Bilateral model performance monitoring

Abstract

Aspects of the present disclosure relate to an apparatus for bilateral model performance monitoring. The apparatus (such as a UE) transmits encoded data to a second device (e.g., a gNB) in a bilateral model, where the encoded data is based at least on the input data and an encoder model of the apparatus, and the apparatus includes a first set of parameters including characterization information of an encoder of the bilateral model. The apparatus performs a calculation that generates at least one model metric associated with performance monitoring of the bilateral model, where the model metric is calculated based on a first set of information characterizing a decoder model of the second device and a second set of decoder parameters. The apparatus sends feedback data to the second device, wherein the feedback data includes the model metric.

Description

Cross Reference to Related Applications

[0001] This application claims priority to U.S. Provisional Application Serial No. 63 / 494,720, filed April 6, 2023, entitled “Two-Sided Model Performance Monitoring,” the entirety of which is incorporated herein by reference. This application also claims priority to U.S. Provisional Application Serial No. 63 / 494,722, filed April 6, 2023, entitled “Two-Sided Model Training,” the entirety of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present disclosure relates to wireless communications, and more specifically to two-sided models. BACKGROUND

[0003] A wireless communication system can include one or more network communication devices (such as base stations), which can be referred to as eNodeBs (eNBs), next generation NodeBs (gNBs), or other suitable terminology. Each network communication device (such as a base station) can support wireless communication for one or more user communication devices, which can also be referred to as user equipment (UE) or other suitable terminology. A wireless communication system can support wireless communication with one or more user communication devices by utilizing resources of the wireless communication system, such as time resources (e.g., symbols, time slots, subframes, frames, etc.) or frequency resources (e.g., subcarriers, carriers). In addition, a wireless communication system can support wireless communication across various radio access technologies, including third generation (3G) radio access technologies, fourth generation (4G) radio access technologies, fifth generation (5G) radio access technologies, and other suitable radio access technologies beyond 5G (e.g., sixth generation (6G)).

[0004] In a wireless communication system, a two-sided model includes a UE side and a network (e.g., gNB) side. A common technique for determining whether the two-sided model is performing properly includes comparing the output of the gNB side of the model to the input of the UE side of the model using a traditional feedback mechanism. However, even with the traditional feedback mechanism, feedback of the correct output of the gNB side of the model is not a simple task and the transmission incurs signaling overhead. SUMMARY

[0005] The present disclosure relates to methods, apparatuses, and systems that support bilateral model performance monitoring. In aspects of bilateral model performance modeling, the present disclosure describes details of techniques for monitoring performance of a bilateral model, such as by sending expected outputs to a second node for monitoring, using a local decoder model generated during separate training at a first node, using an updated local decoder model that is fine-tuned to better match an actual decoder of the second node (which also provides an iterative approach to improve quality of the local decoder model), and / or the first node constructs a local version of the actual decoder (at the second node) and uses it for model monitoring.

[0006] In further aspects of bilateral model training, the present disclosure describes details of techniques for training and updating a bilateral model. Notably, the described techniques improve performance of a trained bilateral model in an iterative manner. The described techniques compensate for performance loss of a bilateral model if there is a discrepancy between data used or assumed at a different first node or second node of the bilateral model and the model. The training techniques are implemented as first node prioritized and second node prioritized schemes. Moreover, there is no need to share an encoder model of the first node and a decoder model of the second node with each other, and there is no need to train the first node and / or the second node(s) simultaneously.

[0007] Notably, advantages of separate training include that the first node and the second node do not need to know the internal structure of the neural network (NN) modules of the other side. In one or more implementations of a bilateral model, the described techniques train and / or update the bilateral model in separate training cycles and eliminate inconsistencies in separate training of the bilateral model.

[0008] In some implementations of the methods and apparatuses described herein, a UE transmits encoded data to a second device (e.g., a gNB) in a bilateral model, where the encoded data is based at least on input data and an encoder model of the UE, and the UE includes a first set of parameters that includes characterization information of an encoder of the bilateral model. The UE performs a computation that generates at least one model metric associated with performance monitoring of the bilateral model, the at least one model metric computed based on a first set of information that characterizes a decoder model of the second device and a second set of decoder parameters. The UE transmits feedback data to the second device, where the feedback data includes the at least one model metric.

[0009] Some implementations of the methods and apparatuses described in this document can also include that the first set of information includes a set of channel data representations during a first time-frequency space region. The first set of parameters is received from the second device. The first set of parameters includes at least one of: a threshold value or scheduling information associated with transmitting feedback data; and the scheduling information includes a first indication of periodically or semi-periodically transmitting feedback data, and a second indication of a transmission interval. The computation of the at least one model metric is performed based on at least one of: an instruction received from the second device, an internal process event, or a scheduled periodic event. The first set of parameters is received via higher layer signaling as at least one of: a radio resource control (RRC) message or a medium access control control element (MAC-CE) message. The at least one model metric is based on the received threshold value. The UE receives, from at least one of the second device or an alternative device, a second set of decoder parameters characterizing a decoder model. The second set of decoder parameters is determined based on a second set of information including data samples, the data samples each representing an input and an expected output of the decoder model. The second set of decoder parameters is determined based on the data samples of the second set of information to minimize a difference between the expected output and an actual output of the decoder model. The second set of information is received from at least one of the second device or the alternative device. The second set of decoder parameters is determined during training of an encoder model. The second set of decoder parameters is updated based at least in part on update information received from at least one of the second device or the alternative device. The first set of information includes data samples, the data samples each representing: an input of an encoder of a bilateral model, and an expected output of a decoder model of the bilateral model. The at least one model metric is determined based on a comparison of the expected output of the data samples of the first set of information and an actual output of the bilateral model constructed by the encoder and the decoder based on the input of the data samples of the first set of information. The comparison is performed by finding a mean Euclidean distance, a generalized cosine similarity.

[0010] In some implementations of the methods and apparatuses described in this document, a gNB receives, from a first device (e.g., a UE), encoded data in a bilateral model, the encoded data encoded using an encoder model of the first device, the gNB including a first set of parameters including characterization information of a decoder of the bilateral model. The gNB transmits, to the first device, a first set of information, where the first set of information includes at least characterization information of a decoder model of the gNB. The gNB receives, from the first device, feedback data, and initiates an update procedure based at least in part on the feedback data.

[0011] Some implementations of the methods and apparatuses described in this document can further include that the first set of parameters includes at least the threshold. The first set of parameters includes scheduling information associated with receiving the feedback data. The first set of information includes an indication of a structure of the decoder model. The first set of information includes data samples each representing an input and an output of the decoder model. The process to initiate the update process is based on at least one of the feedback data, the output of the bilateral model, or the threshold. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 FIG. illustrates an example of a wireless communications system that supports bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0013] Figure 2 FIG. illustrates an example of a wireless network including a gNB (e.g., base station) and multiple UEs related to bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0014] Figure 3 FIG. illustrates an example of a high-level structure of a bilateral model related to bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0015] Figure 4 FIG. illustrates another example of a high-level structure of a bilateral model with multiple encoders related to bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0016] Figure 5a and Figure 5b FIG. illustrates an example of a channel state information (CSI) system with a UE subsystem and a network subsystem supporting operation of a bilateral model related to bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0017] Figure 6 FIG. illustrates an example of a bilateral model and training for performance monitoring related to bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0018] Figure 7 and Figure 8 FIG. illustrates an example of a block diagram of a device that supports bilateral model performance monitoring in accordance with aspects of the present disclosure.

[0019] Figures 9-11 FIG. illustrates a flow diagram of a method that supports bilateral model performance monitoring in accordance with aspects of the present disclosure. DETAILED DESCRIPTION

[0020] A wireless communication system includes a two-sided model with a UE side and a network (e.g., gNB) side. Common techniques to determine whether the two-sided model is performing properly include comparing the output of the gNB side of the model with the input of the UE side of the model using a traditional feedback mechanism. However, even with the traditional feedback mechanism, the feedback of the correct output of the gNB side of the model is not a simple task and the transmission incurs signaling overhead.

[0021] In aspects of two-sided model performance modeling, the present disclosure describes details of techniques for monitoring the performance of a two-sided model, such as by sending expected outputs to a second node for monitoring, using a local decoder model generated during separate training at a first node, using an updated local decoder model that is fine-tuned to better match an actual decoder of the second node (which also provides an iterative approach to improve the quality of the local decoder model), and / or the first node constructs a local version of the actual decoder (at the second node) and uses it for model monitoring.

[0022] Further, as described above, the parameters of the two-sided model are trained before the two-sided model is able to effectively feedback information to the gNB. The model is trained by being presented with a set of inputs and desired outputs (a training dataset), the model uses these datasets to find statistics of the inputs and input / output relationships and captures them in the parameters of the model. The trained model can be used as long as the statistics of the inputs and / or input / output relationships remain similar to the statistics presented to the model during training, or the model can be well generalized to new situations. In other aspects, the described techniques involve how the system (e.g., gNB and / or UE) of the two-sided model determines that the model is not performing properly. For example, the techniques account for the amount of data and signaling needed to implement the monitoring process.

[0023] In aspects of two-sided model training, the present disclosure describes details of techniques for training and updating a two-sided model. Notably, the described techniques improve the performance of the trained two-sided model in an iterative manner. If there is a discrepancy between the data used or assumed at a different first node or second node of the two-sided model and the model, the described techniques compensate for the loss in performance of the two-sided model. The training techniques are implemented as a first node-first and second node-first approach. Further, the encoder model of the first node and the decoder model of the second node do not need to be shared with each other and the first node and / or second node(s) do not need to be trained simultaneously.

[0024] In other aspects of bilateral model training, the described techniques provide for training and iteratively updating the bilateral model in separate training cycles. The described techniques target separate training and model updating where the NN modules of the first node (e.g., UE) and the second node (e.g., gNB) are trained in different training sessions without a forward or backward propagation path between the UE and the gNB. Notably, a benefit of separate training is that the first node and the second node do not need to know the internal structure of the NN module of the other side. In one or more implementations of the bilateral model, the described techniques train and / or update the bilateral model in separate training cycles and eliminate inconsistencies in separate training of the bilateral model.

[0025] Aspects of the disclosure are described in the context of a wireless communication system. Aspects of the disclosure are further illustrated by and described with reference to apparatus diagrams and flowcharts.

[0026] Figure 1 An example of a wireless communication system 100 that supports bilateral model performance monitoring is illustrated in accordance with aspects of the present disclosure. The wireless communication system 100 can include one or more network entities 102, one or more UEs 104, a core network 106, and a packet data network 108. The wireless communication system 100 can support various radio access technologies. In some implementations, the wireless communication system 100 can be a 4G network, such as an LTE network or a LTE-Advanced (LTE-A) network. In some other implementations, the wireless communication system 100 can be a 5G network, such as an NR network. In other implementations, the wireless communication system 100 can be a combination of 4G and 5G networks, or other suitable radio access technologies, including Institute of Electrical and Electronics Engineers (IEEE) 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802.20. The wireless communication system 100 can support radio access technologies other than 5G. Further, the wireless communication system 100 can support technologies such as time division multiple access (TDMA), frequency division multiple access (FDMA), or code division multiple access (CDMA).

