Improving ai / ml inference reliability

Abstract

Embodiments of the present disclosure relate to improving AI / ML inference reliability. A solution for delivering AI explainability information over a radio interface in a cellular network is provided. According to one aspect, a method comprises: transmitting, by a user equipment, capability information to a network node of a cellular network, the capability information indicating a capability of the user equipment to explain an output of an artificial intelligence or machine learning, AI / ML, operation; receiving, from the network node, a configuration for reporting the explanation of the output; transmitting, by the user equipment, a message comprising an information element indicating an approval or a rejection of the configuration to the network node; and based on the approval of the configuration, performing, by the user equipment: performing the AI / ML operation, and reporting, to the network node, the output of the AI / ML operation and, based on the configuration and the AI / ML operation, reporting, to the network node, an explanation for the output.

Description

Technical Field

[0001] Various example embodiments generally relate to the reliability or trustworthiness of artificial intelligence / machine learning (AI / ML) implemented in a user equipment (UE) of a wireless network to improve the functionality of connectivity between the UE and the access node of the wireless network. Background Technology

[0002] Explainable Artificial Intelligence (XAI) has been developed for testing and validating the operation of artificial intelligence / machine learning (AI / ML) models. 3GPP has also studied XAI. Summary of the Invention

[0003] The subject matter of the independent claims is provided according to several aspects. Additional aspects are defined in the dependent claims. Embodiments that do not fall within the scope of the claims will be interpreted as examples useful for understanding this disclosure. Attached Figure Description

[0004] The invention will now be described in more detail with reference to embodiments and accompanying drawings, in which: Figure 1 An example of a communication network to which the examples disclosed herein can be applied is shown; Figure 2 An example of a signaling diagram illustrating a process according to some embodiments is shown; Figure 3 An embodiment of a signaling diagram illustrating a process according to some embodiments is described; Figure 4 It shows Figure 2 Another embodiment of the process; Figure 5 The report includes a detailed signaling diagram illustrating the interpretability information of the AI / ML output for the AI / ML model; Figure 6 A flowchart illustrating the process of using auxiliary information when validating AI / ML interpretability; Figure 7 A flowchart illustrating the process for performing reconfiguration actions based on AI / ML interpretability information reported by the user device; and Figure 8 and Figure 9 Some embodiments of apparatus configured to perform the processes described herein are illustrated. Detailed Implementation

[0005] The following embodiments are exemplary. Although the specification may refer to "a," "an," or "some" embodiments in several places in the text, this does not necessarily mean that each reference is made for the same embodiment(s), or that a particular feature applies only to a single embodiment. Individual features of different embodiments may also be combined to provide other embodiments. Furthermore, when a particular feature, structure, or characteristic is described in connection with an embodiment, whether explicitly described or not, the application of such a feature, structure, or characteristic in connection with other embodiments is within the knowledge of those skilled in the art. It should be understood that although the terms "first," "second," etc., may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another.

[0006] For the purposes of this disclosure, the phrases “at least one of A or B,” “at least one of A and B,” and “A and / or B” mean (A), (B), or (A and B). For the purposes of this disclosure, the phrases “A, B, and / or C” mean (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C).

[0007] The embodiments described herein are designed for next-generation cellular communication systems, such as 6G, but they can be implemented in other cellular networks or more generally in other wireless networks, such as those employing any of the following radio access technologies (RATs): Long Term Evolution (LTE), Advanced LTE and Enhanced LTE (eLTE), and 5G (also known as NR). Furthermore, communication within the communication network can utilize any suitable wireless communication technology, including but not limited to: Code Division Multiple Access (CDMA), Frequency Division Multiple Access (FDMA), Time Division Multiple Access (TDMA), Frequency Division Duplex (FDD), Time Division Duplex (TDD), Multiple Input Multiple Output (MIMO), Orthogonal Frequency Division Multiple Access (OFDM), and / or Discrete Fourier Transform Extended (DFT-s-OFDM).

[0008] As used herein, the term "network device" or "network node" refers to a node in a communication network through which a user equipment (UE) can access the network and / or control radio communications and manage radio resources within a cell. A network node or network device may be referred to as a base station (BS), access point (AP), or access node. Depending on the technology applied, a network device may be, for example, a Node B (NodeB or NB), an evolved Node B (eNodeB or eNB), an NR NB (also known as a gNB), a remote radio unit (RRU), a radio headend (RH), a remote radio headend (RRH), a relay, an integrated access and backhaul (IAB) node, a low-power node, a non-terrestrial network (NTN) or non-terrestrial network equipment (such as satellite network equipment, low Earth orbit (LEO) satellites, and geostationary Earth orbit (GEO) satellites), or an aircraft network device.

[0009] Furthermore, for split radio access networks (RANs), network equipment can refer to a centralized unit (CU) and / or a distributed unit (DU) of a base station. The interface between the CU and the DU may be referred to as the F1 interface in NR. In a split RAN architecture, node operations can be performed at least partially in a central / centralized unit (CU, e.g., a server, host, or node) that is operatively coupled to a DU (e.g., a radio head / node). A CU can control one or more DUs, at least acting as a transmit / receive (Tx / Rx) node. In some embodiments, a DU may include, for example, a Radio Link Control (RLC), a Media Access Control (MAC) layer, and a Physical (PHY) layer, while a CU may include layers above the RLC layer, such as the Packet Data Convergence Protocol (PDCP) layer, Radio Resource Control (RRC), and Internet Protocol (IP) layer. Other functional splitting is also possible. In practice, any processing task can be performed in a CU or a DU, and the boundaries of responsibility transfer between the CU and the DU can depend on the implementation applied.

[0010] In some embodiments, network devices or network nodes may be deeper into the network and do not necessarily provide radio access to the network to user equipment. Any network device that manages non-access stratum (NAS) functions can form such a network device. Network devices in the core network are other examples of such network devices.

[0011] The term "terminal device" refers to any terminal device capable of wireless communication. For example, a terminal device can be referred to as a communication device, user equipment (UE), subscriber station (SS), or mobile station (MS). Terminal devices can include mobile phones, cellular phones, smartphones, Voice over IP (VoIP) phones, wireless local loop phones, tablets, wearable terminal devices, personal digital assistants (PDAs), portable computers, desktop computers, image capture terminal devices (such as digital cameras), gaming terminal devices, music storage and recycle bins, in-vehicle wireless terminal devices, USB dongles, Internet of Things (IoT) devices, watches or other wearable devices, head-mounted displays (HMDs), vehicles, drones, medical devices and applications (e.g., remote surgery), industrial devices and applications (e.g., robots and / or other wireless devices operating in industrial and / or automated processing chain environments), consumer electronics devices, devices operating on commercial and / or industrial wireless networks, and so on.

[0012] Figure 1 Examples of communication networks to which the examples disclosed herein can be applied are illustrated. A communication network or cellular communication network may include a network node 110 providing one or more cells (such as cell 100) and a network node 112 providing one or more other cells (such as cell 102). For example, each cell may be a macrocell, microcell, femtocell, or picocell. A cell may define the coverage area or service area of ​​a corresponding access node.

[0013] Network node 110 can provide radio access to a communication network to user equipment (UE) 120 (one or more UEs). Radio access may include downlink (DL) communication from the network node to UE 120 and uplink (UL) communication from UE 120 to the network node. Examples of uplink channels include the Physical Uplink Control Channel (PUCCH) for transmitting control information and the Physical Uplink Shared Channel (PUSCH) for transmitting data to the network. Examples of downlink channels include the Physical Downlink Control Channel (PDCCH) for transmitting control information and the Physical Downlink Shared Channel (PDSCH) for transmitting data to the user equipment.

[0014] There can be multiple UEs 120 and 122 in the system. Each of them can be served by the same or different network nodes 110 and 112. UEs can be configured with dual connectivity (DC), where a UE (e.g., UE 120) can connect to multiple network nodes 110 and 112. UEs 120 and 122 can communicate with each other when a device-to-device (D2D) communication interface is established between them via a so-called side link (SL). For example, such D2D communication can be referred to as machine-to-machine, peer-to-peer (P2P) communication, or vehicle-to-vehicle (V2V) communication.

[0015] In a communication network with multiple network nodes, these nodes can be connected to each other via an interface. The LTE specification refers to this interface as the X2 interface. The interface between an LTE node and a 5G node, or between two 5G nodes, can be called the Xn interface.

[0016] Network nodes 110 and 112 can also connect to the core network 116 of the communication network via another interface. The LTE specification designates the core network as the Evolved Packet Core (EPC), and the core network may include, for example, a Mobility Management Entity (MME) and gateway nodes. The MME can handle the mobility of terminal devices in a tracking area containing multiple cells and handle signaling connections between the terminal devices and the core network. Gateway nodes can handle data routing in the core network to / from the terminal devices. The 5G specification designates the core network as the 5G Core (5GC). The 5G Core may include, for example, Access and Mobility Management Functions (AMF) and User Plane Functions / Gateways (UPF), and other functions. The AMF can handle the termination of Non-Access Stratum (NAS) signaling, NAS encryption and integrity protection, registration management, connection management, mobility management, access authentication and authorization, and security context management. For example, a UPF node can support packet routing and forwarding, packet inspection, and Quality of Service (QoS) processing.

