Gradient Dataset-Aware Configuration for Over-the-Air (OTA) Model Aggregation in Federated Learning

Abstract

A method performed by a user equipment (UE) generates local gradients for a federated learning task. The method calculates a gradient sum power level based on the local gradients. The method receives a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal. The method also determines the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level. The method also applies an analog modulation and the channel inversion coefficient to the data block to form an uncoded uplink signal. The method also transmits the uncoded uplink signal on a shared uplink resource to the network for air computation of global gradients for the federated learning task.

Description

Technical Field

[0001] Aspects of the present disclosure generally relate to wireless communication, and more specifically, to techniques and apparatus for 5G New Radio (NR) gradient dataset sensing configurations for over-the-air (OTA) model aggregation in federated learning. Background Art

[0002] Wireless communication systems are widely deployed to provide various telecommunication services such as telephony, video, data, messaging, and broadcasting. A typical wireless communication system may employ a multiple access technology capable of supporting communication with multiple user equipments (UEs) by sharing available system resources (e.g., bandwidth, transmit power, etc.). Examples of such multiple access technologies include Code Division Multiple Access (CDMA) systems, Time Division Multiple Access (TDMA) systems, Frequency Division Multiple Access (FDMA) systems, Orthogonal Frequency Division Multiple Access (OFDMA) systems, Single Carrier Frequency Division Multiple Access (SC-FDMA) systems, Time Division Synchronous Code Division Multiple Access (TD-SCDMA) systems, and Long Term Evolution (LTE). LTE / LTE-Advanced is a set of enhancements to the Universal Mobile Telecommunications System (UMTS) mobile standard released by the Third Generation Partnership Project (3GPP).

[0003] A wireless communication network may include multiple base stations (BSs) capable of supporting communication with multiple user equipments (UEs). A user equipment (UE) may communicate with a base station (BS) via a downlink and an uplink. The downlink (or forward link) refers to the communication link from the BS to the UE, and the uplink (or reverse link) refers to the communication link from the UE to the BS. As will be described in more detail herein, a BS may be referred to as a Node B, gNB, access point (AP), radio head, transmit receive point (TRP), New Radio (NR) BS, 5G Node B, etc.

[0004] The above multiple access technologies have been adopted in various telecommunication standards to provide a common protocol that enables different user equipments to communicate at the urban, national, regional, and even global levels. New Radio (NR) (which may also be referred to as 5G) is a set of enhancements to the LTE mobile standard released by the Third Generation Partnership Project (3GPP). NR is designed to better support mobile broadband Internet access by improving spectral efficiency, reducing costs, improving services, utilizing new spectrums, using Orthogonal Frequency Division Multiplexing (OFDM) with Cyclic Prefix (CP) (CP-OFDM) on the downlink (DL), using CP-OFDM and / or SC-FDM (e.g., also referred to as Discrete Fourier Transform Spread OFDM (DFT-s-OFDM)) on the uplink (UL), better integrating with other open standards, and supporting beamforming, multiple-input multiple-output (MIMO) antenna technology, and carrier aggregation.

[0005] An artificial neural network may include interconnected groups of artificial neurons (e.g., neuron models). An artificial neural network may be a computing device or represent a method executed by a computing device. A convolutional neural network, such as a deep convolutional neural network, is a feedforward artificial neural network. A convolutional neural network may include neuron layers configurable in a tiled receptive field. It is desirable to apply neural network processing to wireless communication for higher efficiency. SUMMARY

[0006] According to one aspect of the present disclosure, a method performed by a user equipment (UE) generates local gradients for a federated learning task. The method calculates a gradient sum-power level based on the local gradients. The method receives a mapping between the gradient sum-power level and a scaling factor for a channel inversion coefficient to process data blocks into uncoded uplink signals. The method also determines the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum-power level. The method further applies analog modulation and the channel inversion coefficient to the data blocks to form uncoded uplink signals. The method also transmits the uncoded uplink signals on shared uplink resources to the network for air computing of global gradients for the federated learning task.

[0007] In another aspect of the present disclosure, an apparatus for wireless communication performed by a user equipment (UE) includes a processor and a memory coupled to the processor. Instructions stored in the memory are operable, when executed by the processor, to cause the apparatus to generate local gradients for a federated learning task. The apparatus may also calculate a gradient sum-power level based on the local gradients. The apparatus may receive a mapping between the gradient sum-power level and a scaling factor for a channel inversion coefficient to process data blocks into uncoded uplink signals. The apparatus may determine the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum-power level. The apparatus may further apply analog modulation and the channel inversion coefficient to the data blocks to form uncoded uplink signals. The apparatus may also transmit the uncoded uplink signals on shared uplink resources to the network for air computing of global gradients for the federated learning task.

[0008] In another aspect of the present disclosure, a user equipment (UE) includes: a unit for generating local gradients for a federated learning task. The UE includes: a unit for calculating a gradient sum power level based on the local gradients. The UE includes: a unit for receiving a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal. The UE further includes: a unit for determining the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level. The UE further includes: a unit for applying analog modulation and the channel inversion coefficient to the data block to form an uncoded uplink signal. The UE further includes: a unit for transmitting the uncoded uplink signal on a shared uplink resource to the network for air computing of global gradients for the federated learning task.

[0009] In another aspect of the present disclosure, a non-transitory computer-readable medium having program code recorded thereon is disclosed. The program code is executed by a user equipment (UE) and includes: program code for generating local gradients for a federated learning task. The UE includes: program code for calculating a gradient sum power level based on the local gradients. The UE includes: program code for receiving a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal. The UE includes: program code for determining the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level. The UE further includes: program code for applying analog modulation and the channel inversion coefficient to the data block to form an uncoded uplink signal. The UE further includes: program code for transmitting the uncoded uplink signal on a shared uplink resource to the network for air computing of global gradients for the federated learning task.

[0010] Aspects generally include methods, apparatuses, systems, computer program products, non-transitory computer-readable media, user equipment, base stations, wireless communication devices, and processing systems as substantially described with reference to the figures and as illustrated in the figures and the description.

[0011] The features and technical advantages of examples in accordance with the present disclosure have been outlined rather broadly above so that the detailed description that follows may be better understood. Additional features and advantages will be described. The disclosed concepts and specific examples may be readily utilized as a basis for modifying or designing other structures for achieving the same purposes of the present disclosure. Such equivalent structures do not depart from the scope of the appended claims. When considered in conjunction with the accompanying figures, the features of the concepts disclosed herein, their organization and method of operation, and related advantages will be better understood from the following description. Each of the figures is provided for purposes of illustration and description and is not a definition of the limits of the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To understand the features of the present disclosure in detail, a specific description can be made by referring to various aspects (some of which are shown in the drawings). However, it should be noted that the drawings only show certain aspects of the present disclosure and should not be considered as a limitation of its scope, because the description may allow other equivalent aspects. The same reference numerals in different drawings may identify the same or similar elements.

[0013] Figure 1 is a block diagram conceptually showing an example of a wireless communication network according to various aspects of the present disclosure.

[0014] Figure 2 is a block diagram conceptually showing an example of communication between a base station and a user equipment (UE) in a wireless communication network according to various aspects of the present disclosure.

[0015] Figure 3 shows an example implementation of designing a neural network using a system-on-chip (SOC) including a general-purpose processor according to certain aspects of the present disclosure.

[0016] Figure 4A 、 4B and 4C are schematic diagrams showing neural networks according to various aspects of the present disclosure.

[0017] Figure 4D is a schematic diagram showing an exemplary deep convolutional network (DCN) according to various aspects of the present disclosure.

[0018] Figure 5 is a block diagram showing an exemplary deep convolutional network (DCN) according to various aspects of the present disclosure.

[0019] Figure 6 is a block diagram showing a federated learning technique supporting a gradient dataset-aware configuration for aerial model aggregation according to various aspects of the present disclosure.

[0020] Figure 7 shows an example of an aerial computing technique supporting a gradient dataset-aware configuration for aerial model aggregation in federated learning according to various aspects of the present disclosure.

[0021] Figure 8 is a timing diagram showing an example federated learning process with a gradient dataset-aware configuration according to various aspects of the present disclosure.