[0027] The one or more network entities 102 can be dispersed throughout the geographic region to form the wireless communication system 100. One or more of the network entities 102 described herein can be or include or can be referred to as a network node, base station, network element, radio access network (RAN), base transceiver station, access point, NodeB, eNodeB (eNB), next generation NodeB (gNB), or other suitable terminology. The network entities (NEs) 102 and the UEs 104 can communicate via communication links 110, which can be wireless or wired connections. For example, the network entities 102 and the UEs 104 can perform wireless communication (e.g., receive signaling, transmit signaling) over a Uu interface.

[0028] The network entity 102 can provide a geographic coverage area 112 for which the network entity 102 supports service (e.g., voice, video, packet data, messaging, broadcast, etc.) to one or more UEs 104 within the geographic coverage area 112. For example, the network entity 102 and UE 104 can support wireless communication of signals associated with service (e.g., voice, video, packet data, messaging, broadcast, etc.) in accordance with one or more wireless access technologies. In some implementations, the network entity 102 can be mobile, such as a satellite (e.g., a non-terrestrial station (NTS)) associated with a non-terrestrial network. In some implementations, different geographic coverage areas 112 associated with the same or different radio access technologies can overlap, although different geographic coverage areas 112 can be associated with different network entities 102. Information and signals described herein can be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that can be referenced throughout the description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0029] The one or more UEs 104 can be dispersed throughout the geographic region of the wireless communication system 100. A UE 104 can include or can be referred to as a mobile device, wireless device, remote device, remote unit, handset, subscriber device, or some other suitable terminology. In some implementations, a UE 104 can be referred to as a unit, a station, a terminal, or a client, among other examples. Additionally or alternatively, a UE 104 can be referred to as an Internet of Things (IoT) device, an Internet of Everything (IoE) device, or a machine type communications (MTC) device, among other examples. In some implementations, a UE 104 can be a stationary unit. In some other implementations, a UE 104 can be a mobile unit.

[0030] The one or more UEs 104 can be devices in different forms or having different capabilities. Some examples of UEs 104 are illustrated in FIG. 1. As shown in FIG. 1, UEs 104 can include or can be referred to as a machine-to-machine (M2M) device, a machine type communications (MTC) device, an Internet of Things (IoT) device, an Internet of Everything (IoE) device, or some other suitable terminology. Figure 1 Figure 1 As shown, UEs 104 can be able to communicate with various types of devices, such as the network entity 102, other UEs 104, or network devices (e.g., core network 106, packet data network 108, relay devices, integrated access and backhaul (IAB) nodes, or another network device), as shown in FIG. 1. Additionally or alternatively, UEs 104 can support communication with other UEs 104 or network entities 102 that can act as relays in the wireless communication system 100. Figure 1

[0031] ​​The UEs 104 can be configured to connect directly to one another via a device-to-device (D2D) communication link 110. In some implementations, the D2D communication link 110 can be a cellular D2D communication link 110 that utilizes a licensed spectrum (e.g., compared to a wireless local area network (WLAN) communication link that utilizes an unlicensed spectrum). In some implementations, the D2D communication link 110 can be a sidelink communication link 110 that utilizes a shared spectrum (e.g., licensed, unlicensed, or international mobile telecommunications (IMT) frequency spectrum allowed to be used by the cellular industry). In some implementations, the D2D communication link 110 can be a vehicle-to-everything (V2X) communication link 110 that utilizes a shared spectrum (e.g., licensed, unlicensed, or IMT frequency spectrum allowed to be used by the cellular industry). In some implementations, the D2D communication link 110 can be a vehicle-to-vehicle (V2V) communication link 110, a vehicle-to-infrastructure (V2I) communication link 110, or a vehicle-to-pedestrian (V2P) communication link 110 that utilizes a shared spectrum (e.g., licensed, unlicensed, or IMT frequency spectrum allowed to be used by the cellular industry). In some implementations, the D2D communication link 110 can be a cellular V2X (C-V2X) communication link 110 that utilizes a shared spectrum (e.g., licensed, unlicensed, or IMT frequency spectrum allowed to be used by the cellular industry).

[0032] The network entities 102 can support communication with the core network 106 or with another network entity 102, or both. For example, the network entities 102 can interface with the core network 106 through one or more backhaul links 116 (e.g., via an SI, N2, N6, or another network interface). The network entities 102 can communicate with each other over backhaul links 116 (e.g., via an X2, Xn, or another network interface). In some implementations, the network entities 102 can communicate directly with each other (e.g., between network entities 102). In some other implementations, the network entities 102 can communicate with or indirectly through (e.g., via the core network 106) each other. In some implementations, one or more of the network entities 102 can include subcomponents, such as an access network entity, which can be an example of an access node controller (ANC). The ANC can communicate with one or more UEs 104 through one or more other access network transmission entities (which can be referred to as radio heads, smart radio heads, or transmission reception points (TRPs)).

[0033] In some implementations, the network entities 102 can be configured in a disaggregated architecture that can be configured to utilize a protocol stack that is physically or logically distributed among two or more network entities 102, such as an integrated access backhaul (IAB) network, an open RAN (O-RAN) (e.g., a network configuration sponsored by the O-RAN Alliance), or a virtualized RAN (vRAN) (e.g., a cloud RAN (C-RAN)). For example, the network entities 102 can include one or more of the following: a central unit (CU), a distributed unit (DU), a radio unit (RU), a RAN intelligent controller (RIC) (e.g., a near real-time RIC (near-RT RIC), a non-real-time RIC (non-RT RIC)), a service management and orchestration (SMO) system, or any combination thereof.

[0034] A RU can also be referred to as a radio head, intelligent radio head, remote radio head (RRH), remote radio unit (RRU), or transmission reception point (TRP). In a disaggregated RAN architecture, one or more components of the network entity 102 can be collocated, or one or more components of the network entity 102 can be located at distributed locations (e.g., separate physical locations). In some implementations, one or more network entities 102 of a disaggregated RAN architecture can be implemented as virtual units (e.g., virtual CUs (VCUs), virtual DUs (VDUs), virtual RUs (VRUs)).

[0035] The functional split between the CU, the DU, and the RU can be flexible and can support different functions depending on the functions (e.g., network layer functions, protocol layer functions, baseband functions, radio frequency functions, and any combination thereof) performed at the CU, the DU, or the RU. For example, a functional split of a protocol stack can be employed between the CU and the DU, such that the CU can support one or more layers of the protocol stack and the DU can support one or more different layers of the protocol stack. In some implementations, the CU can host upper layer protocol layer (e.g., layer 3 (L3), layer 2 (L2)) functions and signaling (e.g., radio resource control (RRC), service data adaptation protocol (SDAP), packet data convergence protocol (PDCP)). The CU can be connected to one or more DUs or RUs, and the one or more DUs or RUs can host lower protocol layers, such as layer 1 (LI) (e.g., physical (PHY) layer) or L2 (e.g., radio link control (RLC) layer, medium access control (MAC) layer) functions and signaling, and each can be controlled at least in part by the CU.

[0036] Additionally or alternatively, a functional split of a protocol stack can be employed between the DU and the RU, such that the DU can support one or more layers of the protocol stack and the RU can support one or more different layers of the protocol stack. The DU can support one or more different cells (e.g., via one or more RUs). In some implementations, the functional split between the CU and the DU or between the DU and the RU can be within a protocol layer (e.g., some functions of a protocol layer can be performed by one of the CU, the DU, or the RU, while other functions of the protocol layer are performed by a different one of the CU, the DU, or the RU).

[0037] A CU can be further functionally split into a CU control plane (CU-CP) and a CU user plane (CU-UP) function. A CU can be connected to one or more DUs via a midhaul communication link (e.g., Fl, Fl-c, Fl-u), and a DU can be connected to one or more RUs via a front-haul communication link (e.g., a front-haul (FH) interface). In some implementations, a midhaul or front-haul communication link can be implemented according to an interface (e.g., a channel) between layers of a protocol stack supported by respective network entities 102 communicating via such a communication link.

[0038] The core network 106 can support user authentication, access authorization, tracking, connection, and other access, routing, or mobility functions. The core network 106 can be an evolved packet core (EPC) or 5G core (5GC), which can include control plane entities (e.g., a mobility management entity (MME), an access and mobility management function (AMF)) to manage access and mobility, and user plane entities (e.g., a serving gateway (S-GW), a packet data network (PDN) gateway (P-GW), or a user plane function (UPF)) to route packets or data to a packet data network 108. In some implementations, the control plane entities can manage non-access stratum (NAS) functions such as mobility, authentication, and bearer management (e.g., data bearer, signaling bearer, etc.) for one or more UEs 104 served by the one or more network entities 102 associated with the core network 106.

[0039] The core network 106 can communicate with a packet data network 108 through one or more backhaul links 116 (e.g., via an SI, N2, N6, or another network interface). The packet data network 108 can include an application server 118. In some implementations, the one or more UEs 104 can communicate with the application server 118 through the core network 106 via the network entities 102. A UE 104 can establish a session (e.g., a protocol data unit (PDU) session, etc.) with the core network 106 via a network entity 102. The core network 106 can route traffic (e.g., control information, data, etc.) between the UE 104 and the application server 118 using the established session (e.g., the established PDU session). A PDU session can be an example of a logical connection between a UE 104 and the core network 106 (e.g., one or more network functions of the core network 106).

[0040] In the wireless communication system 100, network entity 102 and UE 104 can use the resources of the wireless communication system 100 (such as time resources (e.g., symbols, time slots, subframes, frames, etc.) or frequency resources (e.g., subcarriers, carriers)) to perform various operations (e.g., wireless communication). In some implementations, network entity 102 and UE 104 can support different resource structures. For example, network entity 102 and UE 104 can support different frame structures. In some implementations, such as in 4G, network entity 102 and UE 104 can support a single frame structure. In some other implementations, such as in 5G and other suitable radio access technologies, network entity 102 and UE 104 can support various frame structures (i.e., multiple frame structures). Network entity 102 and UE 104 can support various frame structures based on one or more sets of parameters.

[0041] The wireless communication system 100 may support one or more parameter sets, and the parameter sets may include subcarrier spacing and cyclic prefixes. The first parameter set (e.g., μ =0) can be associated with the first subcarrier spacing (e.g., 15 kHz) and the normal cyclic prefix. The first set of parameters (e.g., ) associated with the first subcarrier spacing (e.g., 15 kHz) μ =0) can utilize one time slot per subframe. The second parameter set (e.g., μ =1) can be associated with the second subcarrier spacing (e.g., 30kHz) and the normal cyclic prefix. The third parameter set (e.g., μ =2) can be associated with the third subcarrier spacing (e.g., 60 kHz) and a normal cyclic prefix or an extended cyclic prefix. The fourth parameter set (e.g., μ =3) can be associated with the fourth subcarrier spacing (e.g., 120 kHz) and the normal cyclic prefix. The fifth parameter set (e.g., μ =4) can be associated with the fifth subcarrier spacing (e.g., 240 kHz) and the normal cyclic prefix.

[0042] The time intervals of resources (e.g., communication resources) can be organized according to frames (also called radio frames). Each frame can have a duration, for example, 10 milliseconds (ms). In some implementations, each frame can include multiple subframes. For example, each frame can include 10 subframes, and each subframe can have a duration, for example, 1 ms. In some implementations, each frame can have the same duration. In some implementations, each subframe of a frame can have the same duration.