[0017] Artificial intelligence (AI) can be broadly defined as enabling computers to perform tasks that mimic the human brain. Machine learning (ML) is a class of AI technologies: computer algorithms that can automatically improve their performance without explicit programming. AI algorithms were first conceived in the 1950s, but only in recent years has AI / ML become available for a wide range of real-world applications, thanks in part to advancements in computing power and the capacity to store data.

[0018] AI / ML can help adjust and optimize radio access network (RAN) parameters and settings by using (real-time) monitoring and prediction of network performance, quality, and demand. Additionally, AI / ML can identify and diagnose degradations in network performance and provide protection against network attacks. AI / ML can be used for energy saving, load balancing, mobility optimization, link adaptation, and security, to name just a few.

[0019] It is conceivable that AI / ML will enable real-time analytics and automated operation and control both within and outside of 5G RAN. This requires the availability of data streaming from wireless devices in a timely manner, especially in extremely time-critical applications such as real-time video surveillance and extended reality (XR). This can be reflected in network architecture, such as by placing and moving ML agents to desired locations within the network, for example, for data collection. User equipment (mobile devices) can assist network decision-making in resource management, thus acting as an infrastructure resource.

[0020] As networks evolve to programmable and flexible cloud-native implementations, AI / ML-based network automation will be used to simplify network management and optimization. It is anticipated that parts of the air interface, particularly signal processing algorithms, will be supported and eventually even replaced using machine learning models. Therefore, 6G wireless communication standards will natively support AI-based air interfaces.

[0021] Machine learning algorithms are generally categorized into four different types: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. In supervised learning, the algorithm learns from labeled data. For training, the algorithm receives input data and corresponding correct output labels. The algorithm is trained to predict accurate labels for new data. In unsupervised learning, the algorithm analyzes unlabeled data. The goal is to discover patterns, relationships, or structures within the data; for example, unsupervised learning algorithms form groups of similar data points. Semi-supervised learning is a hybrid machine learning approach that combines labeled and unlabeled data for training. A limited amount of labeled data and a larger set of unlabeled data are used to improve training. This approach is useful when acquiring labeled data is expensive or time-consuming, as is the case in many real-world applications. Semi-supervised learning techniques can be applied to a variety of tasks, such as classification, regression, and anomaly detection, allowing models to make more accurate predictions and generalize better in real-world scenarios. Reinforcement learning is a machine learning algorithm that learns from trial and error. ML agents interact with their environment and learn from experiences aimed at maximizing cumulative rewards. ML agents receive feedback based on their actions through rewards or penalties. The agent learns to take actions that lead to the most favorable outcome over time. The algorithm adapts to changing environments and achieves long-term goals through a series of actions.

[0022] An example of a machine learning algorithm that can be applied to tuning and optimizing radio access network (RAN) parameters and settings is deep learning. Deep learning is a subset of machine learning algorithms that use neural networks. Neural networks are also known as artificial neural networks (ANNs) or simulated neural networks (SNNs). Deep learning can be based on supervised, semi-supervised, or unsupervised learning.

[0023] Artificial neural networks (ANNs) consist of an input layer, one or more hidden layers, and an output layer. Each node in a layer, or artificial neuron, is connected to each other and has associated weights and a threshold. If the output of a single node exceeds the threshold assigned to that node, the node is activated and sends or passes data to the next layer of the neural network.

[0024] When supervised learning is applied to the training of a neural network, training is performed using examples, each containing a known "input" and "output," thus establishing probabilistically weighted associations between them. Training involves determining the difference between the neural network's output (prediction) for the input and the target output for the same input. This difference is called the error value. The neural network then adjusts its weighted associations according to a learning rule and using this error value. Continuous adjustments cause the neural network to produce an output that closely approximates the target output. After a sufficient number of these adjustments, training can be terminated based on one or more criteria.

[0025] 3GPP has begun studying AI / ML use cases in radio access networks in Release 18, and work continues in Release 19. The following use cases are considered: enhanced Channel State Information (CSI) estimation and feedback; enhanced beam management (beam prediction and beam selection in the temporal and / or spatial domains); and enhanced positioning (e.g., under non-line-of-sight conditions). The use cases are then broken down into several functions that enable them. Functions may include, for example, using a specific reference signal configuration for measurement, performing measurements on the reference signal configuration, and measuring based on the measurement output. These functions can be applied to all three use cases listed above. To implement these functions, AI / ML models can be used. For example, AI / ML models (AI / ML algorithms) can be used to interpret measurement data and output metrics as AI / ML inference outputs. The metrics can then be used in use cases such as UE positioning, beam selection, or CSI feedback.

[0026] One or more AI / ML models can exist for each function. Lifecycle management of AI / ML models is studied within 3GPP, and it can be used to control whether AI / ML models are enabled or disabled to output inference metrics for functions and use cases. Lifecycle management typically refers to controlling the operation of AI / ML models for functions within a use case. The lifecycle can include the following states: training the AI / ML model, testing the AI / ML model, and applying the AI / ML model to functions within the use case. Lifecycle management can then control the switching between states. Some embodiments described herein focus on the UE executing the AI / ML model and the network device controlling lifecycle management. Therefore, the network device can signal lifecycle management commands to the UE to execute state changes. The UE can also monitor the performance of the AI / ML model and perform lifecycle management. Depending on the embodiment, state changes can be entirely under the control of the network device, or the UE can propose state changes, or both the UE and the network device can perform and determine lifecycle changes.

[0027] As is known in the art, AI / ML models have great potential to improve the performance of cellular networks and connectivity. However, undesirable or unstable behavior is also associated with AI / ML models. From a network performance perspective, distinguishing such behavior is important. Therefore, the lifecycle of the AI / ML model can then be controlled. Another consequence could be reconfiguring the AI / ML model, such as debugging potential errors in the AI / ML model. Inference outputs can be analyzed, and unstable behavior can be distinguished by differentiating erroneous inference outputs. However, it may be difficult to directly distinguish erroneous outputs without knowing the basis for the inference outputs (e.g., the measurement data used as the basis for inference). Transmitting this measurement data between the UE and network equipment via the radio interface may be infeasible in terms of signaling overhead. Therefore, there is a need to improve the reliability of AI / ML models to improve the performance of networks and connectivity within them.

[0028] AI / ML interpretability and interpretable AI / interpretable ML are terms used in the literature to characterize the logical reasoning behind inferences made by AI / ML models. Interpretability can be understood as providing a quantified layer of additional information about how an AI / ML model arrives at its inference output. AI / ML interpretability can be categorized into two types: local interpretability and global interpretability. Local interpretability aims to explain individual inferences made by the AI / ML model; that is, it focuses on explaining why a particular inference was made by the AI / ML model for a particular input. Global interpretability aims to explain the behavior of the AI / ML model more generally; that is, it focuses on explaining how the AI / ML model works generally on several / all possible inputs. The embodiments described herein focus on local interpretability because the purpose is network-assisted performance monitoring of the AI / ML model(s) performed by the UE; however, the embodiments can be equally applied to situations where interpretability is quantified as global interpretability.

[0029] Regarding the quantification of interpretability, the literature provides certain quantitative measures, such as: Feature importance diffusion measures the range or variability of the importance of one or more features in an AI / ML model. High diffusion indicates that some features are much more important than others, while low diffusion suggests that feature importance is more evenly distributed. Feature importance stability assesses the consistency of the importance of a feature or a set of features across different runs or variations of the model. High stability means that the importance of the same feature is consistent, while low stability indicates that the importance of the feature changes significantly. The prediction set comparison evaluates the differences in model predictions across different datasets. It helps identify whether AI / ML models perform differently on different subsets of data. Algebraic feature importance combines feature importance with a regularization parameter (α) to balance the trade-off between model complexity and feature importance. This helps in selecting features that are both important and contribute to a simpler model; The interpretability and usability score measures how easy it is to understand and interpret the predictions of an AI / ML model. Higher scores indicate that the model's behavior is more transparent and easier to interpret; The surrogacy efficacy score assesses how well a simpler surrogacy model (such as a decision tree) can mimic the predictions of a more complex model. A high surrogacy efficacy score means that the surrogacy model provides a good approximation of the behavior of a complex model; Causality measures describe the extent to which the explanation reflects the causal relationships in the data / model; The interaction index indicates the extent to which interactions between features that contribute to AI / ML model decision-making are captured.

