Client independent learning and zero sample adaptation for federated domain generalization

Abstract

Example implementations include methods, apparatuses, and computer-readable media including federated learning by a federated client device, the example implementations including identifying client-invariant information of a neural network for performing machine learning (ML) tasks in a first domain known to a federated server. The specific implementations further include sending the client invariant information to the federated server, the federated server is configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client invariant information and other client invariant information of another neural network for performing the ML task in a second domain known to the federated server.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims U.S. patent application serial number 18 / 238,998, entitled “CLIENT-AGNOSTICLEARNING AND ZERO-SHOT ADAPTATION FOR FEDERATED DOMAIN GENERALIZATION,” filed by Yang et al. on August 28, 2023, and U.S. patent application serial number 63 / 411,034, entitled “SYSTEMS AND METHODS FOR CLIENT-AGNOSTICLEARNING AND ZERO-SHOT ADAPTATION FOR FEDERATED DOMAIN GENERALIZATION,” filed by Fee et al. on September 28, 2022, which are hereby expressly incorporated herein by reference in their entirety. Background Art Technical Field

[0003] The present disclosure generally relates to federated learning.

[0004] introduction

[0005] Wireless communication systems are widely deployed to provide a variety of telecommunication services, such as telephony, video, data, messaging, and broadcast. Typical wireless communication systems may employ multiple access technologies capable of supporting communication with multiple users by sharing available system resources. 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, and time division synchronous code division multiple access (TD-SCDMA) systems.

[0006] These multiple access technologies have been adopted in various telecommunication standards to provide a common protocol that enables different wireless devices to communicate at a city, country, region, and even global level. An example telecommunication standard is 5G New Radio (NR). 5G NR is part of the continuous mobile broadband evolution promulgated by the Third Generation Partnership Project (3GPP) to meet new requirements associated with latency, reliability, security, scalability, such as with the Internet of Things (IoT), and other requirements. 5G NR includes services associated with enhanced mobile broadband (eMBB), massive machine type communication (mMTC), and ultra-reliable low latency communication (URLLC). Certain aspects of 5G NR may be based on the 4G Long Term Evolution (LTE) standard. Further improvements to 5G NR technology are needed. Summary of the invention

[0007] The following presents a simplified summary of one or more aspects in order to provide a basic understanding of these aspects. This summary is not an extensive overview of all contemplated aspects, and is neither intended to identify key or important elements of all aspects, nor to describe the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to a more detailed description presented later.

[0008] In some aspects, the technology described herein relates to a federated client device, comprising: a memory; and a processor, the processor being coupled to the memory and configured to: identify client-invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server; and send the client-invariant information to the federated server, the federated server being configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client-invariant information and other client-invariant information of another neural network for performing the ML task in a second domain known to the federated server.

[0009] In some aspects, the technology described herein relates to a method of wireless communication by a user equipment (UE), including: identifying client-invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server; and sending the client-invariant information to the federated server, the federated server being configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client-invariant information and other client-invariant information of another neural network for performing the ML task in a second domain known to the federated server.

[0010] In some aspects, the technology described herein relates to a federated client device, comprising: a memory; and a processor, the processor being coupled to the memory and configured to: receive a neural network for performing a machine learning task from a federated server, the neural network being generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server; and generate inferred information in a domain unknown to the federated server via the neural network based on an alpha generator added to a batch normalization layer of the neural network.

[0011] In some aspects, the technology described herein relates to a method comprising: receiving a neural network for performing a machine learning task from a federated server, the neural network generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server; and generating inferred information in a domain unknown to the federated server via the neural network based on an alpha generator added to a batch normalization layer of the neural network.

[0012] To achieve the foregoing and related ends, one or more aspects include the features fully described below and particularly pointed out in the claims. The following description and the accompanying drawings set forth in detail some illustrative features of one or more aspects. However, these features are merely indicative of some of the various ways in which the principles of the various aspects may be employed, and this specification is intended to include all such aspects and their equivalents. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The disclosed aspects will be described below in conjunction with the accompanying drawings, which are provided to illustrate and not to limit the disclosed aspects, wherein like designations represent like elements, wherein dashed lines may indicate optional elements, and wherein:

[0014] Figure 1 is a diagram illustrating an example of a wireless communication system and access network according to some aspects of the present disclosure.

[0015] Figure 2A is a diagram illustrating an example of a first 5G / NR frame according to some aspects of the present disclosure.

[0016] Figure 2B is a diagram illustrating an example of DL channels within a 5G / NR subframe according to some aspects of the present disclosure.

[0017] Figure 2C is a diagram illustrating an example of a second 5G / NR frame according to some aspects of the present disclosure.

[0018] Figure 2Dis a diagram illustrating an example of UL channels within a 5G / NR subframe according to some aspects of the present disclosure.

[0019] Figure 3 is a diagram illustrating an example of a base station and a UE in an access network according to some aspects of the present disclosure.

[0020] Figure 4 is a diagram illustrating an example of a decomposed base station architecture according to some aspects of the present disclosure.

[0021] Figure 5 is a diagram illustrating an example of communications of network entities and devices according to some aspects of the present disclosure.

[0022] Figure 6 is a diagram illustrating an example of a Federated Domain Generalization (FDG) architecture in accordance with some aspects of the present disclosure.

[0023] Figure 7 is a diagram illustrating example communications and components of a federated server and a federated client device.

[0024] Figure 8 is a diagram illustrating an example of client-agnostic learning with data augmentation according to some aspects of the present disclosure.

[0025] Fig. 9 is a diagram illustrating an example of an alpha generator according to some aspects of the present disclosure.

[0026] Fig.10 is a diagram illustrating an example of a hardware implementation of a federated client employing a processing system according to some aspects of the present disclosure.

