication;evaluating the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the second values of the performance; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to evaluating that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined second threshold.

23. The method according to any of claims 17 to 21 , further comprising requesting each of the candidate nodes to provide respective evaluation data for determining the performance of the machine learning model; receiving, from each of the candidate nodes, the respective evaluation data; determining the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration based on the received evaluation data; inhibiting the at least one of the plural candidate nodes to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the performance of the machine learning model with the adaptive stage of the at least one candidate node having the improved configuration is poorer than a predefined third threshold.

24. The method according to any of claims 17 to 23, further comprising instructing each of the plural candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.

25. The method according to any of claims 17 to 24, wherein the machine learning model comprises a frozen stage and the adaptive stage, and each of plural candidate nodes is instructed to train the machine learning model with the adaptive stage of the machine learning model having a respective target configuration, wherein the configuration of the frozen stage is the same for all of the candidate nodes, and the parameters of the frozen stage are the same for all of the candidate nodes.

26. Method comprising:receiving, by a first candidate node of plural candidate nodes, an instruction to train a machine learning model comprising an adaptive stage having a target configuration; determining a first value of a performance of the machine learning model with the adaptive stage having the target configuration being trained by the first candidate node; receiving, for each of the plural candidate nodes different from the first candidate node, a performance indication indicating the respective first value of the performance of the machine learning model; receiving, for each of the plural candidate nodes different from the first candidate node, a configuration information informing on a respective target configuration of the adaptive stage of the machine learning model trained by the respective candidate node; identifying an improved configuration of the adaptive stage, wherein the improved configuration is the target configuration of the adaptive stage of the machine learning model trained by a top candidate node among the plural candidate nodes, and the first value of the performance for the top candidate node is better than the first value of the performance for the one or more candidate nodes different from the top candidate node; instructing at least one of the candidate nodes to train and run inference from the machine learning model with the adaptive stage having the improved configuration.

27. The method according to claim 26, further comprising informing each of the plural candidate nodes different from the first candidate node on the first value of the performance of the machine learning model with the adaptive stage having the target configuration trained by the first candidate node and on the target configuration.

28. The method according to any of claims 26 to 27, further comprising determining a second value of the performance of the machine learning model with the adaptive stage having the improved configuration trained by the first candidate node; inhibiting the first candidate node to run inference from the machine learning model with the adaptive stage having the improved configuration in response to determining that the second value of the performance of the machine learning model with the adaptive stage having the improved configuration is poorer than a predefined first threshold.

29. The method according to any of claims 26 to 28, wherein the machine learning model comprises a frozen stage and the adaptive stage, and the configuration of the frozen stage is the same for all the candidate nodes and the parameters of the frozen stage are the same for all the candidate nodes.

30. The method according to any of claims 26 to 29, further comprising instructing all the candidate nodes to train and run inference from the machine learning model such that the adaptive stage has the improved configuration.31 . A computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to carry out the method according to any of claims 17 to 30.

32. The computer program product according to claim 31 , embodied as a computer-readable medium or directly loadable into a computer.

Citation Information

Patent Citations

  • Information reporting method, apparatus and device, and storage medium

    US20220342713A1