[0030] The embodiments described below employ feature importance diffusion and / or feature importance stability. However, as technology and standardization evolve, other metrics may be employed. The embodiments described herein focus on how to perform quantified configurations (e.g., multiple interpretability metrics) such that interpretability information can be delivered to the network across the radio interface in a manner understandable to both the UE and the network. Therefore, the protocols and devices described herein can be configured to deliver one or more of the other metrics listed above.

[0031] Figure 2 An embodiment of a process for providing AI / ML interpretability information from UE 120 to a network node is illustrated. The network node may be a node of the core network 116, such as a Location Management Function (LMF) or Authentication Management Function (AMF), or a node of a radio access network, such as an access node (gNB, DU, CU) or an Operation, Management and Maintenance (OAM) node. These terms conform to 3GPP specifications. According to one aspect, UE 120 or a device suitable for UE performs the following: transmitting (200) capability information instructing the network node of the cellular network that UE 120 can interpret the output of an AI / ML operation; receiving (202) a configuration from the network node for reporting the interpretation of the output; performing (204) the AI / ML operation; and reporting (208) the output of the AI / ML operation to the network node, and reporting an interpretation of the output to the network node based on the configuration and the AI / ML operation.

[0032] According to another aspect, the network node performs the following: receiving (200) capability information from UE 120 instructing the UE to interpret the output of the AI / ML operation; transmitting (202) a configuration for reporting the interpretation of the output to UE 120; receiving (208) the output of the AI / ML operation from UE 120, and receiving an interpretation for the output based on the configuration and the AI / ML operation; and performing (210) a reconfiguration action based on the received output and interpretation.

[0033] The aforementioned processes synergistically provide the technical effect of generating interpretability information to explain the output of AI / ML operations performed in UE 120 and delivering this interpretability information to network nodes to provide an additional information layer for reconfiguration actions performed in 210. Therefore, reconfiguration actions are improved through more available information. This improves the overall performance of the radio connection between UE 120 and the network when reconfiguration actions are performed on the radio connection, and / or the network performance when reconfiguration actions are performed on the network.

[0034] In an embodiment, the AI / ML operation performed in block 204 is one of monitoring the performance of an AI / ML model or AI / ML function, or performing AI / ML inference 206. As is known in the art, monitoring and inference are states in the lifecycle of an AI / ML model. After training an AI / ML model, the performance (output) of the AI / ML model can be monitored using real inputs (without applying the outputs to practice) until the performance is determined to be satisfactory, allowing the state to be changed to inference using the output of the AI / ML model in practice. Similarly, an AI / ML function can be monitored via the output of the AI / ML model. For example, a particular configuration (such as a reference signal configuration) in an AI / ML function for an AI / ML use case may be determined to be suboptimal, which can be monitored and eventually reconfigured.

[0035] Figure 3 The process according to some embodiments is described. Like... Figure 2 The same as in the middle, Figure 3 The signaling diagram illustrates two interrelated processes, one executed by UE 120 and the other by a network node. The network node can be a combination of the above... Figure 2 Any network node described. In Figure 3 In, with Figure 2 The same reference numerals in the figures indicate the same or substantially similar functions or actions.

[0036] refer to Figure 3 The process performed by UE 120 includes: transmitting (200) capability information to the network node indicating the ability of UE 120 to interpret the output of an AI / ML operation; receiving (202) a configuration for reporting the interpretation of the output from the network node; transmitting (300) a message to the network node including an information element indicating approval or rejection of the configuration; and, based on approval of the configuration, performing the following operations: performing (204) the AI / ML operation, and reporting (208) the output of the AI / ML operation to the network node, and, based on the configuration and the AI / ML operation, reporting the interpretation of the output.

[0037] Similarly, the network node performs the following process: receiving (200) from UE 120 an instruction to the user equipment to interpret the output of the AI / ML operation; transmitting (202) to the user equipment a configuration for reporting the interpretation of the output; receiving (300) from the user equipment a message including an information element indicating approval or rejection of the configuration; based on the user equipment's approval of the configuration, performing the following operations: receiving (208) from UE 120 the output of the AI / ML operation, and receiving an interpretation of the output based on the configuration and the AI / ML operation; and performing (210) a reconfiguration action based on the received output and interpretation.

[0038] Figure 3 The embodiments provided by this example offer the advantage of UE 120 confirming in step 202 whether it accepts the configuration offered by the network node. Even if the configuration conforms to the capabilities exchanged in step 200, there are various reasons why the configuration might be unacceptable to the UE. For example, the capabilities might not provide a complete picture of the UE's capabilities. UE 120 might be configured not to expose all its capabilities to the network node, for example, due to confidentiality. Alternatively, the network node might consider various other parameters of the network and / or other UEs when selecting a configuration. Therefore, the configuration might be unacceptable to UE 120. Allowing UE 120 to confirm whether the configuration is acceptable thus improves the performance of both UE 120 and the network.

[0039] Let us now describe the configuration through various embodiments below. In these embodiments, the configuration specifies how the user equipment should quantize the interpretation for the output. Since radio interfaces are typically capacity-limited, the UE 120 may not have sufficient capacity to freely describe the interpretation; however, the interpretation must be quantized so that both the UE 120 and the network node have a common understanding of the meaning of the transmitted interpretation bits(s). For this purpose, the configuration may specify the quantization of the interpretation.

[0040] In this embodiment, the capability information indicates multiple quantization types supported by the user equipment for quantizing the interpretation of the output. Regarding this embodiment, the network node can select one of the indicated quantization types for configuration. The multiple quantization types can include at least one of feature importance dispersion or feature importance stability, where the feature refers to at least one characteristic of the radio channel between the user equipment and the cellular network. The at least one feature can include multiple features. Examples of features include the presence / absence of line-of-sight (a binary feature or a soft feature indicating the line-of-sight) between the UE 120 and the transmitter of the reference signal and the signal strength of the reference signal (e.g., reference signal received power or signal-to-interference-plus-noise ratio). Depending on the use case and the selected AI / ML function, the reference signal can be any one or more of a positioning reference signal, demodulation reference signal, primary reference signal, secondary reference signal, synchronization signal, channel state information (CSI) reference signal, or probe reference signal. For example, a positioning reference signal may be suitable for a positioning use case, and a particular positioning reference signal configuration can be selected as part of the AI / ML function. As another example, a CSI reference signal may be suitable for a CSI feedback use case. However, other reference signals that are not directly related to the use case can also be used to improve performance or for exploration to be performed by the AI / ML model.

[0041] Following the literature definitions of these two quantification types, feature importance distribution measures the range or variability of the importance of one or more selected features in an AI / ML model. High expansion indicates that some features are much more important than others, while low expansion indicates that feature importance is more evenly distributed. Feature importance stability measures the degree of consistency in feature importance across different runs or variations of the AI / ML model. High stability means that the importance of the same feature is consistent, while low stability indicates that feature importance varies significantly.

[0042] Let's now describe some examples of interpretability information. For the UE positioning use case, a local interpretation might indicate that the AI / ML model estimates the UE to be in a specific location (= AI / ML model output) because the reference signal strength value from transmitter #1 is high (indicating proximity to transmitter #1), and the reference signal strength value from TRP #2 is low (indicating greater distance from transmitter #2). If the local interpretation is contradictory—for example, transmitter #1 should provide a better signal strength than transmitter #2 at a specific location—it indicates that the AI / ML model is not performing optimally. On the other hand, if the interpretability information supports the output, it can be determined that the AI / ML model is performing as expected. Network nodes can store benchmark truth information, allowing them to verify (in)mismatches between the output and interpretability, such as the expected reference signal strength at different locations in the examples above. Benchmark truth information can be collected via measurements, historical data, or various other means. As another example related to beam selection, it can be expected that the AI / ML model will select based on the reference signal strength value, and as long as the interpretability indicates that the beam selection is based on the reference signal strength value, it can be determined that the AI / ML model is operating as expected. However, if the interpretation begins to indicate that another criterion dominates the reference signal strength value (e.g., the time of arrival value), it can be determined that the AI / ML model is not operating as expected and is resulting in poor beam selection. A similar approach can be applied to other use cases, generalizing to examples such that if interpretability indicates that the cause or condition leading to the AI / ML model's output does not match the expected cause or condition, the AI / ML model can undergo a state change. Similar to monitoring the AI / ML model to switch it to the inference state, if interpretability indicates appropriate logical inference and conditions for the AI / ML model's output, the AI / ML model can be switched to the inference state, as described in this document. Other use cases include radio link failure detection or prediction and switching.

[0043] exist Figure 3 In this embodiment, UE 120 is configured to determine whether UE 120 supports the configuration received in 202, and when it is determined that UE 120 supports the received configuration, transmits an information element (ACK) indicating approval of the received configuration in 300. On the other hand, when it is determined that UE 120 does not support the received configuration, UE 120 transmits an information element (NAK) indicating rejection of the received configuration in 300. In this embodiment, the capability information indicates only a subset of the capabilities of UE 120 for interpreting the output of AI / ML operations, and UE 120 is configured to determine whether to approve or reject the configuration based on the received configuration and at least one capability not indicated in the capability information. As mentioned above, UE 120 may not reveal all its capabilities to the network node, and therefore, a separate approval process may be beneficial.