[0027] Fig.11 is a flow chart of an example of a method of federated learning performed by a federated learning system.

[0028] Fig.12 is a flow chart of an example of a method of federated learning performed by a federated client.

[0029] Fig.13 is a diagram illustrating an example of a hardware implementation of a federated server employing a processing system according to some aspects of the present disclosure. DETAILED DESCRIPTION

[0030] Various aspects are now described with reference to the accompanying drawings. In the following description, for the purpose of explanation, in order to provide a thorough understanding of one or more aspects, numerous specific details are described. However, it is apparent that such aspects can be practiced without these specific details.

[0031] The detailed description set forth below in conjunction with the accompanying drawings is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described herein may be practiced. In order to provide a thorough understanding of the various concepts, the detailed description includes specific details. However, it will be apparent to one of ordinary skill in the art that these concepts may be practiced without these specific details. In some instances, structures and components are shown in block diagram form in order to avoid diluting such concepts.

[0032] Several aspects of telecommunication systems will now be presented with reference to various apparatus and methods. These apparatus and methods will be described in the following detailed description and illustrated in the accompanying drawings by various blocks, components, circuits, processes, algorithms, etc. (collectively referred to as "elements"). These elements may be implemented using electronic hardware, computer software, or any combination thereof. Whether such elements are implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system.

[0033] As an example, an element, or any part of an element, or any combination of elements may be implemented as a "processing system", which includes one or more processors. Examples of processors include microprocessors, microcontrollers, graphics processing units (GPUs), central processing units (CPUs), application processors, digital signal processors (DSPs), reduced instruction set computing (RISC) processors, systems on chip (SoCs), baseband processors, field programmable gate arrays (FPGAs), programmable logic devices (PLDs), state machines, gating logic, discrete hardware circuits, and other suitable hardware configured to perform various functionalities described throughout the present disclosure. One or more processors in a processing system may execute software. Software should be broadly interpreted as meaning instructions, instruction sets, codes, code segments, program codes, programs, subroutines, software components, applications, software applications, software packages, routines, subroutines, objects, executables, threads of execution, procedures, functions, etc., whether it is described in software, firmware, middleware, microcode, hardware description languages, or other terms. As used herein, a processor, at least one processor, and / or one or more processors configured to perform or operable to perform multiple actions (alone or in combination) is intended to include at least two different processors capable of performing different subsets, overlapping subsets, or non-overlapping subsets of the multiple actions, or a single processor capable of performing all of the multiple actions. In a non-limiting example of multiple processors capable of performing different actions combined in the multiple actions, a description of a processor, at least one processor, and / or one or more processors configured to perform or operable to perform actions X, Y, and Z may include at least a first processor configured to perform or operable to perform a first subset of X, Y, and Z (e.g., perform X) and at least a second processor configured to perform or operable to perform a second subset of X, Y, and Z (e.g., perform Y and Z). Alternatively, the first processor, the second processor, and the third processor may be configured to perform or operable to perform corresponding actions in actions X, Y, and Z, respectively. It should be understood that any combination of one or more processors may each be configured to perform or operable to perform any one of the multiple actions or any combination of the multiple actions.

[0034] Accordingly, in one or more examples, the functions described may be implemented in hardware, software, or any combination thereof. If implemented in software, the functions may be stored or encoded on a computer-readable medium as one or more instructions or codes. Computer-readable media include computer storage media, which may be referred to as non-transitory computer-readable media. Non-transitory computer-readable media may not include transient signals. The storage medium may be any available medium that can be accessed by a computer. By way of example and not limitation, such computer-readable media may include random access memory (RAM), read-only memory (ROM), electrically erasable programmable ROM (EEPROM), optical disk storage, magnetic disk storage, other magnetic storage devices, a combination of the aforementioned types of computer-readable media, or any other medium that can be used to store computer executable code in the form of instructions or data structures that can be accessed by a computer.

[0035] Federated learning (FL) refers to a distributed machine learning technique in which multiple decentralized nodes storing local data samples can train a global machine learning (ML) model without exchanging the data samples themselves between the nodes to perform the training. The FL framework may include a centralized aggregation server and participating FL client devices (i.e., participants or nodes such as UEs). In one aspect or specific implementation, the FL framework enables FL client devices to learn a global ML model by passing the learned information to a central aggregation server or coordinator, which can be configured to communicate with various FL devices and coordinate the learning framework. In some aspects, FL has been studied to obtain the benefits of indirectly training a global model using a distributed dataset while ensuring the privacy of users and reducing communication overhead issues.

[0036] In addition, Federated Domain Generalization (FDG) has been explored to advance FL by collaboratively learning a generalized global model from distributed source domains and deploying the global model directly to new clients in unseen domains (i.e., domains unknown to the central coordinator). However, in some aspects, FDG implementations have shown unsatisfactory accuracy when building a global model with local client models from different domains while keeping data private and providing low generalization within unseen domains. Furthermore, many of the techniques proposed to address the low generalization of FDG implementations conflict with the privacy benefits and privacy goals of federated learning.

[0037] Thus, in some aspects, federated client devices may be configured to achieve federated domain generalization without compromising user privacy. As described in detail herein, federated client devices (e.g., UEs) may collaboratively learn a generalized global model under the coordination of a federated server device, and the global model may be used by other federated client devices associated with domains unknown to the federated server device. Accordingly, in some aspects, a federated domain system may be configured to learn a global model with improved accuracy for performing machine learning tasks in unknown domains without requiring personally identifiable information and / or other private / confidential information to be shared by the federated client devices used to learn the global model.