[0022] Figure 9 is a schematic diagram showing an example federated learning process with a gradient dataset-aware configuration, such as executed by a user equipment (UE), according to various aspects of the present disclosure. Detailed implementation manners

[0023] Aspects of the present disclosure are described more fully hereinafter with reference to the accompanying drawings. However, the present disclosure may be embodied in many different forms and should not be construed as limited to any specific structure or function presented throughout the present disclosure. Rather, these aspects are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art. Based on these teachings, those skilled in the art should understand that the scope of the present disclosure is intended to cover any aspect of the present disclosure, whether implemented independently of or in combination with any other aspect of the present disclosure. For example, any number of the aspects set forth may be used to implement an apparatus or practice a method. Additionally, the scope of the present disclosure is intended to cover such an apparatus or method practiced using other structures, functions, or structures and functions in addition to or different from the aspects of the present disclosure set forth. It should be understood that any aspect of the disclosed present disclosure may be embodied by one or more elements of the claims.

[0024] Several aspects of a telecommunications system will now be presented with reference to various apparatuses and techniques. These apparatuses and techniques will be described in the following detailed description and are illustrated in the accompanying drawings by various boxes, modules, components, circuits, steps, processes, algorithms, etc. (collectively referred to as "elements"). These elements can be implemented using hardware, software, or a combination thereof. Whether these elements are implemented as hardware or software depends on the particular application and the design constraints imposed on the overall system.

[0025] It should be noted that although terms typically associated with 5G and later wireless technologies may be used to describe the aspects, the aspects of the present disclosure can be applied to communication systems based on other generations, such as and including 3G and / or 4G technologies.

[0026] In some wireless communication systems, a user equipment (UE) may send data to a network device (e.g., an edge server, a remote parameter server, a base station, etc.). In such a system, the data may include gradients or parameters locally generated by the UE using a local data model (e.g., an artificial intelligence or machine learning model). The network device may aggregate data from multiple UEs to generate a global or general data model. This technique is referred to as federated learning.

[0027] In some cases, as part of an over-the-air (OTA) computing process that supports data aggregation, multiple UEs may send data to a network device via a shared channel (e.g., a multiple access channel (MAC)). As part of the OTA computing process, a UE may determine pre-equilibration parameters (e.g., channel inversion coefficients, pre-equilibration coefficients, transmit power, etc.). However, a UE may lack the coordination or configuration ability to accurately or efficiently determine the pre-equilibration parameters. For example, traditional OTA aggregation only considers channel inversion on the UE side based on channel state information (CSI). CSI inversion may be sufficient for cases where different UEs upload parameters with a similar range (e.g., the UEs are stationary). When OTA aggregation is applied to gradients for federated learning, the gradients sent by different UEs are non-stationary.

[0028] UEs with larger absolute gradient values may observe training data that is beneficial for faster model convergence. The absolute value of the gradient generally decreases as training converges, but different UEs may experience different rates of decrease. Only channel state information (CSI) inversion may average out the contributions from multiple UEs with gradients having larger absolute values from the final gradient averaging. This may prevent the network device from effectively aggregating the gradients from these UEs, which may degrade system performance.

[0029] According to aspects of the present disclosure, a network device may configure a UE to identify or determine one or more parameters related to the transmission of gradients to the network as part of a federated learning task. The parameters may relate to channel inversion coefficients. For example, the network device may send a control message (e.g., a radio resource control (RRC) message, a media access control - control element (MAC-CE), a downlink control information (DCI) message, etc.) to the UE, and the control message may configure the UE to determine a channel inversion coefficient for the transmission of gradients. In some aspects, the control message may include a mapping between the gradient sum power level and an associated scaling factor for determining the channel inversion coefficient. In some cases, the mapping is pre-determined in a telecommunications standard rather than signaled to the UE.

[0030] According to additional aspects of the present disclosure, a UE reports its gradient sum power for each round or set of rounds of a federated learning task. Before a particular round of gradient OTA aggregation, the UE may receive a configuration from the network (e.g., a base station) for reporting the gradient sum power identified in that round. The configuration may be monitored for each round of OTA aggregation or a set of OTA aggregation rounds, or once for the entire training task. Alternatively, the reporting configuration may be defined in a telecommunications standard. The configuration may include information such as the quantization level of the reported gradient sum power, and / or information for scheduling uplink resources to be used for such reporting.

[0031] In yet another aspect according to the present disclosure, a UE may receive an indication from a network of a scaling factor for determining a channel inversion coefficient. The indication may be received prior to a certain round of gradient OTA aggregation. The indication of the scaling factor may be associated with the reported gradient total power. A UE-group specific scaling factor associated with the gradient total power may be configured to a UE group via group common DCI, MAC-CE, or RRC signaling. Multiple scaling factors may be configured for the group, such as using the mapping described previously. The scaling factor may be a power measurement, such as in dB value, for a specific UE group based on a reference signal received power (RSRP) calculated from a reference signal when determining the channel inversion coefficient.

[0032] Figure 1 FIG. 1 is a schematic diagram of a network 100 in which aspects of the present disclosure may be practiced. Network 100 may be a 5G or NR network or some other wireless network, such as an LTE network. Wireless network 100 may include multiple BSs 110 (shown as BS 110a, BS 110b, BS 110c, and BS 110d) and other network entities. A BS is an entity that communicates with user equipment (UE) and may also be referred to as a base station, NR BS, Node B, gNB, 5G Node B (NB), access point, transmit and receive point (TRP), etc. Each BS may provide communication coverage for a specific geographical area. In 3GPP, the term "cell" may refer to the coverage area of a BS and / or the BS subsystem serving that coverage area, depending on the context in which the term is used.

[0033] A BS may provide communication coverage for a macro cell, a pico cell, a femto cell, and / or another type of cell. A macro cell may cover a relatively large geographical area (e.g., with a radius of several kilometers) and may allow unrestricted access by UEs having a service subscription. A pico cell may cover a relatively small geographical area and may allow unrestricted access by UEs having a service subscription. A femto cell may cover a relatively small geographical area (e.g., a home) and may allow restricted access by UEs associated with the femto cell (e.g., UEs in a closed subscriber group (CSG)). A BS for a macro cell may be referred to as a macro BS. A BS for a pico cell may be referred to as a pico BS. A BS for a femto cell may be referred to as a femto BS or a home BS. In Figure 1In the example shown, BS 110a can be a macro BS for macro cell 102a, BS 110b can be a pico BS for pico cell 102b, and BS 110c can be a femto BS for femto cell 102c. A BS can support one or more (e.g., three) cells. The terms “eNB”, “base station”, “NR BS”, “gNB”, “TRP”, “AP”, “Node B”, “5G NB” and “cell” can be used interchangeably.

[0034] In some aspects, a cell may not necessarily be stationary, and the geographical area of a cell can move according to the location of a mobile BS. In some aspects, BSs can use any suitable transport network and be interconnected with each other and / or interconnected to one or more other BSs or network nodes (not shown) in the wireless network 100 through various types of backhaul interfaces such as direct physical connections, virtual networks, etc.

[0035] The wireless network 100 may also include relay stations. A relay station is an entity that can receive a data transmission from an upstream station (e.g., a BS or a UE) and send the data transmission to a downstream station (e.g., a UE or a BS). A relay station can also be a UE that can relay the transmissions of other UEs. In Figure 1 the example shown, the relay station 110d can communicate with the macro BS 110a and the UE 120d to facilitate communication between the BS 110a and the UE 120d. A relay station can also be referred to as a relay BS, relay base station, relay, etc.

[0036] The wireless network 100 can be a heterogeneous network including different types of BSs (e.g., macro BSs, pico BSs, femto BSs, relay BSs, etc.). These different types of BSs can have different transmit power levels, different coverage areas, and different impacts on interference in the wireless communication network 100. For example, a macro BS can have a high transmit power level (e.g., 5 to 40 watts), while pico BSs, femto BSs, and relay BSs can have lower transmit power levels (e.g., 0.1 to 2 watts).

[0037] The network controller 130 can be coupled to a group of BSs and provide coordination and control for these BSs. The network controller 130 can communicate with the BSs via the backhaul. The BSs can also communicate with each other directly or indirectly, e.g., through wireless or wired backhaul.

[0038] UE 120 (e.g., 120a, 120b, 120c) can be dispersed throughout the wireless network 100, and each UE can be stationary or mobile. A UE can also be referred to as an access terminal, terminal, mobile station, user unit, station, etc. A UE can be a cellular phone (e.g., a smart phone), a personal digital assistant (PDA), a wireless modem, a wireless communication device, a handheld device, a laptop computer, a cordless phone, a wireless local loop (WLL) station, a tablet computer, a camera, a gaming device, a netbook, a smartbook, an ultrabook, a medical device or apparatus, a biometric sensor / device, a wearable device (smart watch, smart clothing, smart glasses, smart bracelet, smart jewelry (e.g., smart ring, smart bracelet)), an entertainment device (e.g., a music or video device, a satellite radio device), a vehicle component or sensor, a smart meter / sensor, an industrial manufacturing device, a global positioning system device, or any other suitable device configured to communicate over a wireless or wired medium.