[0044] In response to a transmission rejection, UE 120 may receive one of the following from the network node: a new configuration for reporting the interpretation of the output; or a message disabling the reporting of the interpretation of the output.

[0045] Figure 4 It shows Figure 3 An example of the process. With Figure 3 The same reference numerals in the accompanying drawings refer to the same function or action. Figure 4 The diagram illustrates the action breakdown resulting from a configuration received due to UE approval (ACK) and a configuration received due to rejection (NAK). As described above, approval can lead to the execution of steps 204, 208, and 210 (in some embodiments, 206 is also an embodiment of 204). When UE 120 rejects the configuration, the network node can perform... Figure 3 Option A) or Option B). Option A) or B) can be hardcoded to reject, or the network node can perform step 400, where the network node determines which of options A) and B) to execute. In option A), the network node transmits (402) a new configuration for reporting the interpretation of the output to the UE 120. Thus, this process can be repeated by the UE 120 evaluating the proposed configuration and approving or rejecting (404) the configuration. Step 400 may include one or more criteria for the network node to retry different configurations before selecting option B. For example, the network node can be configured to switch to option B immediately upon the first rejection by the UE 120. In another embodiment, the network node may attempt one or more alternative configurations that satisfy the capabilities of the UE 120 before switching to option B. In option B), the network node transmits (406) a command to the UE 120 to disable reporting of the interpretation of the output.

[0046] In an embodiment, the information element in step 300 indicates a configuration rejection, and the message exchanged in step 300 further indicates the reason for the rejection. The network node can then be configured to determine, based on the received rejection and reason, whether to transmit a new configuration for reporting the explanation of the output or a report disabling the explanation of the output to the user equipment. Some reasons may directly trigger option B), for example, if the UE reports low battery power as a reason for the rejection. Similarly, some reasons may trigger option A), for example, the UE indicating a mismatch between UE capabilities and the provided configuration as a reason.

[0047] Figure 5 It shows Figure 2 Detailed signaling diagrams of an embodiment of the process. Figure 5 The embodiments can be directly related to Figure 4The embodiments are combined, but for simplicity, a detailed description of the approval process is omitted.

[0048] A possible scenario for this process is as follows: UE 120 trains one or more AI / ML models for one or more AI / ML functions that support interpretability. Alternatively, the UE may obtain the AI / ML models(s) from a network node or from a server of a UE vendor outside the cellular network. The training of the AI / ML models is not disclosed herein, as it is beyond the focus of embodiments of the invention. As described above, UE 120 informs the network node of its ability to interpret the output of the AI / ML models. As described above, the supported capabilities may include or consist of local interpretability metrics.

[0049] In step 200, UE 120 may indicate support for interpretability for each configured AI / ML use case, AI / ML function, or even AI / ML model (where the network selects the AI / ML model that UE 120 should use). For example, indications provided in the table below may be provided.

[0050] In an embodiment, the UE may additionally indicate one or more AI / ML models that can be used for use cases and / or functions, and identify the AI / ML models(s) to the network nodes. In conjunction with each AI / ML model, the UE 120 may indicate the ability to support interpretability for the corresponding AI / ML model. For example, binary indications of the interpretability supported for each AI / ML model may be provided, as shown in the table below.

[0051] UE 120 may provide the information as part of the capability information in step 200, when a request is received from a network node, or in an unsolicited manner (e.g., whenever there is an update to the information).

[0052] Based on the capability information received in box 200 and optional other requirements (such as quality of service requirements for use cases), the network node configures AI / ML operations to the UE in box 500. This may include selecting and activating a particular AI / ML function and / or AI / ML model for inference or monitoring. The network node may prefer functions and / or AI / ML models that support interpretability rather than those without such support, or it may prefer functions or models that support specific interpretability metrics. As part of the AI / ML function configuration, the network node may also provide the necessary configuration for the indicated AI / ML operations. For example, the network node may configure and trigger the transmission of necessary reference signals according to the use case, and configure measurements and reporting as functions for the use case. As described above, in box 500, the network node may further select configurations for interpretability.

[0053] In step 502, the network node provides the UE 120 with a configuration for interpretable AI / ML operation. This configuration may include at least one, a subset of, or any combination of the following: AI / ML functionality for the use case and associated configurations, such as reference signal and measurement and reporting configurations, a selected AI / ML model, or a selected AI / ML interpretability configuration.

[0054] In an embodiment, the configuration also indicates availability or includes auxiliary information for providing the interpretation for the AI / ML output(s), and the UE 120 is configured to further generate an interpretation for the output based on the auxiliary information. In step 504, such auxiliary information is delivered to the UE 120. The auxiliary information may include benchmark truth information considered by the network node to be a benchmark truth. Examples of benchmark truth information and its use are described below. It should be noted that the benchmark truth is not necessarily absolutely true. As is known in the context of AI / ML, benchmark truth information may be based on measurements and / or other data collected by the network node from the radio environment, other network nodes, and / or other UEs. Therefore, benchmark truth information may be inaccurate or approximately absolutely true. Upon receiving the auxiliary information in step 504, the UE 120 may use the auxiliary information for the output of the AI / ML model and the determined interpretation(s) for the output.

[0055] Figure 5Two embodiments of AI / ML operations are illustrated: AI / ML model or function monitoring 520 and AI / ML inference 522. As described above, both monitoring and inference represent the lifecycle states of the AI / ML model selected for the AI / ML use case. In monitoring, the AI / ML model is decoupled from the use case, and its output is not directly used for the use case, such as localization or beam prediction. In 520, the UE therefore monitors the performance of the AI / ML model or function (or both) in 506, determines interpretive information for the output of the AI / ML model and / or function, and reports the performance (and / or output) along with the interpretives to the network node in 508.

[0056] In inference, the AI / ML model is attached to the use case, and the inference output is directly used for the use case, such as for locating UE120. In 522, UE120 generates / produces (multiple) inference outputs and (multiple) interpretability outputs in 510, and reports the outputs to the network node in 512.

[0057] In one embodiment, the UE is configured to report interpretability (multiple) at substantially the same rate, along with the outputs (multiple) interpretability outputs, in an unsolicited manner when the outputs and interpretability outputs become available. In another embodiment, the UE 120 is configured to log the interpretability outputs (multiple) along with the inference / monitoring outputs (multiple) and report them to the network at a later stage (e.g., periodically or at a certain event (such as a request from a network node), when link conditions reach a certain quality, or when the number of logged events reaches a certain value).

[0058] In an embodiment of the reconfiguration action in step 210, the network node determines, based on the received output and interpretation, the state of the AI / ML model and / or function used by the UE 120 to perform AI / ML operations (change 514), and transmits (516) commands to the UE 120 to change the state of the AI / ML model and / or function respectively. Therefore, interpretability can be used for the lifecycle management of the AI / ML model. All of the following are examples of reconfiguration actions related to lifecycle management: switching from one AI / ML model to another, activating an AI / ML model from monitoring to inference, deactivating an AI / ML model from inference to monitoring or training, pausing an AI / ML model, and falling back to a non-AI / ML model. Upon receiving the lifecycle management command from the network node in 516 and based on the reported output(s) and interpretation, the UE 120 can change the state of at least one of the AI / ML model or AI / ML function based on the command. In other words, the UE 120 can implement the necessary (multiple) actions to implement the command.

[0059] The explanation reported by UE 120 to the network node to interpret AI / ML operations can be a summary of multiple outputs of the AI / ML operations. For example, multiple outputs may be needed to explain feature importance diffusion or stability.

[0060] Let us then describe some embodiments for providing and using auxiliary information in the context of interpreting the output of AI / ML operations. Network nodes may include baseline truth information about the radio environment within a cell. Examples of such baseline truth information may include line-of-sight information, received signal strength information, and the number of transmitters, or subsets thereof, within the cell. The format of the line-of-sight information may be a metric indicating how many areas within the cell's coverage area are below line-of-sight, or a line-of-sight map indicating areas(s) with line-of-sight and / or areas(s) without line-of-sight. For example, the format of the signal strength information may include information about the range of received signal strength within the cell's coverage area or the expected received signal strength values ​​at different locations within the cell. The format of the number of transmitters may be a numerical value indicating the minimum number of transmitters at any location within the cell's coverage area. These are merely some examples of baseline truth information that can be collected by network nodes.