[0043] Additionally or alternatively, time intervals of resources (e.g., communication resources) can be organized as slots. For example, a subframe can include a number (e.g., quantity) of slots. Each slot can include a number (e.g., quantity) of symbols (e.g., orthogonal frequency division multiplexing (OFDM) symbols). In some implementations, the number (e.g., quantity) of slots of a subframe can depend on the numerology. For a normal cyclic prefix, a slot can include 14 symbols. For an extended cyclic prefix (e.g., applicable to 60 kHz subcarrier spacing), a slot can include 12 symbols. For both the normal cyclic prefix and the extended cyclic prefix, the relationship between the number of symbols per slot, the number of slots per subframe, and the number of slots per frame can depend on the numerology. It should be understood that reference to a first numerology (e.g., = 0) associated with a first subcarrier spacing (e.g., 15 kHz) can be used interchangeably between subframes and slots. μ =0) can be used interchangeably between subframes and slots.

[0044] In the wireless communications system 100, the electromagnetic (EM) spectrum can be split into various classes, bands, frequency channels, and / or the like based on frequency or wavelength. For example, the wireless communications system 100 can support one or more operating bands, such as frequency range designations FR1 (410 MHz - 7.125 GHz), FR2 (24.25 GHz - 52.6 GHz), FR3 (7.125 GHz - 24.25 GHz), FR4 (52.6 GHz - 114.25 GHz), FR4a or FR4-1 (52.6 GHz - 71 GHz), and FR5 (114.25 GHz - 300 GHz). In some implementations, the network entity 102 and the UE 104 can perform wireless communications on one or more operating bands. In some implementations, FR1 can be used by the network entity 102 and the UE 104, and other equipment or devices, for cellular communications traffic (e.g., control information, data). In some implementations, FR2 can be used by the network entity 102 and the UE 104, and other equipment or devices, for short range, high data rate capabilities.

[0045] FR1 can be associated with one or more numerologies (e.g., at least three numerologies). For example, FR1 can be associated with a first numerology (e.g., μ =0) including a subcarrier spacing of 15 kHz; a second numerology (e.g., μ =1) including a subcarrier spacing of 30 kHz; and a third numerology (e.g., μ =2) including a subcarrier spacing of 60 kHz. FR2 can be associated with one or more numerologies (e.g., at least two numerologies). For example, FR2 can be associated with a third numerology (e.g., μ =2) including a subcarrier spacing of 60 kHz; and a fourth numerology (e.g.,μ =3), which includes a subcarrier spacing of 120kHz.

[0046] Depending on the implementation, such as an implementation for two-sided model performance monitoring, one or more of network entity 102 and UE 104 are operable to implement various aspects of the techniques described herein. For example, UE 104 sends encoded data 120 to network entity 102 (e.g., gNB) in a two-sided model, and gNB receives encoded data 120 from UE 104. In at least one implementation, the encoded data is encoded using the UE's encoder model 122, and the UE includes representation information of the encoder for the two-sided model. The UE performs a computation that generates one or more model metrics 124 associated with the performance monitoring of the two-sided model. In at least one implementation, the model metrics 124 are computed based on representation information 126 and a set of decoder parameters representing the decoder model 128 of the gNB. In the implementation, the gNB sends the representation information 126 of the gNB's decoder model 128 to UE 104. The UE also sends feedback data 130 to the gNB, and the feedback data 130 includes one or more model metrics 124. The gNB receives feedback data 130 from UE 104 and can initiate an update process based on the feedback data.

[0047] Figure 2 The illustration depicts an example 200 of a wireless network including a gNB (e.g., a base station) and multiple UEs in relation to two-sided model performance monitoring. In this example 200, the wireless network includes a gNB 202 (e.g., a base station, network entity 102) and multiple UEs. K There are UE104. For example, UEs include UE1, UE2, ..., and UE104. K A base station can be represented as one equipped with M Nodes of antennas B 1 and by U 1. U 2, ... U K Indicated K Each UE has 1 UE, and each UE has N One antenna. In this example, It can be expressed as in B 1 and U k Between in frequency band l ( Time on) t The channel, U k It is the size of N × M A matrix having complex terms, i.e., .

[0048] In timet and frequency bands l It can be assumed that the base station will transmit U K a message where and the base station uses as the precoding vector. U k , The received signal at where denotes the noise vector at the receiver.

[0049] To improve the achievable rate of the link, the gNB 202 can select that maximizes the received signal-to-noise ratio (SNR). Several schemes have been proposed for selecting , some of which rely on some knowledge of . The gNB can acquire knowledge of by direct measurement (e.g., in time division duplex (TDD) mode and assuming reciprocity of the channel) or indirectly using information sent by the UE to the gNB (e.g., in frequency division duplex (FDD) mode). In the latter case, a large amount of feedback can be required to send accurate information about , which is important for a large number of antennas and / or large frequency bands.

[0050] As described herein, the implementation is discussed with reference to a single slot. However, the implementation of the described techniques can be further extended to more than a single slot. Thus, can be expressed as can be defined as a matrix of size N × M × L , which can be composed by stacking of multiple frequency bands (e.g., the entry at can be equal to ). Thus, each UE can feed back to the gNB information of the most recent N × M × L complex numbers.

[0051] ​Several proposed schemes attempt to reduce the required rate of feedback. For example, these schemes group (often referred to as bilateral methods) include two parts, where the first part is deployed at the UE side and the second part is deployed at the gNB side. The UE side and the gNB side include one or more neural network blocks that are trained using data-driven methods. The UE side can compute a latent representation of the input data (e.g., to be transmitted to the gNB), such as with as few number of bits as possible. The gNB can receive the data sent by the UE side and the gNB attempts to reconstruct the information that the UE intended to send to the gNB. There are several methods to train the NN modules of the bilateral model, including centralized training, simultaneous training, and separate training. Likewise, updating the bilateral model can be done centrally on one entity, or simultaneously or separately on different entities.

[0052] In implementations that reduce the required feedback information, the encoding part (at the UE) computes a quantized latent representation of the input data, and the decoding part (at the gNB) takes the latent representation and uses it to reconstruct the desired output. In this case, the input data can be channel measurement based data. For example, it can be the raw channel input or a precoder (e.g., a eigenvector associated with the largest eigenvalue of for each subband) computed from the channel matrix.

[0053] Figure 3 FIG. illustrates an example 300 of a high-level structure of a bilateral model related to bilateral model performance monitoring, in accordance with aspects of the present disclosure. In this example 300, the bilateral model includes a neural network based UE 104 (e.g., first node) and a gNB (network entity 102) (e.g., second node), referred to in this example as M e (an encoder or encoding model) and M d (a decoder or decoding model), respectively. The input to the bilateral model can be based on channel measurements, e.g., raw channel measurements or eigenvectors associated with the measured channel.

[0054] Figure 4 FIG. illustrates an example 400 of another high-level structure of a bilateral model with multiple encoders related to bilateral model performance monitoring, in accordance with aspects of the present disclosure. In an implementation extension of the bilateral model (as shown and described with reference to Figure 3 ), the encoder part or the decoder part of the bilateral model is shared. In this example 400, the bilateral model includes multiple neural network based UEs 104 (e.g., first nodes, sharing the encoder M e ) and a gNB (network entity 102) (e.g., second node, decoderM d ). In implementations, the structure at the UE side and / or the gNB side can vary depending on the specific bi-lateral model and scheme. Alternatively, similar structures can be implemented with multiple decoders (e.g., at multiple gNBs) with a single encoder (e.g., at the UE), or with a structure with multiple encoders and multiple decoders.

[0055] Figure 5a and Figure 5b FIG. illustrates an example of a CSI system 500 with a UE subsystem 500a and a network subsystem 500b supporting operation of a bi-lateral model, in accordance with aspects of the present disclosure. In at least one implementation, the network subsystem 500b is implemented at a network entity 102 (e.g., a gNB). As described in further detail below, in aspects of the described techniques, the UE side (e.g., the UE subsystem 500a) sends the latest In this figure, blocks B1-B6 are multi-layer NNs. The encoder part first generates two latent representations of the input data (i.e., Int_t_1 and Int_t_2). Then, the encoder part quantizes these latent representations using vector and scalar quantization methods, respectively. The results are then sent to the gNB side where the decoder part generates the output, such as the reconstruction of the input data. In this structure, it is assumed that the UE and gNB quantization codebooks (510, 532) are the same, and their values, as well as the weights of the NNs (for all blocks), will be learned in the training phase. In the training phase, other hyperparameters of the quantizer module 520 can also be set, such as the number of bits for the scalar quantization, the number of bits for the vector quantization, and the number of quantization levels. J number of quantization levels (Nq) Q Note that the described structure, input / output, and 3D shape of the blocks are for illustration only, and any other structure can be used for the bi-lateral model.

[0056] According to one or more implementations, two latent representations of the input data are generated. In at least one example, the input data is a channel matrix H and / or is based on the channel matrix, such as a function of the channel matrix, e.g., a channel covariance matrix, an eigen-decomposition, such as at least one eigenvector, a singular value decomposition (SVD), such as at least one of a left singular vector and / or a right singular vector, etc. According to implementations, the latent representations contain “real” numbers, and thus it can not be practical to directly send the latent representations using a finite number of feedback bits.

[0057] Thus, at the lower branch (e.g., a scalar quantization branch), the UE subsystem 500a quantizes the real values of the latent representation and sends the quantized version to the network subsystem 500b (e.g., a network entity such as a gNB). In at least one example, the quantization that occurs in the lower branch is based on linear quantization with Q levels. At the upper branch (e.g., a branch that quantizes using a codebook), the UE subsystem 500a compares the latent representation to the code words of a codebook and then, instead of sending the actual latent representation, the UE subsystem 500a can send an identifier(s) (ID) and / or index(es) of at least one code word based on a measure of relevance or similarity of the indicated code word(s) and the actual latent representation, such as the closest code word(s), a weighted combination of a subset of code words, etc. Note that the code words of the codebook are not fixed and can be learned in a training phase.

[0058] Further, the various blocks of the network subsystem 500b can be trained to generate a desired output using the bits received from the UE subsystem 500a (e.g., feedback CSI bits, such as bits corresponding to two latent representations). In at least some examples, the training objective is to make the output data (e.g., the reconstructed data) as similar as possible to the input data. Alternatively or additionally, other objective functions (e.g., loss functions) can also be used for training.

[0059] In the CSI system 500, different blocks of the system and associated processes for feedback CSI data can be generated at the UE subsystem 500a (e.g., a transmitter node), which are then used by the network subsystem 500b (e.g., a receiver node) for reconstruction of the input data. In the UE subsystem 500a, the input data 502 is input to the neural network 504. One example of the input data 502 is the H matrix as defined above. In implementations, the input data 502 is a three-dimensional matrix that represents the channel between the Tx-Rx antenna pairs (Nt x Nr) L pairs N pairs M ) at the UE in a frequency band x In at least some examples, the frequency band can represent the channel per subcarrier, per subcarrier, per subcarrier group (such as a PRB or sub-PRB or RBG (resource block group)), etc. Further, the input data 502 can be a function of the H matrix (e.g., a vector corresponding to a singular vector associated with the largest singular value of the matrix H).

[0060] The neural network 504 can be implemented as a multi-layer neural network, for example using a convolutional neural network (CNN). In implementations, the neural network 504 can be shared between the upper branch and the lower branch of the UE subsystem 500a. The size of the intermediate tensor output of the neural network 504 (“Int_t_0”) can be c 0 x r 0 xf 0. The neural network 506 (e.g., a multi-layer neural network such as a CNN) receives the output from the neural network 504 and generates an output 508. For example, the output 508 is a 3D intermediate tensor of size c 1 x r 1 x f 1 (i.e., “Int_t_l”), where f 1 represents, for example, a number of filters at the last convolutional layer of the neural network 506 using a CNN. In at least some implementations of each input sample (and based on the weights of the neural network 506), there will be a tensor of size 1 x f 1 x c 1 x r 1 at the output 508. For example, the parameters c 1, r 1, and f 1 are hyperparameters determined at a training phase.