[0039] Some UEs can be considered machine type communication (MTC) or evolved or enhanced machine type communication (eMTC) UEs. MTC and eMTC UEs include, for example, robots, drones, remote devices, sensors, meters, monitors, location tags, etc., which can communicate with a base station, another device (e.g., a remote device), or some other entity. A wireless node can provide connectivity to or from a network (e.g., a wide area network such as the Internet or a cellular network) via a wired or wireless communication link. Some UEs can be considered Internet of Things (IoT) devices and / or can be implemented as narrowband IoT (NB-IoT) devices. Some UEs can be considered customer premises equipment (CPE). UE 120 can be included within an enclosure that houses components of UE 120, such as processor components, memory components, etc.

[0040] Generally, any number of wireless networks can be deployed in a given geographical area. Each wireless network can support a specific RAT and can operate on one or more frequencies. A RAT can also be referred to as a radio technology, air interface, etc. A frequency can also be referred to as a carrier, frequency channel, etc. In a given geographical area, each frequency can support a single RAT to avoid interference between wireless networks of different RATs. In some cases, an NR or 5G RAT network can be deployed.

[0041] In some aspects, two or more UEs 120 (e.g., shown as UE 120a and UE 120e) can communicate directly using one or more sidelink channels (e.g., communicate with each other without using the base station 110 as an intermediate device). For example, the UE 120 can use peer-to-peer (P2P) communication, device-to-device (D2D) communication, vehicle-to-everything (V2X) protocols (e.g., which can include vehicle-to-vehicle (V2V) protocol, or vehicle-to-infrastructure (V2I) protocol, etc.), mesh networks, etc. to communicate. In this case, the UE 120 can perform scheduling operations, resource selection operations, and / or other operations described elsewhere in this document as being performed by the base station 110. For example, the base station 110 can configure the UE 120 via downlink control information (DCI), radio resource control (RRC) signaling, medium access control-control element (MAC-CE), or via system information (e.g., system information block (SIB)).

[0042] As described above, provide Figure 1 merely as an example. Other examples may be different from the examples regarding Figure 1 described examples.

[0043] Figure 2 FIG. 200 shows a block diagram of a design 200 of the base station 110 and the UE 120, and the base station 110 and the UE 120 can be Figure 1 one of the base stations and one of the UEs in. The base station 110 can be equipped with T antennas 234a to 234t, and the UE 120 can be equipped with R antennas 252a to 252r, where generally T≥1 and R≥1.

[0044] At base station 110, transmit processor 220 may receive data for one or more UEs from data source 212, select one or more modulation and coding schemes (MCSs) for a UE at least in part based on channel quality indicators (CQIs) received from each UE, process (e.g., encode and modulate) data for the UE at least in part based on the MCS selected for the UE, and provide data symbols for all UEs. Reducing the MCS reduces throughput but increases transmission reliability. Transmit processor 220 may also process system information (e.g., for semi-static resource partitioning information (SRPI), etc.) and control information (e.g., CQI requests, grants, upper layer signaling, etc.), and provide overhead symbols and control symbols. Transmit processor 220 may also generate reference symbols for reference signals (e.g., cell-specific reference signals (CRSs)) and synchronization signals (e.g., primary synchronization signal (PSS) and secondary synchronization signal (SSS)). Transmit (TX) multiple-input multiple-output (MIMO) processor 230 may perform spatial processing (e.g., precoding) on the data symbols, control symbols, overhead symbols, and / or reference symbols (if applicable), and may provide T output symbol streams to T modulators (MOD) 232a through 232t. Each modulator 232 may process its respective output symbol stream (e.g., for OFDM, etc.) to obtain an output sample stream. Each modulator 232 may further process (e.g., convert to analog, amplify, filter, and upconvert) the output sample stream to obtain a downlink signal. The T downlink signals from modulators 232a through 232t may be transmitted via T antennas 234a through 234t, respectively. According to various aspects described in more detail below, position coding may be utilized to generate synchronization signals to convey additional information.

[0045] At the UE 120, antennas 252a through 252r may receive downlink signals from the base station 110 and / or other base stations, and may provide the received signals to demodulators (DEMOD) 254a through 254r, respectively. Each demodulator 254 may condition (e.g., filter, amplify, down-convert, and digitize) the received signal to obtain an input sample. Each demodulator 254 may further process the input sample (e.g., for OFDM, etc.) to obtain a received symbol. The MIMO detector 256 may obtain the received symbols from all R demodulators 254a through 254r, perform MIMO detection (if applicable) on the received symbols, and provide detected symbols. The receive processor 258 may process (e.g., demodulate and decode) the detected symbols, provide the decoded data for the UE 120 to the data sink 260, and provide the decoded control information and system information to the controller / processor 280. The channel processor may determine the reference signal received power (RSRP), received signal strength indicator (RSSI), reference signal received quality (RSRQ), channel quality indicator (CQI), etc. In some aspects, one or more components of the UE 120 may be included in a housing.

[0046] On the uplink, at the UE 120, the transmit processor 264 may receive and process data from the data source 262 and control information from the controller / processor 280 (e.g., for reporting including RSRP, RSSI, RSRQ, CQI, etc.). The transmit processor 264 may also generate reference symbols for one or more reference signals. The symbols from the transmit processor 264 are precoded (if applicable) by the TX MIMO processor 266, further processed (e.g., for DFT-s-OFDM, CP-OFDM, etc.) by the modulators 254a through 254r, and sent to the base station 110. At the base station 110, the uplink signals from the UE 120 and other UEs may be received by the antenna 234, processed by the demodulator 254, detected (if applicable) by the MIMO detector 236, and further processed by the receive processor 238 to obtain the decoded data and control information sent by the UE 120. The receive processor 238 may provide the decoded data to the data sink 239, and provide the decoded control information to the controller / processor 240. The base station 110 may include a communication unit 244 and communicate with the network controller 130 via the communication unit 244. The network controller 130 may include a communication unit 294, a controller / processor 290, and a memory 292.

[0047] The controller / processor 240 of the base station 110, the controller / processor 280 of the UE 120, and / or Figure 2Any other components may perform one or more techniques associated with gradient-aware configurations, as described in more detail elsewhere. For example, the controller / processor 240 of the base station 110, the controller / processor 280 of the UE 120, and / or Figure 2 Any other components may perform or direct operations of processes such as Figure 8 and 9 and / or other processes described. The memories 242 and 282 may store data and program codes for the base station 110 and the UE 120, respectively. The scheduler 246 may schedule UEs for data transmission on the downlink and / or uplink.

[0048] In some aspects, the UE 120 may include units for receiving, generating, calculating, determining, applying, transmitting, mapping, and / or reporting. Such units may include one or more components of the UE 120 or the base station 110 described in conjunction with Figure 2 description.

[0049] As described above, the provision of Figure 2 is only an example. Other examples may be different from the examples described with respect to Figure 2 description.

[0050] In some cases, different types of devices that support different types of applications and / or services may coexist in a cell. Examples of different types of devices include UE handsets, customer premise equipment (CPE), vehicles, Internet of Things (IoT) devices, etc. Examples of different types of applications include ultra-reliable low-latency communication (URLLC) applications, massive machine type communication (mMTC) applications, enhanced mobile broadband (eMBB) applications, vehicle-to-everything (V2X) applications, etc. Additionally, in some cases, a single device may support different applications or services simultaneously.

[0051] Figure 3FIG. 0 illustrates an example implementation of a system-on-chip (SOC) 300 in accordance with certain aspects of the present disclosure. The SOC 300 may include a central processing unit (CPU) 302 or a multi-core CPU configured with a gradient-aware configuration for neural network training. The SOC 300 may be included in a base station 110 or a UE 120. Variables (e.g., neural signals and synaptic weights), system parameters associated with the computing device (e.g., a neural network with weights), latency, frequency band information, and task information may be stored in a storage block associated with a neural processing unit (NPU) 308, a storage block associated with the CPU 302, a storage block associated with a graphics processing unit (GPU) 304, a storage block associated with a digital signal processor (DSP) 306, a storage block 318, or may be distributed across multiple blocks. Instructions executed at the CPU 302 may be loaded from a program memory associated with the CPU 302 or may be loaded from the storage block 318.