[0061] Figure 6 A flowchart is shown for a process used to evaluate interpretation quality and inherently assess the performance of AI / ML operations through UE120 using benchmark truth information. This process can be performed as part of the monitoring or inference of the AI / ML operation. (Reference) Figure 6In step 600, UE 120 may obtain baseline truth information from the network node, and in step 602, it may obtain the output(s) of the AI / ML operation and associated interpretation metrics. In step 604, UE 120 compares the interpretation of the output for the AI / ML operation with the baseline truth information. The content of the report in step 208 may then be based on the result of the comparison. In different embodiments, the content may be different or even opposite, and therefore steps 608 and 610 are described at this general level. In any case, different results of the comparison in step 606 may lead to different content in the report. In some cases, it may even be possible to exclude the report of the interpretation based on the comparison. For example, if the comparison indicates that the interpretation(s) of the output for the AI / ML model matches the baseline truth information, the result of the comparison may be reported, for example, the interpretation indicating high reliability for positioning, while the baseline truth information indicates good line-of-sight conditions in the cell. This is to indicate to the network node that the AI / ML action is being performed in the desired manner. In another embodiment, if a comparison indicates that the (multiple) interpretations of the output for the AI / ML model do not match the baseline truth information, the result of the comparison can be reported. In this case, UE 120 can report the mismatch between the output interpretation and the baseline truth information as an explanation. An example of such a mismatch is poor beam prediction performance when the baseline truth information indicates high signal strength around the cell coverage area. In this embodiment, in the case of a match, the reporting of the explanation can be omitted to implicitly indicate that the AI / ML operation is being performed in the desired manner. The (multiple) outputs of the AI / ML operation can still be reported to perform functions and use cases. In this embodiment, the comparison is not binary but can include values ​​describing the correlation between the explanation and the baseline truth information. Such values ​​can then be reported to the network node as an explanation.

[0062] In this embodiment, the assistance information is provided by the network node based on a request from the UE. Step 504 can be used to indicate the availability of the assistance information, allowing the UE 120 to request it.

[0063] Figure 7 The procedure for performing a reconfiguration action on a network node in response to multiple interpreted outputs reported by UE 120 is illustrated. In this embodiment, the reconfiguration action is the aforementioned lifecycle management action. (See reference...) Figure 7In step 700, the network node can receive a report from the UE 120 and analyze its contents. Based on this analysis, the network node can then determine in step 702 whether to command a state change for the AI / ML model or function reported by the UE 120. For example, if the current state of the AI / ML model is inference, and if the network node determines that the UE report indicates that the interpretation(s) for the AI / ML output do not match the baseline truth information, the network node can issue a command to pause or deactivate the inference state in step 704. On the other hand, if the current state is being monitored and under the same conditions, the network node can command the UE 120 to maintain the monitoring state or not perform any action (706), or even command the UE 120 to switch to the training state (704). Alternatively, if the current state is being monitored and if the report indicates that the interpretation matches the baseline truth information, the network node can transmit a command to the UE to enable the AI / ML model for AI / ML inference.

[0064] Figure 8 A block diagram of apparatus 10 is shown by way of example. Apparatus 10 includes, for example, at least one processor 12 and at least one memory 14 storing instructions 15, which, when executed by the at least one processor, cause apparatus 10 to perform at least one or more methods and any embodiments thereof as disclosed herein and performed by UE 120. In the example, the at least one memory and instructions (e.g., computer program code, software) are configured to utilize the at least one processor to cause apparatus 10 to perform one or more methods and any embodiments thereof as disclosed herein and performed by UE 120.

[0065] For example, device 10 is a terminal device, such as Figure 1 UE 120. As another example, the device is included in a terminal device, for example, as a chipset configured to control the terminal device. Device 10 can be made or configured to perform the functions performed by UE 120 and / or any one or more embodiments described in the embodiments. Figure 2 The method involves at least a portion thereof. For this purpose, processor(s) 12 may include modules 20 for implementing one or more configured AI / ML functions using one or more AI / ML models. If configured according to the embodiments described herein, the AI / ML model may be interpretable and is an XAI / ML model. Processors(s) 12 may also implement a protocol stack 22 according to 3GPP specifications, for example, for communicating with network nodes described herein. Protocol stack 22 may form at least a portion of a physical layer, a media access control layer, and higher layers such as a radio resource control layer. Any layer may be configured to deliver capability information and interpretability information, as described herein, and receive configuration from network nodes.

[0066] Device 10 includes a radio interface 16. The radio interface 16 can provide communication capabilities to device 10. The radio interface 16 may include a receiver configured to receive information according to at least one cellular or non-cellular standard. The radio interface 16 may include a transmitter configured to transmit information according to at least one cellular or non-cellular standard. The receiver may include more than one receiver. The transmitter may include more than one transmitter. The radio interface 16 may include a transceiver configured to receive and transmit information according to at least one cellular or non-cellular standard. The transceiver may include more than one transceiver.

[0067] Device 10 may include a user interface 18, which includes at least one of, for example, a keypad, microphone, touch display, monitor, speaker, etc. User interface 18 can be used to control the device by a user. User interface 18 may be external to device 10. For example, device 10 may be connected to another device, such as a computer, via a wireless or wired connection, and device 10 may be controlled by a user via the computer.

[0068] Figure 9 A block diagram of another device 30 is shown by way of example. Device 30 includes, for example, at least one processor 32 and at least one memory 34 storing instructions 35, which, when executed by the at least one processor, cause device 30 to perform at least one or more methods and any embodiments thereof as disclosed herein and performed by a network node. In the example, the at least one memory and instructions (e.g., computer program code, software) are configured to utilize the at least one processor to cause device 30 to perform one or more methods and any embodiments thereof as disclosed herein and performed by a network node.

[0069] The apparatus may include a communication interface 36 that implements a protocol stack and digital and analog components required to enable communication with a UE including UE 120. Depending on the implementation, for example, if the network node is an access node, the network node may communicate directly with UE 120, or if the network node is in the core network, the network node may communicate with UE 120, for example, via an access node via one or more other network nodes.

[0070] Device 30 can form the aforementioned network node. In another embodiment, device 30 is included in such a network node, for example, as a chipset configured to control a network node. Device 30 can be made or configured to... Figure 2The network node functions are performed in any one or more of the methods and / or embodiments described herein. For this purpose, processor(s) 32 may include module 40 for configuring AI / ML functions or AI / ML models in a UE (e.g., UE 120) controlled by the network node. If configured according to the embodiments described herein, the AI / ML model may be interpretable and is an XAI / ML model. The configuration may include providing an initial configuration based on received UE capability information according to step 202. The configuration may also include lifecycle management functions to modify the lifecycle of the AI / ML model. The configuration may include modifying AI / ML functions as described herein. Processor(s) 12 may also include an XAI / ML monitoring module 42 configured to receive the output of the XAI / ML model and associated interpretation(s), evaluate the interpretation(s) and determine whether the XAI / ML model is operating as desired, and output a corresponding LCM state change command to module 40 for configuration to the corresponding UE.

[0071] Processors 12, 32 may include circuitry, or be configured as one or more circuits configured to perform stages of the methods according to the exemplary embodiments described herein. As used herein, the term "circuit" may refer to one or more of the following: (a) a hardware circuit implementation only, such as an implementation in analog and / or digital circuitry only; and (b) a combination of hardware circuitry and software (where applicable): (i) a combination of (multiple) analog and / or digital hardware circuitry with software / firmware; and (ii) any portion of (multiple) hardware processors having software (including (multiple) digital signal processors, software, and (multiple) memories that work together to enable a device (such as a user equipment) to perform various functions); and (c) (multiple) hardware circuitry and / or (multiple) processors that require software (e.g., firmware) to operate, such as (multiple) microprocessors or portions thereof, but where the software may be absent when operation does not require it. This definition of "circuit" applies to all uses of the term herein, including in any claim. As another example, as used herein, the term "circuit" also encompasses implementations of hardware circuitry or processors (or processors) or a portion thereof and their accompanying software and / or firmware. For instance, where applicable to specific claim elements, the term "circuit" also encompasses baseband integrated circuits or processor integrated circuits for mobile devices or similar integrated circuits in servers, cellular network devices, or other computing or network devices.

[0072] Memory 14, 34 can be implemented using any suitable data storage technology. The memory may include a database for storing data. Memory 14, 34 may be at least partially external to device 10, 30, but device 10, 30 may access them.

[0073] Instructions 15 and 35 may be included in a computer-readable medium or a non-transitory computer-readable medium. As used herein, the term non-transitory refers to a limitation on the medium itself (i.e., tangible rather than signaling), rather than a limitation on the persistence of data storage (e.g., random access memory, RAM versus read-only memory, ROM).

[0074] In embodiments, at least some of the processes described herein may be performed by means including components for performing at least some of the processes. Components for performing the method steps disclosed herein may include software and / or hardware components of means 10, 30. For example, the at least one processor 12, 32, memory 14, 34, and computer program code form components for performing one or more methods disclosed herein and any embodiments thereof. As used herein, the term “component” should be interpreted in the singular (i.e., referring to a single element) or in the plural (i.e., referring to a combination of single elements). Thus, the term “component for [performing A, B, C]” should be interpreted to encompass means in which only one component for performing A, B, and C exists, or where separate components for performing A, B, and C exist, or where components for performing A, B, and C partially or completely overlap. Furthermore, the terms "component for performing A, component for performing B, component for performing C" should be interpreted as covering an apparatus in which only one component for performing A, B, and C exists, or in which separate components for performing A, B, and C exist, or in which some or all components for performing A, B, and C overlap.