[0061] According to one or more implementations, the UE subsystem 500a uses a quantized codebook 510 to transmit a representation of the output 508 to the network subsystem 500b. For example, the quantized codebook 510 is composed of f 1 x J 1 tensors (codewords) of size 1 x 1 x J 1. Each of these tensors has an ID or index that can be represented using c 1 x r 1 x c 1 tensors, the mapper module 512 generates at least one ID (between 0 and J) that shows the ID of the codeword (from the quantized codebook 510) that has the closest and / or greatest correlation to the output 508. For example, for the output 508, the mapper module 512 maps the input tensor of size r 1 x f 1 x c 1 x r 1 x 1 x

[0062] The UE subsystem 500a also includes a neural network 516, which can be implemented as a multi-layer neural network (e.g., using a CNN). The neural network 516 receives the output from the neural network 504 (e.g., the intermediate tensor output "Int_t_0") and generates an output 518. For example, the output 518 represents a 3D intermediate tensor of size c 2 x r 2 x f 2, where f 2 is the number of filters at the last convolutional layer of the neural network 516 implemented, e.g., using a CNN. Further, the parameters c 2, r 2 and f 2 are hyperparameters determined at a training phase. The output 518 is not necessarily 3D shaped, and can optionally be a 1D or 2D tensor, such as depending on the structure of the neural network 516.

[0063] To enable the UE subsystem 500a to transmit the output 518 and / or some representation thereof to the network subsystem 500b, and to reduce communication overhead, it can first pass the output 518 through a quantizer module 520, which represents a scalar quantizer in at least some implementations. In at least one example, the quantizer module 520 quantizes each value of the output 518 to 2 Q levels (e.g., each quantized value can be represented using Q bits). Q The values of c 2 x r 2 x f 2, where each entry takes only one value out of 2 Q possible values.

[0064] Thus, the UE subsystem 500a transmits a representation of the outputs 514, 522 (e.g., an encoded representation of the outputs 514, 422) to the network subsystem 500b via the feedback link 524. The outputs 514, 522 and / or the representation thereof are transmitted (e.g., with source and / or channel coding and modulation) to the network subsystem 500b, e.g., with the feedback CSI information bits. Depending on the implementation, the outputs 514, 522 can be transmitted to the network subsystem 500b using bits (information bits). For example, is the number of potential vectors of the upper branch, J is the number of codewords in the upper branch quantization codebook 510, the size of the potential representation of the lower branch is shown, Q is the number of stages used in the scalar quantizer of the lower branch.

[0065] At network subsystem 500b, gNB-side receives input 525 and input 526 representing output 514 and output 522, respectively, via feedback link 524. Network subsystem 500b feeds input 525 to de-mapper module 528 (e.g., in the upper branch) and feeds input 526 to neural network 530 (e.g., in the lower branch). De-mapper module 528 outputs output 534, which represents a 3D tensor (e.g., “Int_t_3”) of size c 1 x r 1, as input, and replaces and / or maps it to a corresponding codeword in quantization codebook 532 of size f 1 x f 1 x J 1 tensor (codeword). De-mapper module 528 outputs output 534, which represents a 3D tensor (e.g., “Int_t_3”) of size c 1 x r 1 x f 1 in at least one implementation. Quantization codebook 532 can be the same or different from quantization codebook 510 of UE subsystem 500a.

[0066] Network subsystem 500b also includes neural network 536, which can be implemented as a multi-layer neural network (e.g., using a CNN). Neural network 536 takes output 534 as input and generates output 538 (“Int_t_4”). For example, output 538 is a 3D tensor of size c 4 x r 4 x f 4. Furthermore, parameters c 4, r 4 and f 4 are hyperparameters determined at a training stage. As mentioned above, neural network 530 takes input 526 as input. Thus, neural network 530 generates output 540 (“Int_t_5”). For example, output 540 is a 3D tensor of size c 5 x r 5 x f 5. Parameters c 5, r 5 and f 5 are hyperparameters determined at a training stage. In an example, c 4 = 5 and c 4 = 5. r r To assist concatenation of outputs 538, 540, parameter

[0067] is used.​c 5 and r 5 can each equal c 4 and r 4. Given these design parameters, c g and r g may be used as the first two dimensions of the output 538, 540, e.g., the output 538 can have a size of c g x r g x f 4 and the output 540 can have a size of c g x r g x f 5. Thus, the concatenator module 542 concatenates the outputs 538, 540 along a third dimension (e.g., a filter dimension) and constructs “Int_t_6”. Thus, “Int_t_6” can be a three-dimensional tensor of size c g x r g x f 4 + 5). f

[0068] The network subsystem 500b also includes a neural network 544, e.g., a multi-layer neural network such as implemented using a CNN. The neural network 544 takes “Int_t_6” (the output of the concatenator module 542) as input and generates output data 546. For example, the output data 546 represents a reconstructed data representation of the input data 502 that was previously input to the UE subsystem 500a. The output data 546 can be shared between the upper branch and the lower branch of the network subsystem 500b. In at least some implementations, in order to enable reconstruction of the original input data 502, the output data 546 has a size of N x M x L .

[0069] The following sections describe implementation details of the system 500. In these sections, “UE” can refer to the UE subsystem 500a and “network”, “network entity”, and / or “gNB” can refer to the network subsystem 500b.

[0070] Considerations regarding network structure: a. In one example, the output of the neural network 516 is designed to be in the [1, 1] range. This can be enabled, e.g., by applying an appropriate activation function (e.g., “tanh”) for the last layer of the neural network 516. ​b. Assuming an ideal feedback channel, inputs 525 and 526 can equal outputs 514 and 522, respectively. In the case of a non-ideal feedback channel, they can differ, e.g., some elements of inputs 525, 526 are received in error, omission of some elements of outputs 514, 522 in the feedback CSI, etc. Such effects can be appropriately modeled in the network structure of system 500. c. The neural network structures of the different neural networks of system 500 can be hyperparameters and can be determined at the training phase. Note that they can be fixed at the inference phase. d. The total available feedback rate can be divided among the data used for transmission of outputs 514 and 522. For example, when choosing , , the number of codewords in quantization codebook 510 (e.g., Nc) and the number of scalar quantization levels (e.g., Ns) are divided among the total available feedback rate. J Q e. System 500 can be reduced to: • Using only the codebook-based quantization branch: In this case, the lower branch of UE subsystem 500a can be turned off or not used. In addition, neural network 530 and concatenator module 542 can be removed from network subsystem 500b. • Using only the scalar quantization branch: In this case, the upper branch of UE subsystem 500a can be turned off or not used. In addition, demapper module 528, neural network 536, and concatenator module 542 can be removed from network subsystem 500b. Codebook 532 can optionally not be implemented and / or used. • In some examples, network subsystem 500b (e.g., gNB) can instruct UE subsystem 500a (e.g., UE) to use at least one of: only the codebook-based quantization branch, only the scalar quantization branch, or both the codebook-based quantization branch and the scalar quantization branch. • In some examples, UE subsystem 500a can determine to feedback outputs of at least one of: only the codebook-based quantization branch, only the scalar quantization branch, or both the codebook-based quantization branch and the scalar quantization branch. Such determination can be based on input data 502, e.g., channel matrix H. UE subsystem 500a can indicate an indication of such determination to network subsystem 500b, e.g., feedback CSI is based on: only the codebook-based quantization branch, only the scalar quantization branch, or both the codebook-based quantization branch and the scalar quantization branch. ​​f. A similar framework can be used when the input data 502 is not directly equal to the matrix H. Optionally, the input data 502 can be input data represented as a 3D matrix, such as a DFT transformed version of a channel matrix, or a matrix representing one / several eigenvectors and / or eigenvalue(s) of a channel matrix in different frequency bands. Alternatively or additionally, the input data 502 can correspond to a set of at least one precoding vector associated with a downlink transmission from a network node to a UE. As some other examples, the size of H can be NxMxT, where the third dimension represents values at different time symbols and / or slots, or NxMxZ, where the third dimension represents a composite time / frequency domain (e.g., stacked or concatenated frequency and time domain vectors). g. The entries of H can be complex numbers, and since most neural network methods work with real numbers, a transformation from the complex domain to the real domain can be made. For example, the real and imaginary parts of the input data 502 (size N × M × L ) can be separated and then concatenated together to generate input data of size 2 N × M × L with only real values. The concatenation can also occur in other dimensions. In another example, the system 500 can extend the channel matrix with its conjugate and then transform the extended data using an inverse fast Fourier transform (IFFT). For example, the result will be real numbers and can be used for the neural network. Some tensors can have a reduced dimension. For example, the second dimension 3D tensor and / or the third dimension 3D tensor described above can have a value of 1 to reduce to a 1D or 2D tensor.

[0071] Considerations regarding the training and / or inference phase: a. The input data 502 can be collected at the UE subsystem 500a and then used at the UE subsystem 500a or transferred to the network subsystem 500b depending on where the model will be trained. b. The neural network weights are randomly initialized for training. The neural network weights can be changed during the training phase in a way that reduces a loss function. The neural network weights can be fixed during inference time. c. The tensors of the quantization codebook can not be fixed and can be determined during the training process. They can also be fixed during the inference phase. d. It can be considered that the quantized codebook 532 of the network subsystem 500b is the same as the quantized codebook 510 of the UE subsystem 500a. For example, after the model is trained, there can be one quantized codebook that will be used by both the subsystem 500a and the subsystem 500b. For example, assume that the full model has been trained at the network subsystem 500b, then the resulting quantized codebook can be transmitted to the UE subsystem 500a together with other weights of the neural network blocks for the UE subsystem 500a. If the training phase occurs at the UE subsystem 500a, then the quantized codebook 510 and the weights of the network subsystem 500b blocks will be transmitted to the network subsystem 500b.

[0072] Considerations regarding the network loss function: a. One example of an objective function is to minimize the mean squared error between the input data 502 and the output data 546 (reconstructed data). b. One approach for having an end-to-end differentiable loss function is to consider that the input 525 is equal to the output 508 and the input 526 is equal to the output 518 in the backpropagation phase.

[0073] Considerations regarding the communication requirements: a. If the model is trained at the network subsystem 500b, then some mechanisms for exchanging certain information between the UE subsystem 500a and the network subsystem 500b can be provided, such as: a method for sending the input data to the network subsystem 500b (e.g., channel measurements (or any desired transformation thereof). A method for sending the final neural network weights on the UE subsystem 500a side to the UE subsystem 500a. A method for sending the quantized codebook (e.g., learned codebook) to the UE subsystem 500a. A method for sending the number of quantization levels Q of the quantizer module 520 to the UE subsystem 500a. A method for sending the output 514 and the output 522 (and / or representations thereof) to the network subsystem 500b. b. If the model is trained at the UE subsystem 500a, then some mechanisms for exchanging certain information between the UE subsystem 500a and the network subsystem 500b can be provided, such as: a method for sending the final neural network weights on the network subsystem 500b side to the network subsystem 500b. A method for sending the quantized codebook (e.g., learned codebook) to the network subsystem 500b. A method for sending the output 514 and the output 522 (and / or representations thereof) to the network subsystem 500b.