[0038] Figure 1 1 is a diagram illustrating an example of a wireless communication system and an access network 100. The wireless communication system (also referred to as a wireless wide area network (WWAN)) includes a base station 102, a UE 104, an evolved packet core (EPC) 160, and another core network 190 (e.g., a 5G core (5GC)). The base station 102 may include a macro cell (a high-power cellular base station) or a small cell (a low-power cellular base station). A macro cell includes a base station. A small cell includes a femto cell, a pico cell, and a micro cell.

[0039] In some aspects, UE 104 may be a federated client device within a federated domain generalization (FDG) system. For example, UE 104 may include a federated learning component 140 configured to perform client-independent learning to determine client-invariant parameters of a model, and to send the client-invariant parameters to a federated server device to be aggregated with other client-invariant parameters to generate a global model to be used in a domain unknown to the federated server device. In addition, the federated learning component 140 may be configured to receive the global model from the federated server device. Furthermore, in some aspects, the federated learning component 140 may be associated with a domain unknown to the federated server device and applied to zero-shot adaptation to determine inferences using the global model.

[0040] The base station 102 configured for 4G LTE (collectively referred to as the Evolved Universal Mobile Telecommunications System (UMTS) Terrestrial Radio Access Network (E-UTRAN)) can interface with the EPC 160 via a first backhaul link 132 (e.g., an S1 interface). The base station 102 configured for 5G NR (collectively referred to as the Next Generation RAN (NG-RAN)) can interface with the core network 190 via a second backhaul link 184. Among other functions, the base station 102 can also perform one or more of the following functions: delivery of user data, radio channel encryption and decryption, integrity protection, header compression, mobility control functions (e.g., handover, dual connectivity), inter-cell interference coordination, connection setup and release, load balancing, distribution of non-access stratum (NAS) messages, NAS node selection, synchronization, radio access network (RAN) sharing, multimedia broadcast multicast service (MBMS), subscriber and equipment tracking, RAN information management (RIM), paging, positioning, and delivery of warning messages. The base stations 102 may communicate with each other directly or indirectly (eg, via the EPC 160 or the core network 190) over a third backhaul link 134 (eg, an X2 interface). The third backhaul link 134 may be wired or wireless.

[0041] Base station 102 can communicate wirelessly with UE 104. Each of base stations 102 can provide communication coverage for a corresponding geographic coverage area 110. There may be overlapping geographic coverage areas 110. For example, a small cell 102a can have a coverage area 110a that overlaps with the coverage area 110 of one or more macro base stations 102. A network including both small cells and macro cells can be referred to as a heterogeneous network. A heterogeneous network can also include a home evolved Node B (eNB) (HeNB), which can provide services to a restricted group called a closed subscriber group (CSG). A communication link 120 between base station 102 and UE 104 can include an uplink (UL) (also known as a reverse link) transmission from UE 104 to base station 102 or a downlink (DL) (also known as a forward link) transmission from base station 102 to UE 104. The communication link 120 can use multiple input multiple output (MIMO) antenna technology, including spatial multiplexing, beamforming, or transmit diversity. The communication link can be through one or more carriers. The base station 102 / UE 104 may use spectrum of up to Y MHz (e.g., 5 MHz, 10 MHz, 15 MHz, 20 MHz, 100 MHz, 400 MHz, etc.) bandwidth for each carrier allocated in the carrier aggregation for a total of up to Yx MHz (x component carriers) for transmission in each direction. These carriers may or may not be adjacent to each other. The allocation of carriers may be asymmetric with respect to DL and UL (e.g., more or fewer carriers may be allocated for DL ​​compared to UL). The component carriers may include a primary component carrier and one or more secondary component carriers. The primary component carrier may be referred to as a primary cell (PCell) and the secondary component carrier may be referred to as a secondary cell (SCell).

[0042] Some UEs 104 may communicate with each other using a device-to-device (D2D) communication link 158. The D2D communication link 158 may use DL / UL WWAN spectrum. The D2D communication link 158 may use one or more sidelink channels, such as a physical sidelink broadcast channel (PSBCH), a physical sidelink discovery channel (PSDCH), a physical sidelink shared channel (PSSCH), and a physical sidelink control channel (PSCCH). The D2D communication may be through various wireless D2D communication systems, such as, for example, FlashLinQ, WiMedia, Bluetooth, ZigBee, Wi-Fi based on IEEE 802.10 standards, LTE, or NR.

[0043] The wireless communication system may also include a Wi-Fi access point (AP) 150 that communicates with a Wi-Fi station (STA) 152 via a communication link 154 in the 5 GHz unlicensed spectrum. When communicating in the unlicensed spectrum, the STA 152 / AP 150 may perform a clear channel assessment (CCA) to determine whether the channel is available prior to communication.

[0044] The small cell 102a may operate in a licensed or unlicensed spectrum. When operating in an unlicensed spectrum, the small cell 102a may employ NR and use the same 5 GHz unlicensed spectrum as that used by the Wi-Fi AP 150. The small cell 102a employing NR in the unlicensed spectrum may boost the coverage of the access network or increase the capacity of the access network.

[0045] Whether a small cell 102a or a large cell (e.g., a macro base station), the base station 102 may include and / or be referred to as an eNB, a gNodeB (gNB), or another type of base station. Some base stations (such as gNB 180) may operate in one or more frequency bands within the electromagnetic spectrum. The electromagnetic spectrum is typically subdivided into various categories, frequency bands, channels, etc. based on frequency / wavelength. In 5GNR, two initial operating bands have been identified as frequency ranges designated FR1 (416 MHz to 7.125 GHz) and FR2 (24.25 GHz to 52.6 GHz). The frequencies between FR1 and FR2 are typically referred to as mid-band frequencies. Although a portion of FR1 is greater than 6 GHz, in various documents and articles, FR1 is typically (interchangeably) referred to as the "below 6 GHz" band. A similar naming issue sometimes arises with FR2, which is often referred to (interchangeably) as the “millimeter wave” band in various documents and articles, although it is different from the extremely high frequency (EHF) band (30GHz to 300GHz) that the International Telecommunication Union (ITU) has identified as the “millimeter wave” (mmW) band.