[0075] The following is a list of some aspects of the present invention.

[0076] According to a first aspect, a user equipment is provided, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the user equipment to at least: transmit capability information to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation; receive from the network node a configuration for reporting the interpretation of the output; perform the AI / ML operation; report the output of the AI / ML operation to the network node, and report an interpretation of the output based on the configuration and the AI / ML operation.

[0077] In this embodiment, the configuration specifies how the user equipment should quantify the interpretation of the output.

[0078] In an embodiment, capability information indicates multiple quantization types supported by the user equipment for the interpretation of the output.

[0079] In an embodiment, multiple quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0080] In an embodiment, the configuration further indicates at least one of an AI / ML model or an AI / ML function, and the user equipment is configured to use the indicated AI / ML model or AI / ML function for AI / ML operations.

[0081] In one embodiment, the configuration also indicates availability or includes auxiliary information for providing the explanation, and wherein the user equipment is configured to further generate an explanation for the output based on the auxiliary information.

[0082] In an embodiment, the user equipment is also configured to perform one of the following: report the output together with an explanation for the output; and record and report the explanation in association with the output after reporting the output.

[0083] In an embodiment, the user equipment is further configured to: receive lifecycle management commands from a network node based on reported output and interpretation to change the state of at least one of an AI / ML model or AI / ML function configured to perform AI / ML operations; and change the state of the at least one of the AI / ML model or AI / ML function based on the commands.

[0084] In the embodiments, the explanation is a summary of multiple outputs of the AI / ML operation.

[0085] Still according to the first aspect, a network node for a cellular network is provided, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the network node to at least: receive capability information from a user equipment of the cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; transmit to the user equipment a configuration for reporting the interpretation of the output; receive the output of the AI / ML operations from the user equipment, and, based on the configuration and the AI / ML operations, receive an interpretation of the output; and perform a reconfiguration action based on the received output and interpretation.

[0086] In an embodiment, the reconfiguration action includes determining, based on the received output and interpretation, to change the state of the AI / ML model used by the user device to perform AI / ML operations, and transmitting a command to the user device to change the state of the AI / ML model.

[0087] In an embodiment, the network node is further configured to store the reference truth information on the cell where the user equipment is served, and to transmit the reference truth information as auxiliary information to the user equipment, wherein the received interpretation is based on the reference truth information.

[0088] In this embodiment, the reference truth information includes information about the radio environment in the cell.

[0089] Still according to the first aspect, a method is provided, comprising: transmitting capability information from a user equipment to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation; receiving from the network node a configuration for reporting the interpretation of the output; performing the AI / ML operation by the user equipment; reporting the output of the AI / ML operation to the network node by the user equipment, and reporting an interpretation of the output to the network node based on the configuration and the AI / ML operation.

[0090] In this embodiment, the configuration specifies how the user equipment should quantify the interpretation of the output.

[0091] In one embodiment, capability information indicates multiple quantization types supported by the user equipment for quantizing the interpretation of the output. In another embodiment, the multiple quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0092] In an embodiment, the configuration further indicates at least one of an AI / ML model or an AI / ML function, and the user equipment is configured to use the indicated AI / ML model or AI / ML function for AI / ML operations.

[0093] In one embodiment, the configuration also indicates availability or includes auxiliary information for providing the explanation, and wherein the user equipment is configured to further generate an explanation for the output based on the auxiliary information.

[0094] In one embodiment, the method further includes the user equipment: reporting the output and an explanation for the output together; and recording and reporting the explanation in association with the output after reporting the output.

[0095] In an embodiment, the method further includes the user equipment: receiving a lifecycle management command from a network node based on reported output and interpretation to change the state of at least one of an AI / ML model or AI / ML function configured to perform AI / ML operations; and changing the state of at least one of the AI / ML model or AI / ML function based on the command.

[0096] In the embodiments, the explanation is a summary of multiple outputs of the AI / ML operation.

[0097] Still according to the first aspect, a method is provided, comprising: receiving capability information from a user equipment in a cellular network by a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; transmitting a configuration for reporting the interpretation of the output to the user equipment by the network node; receiving the output of the AI / ML operations from the user equipment by the network node, and receiving an interpretation of the output based on the configuration and the AI / ML operations; and performing a reconfiguration action by the network node based on the received output and interpretation.

[0098] In an embodiment, the reconfiguration action includes determining, based on the received output and interpretation, to change the state of the AI / ML model used by the user device to perform AI / ML operations, and transmitting a command to the user device to change the state of the AI / ML model.

[0099] In one embodiment, the method further includes a network node storing reference truth information on the cell served by the user equipment, and transmitting the reference truth information as auxiliary information to the user equipment, wherein the received interpretation is based on the reference truth information. In another embodiment, the reference truth information includes information about the radio environment in the cell.

[0100] Still according to the first aspect, a computer program product is provided, embodied on a distribution medium that is computer readable and includes program instructions that, when executed by a device, cause the device to perform any of the methods according to the first aspect above.

[0101] According to a second aspect, a user equipment is provided, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the user equipment to at least: transmit capability information to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation; receive from the network node a configuration for reporting the interpretation of the output; transmit a message to the network node, the message including information elements indicating approval or rejection of the configuration; and, based on the approved configuration, perform the following operations: execute the AI / ML operation, and report the output of the AI / ML operation to the network node, and, based on the configuration and the AI / ML operation, report an interpretation of the output to the network node.

[0102] In this embodiment, the configuration specifies how the user equipment should quantify the interpretation of the output.

[0103] In one embodiment, capability information indicates multiple quantization types supported by the user equipment for quantizing the interpretation of the output. In another embodiment, the multiple quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0104] In one embodiment, the user equipment is configured to determine whether the user equipment supports the received configuration, and when it is determined that the user equipment supports the received configuration, to transmit an information element indicating approval of the received configuration.

[0105] In an embodiment, the user equipment is configured to: transmit an information element indicating rejection of the received configuration when it is determined that the user equipment does not support the received configuration; and in response to the transmission rejection, receive from a network node one of the following: a new configuration for reporting the interpretation of the output; or a message disabling the reporting of the interpretation of the output.

[0106] In this embodiment, the user equipment is configured to indicate the reason for the rejection in a message when an information element indicates a rejection of the configuration.

[0107] In one embodiment, the capability information indicates only a subset of the user equipment's capabilities to interpret the output of the AI / ML operation, and the user equipment is configured to determine whether to approve or reject the configuration based on the received configuration and at least one capability not indicated in the capability information.

[0108] Still according to the second aspect, a network node for a cellular network is provided, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the network node to at least: receive capability information from a user equipment of the cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; transmit to the user equipment a configuration for reporting the interpretation of the output; receive from the user equipment a message including information elements indicating approval or rejection of the configuration; and, based on the user equipment's approval of the configuration, perform the following operations: receive the output of the AI / ML operation from the user equipment and interpret the output based on the configuration and the AI / ML operation; and perform a reconfiguration action based on the received output and interpretation.

[0109] In one embodiment, the network node is configured to transmit a new configuration to the user equipment for reporting the explanation of the output based on the user equipment rejection configuration.

[0110] In one embodiment, the network node is configured to transmit a message to the user equipment disabling the reporting of the explanation of the output based on the user equipment's rejection configuration.

[0111] In an embodiment, the information element indicates a rejection of the configuration, and the message also indicates the reason for the rejection, wherein the network node is further configured to: determine whether to transmit a new configuration to the user equipment for reporting the interpretation of the output or disabling reporting of the interpretation of the output; and based on the determination, transmit to the user equipment a corresponding one of the following: a new configuration for reporting the interpretation of the output; or a message disabling reporting of the interpretation of the output.

[0112] Still according to the second aspect, a method is provided, comprising: transmitting capability information from a user equipment to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation; receiving from the network node a configuration for reporting the interpretation of the output; transmitting a message from the user equipment to the network node, the message including an information element indicating approval or rejection of the configuration; and, based on approval of the configuration, performing the following operations by the user equipment: performing the AI / ML operation, and reporting the output of the AI / ML operation to the network node, and reporting an interpretation of the output to the network node based on the configuration and the AI / ML operation.

[0113] In this embodiment, the configuration specifies how the user equipment should quantify the interpretation of the output.

[0114] In an embodiment, capability information indicates multiple quantization types of the interpretation of the output supported by the user equipment.

[0115] In an embodiment, multiple quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0116] In one embodiment, the method includes having the user equipment determine whether the user equipment supports the received configuration, and, upon determining that the user equipment supports the received configuration, transmitting an information element indicating approval of the received configuration. In another embodiment, the method further includes having the user equipment: upon determining that the user equipment does not support the received configuration, transmitting an information element indicating rejection of the received configuration; and, in response to the transmission rejection, receiving from a network node one of: a new configuration for reporting the interpretation of the output; or a message disabling the reporting of the interpretation of the output.