[0074] It should be noted that in at least some implementations, the UE subsystem 500a will have sufficient computational resources to perform the training. Moreover, the UE subsystem 500a can have access to sufficient training samples of the environment to create a model with proper generalization (e.g., in scenarios where different UEs will use the same model). In at least some examples, the UE performing the training can be a high-performance UE and / or an artificial intelligence / machine learning (AI / ML) model training source that has the capability for model training (e.g., sufficient computational and / or storage resources) and model transfer to gNBs (e.g., via a Uu interface) and / or UEs (e.g., via a sidelink channel).

[0075] With respect to monitoring the performance of the dual-sided model, one common approach to determine that the model is not performing well is to compare the output at the gNB side with what it should have fed back to it or the input at the UE side. For example, if the output should be One approach is for the UE to send the correct to the gNB (e.g., using a conventional feedback mechanism). The gNB would then compare the it estimates using the dual-sided model with the correct it receives using the second mechanism. If the difference between them is large, then the gNB can determine that the model is not well-calibrated and initiate an update procedure.

[0076] However, a challenge with this approach is that transmitting the correct is not a simple task (even using conventional feedback schemes) and requires a new approach. Moreover, the signaling overhead required to transmit the correct data can also be an issue. An alternative is for the gNB to feed back its estimate to the UE, e.g., using some modified conventional approach, and then the UE compares the gNB estimate with the actual value (e.g., the input data at the UE side of the model). If the difference is larger than a threshold, then the UE determines that the model is not performing well. However, this also has the same drawbacks as the previous approach and requires a new approach to be designed to transmit the gNB estimate output to the UE and requires signaling overhead.

[0077] With respect to training a bilateral model, the reason for separate model training is that the first node and the second node want to use their own designed and optimized model, rather than just running a model provided by another vendor. This also provides protection for the design of each node's NN model, as it does not need to be shared with another party or entity. There are several approaches that can be used for separate training of a bilateral model. For example, separate training of a bilateral model can start with training the model at the first node (e.g., UE side), and then training the model at the second node (e.g., gNB side), which is referred to herein as first node prioritized training. Alternatively, training a bilateral model can first train at the second node (e.g., gNB side), and then train the model at the first node (e.g., UE side), referred to herein as second node prioritized training. There can be other training alternatives as well.

[0078] The concept of first node prioritized training is that the first node (e.g., UE) first trains the encoder portion by assuming a model for the decoder portion. Note that the assumed decoder portion can be different from the actual model of the actual decoder portion at the second node. It should also be noted that the training dataset at the first node can be collected and / or measured by the first node, or can be received from other nodes. After the completion of the training, the first node generates a sample dataset from the output of the encoder (e.g., which is also considered as the input of the decoder) and the expected output of the decoder. This dataset is sent to the second node (e.g., gNB), which can train the decoder portion based on a model structure determined by itself. Another approach for first node prioritized separate training can be that, instead of generating a dataset for the second node, the first node sends information about the trained "assumed decoder" so that the second node can train the actual decoder to match the input / output relationship of the assumed decoder.

[0079] A similar procedure can be used for second node prioritized training, where the second node (e.g., gNB) first trains the decoder portion by assuming a model for the encoder portion. Note that the assumed encoder portion can be different from the actual model of the actual encoder portion at the first node. It should also be noted that the training dataset at the second node can be collected and / or measured by the second node, or can be received from other nodes. After the completion of the training, the second node generates a sample dataset from the nominal input of the assumed encoder (e.g., which is also considered as the input of the encoder) and the output of the assumed encoder (e.g., which is also considered as the input of the encoder, and which can also be considered as the expected encoder output). In sending this dataset to the first node (e.g., UE), the first node can train the encoder portion based on a model structure determined by itself.

[0080] Alternatively, instead of generating a dataset for the first node, the second node sends information about the trained "assumed encoder" to the first node, so that the first node can train the actual encoder to match the input / output relationship of the assumed encoder. One drawback of the above approach is that, for example, in the first node prioritized training approach, the first node designs its encoder part based on the assumption that the performance of the "actual decoder" is similar to that of the "assumed decoder." However, after the "actual decoder" is trained at the second node, it can not behave exactly like the "assumed decoder," which can be due to many reasons, such as lack of training samples, overfitting, and differences in the NN structure of the "assumed decoder" and the "actual decoder." A similar problem exists due to the possible difference between the "assumed encoder" and the "actual encoder" in the second node prioritized training approach. Aspects of the technology described herein have the advantage that such inconsistency in the separate training of the bilateral model is eliminated.

[0081] In aspects of the described technology for bilateral model performance monitoring, the described technology is directed to separate training and model updating, where the NN modules of the first node (e.g., UE) and the second node (e.g., gNB) are trained in different training sessions without a forward propagation path or a backward propagation path between the UE and the gNB. Notably, an advantage of separate training is that the first node and the second node do not need to know the internal structure of the NN module of the other side. In one or more implementations of the bilateral model, the described technology trains and / or updates the bilateral model in separate training loops, which, as described above, eliminates the inconsistency in the separate training of the bilateral model.

[0082] Further, as described above, the parameters of the bilateral model are trained before the bilateral model can effectively feedback information to the gNB. A model is trained by presenting it with a set of inputs and desired outputs (a training dataset), which the model uses to find the statistics of the inputs and the input / output relationship and capture it in the parameters of the model. A trained model can be used as long as the statistics of the inputs and / or the input / output relationship remain similar to the statistics presented to the model during training, or the model can be well generalized to new cases. In other aspects, the described technology is directed to how the system (e.g., gNB and / or UE) of the bilateral model determines that the model is not performing properly. For example, these techniques consider the amount of data and signaling needed to implement the monitoring process.

[0083] With respect to model monitoring in the bilateral model, is used to refer to the complete model, while M e and M drespectively refer to the UE-side and gNB-side of the model. Note that the proposed implementation also applies to the case where the roles of the UE and gNB are reversed by the dual-sided model, i.e., the encoder M e The model is executed at the gNB, and the decoder M d The model is executed at the UE.

[0084] It is assumed that the dual-sided model has been trained (e.g., at the gNB, network node, or UE). For the system of the dual-sided model system, several implementations can be considered, such as: a) the UE group uses the same model (i.e., the same M e and M d for encoding and decoding of the input data, b) the modules at the gNB are the same, but each UE in the group can have a different UE part, i.e., M d The modules are the same for the UE group, but each UE in the group can have a different UE part, i.e., M e c) each UE has its own model, i.e., each user has a pair of M e and M d In all cases, we assume that the parameters of the UE-side (i.e., the M e ) are known to or transmitted to each UE; and the parameters of the gNB-side (i.e., the M d ) are known to or transmitted to the gNB.

[0085] It is expected that the dual-sided model will have satisfactory results after training. However, if the statistics of the input or input / output relationship change over time, the performance of the trained model can decrease. Consider a dual-sided model where there are K a first node (e.g., UE) with an encoder part, and there are T a second node (e.g., gNB) with a decoder part. It is assumed that the dual-sided model is trained using the dual-sided model. For example, the i first node determines the NN structure for its own encoder part ( ), and assumes the NN structure for the decoder part ( ). Furthermore, the i first node has access to the i first training dataset (where is the input of the dual-sided model, This is the expected output of the two-sided model. The training dataset (e.g., the CSI dataset) can be collected and / or measured from the environment, or it can be used in the second... i The first node receives data from another node.

[0086] Then, the i The first node uses the training dataset. Come to train and This is called a locally bilateral model. Then, the... i Each first node generates a training dataset using a trained two-sided model. ,in It indicates that it is based on a trained model. Input data The potential representation of . As step 1, It is the input of the two-sided model. These are the expected outputs of the two-sided model. They can be sets. The same sample can also be different samples.

[0087] Then, the i The first node to the first j Send by the second node . No. j The second node receives from all the first nodes. And use all of these datasets to create training sets. . No. j Each second node is its own decoder part ( Determine the NNN structure. j The second node uses the training dataset. Come to train (Right now, act as The input of the input, (Serves as the expected output of the model).

[0088] Instead, not by the first i The first node generates the training dataset. Instead, the first i The first node can send to the... j The second node sends about Information along with The latent representation set constructed, i.e. The second node can use the set of all potential representations from all first nodes and their... Come to train At this stage, there are three types of trained two-sided models: the first-node side... Encoder, second node side decoder, is the i first node to understand the way the actual decoder at the second node works.

[0089] In aspects of the described bilateral model performance monitoring, the monitoring technique utilizes . Specifically, for a given sample (input and expected output pair), the i first node can first find the latent representation of the input data, and then use its own decoder local version (i.e., the ) to generate an output. This output should be similar to the expected output. Moreover, it should have a high similarity to the actual output of the second node, because is also trained to generate an output similar to the expected output.

[0090] In one or more implementations, a monitoring technique is to send an “expected output” or a set of “expected outputs” (associated with an input data or a set of input data) to the second node, and then determine the similarity between the expected output and the output of generated from the latent representations corresponding to the associated input data or set of input data. In case of large difference (or average difference), the system will flag possible problems in the trained model and possible need for model update. Note that there are several ways to compute the difference between the output and the expected output, such as by Euclidean distance, cosine similarity, etc. Which one to use depends on the application. The main problem of the previous monitoring scheme is the overhead of sending the “expected output” to the second node, and the possible delay of the decision happening at the second node instead of the first node.

[0091] In one or more implementations, a technique utilizes the local decoder after separate training, where is used instead of . More precisely, the performance of the system is monitored by monitoring the difference between an “expected output” or a set of “expected outputs” (associated with an input data or a set of input data) and the output of generated from the latent representations corresponding to the associated input data or set of input data. In case of large difference (or average difference), the system will flag possible problems in the trained model and possible need for model update. Note that since there is no need to transmit the “expected output” to the second node, this proposed technique has lower overhead compared to the first described technique. Moreover, the monitoring is performed at the first node, which can reduce the delay. Note also that in some cases, the expected output can be the input data (e.g., the goal is to reconstruct the input).

[0092] In the implementation scenario, the local decoder can differ from the actual decoder at the second node. Therefore, compared to decisions made based on the actual output at the second node, the monitoring scheme is based on... The accuracy of the decisions made may be lower. This discrepancy could be due to various reasons, such as when the model structures of the local decoder and the actual decoder are different, or when the second node is used for generation. The model and the first i When the available samples at a first node are different (e.g., this could be due to differences in data from different first node observers).

[0093] In one or more implementations, a technique utilizes an updated local decoder based on samples fed back from second nodes, wherein, to improve performance, each second node sends feedback information to each first node and characterizes its development. This representation can be shown by sending... Some input / output samples or sent directly to the first node This is how it's implemented. Then, the first node can use this information to update its local decoder model. This is to make it better match the actual decoder.

[0094] For example, in After training, each second node (e.g., the first...) j Construct and extend the second node) to the first i Send by the first node Note that the sample Can be with The samples in the same place can also be the same. A subset of the samples in, and may include those not included Some samples Then, the i-th first node receives data from all second nodes in the second node and creates a combined dataset. It is used to train a local copy of the decoder (i.e., And the process continues as before.

[0095] Additionally, note that during the update... After that, the first node can attempt to update. Without change In updating Afterwards, the first node can generate another training dataset for the second node, just as before, and iterate as an update process to improve the accuracy of the bilateral model and the similarity between the local decoder model and the actual decoder model. Then, the updated... The iteration can be used in the monitoring scheme described above. Note that the local decoder is now more consistent with the actual decoder, which reduces the likelihood of mismatch between the output of the local decoder and the actual decoder. This will improve the accuracy of the monitoring scheme.