[0052] The SOC 300 may further include additional processing blocks customized for specific functions, such as a GPU 304, a DSP 306, a connectivity block 310 (which may include fifth-generation (5G) connectivity, fourth-generation long-term evolution (4G LTE) connectivity, Wi-Fi connectivity, USB connectivity, Bluetooth connectivity, etc.), and a multimedia processor 312 that may detect and identify poses, for example. In one implementation, the NPU is implemented in the CPU, DSP, and / or GPU. The SOC 300 may further include a sensor processor 314, an image signal processor (ISP) 316, and / or a navigation module 320, and the navigation module 320 may include a global positioning system.

[0053] The SOC 300 may be based on the ARM instruction set. In one aspect of the present disclosure, the instructions loaded into the general-purpose processor 302 may include: code for generating local gradients for a federated learning task. The general-purpose processor 302 may include: code for calculating a gradient sum power level based on the local gradients. The general-purpose processor 302 may include: code for receiving a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal. The general-purpose processor 302 may further include: code for determining a channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level. The general-purpose processor 302 may further include: code for applying an analog modulation and the channel inversion coefficient to the data block to form an uncoded uplink signal. The general-purpose processor 302 may further include: code for transmitting the uncoded uplink signal on a shared uplink resource to a network for air computing of global gradients for a federated learning task.

[0054] Deep learning architectures can perform object recognition tasks by learning to represent the input at successive higher levels of abstraction in each layer, thereby constructing a useful feature representation of the input data. In this way, deep learning addresses the main bottlenecks of traditional machine learning. Before the advent of deep learning, machine learning methods for object recognition problems may have relied heavily on human-engineered features and may have been combined with shallow classifiers. A shallow classifier can be a two-class linear classifier. For example, in a two-class linear classifier, the weighted sum of the feature vector components can be compared with a threshold to predict which class the input belongs to. Human-engineered features can be templates or kernels customized by engineers with domain expertise for a specific problem domain. In contrast, deep learning architectures can learn to represent features similar to those that human engineers might design, but through training. Additionally, deep networks can learn to represent and recognize new types of features that humans may not have considered yet.

[0055] Deep learning architectures can learn a hierarchy of features. For example, if provided with visual data, the first layer can learn to recognize relatively simple features in the input stream, such as edges. In another example, if provided with auditory data, the first layer can learn to recognize spectral power in specific frequencies. The second layer, which takes the output of the first layer as input, can learn to recognize combinations of features, such as simple shapes for visual data or combinations of sounds for auditory data. For example, higher layers can learn to represent complex shapes in visual data or words in auditory data. Even higher layers can learn to recognize common visual objects or spoken phrases.

[0056] When applied to problems with a natural hierarchical structure, deep learning architectures can perform particularly well. For example, the classification of motor vehicles can benefit from first learning to recognize wheels, windshields, and other features. These features can be combined in different ways at higher levels to recognize cars, trucks, and airplanes.

[0057] Neural networks can be designed with various connection patterns. In a feedforward network, information passes from lower layers to higher layers, where each neuron in a given layer communicates with neurons in the higher layer. As described above, hierarchical representations can be constructed in successive layers of a feedforward network. Neural networks can also have recurrent or feedback (also known as top-down) connections. In recurrent connections, the output from a neuron in a given layer can be communicated to another neuron in the same layer. Recurrent architectures can help with the recognition of patterns that span more than one block of input data delivered sequentially to the neural network. Connections from neurons in a given layer to neurons in lower layers are called feedback (or top-down) connections. Networks with many feedback connections can be helpful when the recognition of high-level concepts can assist in distinguishing specific low-level features of the input.

[0058] The connections between the layers of a neural network can be fully connected or locally connected. Figure 4A An example of a fully connected neural network 402 is shown. In the fully connected neural network 402, the neurons in the first layer can convey their outputs to each neuron in the second layer such that each neuron in the second layer will receive inputs from each neuron in the first layer. Figure 4B An example of a locally connected neural network 404 is shown. In the locally connected neural network 404, the neurons in the first layer can be connected to a limited number of neurons in the second layer. More generally, the locally connected layer of the locally connected neural network 404 can be configured such that each neuron in one layer will have the same or a similar connectivity pattern, but with connection strengths (e.g., 410, 412, 414, and 416) that can have different values. The locally connected connectivity pattern can give rise to spatially distinct receptive fields in higher layers because the higher layer neurons in a given region can receive inputs that are conditioned by training on the attributes of a restricted portion of the total input to the network.

[0059] An example of a locally connected neural network is a convolutional neural network. Figure 4C An example of a convolutional neural network 406 is shown. The convolutional neural network 406 can be configured such that the connection strengths associated with the inputs to each neuron in the second layer are shared (e.g., 408). Convolutional neural networks can be well-suited to problems where the spatial location of the input is meaningful.

[0060] One type of convolutional neural network is a deep convolutional network (DCN). Figure 4D A detailed example of a DCN 400 is shown, which is designed to identify visual features in an image 426 input from an image capture device 430 (such as an on-vehicle camera). The DCN 400 of the current example can be trained to identify traffic signs and the numbers provided on the traffic signs. Of course, the DCN 400 can be trained for other tasks, such as identifying lane markings or identifying traffic lights.

[0061] Supervised learning can be used to train the DCN 400. During training, images such as the image 426 of a speed limit sign can be provided to the DCN 400, and then a forward pass can be computed to produce the output 422. The DCN 400 can include a feature extraction part and a classification part. After receiving the image 426, the convolutional layer 432 can apply convolutional kernels (not shown) to the image 426 to generate a first set of feature maps 418. As an example, the convolutional kernels of the convolutional layer 432 can be 5×5 kernels that generate 28×28 feature maps. In this example, since four different feature maps are generated in the first set of feature maps 418, four different convolutional kernels are applied to the image 426 at the convolutional layer 432. Convolutional kernels can also be referred to as filters or convolutional filters.

[0062] The first set of feature maps 418 can be subsampled by a max pooling layer (not shown) to generate a second set of feature maps 420. The max pooling layer reduces the size of the first set of feature maps 418. That is, the size of the second set of feature maps 420 (such as 14×14) is smaller than the size of the first set of feature maps 418 (such as 28×28). The reduced size provides similar information to subsequent layers while reducing memory consumption. The second set of feature maps 420 can be further convolved via one or more subsequent convolutional layers (not shown) to generate one or more subsequent sets of feature maps (not shown).

[0063] In Figure 4D the example, the second set of feature maps 420 is convolved to generate a first feature vector 424. Additionally, the first feature vector 424 is further convolved to generate a second feature vector 428. Each feature of the second feature vector 428 can include numbers corresponding to possible features of the image 426 (e.g., "sign", "60", and "100"). A softmax function (not shown) can convert the numbers in the second feature vector 428 into probabilities. Thus, the output 422 of the DCN 400 is the probability that the image 426 includes one or more features.

[0064] In this example, the probabilities for "sign" and "60" in the output 422 are higher than the probabilities of other outputs in the output 422, such as "30", "40", "50", "70", "80", "90", and "100". Before training, the output 422 produced by the DCN 400 may be incorrect. Therefore, the error between the output 422 and the target output can be computed. The target output is the ground truth of the image 426 (e.g., "sign" and "60"). Then the weights of the DCN 400 can be adjusted such that the output 422 of the DCN 400 is more closely aligned with the target output.

[0065] To adjust the weights, the learning algorithm can compute the gradient vector of the weights. The gradient can indicate the amount by which the error will increase or decrease if the weights are adjusted. At the top layer, the gradient can directly correspond to the values of the weights connecting the activated neurons in the penultimate layer and the neurons in the output layer. In the lower layers, the gradient can depend on the values of the weights and the computed error gradients of the higher layers. The weights can then be adjusted to reduce the error. This way of adjusting the weights can be called "backpropagation" because it involves a "backward pass" through the neural network.

[0066] In practice, the error gradient of the weights can be computed on a small number of examples such that the computed gradient approximates the true error gradient. This approximation method can be called stochastic gradient descent. Stochastic gradient descent can be repeated until the achievable error rate of the entire system has stopped decreasing or until the error rate has reached a target level. After learning, a new image (e.g., the speed limit sign of image 426) can be provided to the DCN, and a forward pass through the network can produce an output 422, which can be considered as an inference or prediction of the DCN.