[0117] In one embodiment, when an information element indicates a configuration rejection, the user equipment indicates the reason for the rejection in the message.

[0118] In one embodiment, the capability information indicates only a subset of the capabilities of the user equipment to interpret the output of the AI / ML operation, and the user equipment determines whether to approve or reject the configuration based on the received configuration and based on at least one capability not indicated in the capability information.

[0119] Still according to the second aspect, a method is provided, comprising: receiving capability information from a user equipment in a cellular network by a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; transmitting a configuration for reporting the interpretation of the output to the user equipment by the network node; receiving a message from the user equipment by the network node, the message including information elements indicating approval or rejection of the configuration; based on the user equipment's approval of the configuration, performing the following actions by the network node: receiving the output of the AI / ML operations from the user equipment, and receiving an interpretation of the output based on the configuration and the AI / ML operations; and performing a reconfiguration action based on the received output and interpretation.

[0120] In one embodiment, the network node transmits a new configuration to the user equipment for reporting the explanation of the output based on the user equipment rejection configuration.

[0121] In one embodiment, the network node transmits a message to the user equipment that disables the reporting of the output based on the user equipment's rejection configuration.

[0122] In an embodiment, the information element indicates a rejection of the configuration, and the message also indicates the reason for the rejection, wherein the network node further performs the following operations: determining whether to transmit a new configuration for reporting the interpretation of the output to the user equipment, or whether to disable reporting of the interpretation of the output; and based on the determination, transmitting to the user equipment a message corresponding to one of the following: a new configuration for reporting the interpretation of the output; or a message disabling reporting of the interpretation of the output.

[0123] Still according to the second aspect, a computer program product is provided, embodied on a distribution medium that is computer readable and includes program instructions that, when executed by a device, cause the device to perform any of the methods described above in the second aspect.

[0124] According to a third aspect, a user equipment is provided, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the user equipment to at least: transmit capability information to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation; receive from the network node a configuration for reporting the interpretation of the output; perform the AI / ML operation; report the output of the AI / ML operation to the network node, and, based on the configuration and the AI / ML operation, report an interpretation of the output to the network node, wherein the AI / ML operation is one of monitoring the performance of an AI / ML model or AI / ML function, or AI / ML inference.

[0125] In this embodiment, the configuration specifies how the user equipment should quantify the interpretation of the output.

[0126] In an embodiment, capability information indicates multiple quantization types supported by the user equipment for quantizing the interpretation of the output.

[0127] In an embodiment, multiple quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0128] In one embodiment, the AI / ML operation involves monitoring the performance of an AI / ML model, and the user equipment is further configured to: receive auxiliary information from a network node for monitoring performance; monitor at least one of the output of the AI / ML model and an interpretation of the output based on the auxiliary information; and perform the reporting based on the monitoring. In another embodiment, the auxiliary information includes benchmark truth information, and the user equipment is configured to: compare the interpretation of the output of the AI / ML model with the benchmark truth information; and report the result of the comparison to a network element. In yet another embodiment, if the comparison indicates that the interpretation of the output of the AI / ML model matches the benchmark truth information, the user equipment is further configured to report the result of the comparison.

[0129] In this embodiment, the AI / ML operation is AI / ML inference performed for an AI / ML function configured between a network node and a user equipment, wherein the output is the inference output for the AI / ML function, and wherein the AI / ML function is used for one of the following use cases: user equipment location, beam prediction, beam selection, and channel state information feedback.

[0130] Still according to a third aspect, a network node for a cellular network is provided, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the network node to at least: receive capability information from a user equipment of the cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation, wherein the AI / ML operation is one of monitoring the performance of an AI / ML model or AI / ML inference; transmit to the user equipment a configuration for reporting the interpretation of the output; receive the output of the AI / ML operation from the user equipment, and receive an interpretation of the output based on the configuration and the AI / ML operation; and perform a reconfiguration action based on the received output and interpretation.

[0131] In this embodiment, the AI / ML operation is to monitor the performance of the AI / ML model, and the network node is also configured to store baseline truth information on the cell where the user equipment is served, and transmit the baseline truth information to the user equipment as auxiliary information for performance monitoring, wherein the received output and the interpretation of the output are received in the performance monitoring report.

[0132] In one embodiment, the report indicates whether the interpretation matches the baseline truth information, and the network node is configured to: transmit a message to the user device enabling the AI / ML model for AI / ML inference if the report indicates the interpretation matches the baseline truth information; and determine to maintain the performance monitoring if the report indicates the interpretation does not match the baseline truth information.

[0133] In this embodiment, the reference truth information includes information about the radio environment in the cell.

[0134] Still according to the third aspect, a method is provided, comprising: transmitting capability information from a user equipment to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation; receiving from the network node a configuration for reporting the interpretation of the output; performing the AI / ML operation by the user equipment; reporting the output of the AI / ML operation to the network node, and reporting an interpretation of the output based on the configuration and the AI / ML operation, wherein the AI / ML operation is one of monitoring the performance of an AI / ML model or AI / ML function, or AI / ML inference.

[0135] In this embodiment, the configuration specifies how the user equipment should quantify the interpretation of the output.

[0136] In one embodiment, capability information indicates multiple quantization types supported by the user equipment for quantizing the interpretation of the output. In another embodiment, the multiple quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0137] In one embodiment, the AI / ML operation is to monitor the performance of an AI / ML model, and the user device further performs the following: receiving auxiliary information from a network node for monitoring performance, and monitoring at least one of the output of the AI / ML model and an interpretation of the output of the AI / ML model based on the auxiliary information, and executing the report based on the monitoring.

[0138] In one embodiment, the auxiliary information includes baseline truth information, and the user device performs the following: comparing the interpretation of the AI / ML model's output with the baseline truth information; and reporting the result of the comparison to the network element.

[0139] In an embodiment, if the comparison indicates that the interpretation of the output of the AI / ML model matches the baseline truth information, the user equipment reports the result of the comparison.

[0140] In this embodiment, the AI / ML operation is AI / ML inference performed for an AI / ML function configured between a network node and a user equipment, wherein the output is the inference output for the AI / ML function, and wherein the AI / ML function is used for one of the following use cases: user equipment location, beam prediction, beam selection, and channel state information feedback.

[0141] Still according to the third aspect, a method is provided, comprising: receiving capability information from a user equipment in a cellular network by a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of an artificial intelligence or machine learning (AI / ML) operation, wherein the AI / ML operation is one of monitoring the performance of an AI / ML model or AI / ML inference; transmitting a configuration for reporting the interpretation of the output to the user equipment by the network node; receiving the output of the AI / ML operation from the user equipment by the network node, and receiving an interpretation of the output based on the configuration and the AI / ML operation; and performing a reconfiguration action by the network node based on the received output and interpretation.

[0142] In this embodiment, the AI / ML operation is to monitor the performance of the AI / ML model, and the network node also stores benchmark truth information on the cell where the user equipment is served, and transmits the benchmark truth information to the user equipment as auxiliary information for performance monitoring, wherein the received output and the interpretation of the output are received in the performance monitoring report.

[0143] In one embodiment, the report indicates whether the interpretation matches the baseline truth information, and wherein the network node performs the following: if the report indicates the interpretation matches the baseline truth information, it transmits a message to the user equipment to enable the AI / ML model for AI / ML inference; and if the report indicates the interpretation does not match the baseline truth information, it determines to maintain performance monitoring.

[0144] In this embodiment, the reference truth information includes information about the radio environment in the cell.

[0145] Still according to the third aspect, a computer program product is provided, which is embodied on a distribution medium that is computer readable and includes program instructions that, when executed by a device, cause the device to perform any of the methods described above.

[0146] Although the invention has been described above with reference to examples in conjunction with the accompanying drawings, it is clear that the invention is not limited thereto, but can be modified in several ways within the scope of the appended claims. Therefore, all words and expressions should be interpreted broadly, and they are intended to illustrate rather than limit the embodiments. It will be apparent to those skilled in the art that the inventive concept can be implemented in various ways as technology advances. Furthermore, it will be apparent to those skilled in the art that the described embodiments can, but are not required to, be combined with other embodiments in various ways.

[0147] Furthermore, the various implementations of this disclosure can be described with reference to the following terms, and their features can be combined in any reasonable manner.

[0148] Clause 1. A user equipment, comprising: at least one processor; and at least one memory, the at least one memory storing instructions that, when executed by the at least one processor, cause the user equipment to at least: transmit capability information to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; receive from the network node a configuration for reporting the interpretation of the output; transmit to the network node a message including information elements indicating approval or rejection of the configuration; and, based on the approved configuration, perform the following operations: perform the AI / ML operation, and report the output of the AI / ML operation to the network node, and, based on the configuration and the AI / ML operation, report the interpretation of the output to the network node.