[0096] In one or more alternative implementations that lack a local decoder but have a construction based on a set of samples received from the gNB side, one technique involves the first node not having a local decoder. In other words, the system is a two-sided model where the encoder model is at the first node and the decoder model is at the second node. In this scenario, there is no local decoder model to be used for monitoring. In this case, it is assumed that the first node... j The second node to the first i Each of the first nodes sends information and represents the generation This representation can be shown by sending... This is achieved using some samples of the input / output (e.g., Note that the sample It can be the second node from the first i Some samples received by one first node or another first node. Another alternative is to send directly to the first node. .

[0097] Then, the first node can use this information to construct a local decoder model. With a local decoder model, the first node can monitor an "expected output" or "expected output set" (associated with an input data or input dataset) and the encoder model generated based on the associated input data or input dataset. The system monitors the model by comparing the differences between its outputs. In cases of large differences (or average differences), the system will flag potential problems in the trained model and the need for model updates. Note that the level of difference indicating potential problems can be based on a threshold set by another node in the network, which also applies to one or more implementations described.

[0098] Referencing the first-node priority method, which has K The training process includes the training of a two-sided model with a first node and a single second node, and the training process includes the training of the first node and a single second node. i Each first node is its own encoder part. Determine the NN structure and define the decoder part ( Assume an N / A structure. Assume the first... i The first node can access the first... i training datasets (in It is the input of the two-sided model. This is the expected output of the two-sided model. The training dataset (e.g., the CSI dataset) can be collected and / or measured from the environment, or it can be used in the second... i The first node receives from another node. Then, the second node... i The first node uses the training dataset. Come to train and This is called a locally bilateral model. Then, the... i Each first node generates a training dataset using a trained two-sided model. ,in It indicates that it is based on a trained model. Input data The potential representation of . As step 1, It is the input of the two-sided model, and These are the expected outputs of the two-sided model. They can be sets. The same sample can also be different samples.

[0099] Then, the i The first node sends to the second node. And the second node receives from all the first nodes. And use all of these datasets to create training sets. The second node is its own decoder part ( The neural network structure is determined. The second node uses the training dataset. train ,Right now, act as The input of the input, This serves as the expected output of the model. Alternatively, it is not determined by the first... i The first node generates the training dataset. Instead, the first i The first node can send information about the second node. Information and by The constructed latent representation set (i.e., The second node can use the set of all potential representations from all first nodes and their... Come to train .

[0100] Because it has a trained decoder section M d Construction of the second node K Datasets Note that the sample Can be with The samples in the same place can also be the same. A subset of the samples in, and may be those not included Some samples The second node leads to the first... i Send by the first node Among them, in receiving Afterwards, the first node can train or retrain the local decoder part. (Right now, act as The input of the input, act as The expected output). Note that the goal of this step is to update the "local decoder section" (i.e., the expected output). i The first node ), so that it is at least in the first i Among the input samples observable at the first node, the decoder portion of the second node should be as similar as possible.

[0101] Instead, the training dataset is not generated by the second node. Instead, the second node can send information about the first node. M d The first node can use the received information. Together with use The latent representation of the generated input data is used for retraining. In updating After that, the i The first node is fixed. The parameters, and using the dataset Retrain the local two-sided model ( and Dataset It can be the same dataset as the initial dataset, or it can be a dataset constructed from newly collected, measured, and / or received samples. Because it has... The updated version, number i The first node then sends the required information (e.g., the updated information) to the second node again. ).

[0102] The process continues at the second nodes by combining the data sets from all first nodes and retraining the decoder part. This process can be repeated. The training process can be done based on different conditions, such as the number of communication rounds between the first round and the second round, the time, the accuracy / loss of the training of each of the NN modules (i.e., the loss of the decoder part, for example). Note that the above process can also be used for the case where there are different data sets at different first nodes, also for the case where the assumed NN structure for the encoder part is different, and for the case where the NN structure of the local decoder part is different from the NN structure of the actual decoder at the second nodes. It is worth noting that this process helps to reduce the impact of such imbalance between the different nodes and assumptions on the final performance of the trained bilateral model.

[0103] With respect to training the bilateral model with respect to multiple first nodes, multiple second nodes, and implementing an extension can not be one decoder part, but multiple decoder parts at different second nodes, such as where there are K first nodes and T second nodes. In this case, a similar process as described above can be used, with the difference that each first node sends the above information (e.g., the ) to all of the second nodes.

[0104] Then, each of the second nodes (e.g., the first j second nodes) combines all received data into a set and trains its using the constructed data set. In addition, it constructs the required information for each of the first nodes based on the trained version. For example, the j second node generates and sends to the i first node . Note that the samples may be the same as in , can be a subset of the samples in , and can include some samples that are not included in . Then, the i first node receives all information from all second nodes and creates a combined data set that uses the local copy of the decoder (i.e., ) is trained, and the process continues as before. The training process can be done based on different conditions, such as the number of communication rounds between the first and second rounds, the time, the accuracy / loss of the training of each of the NN modules (i.e., the loss of the decoder part, for example).

[0105] The above discussion generally follows the first node priority approach. Similar ideas can be used for the second node priority approach with respect to training a bilateral model for multiple first nodes, multiple second nodes. Consider the existence of K first nodes and T second nodes, which is outlined below. The first j second node determines the NN structure for its own encoder part ( ), and then assumes the NN structure for the decoder part ( ). Furthermore, it is assumed that the first j first node has access to the first j training data set (where is the input to the bilateral model, is the desired output of the bilateral model). This training data set (e.g., a CSI data set) can be collected and / or measured from the environment or received at the first j second node from another node (e.g., from a first node). Then, the first j second node uses the training data set to train and , referred to as a local bilateral model.

[0106] Then, the first j second node uses the trained local bilateral model to generate a training data set where represents a latent representation of the input data based on the trained local model of the encoder . The samples in can be the same samples of the set , a subset thereof, or different samples thereof. Then, the first j second node sends i to the first first node. The first i first node receives from all of the second nodes, and creates a combined training set using all of this data. The first i first node determines the NN structure for its own encoding part ( ). The first i first node uses the training data set train (Right now, act as The input of the input, (Serves as the expected output of the encoder model).

[0107] Instead, not by the first j The second node generates the training dataset. Instead, the first j A second node can send information about the first node. Information, and in some cases, it can also send some input vectors. The first node can use all input sample sets along with all Come to train Because it has a trained encoder section. , No. i Construct a first node T Datasets Note that the sample Can be with The same samples in the same way can also be A subset of the samples in, and may include those not included Some samples .

[0108] No. i The first node to the first j Send by the second node The second node receives from all the first nodes in the first node. and create combined datasets Use it to train the encoder The local copy (i.e., act as The input of the input, act as The expected output). Note that the goal of this step is to update the "local encoder section" (i.e., the first...). j The second node ), so that it is at least in Among the observable input samples, the average encoder portion of the first node is made as similar as possible.

[0109] Instead, not by the first i The first node generates the training dataset. Instead, the first i The first node can send information about the second node. The second node can use the received information. Retrain using the same input data In updating After that, the j The second node is fixed. The parameters, and using the dataset Retrain the local two-sided model ( and Dataset It can be the same dataset as the initial dataset, or it can be a dataset constructed from newly collected, measured, and / or received samples. Similarly, due to having The updated version, number j The second node then sends to the first i The first node sends the required information as described above (e.g., updated...). The process continues at the first node by combining the datasets from all the second nodes and retraining the encoder part. This process is repeated, and the training process can be performed based on different conditions, such as the number of communication rounds between the first and second rounds, the time, and / or the accuracy / loss of training each NN module (i.e., the loss of the decoder part, for example).

[0110] Figure 6 The illustration depicts an example system 600 for training a two-sided model for performance monitoring, in accordance with various aspects of this disclosure. In this example system 600, a first information set 602 is used to train the two-sided model and includes the inputs to the second model and the expected outputs of the second node. A second information set 604 includes the inputs to the second model and the expected outputs of the second node. This second information set 604 contributes to inputting a “first dataset” into the first model of the second device. The second device uses the first information set 606 to train its first model, and the second device generates a second information set 608 as output feedback to the first device. A feedback set 610 from the second device is used to train the second model of the first device. The now updated second model at the first device utilizes a third information set 612 to update the first model of the first device. This provides a more accurate output from the UE (e.g., the first device), and the second device (e.g., the gNB) matches the first device better.

[0111] Figure 7FIG. 13 illustrates an example of a diagram 1300 that supports dual-sided model performance monitoring in accordance with aspects of the present disclosure. The diagram 1300 can implement aspects of the wireless communications system 100, the system 900, the device 902, the UE 104, the device 702, or any other system or device described herein. The diagram 1300 can include a first device 1302 and a second device 1304 in communication. The first device 1302 can include a transmitter 1306 and a receiver 1308. The second device 1304 can include a transmitter 1310 and a receiver 1312. The first device 1302 and the second device 1304 can be examples of the first device 702 and the second device 702 described with reference to FIG. 7. The first device 1302 and the second device 1304 can communicate via one or more channels, such as the sidelink channel.

[0112] The processor 704, the memory 706, the transceiver 708, or various combinations or

[0113] In some implementations, the processor 704, the memory 706, the transceiver 708, or various combinations or components thereof can be implemented in hardware (e.g., in communication management circuitry). The hardware can include a processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof configured as or otherwise supporting a means for performing the functions described in the present disclosure. In some implementations, a processor 704 and memory 706 coupled with the processor 704 can be configured to perform one or more functions described herein (e.g., by the processor 704 executing instructions stored in memory 706).

[0114] For example, the processor 704 can support wireless communications at the device 702 in accordance with examples as disclosed herein. The processor 704 can be configured as or otherwise support a means for transmitting, to a second device in a dual-sided model, encoded data, the encoded data based at least on input data and an encoder model of the first device, the first device comprising a first set of parameters including characterization information of an encoder of the dual-sided model, performing a computation that generates at least one model metric associated with performance monitoring of the dual-sided model, the at least one model metric computed based at least in part on a first set of information characterizing a decoder model of the second device and a second set of decoder parameters, and transmitting, to the second device, feedback data, the feedback data comprising the at least one model metric.

[0115] Further, the processor 704 can be configured as or otherwise support a means for the first set of information to include a set of channel data representations during a first time- frequency space region. The first set of parameters is received from the second device. The first set of parameters includes at least one of a threshold value or scheduling information associated with transmitting the feedback data; and the scheduling information includes a first indication of periodically or semi-periodically transmitting the feedback data and a second indication of a transmission interval. The computation of the at least one model metric is performed based on at least one of an instruction received from the second device, an internal process event, or a scheduled periodic event. The first set of parameters is received via higher layer signaling as at least one of an RRC message or a MAC-CE message. The at least one model metric is based at least in part on the received threshold value. The method further includes receiving, from at least one of the second device or an alternative device, a second set of decoder parameters characterizing a decoder model. The second set of decoder parameters is determined based at least in part on a second set of information including data samples each representing an input and an expected output of the decoder model. The second set of decoder parameters is determined based on the data samples of the second set of information to minimize a difference between the expected output and an actual output of the decoder model. The second set of information is received from at least one of the second device or the alternative device. The second set of decoder parameters is determined during training of the encoder model. The second set of decoder parameters is updated based at least in part on update information received from at least one of the second device or the alternative device. The first set of information includes data samples each representing: an input of an encoder of a bilateral model, and an expected output of a decoder model of the bilateral model. The at least one model metric is determined based at least in part on a comparison of the expected output of the data samples of the first set of information and an actual output of the bilateral model constructed by the encoder and the decoder based on the input of the data samples of the first set of information. The comparison is performed by finding a mean Euclidean distance, a generalized cosine similarity.