[0046] In view of the above aspects, unless otherwise specifically stated, it should be understood that if used in this article, the term "below 6 GHz" or the like can broadly represent frequencies that can be less than 6 GHz, can be within FR1, or can include mid-band frequencies. In addition, unless otherwise specifically stated, it should be understood that if the term "millimeter wave" or the like is used in this article, it can broadly represent frequencies that can include mid-band frequencies, can be within FR2, or can be within the EHF band. Communications using mmW radio frequency bands have extremely high path loss and short range. The mmW base station 180 can utilize beamforming 182 with the UE 104 to compensate for this path loss and short range. The base station 180 and the UE 104 can each include multiple antennas, such as antenna elements, antenna panels, or antenna arrays to facilitate beamforming.

[0047] The base station 180 may transmit beamformed signals in one or more transmit directions 182a to the UE 104. The UE 104 may receive beamformed signals from the base station 180 in one or more receive directions 182b. The UE 104 may also transmit beamformed signals to the base station 180 in one or more transmit directions. The base station 180 may receive beamformed signals from the UE 104 in one or more receive directions. The base station 180 / UE 104 may perform beam training to determine the best receive direction and transmit direction for each of the base station 180 / UE 104. The transmit direction and receive direction of the base station 180 may be the same or different. The transmit direction and receive direction of the UE 104 may be the same or different.

[0048] The EPC 160 may include a mobility management entity (MME) 162, other MMEs 164, a serving gateway 166, a multimedia broadcast multicast service (MBMS) gateway 168, a broadcast multicast service center (BM-SC) 170, and a packet data network (PDN) gateway 172. The MME 162 may communicate with a home subscriber server (HSS) 174. The MME 162 is a control node that handles signaling between the UE 104 and the EPC 160. Generally speaking, the MME 162 provides bearer and connection management. All user Internet Protocol (IP) packets are passed through the serving gateway 166, which itself is connected to the PDN gateway 172. The PDN gateway 172 provides UE IP address allocation and other functions. The PDN gateway 172 and the BM-SC 170 are connected to IP services 176. The IP services 176 may include the Internet, an intranet, an IP multimedia subsystem (IMS), a PS streaming service, or other IP services. The BM-SC 170 may provide functions for MBMS user service configuration and delivery. The BM-SC 170 may be used as an entry point for content provider MBMS transmissions, may be used to authorize and initiate MBMS bearer services in a public land mobile network (PLMN), and may be used to schedule MBMS transmissions. The MBMS Gateway 168 may be used to distribute MBMS services to base stations 102 belonging to a multicast broadcast single frequency network (MBSFN) area of ​​a broadcast specific service, and may be responsible for session management (start / stop) and for collecting eMBMS related billing information.

[0049] The core network 190 may include an access and mobility management function (AMF) 192, other AMFs 193, a session management function (SMF) 194, and a user plane function (UPF) 195. The AMF 192 may communicate with a unified data management (UDM) 196. The AMF 192 is a control node that handles signaling between the UE 104 and the core network 190. Generally speaking, the AMF 192 provides QoS flow and session management. All user Internet Protocol (IP) packets are passed through the UPF 195. The UPF 195 provides UE IP address allocation and other functions. The UPF 195 is connected to an IP service 197. The IP service 197 may include the Internet, an intranet, an IP multimedia subsystem (IMS), a PS streaming service, or other IP services.

[0050] Base station 102 may include and / or be referred to as a gNB, Node B, eNB, access point, base transceiver station, radio base station, radio transceiver, transceiver function, basic service set (BSS), extended service set (ESS), transmit receive point (TRP), or some other suitable term. Base station 102 provides an access point to EPC 160 or core network 190 for UE 104. Examples of UE 104 include satellite phones, cellular phones, smart phones, session initiation protocol (SIP) phones, laptops, personal digital assistants (PDAs), satellite radios, global positioning systems, multimedia devices, video devices, digital audio players (e.g., MP3 players), cameras, game consoles, tablet devices, smart devices, wearable devices, vehicles, electric meters, gas pumps, large or small kitchen appliances, health care devices, implants, sensors / actuators, displays, or any other similar functional devices. Some UEs 104 may be referred to as IoT devices (e.g., parking meters, gas pumps, ovens, vehicles, heart monitors, etc.). UE 104 may also be referred to as a station, a mobile station, a subscriber station, a mobile unit, a subscriber unit, a wireless unit, a remote unit, a mobile device, a wireless device, a wireless communication device, a remote device, a mobile subscriber station, an access terminal, a mobile terminal, a wireless terminal, a remote terminal, a handset, a user agent, a mobile client, a client, or some other suitable terminology.

[0051] Although the following description may focus on 5G NR, the concepts described herein may be applicable to other similar areas such as 6G, LTE, LTE-A, CDMA, GSM, and other wireless technologies.