[0067] A deep belief network (DBN) is a probabilistic model that includes multiple layers of hidden nodes. The DBN can be used to extract hierarchical representations of a training data set. The DBN can be obtained by stacking multiple layers of restricted Boltzmann machines (RBMs). An RBM is an artificial neural network that can learn a probability distribution over an input set. Since an RBM can learn a probability distribution without information about the class to which each input should be classified, an RBM is typically used for unsupervised learning. Using a hybrid unsupervised and supervised paradigm, the bottom RBM of the DBN can be trained in an unsupervised manner and can be used as a feature extractor, and the top RBM can be trained in a supervised manner (for the joint distribution of inputs from the target class and the previous layer) and can be used as a classifier.

[0068] A deep convolutional network (DCN) is a network of convolutional networks configured with additional pooling and normalization layers. The DCN has achieved state-of-the-art performance on many tasks. The DCN can be trained using supervised learning, in which the inputs and output targets are known for many samples, and is used to modify the weights of the network by using the gradient descent method.

[0069] The DCN can be a feed-forward network. Additionally, as described above, the connections from the neurons in the first layer of the DCN to a set of neurons in the next higher layer are shared among the neurons in the first layer. The feed-forward and shared connections of the DCN can be used for fast processing. The computational burden of the DCN can be, for example, much smaller than that of a similarly sized neural network that includes recursive or feedback connections.

[0070] The processing of each layer of a convolutional network can be considered as a spatially invariant template or basis projection. If the input is first decomposed into multiple channels, such as the red, green, and blue channels of a color image, the convolutional network trained on this input can be considered three-dimensional, with two spatial dimensions along the axes of the image and a third dimension capturing color information. The output of the convolutional connections can be considered to form feature maps in subsequent layers, where each element of the feature map (e.g., 420) receives inputs from a series of neurons in the previous layer (e.g., feature map 418) as well as from each of the multiple channels. The values in the feature map can be further processed with a non-linearity such as rectification, max(0,x). Values from adjacent neurons can be further pooled (which corresponds to downsampling), and can provide additional local invariance and dimensionality reduction. Normalization corresponding to whitening can also be applied through lateral inhibition between neurons in the feature map.

[0071] The performance of deep learning architectures can increase as more labeled data points become available or as computing power increases. Modern deep neural networks are routinely trained with computational resources that are thousands of times larger than what was available to a typical researcher just fifteen years ago. New architectures and training paradigms can further boost the performance of deep learning. Rectified linear units can reduce the training problem known as vanishing gradients. New training techniques can reduce overfitting, enabling larger models to achieve better generalization. Encapsulation techniques can abstract the data within a given receptive field and further improve overall performance.

[0072] Figure 5 is a block diagram showing a deep convolutional network 550. The deep convolutional network 550 can include multiple different types of layers based on connectivity and weight sharing. As Figure 5 shown, the deep convolutional network 550 includes convolutional blocks 554A, 554B. Each of the convolutional blocks 554A, 554B can be configured with a convolutional layer (CONV) 356, a normalization layer (LNorm) 558, and a max pooling layer (MAX POOL) 560.

[0073] The convolutional layer 556 can include one or more convolutional filters, which can be applied to the input data to generate a feature map. Although only two convolutional blocks 554A, 554B are shown, the present disclosure is not limited thereto. Instead, any number of convolutional blocks 554A, 554B can be included in the deep convolutional network 550 according to design preferences. The normalization layer 558 can normalize the output of the convolutional filters. For example, the normalization layer 558 can provide whitening or lateral inhibition. The max pooling layer 560 can provide spatially downsampled aggregation for local invariance and dimensionality reduction.

[0074] For example, the parallel filter bank of the deep convolutional network can be loaded on the CPU 302 or GPU 304 of the SOC 300 to achieve high performance and low power consumption. In an alternative embodiment, the parallel filter bank can be loaded on the DSP 306 or ISP 316 of the SOC 300. Additionally, the deep convolutional network 550 can access other processing blocks that may be present on the SOC 300, such as the sensor processor 314 and the navigation module 320 dedicated to sensors and navigation, respectively.

[0075] The deep convolutional network 550 may also include one or more fully connected layers 562 (FC1 and FC2). The deep convolutional network 550 may also include a logistic regression (LR) layer 564. Between the various layers 556, 558, 560, 562, 564 of the deep convolutional network 550 are weights (not shown) that are to be updated. The output of each of these layers (e.g., 556, 558, 560, 562, 564) can be used as the input to the subsequent layer among these layers (e.g., 556, 558, 560, 562, 564) in the deep convolutional network 550 to learn a hierarchical feature representation from the input data 552 (e.g., image, audio, video, sensor data, and / or other input data) provided at the first convolutional block 554A. The output of the deep convolutional network 550 is a classification score 566 for the input data 552. The classification score 566 can be a set of probabilities, where each probability is the probability of the input data, including features from a set of features.

[0076] As described above, provided Figures 3 - 5 merely as an example. Other examples may be different from the examples described with respect to Figures 3 - 5 the examples.

[0077] As described above, in some wireless communication systems, a user equipment (UE) may send data to a network device (e.g., an edge server, a remote parameter server, a base station, etc.). In such a system, the data may include gradients or parameters locally generated by the UE using a local data model (e.g., an artificial intelligence or machine learning model). The network device may aggregate data from multiple UEs to generate a global or general data model. This technique is referred to as federated learning.

[0078] In some cases, as part of an over-the-air (OTA) computing process that supports data aggregation, multiple UEs may send data to a network device over a shared channel (e.g., a multiple access channel (MAC)). As part of the OTA computing process, a UE may determine pre-equilibration parameters (e.g., a channel inversion coefficient, a pre-equilibration coefficient, a transmit power, etc.). However, a UE may lack the coordination or configuration ability to accurately or efficiently determine the pre-equilibration parameters. For example, traditional OTA aggregation only considers channel inversion at the UE side based on channel state information (CSI). CSI inversion may be sufficient for cases where different UEs upload parameters with a similar range (e.g., the UEs are stationary). When OTA aggregation is applied to gradients for federated learning, the gradients sent by different UEs are non-stationary.

[0079] UEs with larger absolute gradient values may observe training data that is beneficial for faster model convergence. The absolute value of a gradient typically decreases as training converges, but different UEs may experience different rates of decrease. Only channel state information (CSI) inversion may average out the contribution from UEs with gradients having larger absolute values from the final gradient averaging. This may prevent the network device from effectively aggregating gradients from UEs, which may degrade system performance.

[0080] When a local training is triggered by an edge server, training of an artificial intelligence (AI) model via computing at the edge device / server occurs. Then the local models are distributed, trained, and uploaded to the edge server. The uploaded parameters may be parameters in a machine learning model (such as a recurrent neural network (RNN)), or gradients for deriving a machine learning model. The edge server aggregates information from the local models to update the global model at the edge server. This aggregation may be simple parameter / gradient averaging. Then, the edge server broadcasts the updated global model to each edge device.

[0081] The advantages of federated learning are the fast access to real-time data generated at edge devices for fast training of an AI model. Additionally, there is no need to consume large radio resources / delay for raw data transmission, and there is better privacy because raw data is not necessarily required. For example, gradients may be sent from the edge device instead of raw data.

[0082] Figure 6 An example of a federated learning system 600 that supports gradient dataset-aware configuration for over-the-air model aggregation in accordance with aspects of the present disclosure is shown. In some examples, the federated learning system 600 may implement aspects of the wireless communication system 100 described in reference Figure 1 The federated learning system 600 may include UEs 120a and 120b and a base station 110, which may be reference Figure 1Example of the described UE 120 and base station 110.

[0083] The federated learning system 600 can update the global data model 625 based on the local data models 640. In some cases, the data models can correspond to neural networks, and the global data model can correspond to a general data model. UE 120a can generate a local data model 640a based on the local dataset 650a and send a set of parameters or gradients corresponding to the local data model 640a to the base station 110 via the multiple access channel 660a. UE 120b can generate a local data model 640b based on the local dataset 650b and send a set of parameters or gradients corresponding to the local data model 640b to the base station 110 via the multiple access channel 660b. UE 120a and 120b can modulate the set of parameters or gradients into a symbol sequence, partition the symbol sequence into data blocks, and send each data block via the multiple access channel 660 in an orthogonal frequency division multiplexing (OFDM) symbol, where one parameter or gradient is sent via a subchannel of the multiple access channel 660 during the OFDM symbol. The transmit power of the subchannel can be selected to mitigate channel fading.