[0149] Clause 2. User equipment as specified in Clause 1, wherein the configuration specifies how the user equipment should quantify the interpretation of the output.

[0150] Clause 3. User equipment pursuant to Clause 1 or 2, wherein capability information indicates multiple quantization types supported by the user equipment for quantifying the interpretation of the output.

[0151] Clause 4. User equipment pursuant to Clause 3, wherein multiple quantification types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

[0152] Clause 5. A user equipment pursuant to any of the foregoing clauses is configured to determine whether the user equipment supports a received configuration, and upon determining that the user equipment supports the received configuration, to transmit an information element indicating approval of the received configuration.

[0153] Clause 6. A user equipment pursuant to Clause 5 is configured to: transmit an information element indicating rejection of the received configuration when it is determined that the user equipment does not support the received configuration; and in response to the transmission rejection, receive from a network node one of the following: a new configuration for reporting the interpretation of the output; or a message disabling the reporting of the interpretation of the output.

[0154] Clause 7. User equipment pursuant to any of the foregoing clauses is configured to indicate the reason for the rejection in a message when an information element indicates a rejection of the configuration.

[0155] Clause 8. A user equipment pursuant to any of the foregoing clauses, wherein the capability information indicates only a subset of capabilities that the user equipment uses to interpret the output of AI / ML operations, and wherein the user equipment is configured to determine whether to approve or reject the configuration based on the received configuration and based on at least one capability not indicated in the capability information.

[0156] Clause 9. A network node of a cellular network, comprising: at least one processor; and at least one memory, the at least one memory storing instructions that, when executed by the at least one processor, cause the network node to at least: receive capability information from a user equipment of the cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; transmit a configuration to the user equipment for reporting the interpretation of the output; receive from the user equipment a message including information elements indicating approval or rejection of the configuration; and, based on the user equipment's approval of the configuration, perform the following actions: receive the output of the AI / ML operations from the user equipment, and, based on the configuration and the AI / ML operations, receive an interpretation of the output; and perform a reconfiguration action based on the received output and interpretation.

[0157] Clause 10. A network node pursuant to Clause 9 is configured to transmit a new configuration for interpreting the report output to the user equipment based on the user equipment rejecting the configuration.

[0158] Clause 11. A network node pursuant to Clause 9 is configured to transmit a message to the user equipment reporting that the interpretation of the output is disabled based on the user equipment's refusal to configure.

[0159] Clause 12. A network node pursuant to Clause 9, wherein the information element indicates a rejection of the configuration, and the message further indicates the reason for the rejection, and wherein the network node is further configured to: determine whether to transmit a new configuration for reporting the interpretation of the output to the user equipment or to disable reporting the interpretation of the output; and based on the determination, transmit to the user equipment a message that corresponds to: a new configuration for reporting the interpretation of the output; or a message that disables reporting the interpretation of the output.

[0160] Clause 13. A method for communication, comprising: transmitting capability information from a user equipment to a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; receiving from the network node a configuration for reporting the interpretation of the output; transmitting from the user equipment to the network node a message including information elements indicating approval or rejection of the configuration; and, based on the approved configuration, performing the following actions by the user equipment: performing the AI / ML operation, and reporting the output of the AI / ML operation to the network node, and, based on the configuration and the AI / ML operation, reporting an interpretation of the output to the network node.

[0161] Clause 14. The method of Clause 13, wherein the configuration specifies how the user equipment should quantify the interpretation of the output.

[0162] Clause 15. The method according to Clause 13 or 14, wherein the capability information indicates multiple quantization types supported by the user equipment for quantifying the interpretation of the output.

[0163] Clause 16. The method pursuant to Clause 15 includes the user equipment determining whether the user equipment supports the received configuration, and, upon determining that the user equipment supports the received configuration, transmitting an information element indicating approval of the received configuration.

[0164] Clause 17. The method according to Clause 16 includes the user equipment: upon determining that the user equipment does not support the received configuration, transmitting an information element indicating a rejection of the received configuration; and in response to the transmission rejection, receiving from the network node one of the following: a new configuration for reporting the interpretation of the output; or a message disabling the reporting of the interpretation of the output.

[0165] Clause 18. The method according to Clause 13 or 14, wherein when an information element indicates a rejection of the configuration, the user equipment indicates the reason for the rejection in the message.

[0166] Clause 19. The method according to Clause 13 or 14, wherein the capability information indicates only a subset of capabilities that the user equipment uses to interpret the output of AI / ML operations, and wherein the user equipment determines whether to approve or reject the configuration based on the received configuration and based on at least one capability not indicated in the capability information.

[0167] Clause 20. A method for communication, comprising: receiving capability information from a user equipment of a cellular network by a network node of a cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; transmitting a configuration for reporting the interpretation of the output to the user equipment by the network node; receiving from the user equipment a message including information elements indicating approval or rejection of the configuration; based on the user equipment's approval of the configuration, performing the following actions by the network node: receiving the output of the AI / ML operations from the user equipment, and receiving an interpretation of the output based on the configuration and the AI / ML operations; and performing a reconfiguration action based on the received output and interpretation.

[0168] Clause 21. The method according to Clause 20, wherein the network node transmits a new configuration to the user equipment for interpreting the report output based on the user equipment rejecting the configuration.

[0169] Clause 22. The method according to Clause 20, wherein the network node transmits a message to the user equipment reporting a disallowed interpretation of the output based on the user equipment's rejection configuration.

[0170] Clause 23. The method according to Clause 20, wherein the information element indicates a rejection of the configuration, and the message further indicates the reason for the rejection, and wherein the network node further performs the following: determining whether to transmit a new configuration for reporting the interpretation of the output to the user equipment, or to disable reporting the interpretation of the output; and based on the determination, transmitting to the user equipment a message of the corresponding one of the following: a new configuration for reporting the interpretation of the output; or a message disabling reporting the interpretation of the output.

[0171] Clause 24. A computer program product embodied on a distributed medium, the distributed medium being computer readable and including program instructions that, when executed by a device, cause the device to perform the method according to any one of Clauses 13 to 23.

Claims

1. A user equipment, comprising: At least one processor; as well as At least one memory, the at least one memory storing instructions, the instructions, when executed by the at least one processor, cause the user equipment to at least: Transmit capability information to network nodes in a cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; Receive configuration from the network node for reporting the interpretation of the output; Transmit to the network node a message including information elements indicating approval or rejection of the configuration; and Based on the approved configuration, perform the following operations: Perform the AI / ML operations, and The system reports the output of the AI / ML operation to the network node, and, based on the configuration and the AI / ML operation, reports an explanation of the output to the network node.

2. The user equipment of claim 1, wherein the configuration specifies how the user equipment should quantify the interpretation of the output.

3. The user equipment according to claim 1 or 2, wherein the capability information indicates a plurality of quantization types supported by the user equipment for quantizing the interpretation of the output.

4. The user equipment of claim 3, wherein the plurality of quantization types include at least one of feature importance dispersion or feature importance stability, wherein the feature is a characteristic of the radio channel between the user equipment and the cellular network.

5. The user equipment according to claim 1 or 2, configured to determine whether the user equipment supports the received configuration, and when it is determined that the user equipment supports the received configuration, to transmit the information element indicating approval of the received configuration.

6. The user equipment according to claim 5, configured as follows: When it is determined that the user equipment does not support the received configuration, the information element indicating the rejection of the received configuration is transmitted; and In response to the transmission of the rejection, one of the following is received from the network node: A new configuration for reporting the interpretation of the output; Disable the message in the report that interprets the output.

7. The user equipment according to claim 1 or 2, configured as follows: When the information element indicates a rejection of the configuration, the message indicates the reason for the rejection.

8. The user equipment of claim 1 or 2, wherein the capability information indicates only a subset of the capabilities of the user equipment for interpreting the output of the AI / ML operation, and wherein the user equipment is configured to determine whether to approve or reject the configuration based on receiving the configuration and based on at least one capability not indicated in the capability information.

9. A network node of a cellular network, comprising: At least one processor; as well as At least one memory, the at least one memory storing instructions, the instructions, when executed by the at least one processor, cause the network node to at least: Receive capability information from the user equipment in the cellular network, the capability information indicating the user equipment's ability to interpret the output of artificial intelligence or machine learning (AI / ML) operations; Transmit the configuration for reporting the interpretation of the output to the user equipment; Receive a message from the user equipment including information elements indicating approval or rejection of the configuration; Based on the user equipment approving the configuration, perform the following operations: Receive the output of the AI / ML operation from the user equipment, and receive an interpretation of the output based on the configuration and the AI / ML operation; as well as The reconfiguration action is performed based on the received output and the interpretation.

10. The network node of claim 9, configured to transmit a new configuration to the user equipment for reporting the interpretation of the output based on the user equipment rejecting the configuration.

Images