[0052] FIG. 2A to FIG. 2D Included are example diagrams 200, 230, 250, and 280 illustrating example structures that may be used for wireless communications by a base station 102 and a UE 104 (e.g., for 5G NR communications). Figure 2A200 is a diagram illustrating an example of a first subframe within a 5G / NR frame structure. Figure 2B FIG230 is a diagram illustrating an example of DL channels within a 5G / NR subframe. Figure 2C 250 is a diagram illustrating an example of a second subframe within a 5G / NR frame structure. Figure 2D 280 is a diagram illustrating an example of UL channels within a 5G / NR subframe. The 5G / NR frame structure can be FDD, where for a particular set of subcarriers (carrier system bandwidth), subframes within that subcarrier set are dedicated to DL or UL, or can be TDD, where for a particular set of subcarriers (carrier system bandwidth), subframes within that subcarrier set are dedicated to DL and UL. Figure 2A , Figure 2C In the example provided, the 5G NR frame structure is assumed to be TDD, where subframe 4 is configured with slot format 28 (most of which are DL), where D is DL, U is UL, and X is flexible between DL / UL, and subframe 3 is configured with slot format 34 (most of which are UL). Although subframes 3 and 4 are shown as having slot formats 34 and 28, respectively, any particular subframe may be configured with any of the various available slot formats 0-61. Slot formats 0 and 1 are full DL and full UL, respectively. Other slot formats 2-61 include a mix of DL, UL, and flexible symbols. The UE is configured with the slot format through the received slot format indicator (SFI) (dynamically configured through DL control information (DCI) or semi-statically / statically configured through radio resource control (RRC) signaling). Note that the description presented herein also applies to the 5G / NR frame structure for TDD.

[0053] Other wireless communication technologies may have different frame structures or different channels. A frame (10ms) may be divided into 10 equally sized subframes (1ms). Each subframe may include one or more time slots. A subframe may also include a microslot, which may include 7, 4, or 2 symbols. Each time slot may include 7 or 14 symbols, depending on the time slot configuration. For time slot configuration 0, each time slot may include 14 symbols, and for time slot configuration 1, each time slot may include 7 symbols. The symbols on the DL may be cyclic prefix (CP) OFDM (CP-OFDM) symbols. The symbols on the UL may be CP-OFDM symbols (for high throughput scenarios) or discrete Fourier transform (DFT) spread OFDM (DFT-s-OFDM) symbols (also known as single carrier frequency division multiple access (SC-FDMA) symbols) (for power-limited scenarios; limited to single stream transmission). The number of time slots in a subframe is based on the time slot configuration and parameter set (numerology). For slot configuration 0, different parameter sets μ0 to 5 allow 1, 2, 4, 8, 16 and 32 slots per subframe, respectively. For slot configuration 1, different parameter sets 0 to 2 allow 2, 4 and 8 slots per subframe, respectively. For slot configuration 0 and parameter set μ, there are 14 symbols per slot and 2μ slots per subframe. The subcarrier spacing and symbol length / duration are a function of the parameter set. The subcarrier spacing can be equal to 2 μ *15kHz, where μ is parameter set 0 to 5. Therefore, the subcarrier spacing for parameter set μ=0 is 15kHz, and the subcarrier spacing for parameter set μ=5 is 480kHz. The symbol length / duration is inversely related to the subcarrier spacing. FIG. 2A to FIG. 2D An example is provided for slot configuration 0 with 14 symbols per slot and parameter set μ=0 with 1 slot per subframe. The subcarrier spacing is 15 kHz and the symbol duration is about 66.7 μs.

[0054] A resource grid may be used to represent the frame structure. Each slot includes a resource block (RB) (also referred to as a physical RB (PRB)) extending over 12 consecutive subcarriers. The resource grid is divided into a number of resource elements (REs). The number of bits carried by each RE depends on the modulation scheme.

[0055] like Figure 2A As illustrated, some of the REs carry reference (pilot) signals (RS) for the UE. The RS may include a demodulation RS (DM-RS) (indicated as Rx for a particular configuration, where 100x is the port number, but other DM-RS configurations are also possible) and a channel state information reference signal (CSI-RS) for channel estimation at the UE. The RS may also include a beam measurement RS (BRS), a beam refinement RS (BRRS), and a phase tracking RS (PT-RS).

[0056] Figure 2B Examples of various DL channels within a subframe of a frame are illustrated. The physical downlink control channel (PDCCH) carries DCI within one or more CCEs, each CCE includes nine RE groups (REGs), and each REG includes four consecutive REs in an OFDM symbol. The primary synchronization signal (PSS) may be within symbol 2 of a specific subframe of the frame. The PSS is used by UE 104 to determine the subframe / symbol timing and the physical layer identification. The secondary synchronization signal (SSS) may be within symbol 4 of a specific subframe of the frame. The SSS is used by the UE to determine the physical layer cell identification group number and the radio frame timing. Based on the physical layer identification and the physical layer cell identification group number, the UE can determine the physical cell identifier (PCI). Based on the PCI, the UE can determine the location of the aforementioned DM-RS. The physical broadcast channel (PBCH) carrying the master information block (MIB) can be logically grouped with the PSS and SSS to form a synchronization signal (SS) / PBCH block (SSB). The MIB provides the system frame number (SFN) and the number of RBs in the system bandwidth. The Physical Downlink Shared Channel (PDSCH) carries user data, broadcast system information not sent through the PBCH (such as System Information Blocks (SIBs)), and paging messages.

[0057] like Figure 2C As illustrated, some of the REs carry DM-RS (indicated as R for a particular configuration, but other DM-RS configurations are possible) for channel estimation at the base station. The UE may send DM-RS for a physical uplink control channel (PUCCH) and DM-RS for a physical uplink shared channel (PUSCH). The PUSCH DM-RS may be sent in the first or first two symbols of the PUSCH. Depending on whether a short PUCCH or a long PUCCH is sent and depending on the specific PUCCH format used, the PUCCH DM-RS may be sent in different configurations. Although not shown, the UE may send a sounding reference signal (SRS). The SRS may be used by the base station for channel quality estimation to enable frequency-dependent scheduling of the UL.

[0058] Figure 2D Examples of various UL channels within a subframe of a frame are illustrated. The PUCCH may be located at a position as indicated in one configuration. The PUCCH carries uplink control information (UCI), such as scheduling requests, channel quality indicator (CQI), precoding matrix indicator (PMI), rank indicator (RI), and HARQ ACK / NACK feedback. The PUSCH carries data, and may additionally be used to carry a buffer status report (BSR), a power headroom report (PHR), or UCI.