[0084] The base station 110 can receive a set of aggregated parameters corresponding to the parameters or gradients of the local model 640. The base station 110 can aggregate the received gradients or parameters, for example, by calculating the set of average parameters or average gradients, and use the set of aggregated parameters or gradients to update the global model 625. The base station 110 can broadcast the updated parameters or gradients of the global model 625 to the UE 120 via the broadcast channels 670a and 670b. In some cases, the UE 120 can further train the local model 640 and determine another round of local parameters or gradients based on the training indication received from the base station 110.

[0085] The process of training a neural network at the UE 120, sending the parameters or gradients of the neural network to the base station 110, and receiving an indication of the updated global model from the base station 110 can be considered a communication round. The communication rounds can continue until the base station 110 determines that the global model has converged (e.g., the loss of the global model approaches a minimum with a decreasing trend), and the base station 110 can avoid broadcasting an indication of the updated model to the UE 120 based on determining that the global model has converged. Performing the distributed learning process can improve data security and privacy because the UE 120 can send neural network parameters or gradients to the base station 110 instead of the raw data. Additionally, the air computing process for concurrent analog transmission can utilize the signal superposition property of the shared channel, thereby improving system efficiency.

[0086] Over-the-air (OTA) computation for federated learning (also known as "AirComp") uses a wireless multiple access channel (MAC) for AI model aggregation. Uncoded analog modulation with preequalization is used for each OFDM tone. The preequalization can include channel inversion (e.g., truncated channel inversion) or channel-based power modulation. For OTA computation, all UEs can use fully overlapping radio resources to transmit parameters. That is, all UEs share the same time and frequency resources. The parameters received from different UEs at a network device (e.g., a base station) can have the same amplitude and are thus easier to average. OTA computation can be applied to both neural network weights and gradients.

[0087] Figure 7 An example of an over-the-air computing technique 700 is shown in accordance with aspects of the present disclosure. The over-the-air computing technique 700 supports a gradient dataset-aware configuration for over-the-air model aggregation in federated learning. In some examples, the over-the-air computing technique 700 can implement aspects of the wireless communication system 100 described with respect to Figure 1 The over-the-air computing technique 700 can include: a network device 110 (e.g., a base station, an edge server, a remote parameter server, etc.), which can be an example of the base station 110 described with respect to Figure 1 And UEs 120a, 120b, and 120c, which can be examples of the UE 120 described with respect to Figure 1 Described UE 120.

[0088] Multiple UEs 120 (e.g., UEs 120a, 120b, and 120c) can transmit data to the network device 110 based on a transmitter design 715. The transmitter design 715 can apply analog modulation (e.g., analog amplitude modulation) and preequalization to data blocks to form an uncoded uplink signal and can transmit the uncoded uplink signal to the network device 110 via a shared channel (e.g., a multiple access channel). Transmitting the uncoded uplink signal via the shared channel can support over-the-air computing, which can reduce data transmission latency and reduce the amount of radio resources consumed.

[0089] UE 120a may be associated with radio resources 710a, UE 120b may be associated with radio resources 710b, and UE 120c may be associated with radio resources 710c. The radio resources 710a, 710b, and 710c may partially or fully overlap (e.g., may correspond to the same time and frequency resources) and correspond to a multi-access channel. UE 120 may apply pre-equalization parameters (e.g., channel inversion coefficient, transmit power scaling) to the uncoded uplink signal to improve signal characteristics (e.g., received signal power, signal-to-noise ratio, etc. at network device 110), which may improve the efficiency of aggregating the data received at network device 110.

[0090] UE 120a may process data blocks according to transmitter design 715. Transmitter design 715 may apply analog modulation (e.g., analog amplitude modulation) to the data block at block 720a, perform serial-to-parallel conversion at block 720b, perform truncated channel inversion at block 720c, perform inverse fast Fourier transform (IFFT) at block 720d, and add a cyclic prefix (CP) and perform parallel-to-serial conversion at block 720e. The resulting data may be sent to network device 110 via a carrier (e.g., a multi-access channel). In some cases, UE 120 may send parameters or gradients of a data model (e.g., a neural network) to network device 110. However, such techniques may also be applicable to other scenarios, such as distributed sensor measurements, etc.

[0091] Network device 110 may process the superimposed waveform received from multiple devices via a carrier according to receiver design 725. Network device 110 may remove the CP and perform parallel-to-serial conversion at block 730a, perform fast Fourier transform (FFT) at block 730b, perform parallel-to-serial conversion at block 730c, and average the aggregated parameters or gradients at block 730d (e.g., divide the aggregated parameters and / or gradients by the number of UE 120s (e.g., K)). In this way, network device 110 may receive one or more aggregated values (e.g., aggregated parameters, aggregated edge weights, aggregated gradients, etc.) corresponding to the aggregation of values from UE120, and average the aggregated values by dividing the aggregated values by the number of UE 120s that send data (e.g., parameters and / or gradients) on the shared channel. Network device 110 may update the parameters or gradients of the global data model based on the aggregated values or the average value, and send (e.g., broadcast) the updated parameters and / or gradients to UE 120.

[0092] The network device 110 may configure the UE 120 to identify or determine one or more parameters related to the processing or transmission of gradients carried by the uncoded uplink signal. The parameters may relate to the channel inversion coefficient. For example, the network device 110 may send a control message (e.g., an RRC message, a MAC-CE, DCI, etc.) to the UE 120, and the control message may configure the UE 120 to determine the channel inversion coefficient. The control message may include a mapping between the gradient sum power level and an associated scaling factor for determining the channel inversion coefficient. In some cases, the mapping is pre-determined in the telecommunication standard and not signaled to the UE.

[0093] It has been proposed that when calculating the channel inversion coefficient, in addition to the original channel gain, the virtual channel gain is also considered. The virtual channel gain is proportional to the reciprocal of the sum power value of the gradients. The sum power is a mathematical operation where the sum of the absolute values of each local gradient squared is calculated for each UE, as shown in Equation (1).

[0094]

[0095] where b represents the k-th gradient coefficient, and n is the UE index. Thus, for each UE n , the transmission weight is determined based on the sum power. In the case of CSI and data-aware OTA aggregation, this proposal can accelerate training convergence. For example, for UEs with similar channel gains, the UE with a larger absolute value of the gradient will have a lower virtual channel gain and will receive a larger channel inversion coefficient (e.g., higher transmit power) than other UEs.

[0096] Aspects of the present disclosure introduce signaling enhancements to support the above proposal. According to aspects of the present disclosure, the UE is configured via radio resource control (RRC) signaling, a MAC control element (MAC-CE), or downlink control information (DCI) with a mapping between the gradient sum power level and an associated scaling factor for determining the channel inversion coefficient. That is, the scaling factor for transmitting the gradient is based on the gradient sum power level. When the sum power of the gradients is small, the transmit power is low. The mapping may alternatively be defined in the telecommunication standard instead of being signaled by the network.

[0097] In one example, when the gradient sum power is greater than 5 but less than 10, the channel inversion coefficient is scaled by a factor of 1. When the gradient sum power is greater than 1 but less than 5, the channel inversion coefficient is scaled by a factor of 0.2. In this example, when the gradient sum power is below 1, the channel inversion coefficient is scaled by a factor of 0.05.

[0098] In other aspects of the present disclosure, such mapping can be specifically configured or indicated for different training tasks. For example, a federated learning training task with more gradients (or a more complex NN) can have a larger total power range. In the above embodiment, the first range is from 5 to 10. In the case of a larger number of gradients, the first range of the total power value can be from 20 to 30.

[0099] As described above, multiple rounds of aggregation occur during a federated learning task. According to aspects of the present disclosure, the UE reports its total gradient power level for each round or group of rounds. Before a certain round of gradient OTA aggregation, the UE can receive a configuration from the network (e.g., a base station) for reporting the total gradient power identified in that round. The reporting configuration can be monitored for each round of OTA aggregation or a group of OTA aggregation rounds, or once for the overall training task. Alternatively, the reporting configuration can be defined in a telecommunications standard.

[0100] In some aspects, the reporting configuration is received via a conventional UE-specific downlink channel. For example, the configuration can be received via a unicast physical downlink shared channel (PDSCH) or a physical downlink control channel (PDCCH), MAC-CE, or RRC signaling. In other aspects, the configuration is received via a group common downlink channel (e.g., a broadcast PDSCH or PDCCH).