[0116] Additionally or alternatively, in accordance with examples disclosed herein, the device 702 can include a processor and a memory coupled with the processor, the processor configured to cause the apparatus to transmit, to a second device, encoded data in a bilateral model, the encoded data based at least on input data and an encoder model of the apparatus, the apparatus including a first set of parameters, the first set of parameters including characterization information of an encoder of the bilateral model; perform a computation that generates at least one model metric associated with performance monitoring of the bilateral model, the at least one model metric computed based at least in part on a first set of information characterizing a decoder model of the second device and a second set of decoder parameters; and transmit, to the second device, feedback data, the feedback data including the at least one model metric.

[0117] Further, the wireless communication at the device 702 can include any one or combination of the following: the first set of information includes a set of channel data representations during the first time-frequency space region. The first set of parameters is received from the second device. The first set of parameters includes at least one of: a threshold value or scheduling information associated with transmitting the feedback data; and the scheduling information includes a first indication of periodically or semi-periodically transmitting the feedback data, and a second indication of a transmission interval. The computation of the at least one model metric is performed based on at least one of an instruction received from the second device, an internal process event, or a scheduled periodic event. The first set of parameters is received via at least one of a higher layer signaling, as a RRC message, or a MAC-CE message. The at least one model metric is based at least in part on the received threshold value. The processor is configured to cause the apparatus to receive, from at least one of the second device or an alternative device, a second set of decoder parameters characterizing a decoder model. The second set of decoder parameters is determined based at least in part on a second set of information including data samples, the data samples each representing an input and an expected output of the decoder model. The second set of decoder parameters is determined based on the data samples of the second set of information to minimize a difference between the expected output and an actual output of the decoder model. The second set of information is received from at least one of the second device or the alternative device. The second set of decoder parameters is determined during a training of an encoder model. The second set of decoder parameters is updated based at least in part on update information received from at least one of the second device or the alternative device. The first set of information includes data samples, the data samples each representing: an input of an encoder of a bilateral model, and an expected output of a decoder model of the bilateral model. The at least one model metric is determined based at least in part on a comparison of the expected output of the data samples of the first set of information and an actual output of the bilateral model constructed by the encoder and the decoder based on the input of the data samples of the first set of information. The comparison is performed by finding a mean Euclidean distance, a generalized cosine similarity.

[0118] According to examples disclosed herein, a processor 704 of a device 702, such as a UE 104, can support wireless communication. The processor 704 includes at least one controller coupled with at least one memory and configured or operable to cause the processor to: transmit, to a network entity (NE) in a bilateral model, encoded data based at least on input data and an encoder model of a user equipment (UE), the UE including a first set of parameters including characterization information of an encoder of the bilateral model; perform a computation that generates at least one model metric associated with performance monitoring of the bilateral model, the at least one model metric computed based at least in part on a first set of information characterizing a decoder model of the NE and a second set of decoder parameters; and transmit, to the NE, feedback data including the at least one model metric.

[0119] The processor 704 can include an intelligent hardware device, (e.g., a general- purpose processor, a DSP, a CPU, a microcontroller, an ASIC, an FPGA, a programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or any combination thereof). In some implementations, the processor 704 can be configured to operate a memory array using a memory controller. In some other implementations, a memory controller can be integrated into the processor 704. The processor 704 can be configured to execute computer-readable instructions stored in a memory (e.g., the memory 706) to cause the device 702 to perform various functions in accordance with this disclosure.

[0120] The memory 706 can include random access memory (RAM) and read-only memory (ROM). The memory 706 can store computer-readable computer-executable code including instructions that, when executed by the processor 704, cause the device 702 to perform various functions described herein. The code can be stored in a non-transitory computer-readable medium such as system memory or another type of memory. In some implementations, the code can not be directly executable by the processor 704 but can cause a computer (e.g., when compiled and executed) to perform functions described herein. In some implementations, the memory 706 can include a basic I / O system (BIOS), which can control basic hardware or software operation such as the interaction with peripheral components or devices.

[0121] The I / O controller 710 can manage input and output signals for the device 702. The I / O controller 710 can also manage peripherals not integrated into the device 702. In some implementations, the I / O controller 710 can represent a physical connection or port to the external peripherals. In some implementations, the I / O controller 710 can utilize an operating system such as iOS®, ANDROID®, MS-DOS®, MS WINDOWS®, OS / 2®, UNIX®,

[0122] In some implementations, the device 702 can include a single antenna 712. However, in some other implementations, the device 702 can have more than one antenna 712 (i.e., multiple antennas), including multiple antenna panels or antenna arrays, which can be capable of concurrently sending or receiving multiple wireless transmissions. The transceiver 708 can communicate bi-directionally, via the one or more antennas 712, wired, or wireless links as described herein. For example, the transceiver 708 can represent a wireless transceiver and can communicate bi-directionally with another wireless transceiver. The transceiver 708 can also include a modem to modulate the packets and to demodulate packets received from one or more antennas 712.

[0123] Figure 8 FIG. 8 illustrates an example of a block diagram 800 of a device 802 that supports double-sided model performance monitoring in accordance with aspects of the present disclosure. The device 802 can be an example of a network entity 102 (e.g., a gNB) as described herein. The device 802 can support wireless communication with one or more network entities 102, UEs 104, or any combination thereof. The device 802 can include components for bi-directional communications including components for transmitting and receiving communications, such as a processor 804, a memory 806, a transceiver 808, and an I / O controller 810. These components can be in electronic communication or otherwise coupled (e.g., operatively, communicatively, functionally, electronically, electrically) via one or more interfaces (e.g., buses). The processor 804, the memory 806, the transceiver 808, or various combinations thereof can be examples of means for performing various aspects of the present disclosure as described herein. For example, the processor 804, the memory 806, the transceiver 808, or various combinations

[0124] The processor 804, the memory 806, the transceiver 808, or various combinations thereof or various components thereof can be examples of means for performing various aspects of the present disclosure as described herein. For example, the processor 804, the memory 806, the transceiver 808, or various combinations or components thereof can support a method for performing one or more of the operations described herein.

[0125] In some implementations, the processor 804, the memory 806, the transceiver 808, or various combinations or components thereof can be implemented in hardware (e.g., in communication management circuitry). The hardware can include a processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic, discrete hardware components, or any combination thereof configured as or otherwise supporting a means for performing the functions described in the present disclosure. In some implementations, a processor 804 and memory 806 coupled with the processor 804 can be configured to perform one or more functions described herein (e.g., by the processor 804 executing instructions stored in memory 806).

[0126] For example, the processor 804 can support wireless communication at the device 802 in accordance with examples as disclosed herein. The processor 804 can be configured as or otherwise support a means for receiving, in a bilateral model, encoded data from a first device, the encoded data encoded using an encoder model of the first device, the second device comprising a first set of parameters, the first set of parameters comprising characterization information of a decoder of the bilateral model; transmitting, to the first device, a first set of information, the first set of information comprising at least characterization information of a decoder model of the second device; receiving feedback data from the first device; and initiating an update process based at least in part on the feedback data.

[0127] Further, the processor 804 can be configured as or otherwise support any one or combination of the following: the first set of parameters comprises at least a threshold value. The first set of parameters comprises scheduling information associated with receiving the feedback data. The first set of information comprises an indication of a structure of the decoder model. The first set of information comprises data samples each representing an input and an output of the decoder model. The process for initiating the update process is based on at least one of the feedback data, an output of the bilateral model, or the threshold value.

[0128] Additionally or alternatively, the device 802 can comprise a processor and a memory coupled with the processor, the processor configured to cause the apparatus to receive, in a bilateral model, encoded data from a first device, the encoded data encoded using an encoder model of the first device, the apparatus comprising a first set of parameters, the first set of parameters comprising characterization information of a decoder of the bilateral model; transmit, to the first device, a first set of information, the first set of information comprising at least characterization information of a decoder model of the apparatus; receive feedback data from the first device; and initiate an update process based at least in part on the feedback data, in accordance with examples as disclosed herein.

[0129] Further, wireless communication at the device 802 can comprise any one or combination of the following: the first set of parameters comprises at least a threshold value. The first set of parameters comprises scheduling information associated with receiving the feedback data. The first set of information comprises an indication of a structure of the decoder model. The first set of information comprises data samples each representing an input and an output of the decoder model. The process for initiating the update process is based on at least one of the feedback data, an output of the bilateral model, or the threshold value.

[0130] The processor 804 can include an intelligent hardware device, (e.g., a general- purpose processor, a DSP, a CPU, a microcontroller, an ASIC, an FPGA, a programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or any combination thereof). In some implementations, the processor 804 can be configured to operate a memory array using a memory controller. In some other implementations, a memory controller can be integrated into the processor 804. The processor 804 can be configured to execute computer-readable instructions stored in a memory (e.g., the memory 806) to cause the device 802 to perform various functions in accordance with this disclosure.

[0131] The memory 806 can include random access memory (RAM) and read-only memory (ROM). The memory 806 can store computer-readable computer-executable code including instructions that, when executed by the processor 804, cause the device 802 to perform various functions described herein. The code can be stored in a non-transitory computer-readable medium such as system memory or another type of memory. In some implementations, the code can not be directly executable by the processor 804 but can cause a computer (e.g., when compiled and executed) to perform functions described herein. In some implementations, the memory 806 can include a basic I / O system (BIOS), which can control basic hardware or software operation such as the interaction with peripheral components or devices.

[0132] The I / O controller 810 can manage input and output signals for the device 802. The I / O controller 810 can also manage peripherals not integrated into the device 802. In some implementations, the I / O controller 810 can represent a physical connection or port to the external peripherals. In some implementations, the I / O controller 810 can utilize an operating system such as iOS®, ANDROID®, MS-DOS®, MS WINDOWS®, OS / 2®, UNIX®,

[0133] In some implementations, the device 802 can include a single antenna 812. However, in some other implementations, the device 802 can have more than one antenna 812 (i.e., multiple antennas), including multiple antenna panels or antenna arrays, which can be capable of concurrently sending or receiving multiple wireless transmissions. The transceiver 808 can communicate bi-directionally, via the one or more antennas 812, wired, or wireless links as described herein. For example, the transceiver 808 can represent a wireless transceiver and can communicate bi-directionally with another wireless transceiver. The transceiver 808 can also include a modem to modulate the packets and to demodulate packets received from one or more antennas 812.

[0134] Figure 9 A method 900 that supports bilateral model performance monitoring is illustrated in accordance with aspects of the present disclosure. Operations of the method 900 can be implemented by a device described herein or its components. For example, operations of the method 900 can be performed by the UE 104 described with reference to FIG. 1. In some implementations, the device can execute a set of instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can perform aspects of the described functions using special-purpose hardware. Figures 1-8 The device described with reference to FIG. 1 can perform operations of the method 900. In some implementations, the device can execute a set of instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can perform aspects of the described functions using special-purpose hardware.