[0059] Figure 31 is a block diagram of a base station 310 (e.g., base station 102 / 180) communicating with a user equipment (UE) 350 (e.g., UE 104) in an access network. In the DL, IP packets from EPC 160 may be provided to a controller / processor 375. Controller / processor 375 implements layer 3 and layer 2 functionality. Layer 3 includes a radio resource control (RRC) layer, and layer 2 includes a service data adaptation protocol (SDAP) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, and a medium access control (MAC) layer. The controller / processor 375 provides RRC layer functionality associated with broadcasting of system information (such as MIB, SIB), RRC connection control (such as RRC connection paging, RRC connection establishment, RRC connection modification and RRC connection release), inter-radio access technology (RAT) mobility, and measurement configuration for UE measurement reporting; PDCP layer functionality associated with header compression / decompression, security (encryption, decryption, integrity protection, integrity verification) and handover support functions; RLC layer functionality associated with delivery of upper layer packet data units (PDUs), error correction through ARQ, concatenation, segmentation and reassembly of RLC service data units (SDUs), re-segmentation of RLC data PDUs and reordering of RLC data PDUs; and MAC layer functionality associated with mapping between logical channels and transport channels, multiplexing of MAC SDUs onto transport blocks (TBs), demultiplexing of MAC SDUs from TBs, scheduling information reporting, error correction through HARQ, priority handling and logical channel prioritization.

[0060] The transmit (TX) processor 316 and receive (RX) processor 370 implement layer 1 functionality associated with various signal processing functions. Layer 1 (which includes the physical (PHY) layer) may include error detection on the transport channel, forward error correction (FEC) coding / decoding of the transport channel, interleaving, rate matching, mapping to the physical channel, modulation / demodulation of the physical channel, and MIMO antenna processing. The TX processor 316 handles the mapping to the signal constellation based on various modulation schemes (such as binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), M-phase shift keying (M-PSK), M-quadrature amplitude modulation (M-QAM)). The decoded and modulated symbols can then be divided into parallel streams. Each stream can then be mapped to an OFDM subcarrier, multiplexed with a reference signal (such as a pilot) in the time domain or frequency domain, and then combined together using an inverse fast Fourier transform (IFFT) to produce a physical channel carrying a time domain OFDM symbol stream. The OFDM stream is spatially precoded to produce multiple spatial streams. Channel estimates from a channel estimator 374 may be used to determine coding and modulation schemes, as well as for spatial processing. The channel estimate may be derived from a reference signal or channel condition feedback sent by the UE 104. Each spatial stream may then be provided to a different antenna 320 via a separate transmitter 318TX. Each transmitter 318TX may modulate an RF carrier with a corresponding spatial stream for transmission.

[0061] At the UE 350, each receiver 354RX receives a signal through its corresponding antenna 352. Each receiver 354RX recovers the information modulated onto the RF carrier and provides the information to a receive (RX) processor 356. The TX processor 368 and the RX processor 356 implement layer 1 functionality associated with various signal processing functions. The RX processor 356 can perform spatial processing on the information to recover any spatial stream destined for the UE 350. If multiple spatial streams are destined for the UE 350, they can be combined into a single OFDM symbol stream by the RX processor 356. The RX processor 356 then converts the OFDM symbol stream from the time domain to the frequency domain using a fast Fourier transform (FFT). The frequency domain signal includes a separate OFDM symbol stream for each subcarrier of the OFDM signal. The symbols on each subcarrier and the reference signal are recovered and demodulated by determining the most likely signal constellation point sent by the base station 310. These soft decisions can be based on channel estimates calculated by the channel estimator 358. The soft decisions are then decoded and deinterleaved to recover the data and control signals originally sent on the physical channel by base station 310. The data and control signals are then provided to controller / processor 359, which implements layer 3 and layer 2 functionality.

[0062] The controller / processor 359 may be associated with a memory 360 that stores program codes and data. The memory 360 may be referred to as a computer readable medium. In the UL, the controller / processor 359 provides demultiplexing between transport channels and logical channels, packet reassembly, decryption, header decompression, and control signal processing to recover IP packets from the EPC 160. The controller / processor 359 is also responsible for error detection using an ACK or NACK protocol to support HARQ operations.

[0063] Similar to the functionality described in conjunction with DL transmissions performed by the base station 310, the controller / processor 359 provides RRC layer functionality associated with system information (e.g., MIB, SIB) acquisition, RRC connection, and measurement reporting; PDCP layer functionality associated with header compression / decompression and security (encryption, decryption, integrity protection, integrity verification); RLC layer functionality associated with delivery of upper layer PDUs, error correction through ARQ, concatenation, segmentation and reassembly of RLC SDUs, resegmentation of RLC data PDUs, and reordering of RLC data PDUs; and MAC layer functionality associated with mapping between logical channels and transport channels, multiplexing of MAC SDUs onto TBs, demultiplexing of MAC SDUs from TBs, scheduling information reporting, error correction through HARQ, priority handling, and logical channel prioritization.

[0064] Channel estimates derived by the channel estimator 358 based on a reference signal or feedback sent by the base station 310 may be used by the TX processor 368 to select the appropriate coding and modulation schemes and to facilitate spatial processing. The spatial streams generated by the TX processor 368 may be provided to different antennas 352 via separate transmitters 354TX. Each transmitter 354TX may modulate an RF carrier with a corresponding spatial stream for transmission.

[0065] UL transmissions are processed at the base station 310 in a manner similar to that described in conjunction with the receiver functionality at the UE 350. Each receiver 318RX receives a signal through its respective antenna 320. Each receiver 318RX recovers information modulated onto an RF carrier and provides the information to a RX processor 370.