[0101] The reporting configuration can include information such as quantization levels for the reported total gradient power. For example, the configuration can indicate certain quantization levels with a limited number of bits. The configuration can include information for scheduling uplink resources that can be used for such reporting. The uplink channel can be a conventional UE-specific uplink channel, such as a physical uplink shared channel (PUSCH), a physical uplink control channel (PUCCH), MAC-CE, or RRC signaling.

[0102] According to additional aspects of the present disclosure, the UE can receive an indication from the network of a scaling factor for determining a channel inversion coefficient. The indication can be received before a certain round of gradient OTA aggregation. The indication of the scaling factor can be associated with the reported total gradient power. The indication can be received via a conventional UE-specific downlink channel (e.g., a unicast PDSCH or PDCCH, MAC-CE, or RRC signaling). Alternatively, the indication can be received via a group common downlink channel (e.g., a broadcast PDSCH / PDCCH).

[0103] A UE group-specific scaling factor associated with the gradient sum power level can be configured to the UE group via group common DCI, MAC-CE, or RRC signaling. Multiple scaling factors can be configured for the group, such as using the mapping described previously. The scaling factor can be a power measurement result for a certain UE based on the reference signal received power (RSRP) calculated from reference signals (e.g., CSI-RS and / or synchronization signal block (SSB)) when determining the channel inversion coefficient, such as a dB value. In this case, the scaling factor can be signaled to the UE together with the signaling for triggering the reference signal. For example, a single CSI-RS can be used for a group of UEs with similar RSRP levels. The scaling factor for a UE can be included in the RRC signaling, MAC-CE, or DCI message for triggering the CSI-RS.

[0104] Figure 8 is a timing diagram showing an example federated learning process with gradient dataset awareness configuration according to various aspects of the present disclosure. In federated learning, multiple UEs communicate with a base station (as Figure 6 shown). For ease of explanation, in Figure 8 only a single UE 120 communicating with base station 110 is shown.

[0105] At time t1, base station 110 sends a reporting configuration to UE 120. The reporting configuration enables UE 120 to report the gradient sum power level to base station 110. The reporting configuration can be received via unicast physical downlink shared channel (PDSCH) or physical downlink control channel (PDCCH), MAC-CE, or RRC signaling. The reporting configuration can include information such as the quantization level for the reported gradient sum power and information for scheduling uplink resources that can be used for such reporting.

[0106] At time t2, UE 120 reports its gradient sum power. The reporting can be via a conventional UE-specific uplink channel, such as physical uplink shared channel (PUSCH), physical uplink control channel (PUCCH), MAC-CE, or RRC signaling. In response to receiving the gradient sum power from UE 120 and the gradient sum power values from other UEs participating in the federated learning task, base station 110 determines an appropriate scaling factor for each UE. At time t3, base station 110 sends the scaling factor determined for UE 120 to UE 120. In some aspects, base station 110 sends a mapping of scaling factors to gradient sum power values. In some aspects, base station 110 sends the scaling factor to a group of UEs with similar RSRP levels.

[0107] The UE 120 determines a channel inversion coefficient based on a scaling factor. Then, the UE 120 applies analog modulation and the determined channel inversion coefficient to the locally computed gradient to form an uncoded uplink signal. At time t4, the UE 120 transmits the uncoded uplink signal on the shared uplink resource for air computing of the global gradient for the federated learning task. The uplink resource is shared with other UEs participating in the federated learning task.

[0108] Figure 9 is a schematic diagram showing an example federated learning process with a gradient dataset awareness configuration, for example, performed by a user equipment (UE) according to various aspects of the present disclosure.

[0109] As Figure 9 shown, in some aspects, process 900 may include: generating a local gradient for a federated learning task (block 902). For example, a user equipment (UE) (e.g., using the controller / processor 280 and / or the memory 282) may generate a local gradient. Process 900 may further include: calculating a gradient sum power level based on the local gradient (block 904). For example, the UE (e.g., using the controller / processor 280 and / or the memory 282) may calculate the gradient sum power level. Process 900 may further include: receiving a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal (block 906). For example, the UE (e.g., using the antenna 252, DEMOD / MOD 254, MIMO detector 256, receive processor 258, controller / processor 280 and / or the memory 282) may receive the mapping. Process 900 may further include: determining a channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level (block 908). For example, the UE (e.g., using the controller / processor 280 and / or the memory 282) may determine the channel inversion coefficient.

[0110] In some aspects, process 900 may include: applying analog modulation and the channel inversion coefficient to the data block to form an uncoded uplink signal (block 910). For example, the UE (e.g., using the controller / processor 280 and / or the memory 282) may apply analog modulation and the channel inversion coefficient to the data block. Process 900 may include: transmitting the uncoded uplink signal on the shared uplink resource to the network for air computing of the global gradient for the federated learning task (block 912). For example, the UE (e.g., using the antenna 252, DEMOD / MOD 254, TX MIMO processor 266, transmit processor 264, controller / processor 280 and / or the memory 282) may transmit to the network.

[0111] Examples of embodiments are described in the following numbered clauses.

[0112] 1. A method of wireless communication performed by a user equipment (UE), comprising:

[0113] generating local gradients for a federated learning task;

[0114] calculating a gradient sum power level based on the local gradients;

[0115] receiving a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal;

[0116] determining a channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level;

[0117] applying analog modulation and the channel inversion coefficient to the data block to form the uncoded uplink signal; and

[0118] transmitting the uncoded uplink signal on a shared uplink resource to a network for air computing of global gradients for the federated learning task.

[0119] 2. The method according to clause 1, wherein the mapping is associated with the federated learning task.

[0120] 3. The method according to clause 2, wherein the mapping includes a sum power level range assigned to each of the scaling factors.

[0121] 4. The method according to clause 3, wherein the federated learning task includes a training task having a number of gradients, and an increase in the number of gradients beyond a threshold results in an increase in the sum power level range assigned to each of the scaling factors.

[0122] 5. The method according to any one of the preceding clauses, wherein the mapping is defined in a telecommunication standard.

[0123] 6. The method according to any one of the preceding clauses, further comprising:

[0124] receiving a configuration for reporting the calculated gradient sum power associated with one round of gradient air aggregation; and

[0125] reporting the calculated gradient sum power based on the configuration.

[0126] 7. The method according to clause 6, wherein receiving the configuration is performed via a UE-specific downlink channel.

[0127] 8. The method according to clause 6 or 7, wherein receiving the configuration is performed via a group common downlink channel.

[0128] 9. The method according to clauses 6 - 8, wherein the configuration indicates a quantization level for reporting the calculated gradient sum power.

[0129] 10. The method according to clauses 6 - 9, wherein the configuration includes an allocation of uplink resources for the reporting.

[0130] 11. The method according to clauses 6 - 10, wherein the configuration corresponds to multiple rounds of the gradient air aggregation.

[0131] 12. The method according to clause 11, wherein the multiple rounds include an overall federated learning task.

[0132] 13. The method according to clauses 6 - 12, wherein the reporting is performed via a UE - specific uplink channel.

[0133] 14. The method according to any one of the preceding clauses, wherein receiving the mapping is performed via a group common downlink channel associated with the selected gradient sum power.

[0134] 15. The method according to clause 14, wherein the scaling factor corresponds to a reference signal received power (RSRP) value associated with the channel inversion coefficient.

[0135] 16. The method according to clause 15, further comprising: receiving the scaling factor together with a message for triggering a reference signal associated with the RSRP value.

[0136] 17. An apparatus for wireless communication to be performed by a user equipment (UE), comprising:

[0137] a processor,

[0138] a memory coupled to the processor; and

[0139] instructions stored in the memory and operable, when executed by the processor, to cause the apparatus to:

[0140] generate local gradients for a federated learning task;

[0141] calculate a gradient sum power level based on the local gradients;

[0142] receive a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal;

[0143] Determine a channel inversion coefficient based on a scaling factor obtained from the mapping and the calculated gradient total power level;

[0144] Apply analog modulation and the channel inversion coefficient to the data block to form the uncoded uplink signal; and

[0145] Transmit the uncoded uplink signal on a shared uplink resource to the network for air computing of the global gradient for the federated learning task.

[0146] 18. The apparatus according to clause 17, wherein the mapping is associated with the federated learning task.

[0147] 19. The apparatus according to clause 18, wherein the mapping includes a total power level range assigned to each of the scaling factors.