[0135] At 902, the method can include transmitting, to a second device, encoded data in a bilateral model, the encoded data based at least on input data and an encoder model of a first device, the first device comprising a first set of parameters including characterization information of an encoder of the bilateral model. The operations of 902 can be performed in accordance with examples as described herein. In some implementations, aspects of the operations of 902 can be performed by a device as described with reference to FIG. 1. Figure 1 The device described with reference to FIG. 1 can perform operations of the method 900. In some implementations, the device can execute a set of instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can perform aspects of the described functions using special-purpose hardware.

[0136] At 904, the method can include performing a computation that generates at least one model metric associated with performance monitoring of the bilateral model, the at least one model metric computed based at least in part on a first set of information characterizing a decoder model of the second device and a second set of decoder parameters. The operations of 904 can be performed in accordance with examples as described herein. In some implementations, aspects of the operations of 904 can be performed by a device as described with reference to FIG. 1. Figure 1 The device described with reference to FIG. 1 can perform operations of the method 900. In some implementations, the device can execute a set of instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can perform aspects of the described functions using special-purpose hardware.

[0137] At 906, the method can include transmitting, to the second device, feedback data, the feedback data including the at least one model metric. The operations of 906 can be performed in accordance with examples as described herein. In some implementations, aspects of the operations of 906 can be performed by a device as described with reference to FIG. 1. Figure 1 The device described with reference to FIG. 1 can perform operations of the method 900. In some implementations, the device can execute a set of instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can perform aspects of the described functions using special-purpose hardware.

[0138] Figure 10A flow diagram illustrating a method 1000 that supports bilateral model performance monitoring in accordance with aspects of the present disclosure is shown. The operations of method 1000 can be implemented by a device described herein or its components. For example, the operations of method 1000 can be performed by a network entity 102 (e.g., a gNB) as described with reference to FIGS. 1-2. Figures 1-8 In some implementations, the device can execute instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can use special-purpose hardware to perform aspects of the described functions.

[0139] At 1002, the method can include receiving, from at least one of a second device or an alternative device, a second set of decoder parameters that characterize a decoder model. The operations of 1002 can be performed according to the methods described herein. In some implementations, aspects of the operations of 1002 can be performed by a device as described with reference to FIGS. 1-2. Figure 1 may be performed by the device described with reference to FIGS. 1-2.

[0140] Figure 11 A flow diagram illustrating a method 1100 that supports bilateral model performance monitoring in accordance with aspects of the present disclosure is shown. The operations of method 1100 can be implemented by a device described herein or its components. For example, the operations of method 1100 can be performed by a network entity 102 (e.g., a gNB) as described with reference to FIGS. 1-2. Figures 1-8 In some implementations, the device can execute instructions to control the functional elements of the device to perform the described functions. Additionally or alternatively, the device can use special-purpose hardware to perform aspects of the described functions.

[0141] At 1102, the method can include receiving, from a first device, encoded data in a bilateral model, the encoded data encoded using an encoder model of the first device, the second device comprising a first set of parameters including characterization information of a decoder of the bilateral model. The operations of 1102 can be performed according to the methods described herein. In some implementations, aspects of the operations of 1102 can be performed by a device as described with reference to FIGS. 1-2. Figure 1 may be performed by the device described with reference to FIGS. 1-2.

[0142] At 1104, the method can include transmitting, to the first device, a first set of information including at least characterization information of a decoder model of the second device. The operations of 1104 can be performed according to the methods described herein. In some implementations, aspects of the operations of 1104 can be performed by a device as described with reference to FIGS. 1-2. Figure 1 may be performed by the device described with reference to FIGS. 1-2.

[0143] At 1106, the method can include receiving, from the first device, feedback data. The operations of 1106 can be performed according to the methods described herein. In some implementations, aspects of the operations of 1106 can be performed by a device as described with reference to FIGS. 1-2. Figure 1 may be performed by the device described with reference to FIGS. 1-2.

[0144] At 1108, the method can include initiating an update procedure based at least in part on the feedback data. The operations of 1108 can be performed according to the methods described herein. In some implementations, aspects of the operations of 1108 can be performed by a device as described with reference to FIGs. Figure 1

[0145] It should be noted that the methods described herein describe possible implementations, and that the operations and the steps can be rearranged or otherwise modified and that other implementations can be possible. Further, aspects from two or more of the methods can be combined.

[0146] The various illustrative blocks and components described herein can be implemented with a general purpose processor, a DSP, an ASIC, a CPU, a FPGA or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein and illustrated in the block diagrams. The general purpose processor can be a microprocessor, but in the alternative, the processor can be any processor, controller, microcontroller, or state machine. The processor can also be implemented as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0147] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Other examples and implementations are within the scope of the disclosure and appended claims. For example, due to the nature of software, functions described herein can be implemented using software executed by a processor, hardware, firmware, hardwiring, or combinations of any of these. Features implementing functions can also be physically located at various positions, including being distributed such that portions of functions are implemented at different physical locations.

[0148] Computer-readable media includes both non-transitory computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A non-transitory storage medium can be any available medium that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, non-transitory computer-readable media can comprise RAM, ROM, electrically erasable programmable ROM (EEPROM), flash memory, compact disk (CD) ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to carry or store desired program code means in the form of instructions or data structures and that can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor.

[0149] ​Any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk and disc, as used herein, include CD, laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above are also included within the scope of computer-readable media.

[0150] As used herein, including in the claims “or” as used in a list of items (for example, a list of items prefaced by a phrase such as “at least one of’ or “one or more of’) indicates an inclusive list such that, for example, a list of at least one of A, B, or C means A or B or C or AB or AC or BC or ABC (i.e., A and B and C). Similarly, a list of one or more of A, B, or C means A or B or C or AB or AC or BC or ABC (i.e., A and B and C). Also, as used herein, the phrase “based on” shall not be construed as a reference to a closed set of conditions. For example, an example step that is described as “based on condition A” can be based on both a condition A and a condition B without departing from the scope of the present disclosure. In other words, as used herein, the phrase “based on” shall be construed in the same manner as the phrase “based at least in part on.” Moreover, as used herein, including in the claims “collection” can include one or more elements.

[0151] When referring to a network entity, the terms “transmit,” “receive,” or “communicate” can refer to any portion of the network entity of the RAN (e.g., base station, CU, DU, RU) that communicates with another device (e.g., directly or via one or more other network entities).

[0152] The description set forth herein, in connection with the appended drawings, describes example configurations and does not represent all the examples that can be implemented or that are within the scope of the claims. The term “example” as used herein means “serving as an example, instance, or illustration,” and not “preferred” or “superior” over other examples. The detailed description includes specific details for the purpose of providing an understanding of the described techniques. These techniques, however, can be practiced without these specific details. In some instances, well-known structures and devices are shown in block diagram form in order to avoid obscuring the concepts of the described examples.

[0153] The description herein is presented to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other variations without departing from the scope of the disclosure. Thus, the disclosure is not to be limited to the examples and designs described herein but is to be accorded the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A user equipment (UE) for wireless communication, comprising: At least one memory; as well as At least one processor, coupled to the at least one memory, and configured such that the UE: In a two-sided model, encoded data is sent to a network entity (NE), the encoded data being based at least on input data and the encoder model of the UE, the UE including a first parameter set, the first parameter set including representation information of the encoder of the two-sided model; The calculation generates at least one model metric associated with the performance monitoring of the two-sided model, the at least one model metric being calculated at least in part based on a first information set and a second decoder parameter set characterizing the decoder model of the NE; as well as Feedback data is sent to the NE, the feedback data including the at least one model metric.

2. The UE according to claim 1, wherein the first information set includes: The set of channel data representations during the first time-frequency space region.

3. The UE according to claim 1, wherein the first parameter set is received from the NE.

4. The UE according to claim 1, wherein: The first parameter set includes at least one of a threshold or scheduling information associated with sending the feedback data; and The scheduling information includes: a first indication to send the feedback data periodically or semi-periodically, and a second indication of the transmission interval.

5. The UE of claim 1, wherein the calculation for generating the at least one model metric is performed based on at least one of an instruction received from the NE, an internal process event, or a scheduled periodic event.

6. The UE of claim 1, wherein the first parameter set is received via higher-layer signaling, as at least one of a Radio Resource Control (RRC) message or a Media Access Control Element (MAC-CE) message.

7. The UE of claim 1, wherein the at least one model metric is at least partially based on a received threshold.

8. The UE of claim 1, wherein the second decoder parameter set is determined at least in part based on a second information set including data samples, each data sample representing an input and expected output of the decoder model, and the second decoder parameter set is determined based on the data samples of the second information set to minimize the difference between the expected output and the actual output of the decoder model.

9. The UE of claim 1, wherein the first information set comprises data samples, each data sample representing: the input of the encoder of the two-sided model, and the expected output of the decoder model of the two-sided model.

10. The UE of claim 9, wherein the at least one model metric is determined at least in part based on a comparison of the expected output of the data sample of the first information set with the actual output of the two-sided model constructed by the encoder and decoder based on the input of the data sample of the first information set, and wherein the comparison is performed by looking up the mean Euclidean distance and the generalized cosine similarity.

11. A processor for wireless communication, comprising: At least one controller, coupled to at least one memory, and configured such that the processor: In a two-sided model, encoded data is sent to a network entity (NE), the encoded data being based at least on input data and an encoder model of a user equipment (UE), the UE including a first parameter set, the first parameter set including representation information of the encoder of the two-sided model; The calculation generates at least one model metric associated with the performance monitoring of the two-sided model, the at least one model metric being calculated at least in part based on a first information set and a second decoder parameter set characterizing the decoder model of the NE; as well as Feedback data is sent to the NE, the feedback data including the at least one model metric.

12. A network entity (NE) for wireless communication, comprising: At least one memory; as well as At least one processor, coupled to the at least one memory, and configured such that the NE: In a two-sided model, encoded data is received from a user equipment (UE), the encoded data being encoded using the encoder model of the UE, the NE including a first parameter set, the first parameter set including representation information of the decoder of the two-sided model; Send a first information set to the UE, wherein the first information set includes at least the representation information of the decoder model of the NE; Receive feedback data from the UE; as well as An update process is initiated based at least in part on the feedback data.

13. The network element according to claim 12, wherein the first parameter set includes at least: One or more of the following: a threshold, or scheduling information associated with receiving the feedback data.

14. The NE of claim 12, wherein the first information set comprises: The structure of the decoder model is indicated by one or more of the following data samples, each of which represents the input and output of the decoder model.

15. The NE of claim 12, wherein the process for initiating the update process is based on at least one of the feedback data, the output of the two-sided model, or a threshold.

16. A method performed by a user equipment (UE), the method comprising: In a two-sided model, encoded data is sent to a network entity (NE), the encoded data being based at least on input data and the encoder model of the UE, the UE including a first parameter set, the first parameter set including representation information of the encoder of the two-sided model; The calculation generates at least one model metric associated with the performance monitoring of the two-sided model, the at least one model metric being calculated at least in part based on a first information set and a second decoder parameter set characterizing the decoder model of the NE; as well as Feedback data is sent to the NE, the feedback data including the at least one model metric.

17. The method of claim 16, wherein the first information set comprises: The set of channel data representations during the first time-frequency space region.

18. The method of claim 16, wherein the first parameter set is received from the NE.

19. The method of claim 16, wherein: The first parameter set includes at least one of a threshold or scheduling information associated with sending the feedback data; and The scheduling information includes: a first indication to send the feedback data periodically or semi-periodically, and a second indication of the transmission interval.

20. The method of claim 16, wherein the calculation for generating the at least one model metric is performed based on at least one of an instruction received from the NE, an internal process event, or a scheduled periodic event.

Images