[0066] The controller / processor 375 may be associated with a memory 376 that stores program codes and data. The memory 376 may be referred to as a computer readable medium. In the UL, the controller / processor 375 provides demultiplexing between transport channels and logical channels, packet reassembly, decryption, header decompression, control signal processing to recover IP packets from the UE 350. The IP packets from the controller / processor 375 may be provided to the EPC 160. The controller / processor 375 is also responsible for error detection using an ACK or NACK protocol to support HARQ operations.

[0067] In the UE 350, at least one of the TX processor 368, the RX processor 356, and the controller / processor 359 may be configured to perform operations related to Figure 1 The parameter set management component 140 combines various aspects.

[0068] The deployment of a communication system such as a 5G New Radio (NR) system may be arranged with various components or constituent parts in a variety of ways. In a 5G NR system or network, a network node, a network entity, a mobility element of a network, a radio access network (RAN) node, a core network node, a network element or network equipment such as a base station (BS) or one or more units (or one or more components) performing base station functionality may be implemented in an aggregated or decomposed architecture. For example, a BS such as a Node B (NB), an evolved NB (eNB), an NR BS, a 5G NB, an access point (AP), a transmit receive point (TRP) or a cell, etc. may be implemented as an aggregated base station (also referred to as an independent BS or a monolithic BS) or a decomposed base station.

[0069] A converged base station may be configured to utilize a radio protocol stack physically or logically integrated within a single RAN node. A decomposed base station may be configured to utilize a protocol stack physically or logically distributed between two or more units, such as one or more central or centralized units (CUs), one or more distributed units (DUs), or one or more radio units (RUs). In some aspects, a CU may be implemented within a RAN node, and one or more DUs may be co-located with the CU, or alternatively, may be geographically or virtually distributed in one or more other RAN nodes. A DU may be implemented to communicate with one or more RUs. Each of a CU, a DU, and a RU may also be implemented as a virtual unit, i.e., a virtual central unit (VCU), a virtual distributed unit (VDU), or a virtual radio unit (VRU).

[0070] Base station type operations or network designs may take into account the aggregated nature of base station functionality. For example, a disaggregated base station may be used in an integrated access backhaul (IAB) network, an open radio access network (O-RAN (network configurations such as those initiated by the O-RAN Alliance)), or a virtualized radio access network (vRAN, also known as a cloud radio access network (C-RAN)). Decomposition may include distributing functions across two or more units at various physical locations, as well as virtually distributing the functionality of at least one unit, which may enable flexibility in network design. Various units of a disaggregated base station or disaggregated RAN architecture may be configured for wired or wireless communication with at least one other unit.

[0071] Figure 4A diagram illustrating an example decomposed base station 400 architecture is shown. The decomposed base station 400 architecture may include one or more central units (CUs) 410 that may communicate directly with a core network 420 via a backhaul link, or indirectly with the core network 420 through one or more decomposed base station units, such as a near real-time (near-RT) RAN intelligent controller (RIC) 425 via an E2 link, or a non-real-time (non-RT) RIC 415 associated with a service management and orchestration (SMO) framework 405, or both. The CU 410 may communicate with one or more distributed units (DUs) 430 via respective midhaul links, such as an F1 interface. The DU 430 may communicate with one or more radio units (RUs) 440 via respective fronthaul links. The RU 440 may communicate with respective UEs 104 via one or more radio frequency (RF) access links. In some implementations, a UE 104 may be served simultaneously by multiple RUs 440.

[0072] Each of the units (i.e., CU 410, DU 430, RU 440, and near-RT RIC 425, non-RT RIC 415, and SMO framework 405) may include or be coupled to one or more interfaces configured to receive or send signals, data, or information (collectively referred to as signals) via a wired or wireless transmission medium. Each of the units or an associated processor or controller that provides instructions to the communication interface of these units may be configured to communicate with one or more of the other units via a transmission medium. For example, these units may include a wired interface configured to receive or send signals to one or more of the other units via a wired transmission medium. Additionally, the unit may include a wireless interface that may include a receiver, a transmitter, or a transceiver (such as a radio frequency (RF) transceiver) that is configured to receive or send signals, or both, to one or more of the other units via a wireless transmission medium.

[0073] In some aspects, CU 410 may host one or more higher layer control functions. Such control functions may include radio resource control (RRC), packet data convergence protocol (PDCP), service data adaptation protocol (SDAP), etc. Each control function may be implemented with an interface configured to communicate signals with other control functions hosted by CU 410. CU 410 may be configured to handle user plane functionality (i.e., central unit-user plane (CU-UP)), control plane functionality (i.e., central unit-control plane (CU-CP)), or a combination thereof. In some specific implementations, CU 410 may be logically split into one or more CU-UP units and one or more CU-CP units. When implemented in an O-RAN configuration, the CU-UP unit may communicate bidirectionally with the CU-CP unit via an interface (such as an E1 interface). As needed, CU 410 may be implemented to communicate with DU 430 for network control and signaling.

[0074] DU 430 may correspond to a logical unit that includes one or more base station functions for controlling the operation of one or more RUs 440. In some aspects, DU 430 may host one or more of a radio link control (RLC) layer, a medium access control (MAC) layer, and one or more high physical (PHY) layers (such as modules for forward error correction (FEC) encoding and decoding, scrambling, modulation and demodulation, etc.) depending at least in part on a functional split, such as that defined by the Third Generation Partnership Project (3GPP). In some aspects, DU 430 may also host one or more low PHY layers. Each layer (or module) may be implemented using an interface that is configured to communicate signals with other layers (and modules) hosted by DU 430 or with control functions hosted by CU 410.