[0148] 20. The apparatus according to clause 19, wherein the federated learning task includes a training task having a number of gradients, and the number of gradients exceeds a threshold, increasing the total power level range assigned to each of the scaling factors.

[0149] 21. The apparatus according to any one of the preceding clauses, wherein the mapping is defined in a telecommunication standard.

[0150] 22. The apparatus according to any one of the preceding clauses, wherein the processor causes the apparatus to:

[0151] Receive a configuration for reporting the calculated gradient total power associated with one round of gradient air aggregation; and

[0152] Report the calculated gradient total power based on the configuration.

[0153] 23. The apparatus according to clause 22, wherein the processor causes the apparatus to receive the configuration via a UE-specific downlink channel.

[0154] 24. The apparatus according to clause 22 or 23, wherein the processor causes the apparatus to receive the configuration via a group common downlink channel.

[0155] 25. The apparatus according to clauses 22-24, wherein the configuration indicates a quantization level for reporting the calculated gradient total power.

[0156] 26. The apparatus according to clauses 22-25, wherein the configuration includes an allocation of uplink resources for the reporting.

[0157] 27. The apparatus according to clauses 22-26, wherein the configuration corresponds to multiple rounds of the gradient air aggregation.

[0158] 28. The apparatus according to clause 27, wherein the multiple rounds include an overall federated learning task.

[0159] 29. The apparatus according to clauses 22-28, wherein the processor causes the apparatus to report via a UE-specific uplink channel.

[0160] 30. The apparatus according to any of the preceding clauses, wherein the processor causes the apparatus to receive the mapping via a group common downlink channel associated with the selected gradient sum power.

[0161] The foregoing disclosure provides illustration and description, but is not intended to be exhaustive or to limit the various aspects to the precise forms disclosed. Modifications and variations may be made in light of the above disclosure, or may be obtained from practice of these aspects.

[0162] As used, the term "component" is intended to be broadly construed as hardware, firmware, and / or a combination of hardware and software. As used, a processor is implemented in hardware, firmware, and / or a combination of hardware and software.

[0163] Some aspects are described in connection with a threshold. As used, depending on the context, meeting a threshold may refer to a value greater than the threshold, a value greater than or equal to the threshold, a value less than the threshold, a value less than or equal to the threshold, a value equal to the threshold, a value not equal to the threshold, and so on.

[0164] It will be apparent that the described systems and / or methods may be implemented in different forms of hardware, firmware, and / or combinations of hardware and software. The actual specific control hardware or software code used to implement these systems and / or methods does not limit these aspects. Accordingly, the operations and behavior of the systems and / or methods are described without reference to specific software code - it should be understood that the software and hardware can be designed to implement the systems and / or methods at least in part based on this description.

[0165] Although specific combinations of features are recited in the claims and / or disclosed in the specification, these combinations are not intended to limit the disclosure of each aspect. Indeed, many of these features may be combined in ways not specifically recited in the claims and / or disclosed in the specification. Although each dependent claim listed below may directly depend on only one claim, the disclosure of each aspect includes each dependent claim in combination with every other claim in the set of claims. A phrase referring to "at least one" of a list of items refers to any combination of those items, including a single member. As an example, "at least one of a, b, or c" is intended to cover: a, b, c, a-b, a-c, b-c, and a-b-c, as well as any combination with multiples of the same element (e.g., a-a, a-a-a, a-a-b, a-a-c, a-b-b, a-c-c, b-b, b-b-b, b-b-c, c-c, and c-c-c, or any other ordering of a, b, and c).

[0166] Any element, operation, or instruction used should not be construed as critical or essential unless expressly so stated. Further, as used herein, the articles "a" and "an" are intended to include one or more items and may be used interchangeably with "one or more." Further, as used herein, the terms "set" and "group" are intended to include one or more items (e.g., related items, unrelated items, or a combination of related and unrelated items, etc.) and may be used interchangeably with "one or more." Where only one item is intended, the phrase "only one" or similar language is used. Further, as used herein, the terms "has," "have," "having," etc. are intended to be open-ended terms. Further, unless expressly stated otherwise, the phrase "based on" is intended to mean "at least partially based on."

Claims

1. A method of wireless communication performed by a user equipment (UE), comprising: generating local gradients for a federated learning task; calculating a gradient sum power level based on the local gradients; receiving a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal; determining the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level; applying analog modulation and the channel inversion coefficient to the data block to form the uncoded uplink signal; and transmitting the uncoded uplink signal on a shared uplink resource to a network for air computation of global gradients for the federated learning task.

2. The method according to claim 1, wherein, the mapping is associated with the federated learning task.

3. The method according to claim 2, wherein, the mapping includes a sum power level range assigned to each of the scaling factors.

4. The method according to claim 3, wherein, the federated learning task includes a training task having a number of gradients, and an increase in the number of gradients beyond a threshold results in an increase in the sum power level range assigned to each of the scaling factors.

5. The method according to claim 1, wherein, the mapping is defined in a telecommunication standard.

6. The method according to claim 1, further comprising: receiving a configuration for reporting the calculated gradient sum power associated with one round of gradient air aggregation; and reporting the calculated gradient sum power based on the configuration.

7. The method according to claim 6, wherein, receiving the configuration is performed via a UE-specific downlink channel.

8. The method according to claim 6, wherein, receiving the configuration is performed via a group common downlink channel.

9. The method according to claim 6, wherein, the configuration indicates a quantization level for reporting the calculated gradient sum power.

10. The method according to claim 6, wherein, the configuration includes an allocation of an uplink resource for the reporting.

11. The method according to claim 6, wherein, the configuration corresponds to multiple rounds of the gradient air aggregation.

12. The method according to claim 11, wherein, the multiple rounds include an overall federated learning task.

13. The method according to claim 6, wherein, the reporting is performed via a UE-specific uplink channel.

14. The method according to claim 1, wherein, receiving the mapping is performed via a group common downlink channel associated with a selected gradient sum power.

15. The method according to claim 14, wherein, the scaling factor corresponds to a reference signal received power (RSRP) value associated with the channel inversion coefficient.

16. The method according to claim 15, further comprising: receiving the scaling factor together with a message for triggering a reference signal associated with the RSRP value.

17. An apparatus for wireless communication performed by a user equipment (UE), comprising: a processor, A memory, coupled to the processor; and instructions stored in the memory and operable, when executed by the processor, to cause the device to: generate local gradients for a federated learning task; calculate a gradient sum power level based on the local gradients; receive a mapping between the gradient sum power level and a scaling factor for a channel inversion coefficient to process a data block into an uncoded uplink signal; determine the channel inversion coefficient based on the scaling factor obtained from the mapping and the calculated gradient sum power level; apply analog modulation and the channel inversion coefficient to the data block to form the uncoded uplink signal; and transmit the uncoded uplink signal over a shared uplink resource to a network for air computation of global gradients for the federated learning task.

18. The apparatus according to claim 17, wherein the mapping is associated with the federated learning task.

19. The apparatus according to claim 18, wherein the mapping includes a range of sum power levels assigned to each of the scaling factors.

20. The apparatus according to claim 19, wherein the federated learning task includes a training task having a number of gradients, and an increase in the number of gradients beyond a threshold results in an increase in the range of sum power levels assigned to each of the scaling factors.

21. The apparatus according to claim 17, wherein the mapping is defined in a telecommunication standard.

22. The apparatus according to claim 17, wherein the processor causes the device to: receive a configuration for reporting the calculated gradient sum power associated with one round of gradient air aggregation; and report the calculated gradient sum power based on the configuration.

23. The apparatus according to claim 22, wherein the processor causes the device to receive the configuration via a UE-specific downlink channel.

24. The apparatus according to claim 22, wherein the processor causes the device to receive the configuration via a group common downlink channel.

25. The apparatus according to claim 22, wherein the configuration indicates a quantization level for reporting the calculated gradient sum power.

26. The apparatus according to claim 22, wherein the configuration includes an allocation of uplink resources for the reporting.

27. The apparatus according to claim 22, wherein the configuration corresponds to multiple rounds of the gradient air aggregation.

28. The apparatus according to claim 27, wherein the multiple rounds include an overall federated learning task.

29. The apparatus according to claim 22, wherein the processor causes the device to report via a UE-specific uplink channel.

30. The apparatus according to claim 17, wherein the processor causes the device to receive the mapping via a group common downlink channel associated with a selected gradient sum power.

Images