[0075] The lower layer functionality may be implemented by one or more RUs 440. In some deployments, a RU 440 controlled by a DU 430 may correspond to a logical node that hosts RF processing functions or low PHY layer functions (such as performing Fast Fourier Transform (FFT), Inverse FFT (iFFT), digital beamforming, physical random access channel (PRACH) extraction and filtering, etc.), or both, based at least in part on a functional split (such as a lower layer functional split). In such an architecture, the RU 440 may be implemented to handle over-the-air (OTA) communications with one or more UEs 104. In some implementations, real-time and non-real-time aspects of control plane and user plane communications with the RU 440 may be controlled by the corresponding DU 430. In some scenarios, this configuration may enable the implementation of the DU 430 and the CU 410 in a cloud-based RAN architecture (such as a vRAN architecture).

[0076] The SMO framework 405 may be configured to support RAN deployment and provisioning of non-virtualized network elements and virtualized network elements. For non-virtualized network elements, the SMO framework 405 may be configured to support the deployment of dedicated physical resources for RAN coverage requirements, which may be managed via an operation and maintenance interface (such as an O1 interface). For virtualized network elements, the SMO framework 405 may be configured to interact with a cloud computing platform (such as an open cloud (O-Cloud) 490) to perform network element lifecycle management (such as to instantiate virtualized network elements) via a cloud computing platform interface (such as an O2 interface). Such virtualized network elements may include, but are not limited to, CU 410, DU 430, RU 440, and near-RT RIC 425. In some specific implementations, the SMO framework 405 may communicate with hardware aspects of the 4G RAN (such as an open eNB (O-eNB) 410) via the O1 interface. Additionally, in some specific implementations, the SMO framework 405 may communicate directly with one or more RUs 440 via the O1 interface. The SMO framework 405 may also include a non-RT RIC 415 configured to support the functionality of the SMO framework 405 .

[0077] The non-RT RIC 415 may be configured to include logic functions that enable non-real-time control and optimization of RAN elements and resources, artificial intelligence / machine learning (ML) workflows including model training and updating, or policy-based guidance of applications / features in the near-RT RIC 425. The non-RT RIC 415 may be coupled to or communicate with the near-RT RIC 425 (such as via an A1 interface). The near-RT RIC 425 may be configured to include logic functions that enable near-real-time control and optimization of RAN elements and resources via data collection and actions through an interface (such as via an E2 interface) that connects one or more CUs 410, one or more DUs 430, or both, and the O-eNB with the near-RT RIC 425.

[0078] In some implementations, in order to generate the ML model to be deployed in the near-RT RIC 425, the non-RT RIC 415 may receive parameters or external enrichment information from an external server. Such information may be utilized by the near-RT RIC 425 and may be received from a non-network data source or from a network function at the SMO framework 405 or the non-RT RIC 415. In some examples, the non-RT RIC 415 or the near-RT RIC 425 may be configured to tune RAN behavior or performance. For example, the non-RT RIC 415 may monitor long-term trends and patterns for performance and employ ML models to perform corrective actions via the SMO framework 405 (such as via reconfiguration of O1) or via creation of RAN management policies (such as A1 policies).

[0079] Figure 5 An example of a neural network 500, specifically, a convolutional neural network (CNN), is illustrated. The CNN may be designed to detect objects sensed from a video capture device 502, such as an onboard camera or other sensor. The neural network 500 may initially receive an input 504, such as an image, such as an image of a speed limit sign (or other object or size) of size 32×32 pixels. During a forward pass, the input image is initially passed through a convolutional layer 506 including a plurality of convolutional kernels (e.g., six kernels of size 5×5 pixels or some other number or size) that slide over the image to detect basic patterns or features, such as straight edges and corners. The images output from the convolutional layer 506 (e.g., six images of size 28×28 pixels or some other number or size) are passed through an activation function (e.g., a rectified linear unit (ReLU)) and then passed as input into a subsampling layer 508, which scales down the size of the image, for example, by a factor of two (e.g., producing six images of size 14×14 pixels or some other number or size). These scaled-down images output from the subsampling layer 508 may similarly be passed through an activation function (e.g., ReLU or other function), and similarly passed as input through subsequent convolutional layers, subsampling layers, and activation functions (not shown) to detect more complex features and further scale down the image or kernel size. These outputs are ultimately passed as input to a fully connected (FC) layer 510, where each node output from the previous layer is connected to all neurons in the current layer. The output from this layer may similarly be passed through an activation function and potentially as input through one or more other fully connected layers (not shown). Thereafter, the output is passed as input to an output layer 512, which converts the input into an output 514 such as a probability distribution (e.g., using a softmax function). The probability distribution may include a vector of confidence levels or probability estimates that the input image depicts a predicted feature (such as a sign or speed limit value (or another object)).

[0080] During training, ML models (e.g., classifiers) with weights 516 and biases 518 are initially created for different layers of the neural network 500, respectively. For example, when input from a training image (or other source) enters a given layer of an MLP or CNN, the function of the input and weights summed with the bias can be transformed using an activation function before being passed to the next layer. The probability estimates generated from the final output layer can then be applied to a loss function that measures the accuracy of the ANN, such as a cross entropy loss function. Initially, the output of the loss function may be quite large, indicating that the predicted value is far from the true or actual value. In order to reduce the value of the loss function and produce more accurate predictions, gradient descent can be applied.

[0081] In gradient descent, the gradient of the loss function may be calculated with respect to each weight of the ANN using back propagation, where the gradient is calculated for the last layer back to the first layer of the neural network 500. The gradient may then be used to update each weight to reduce the loss function with respect to that weight until a global minimization of the loss function is achieved, for example using stochastic gradient descent. For example, after each weight adjustment, subsequent iterations of the above training process may be performed using the same or new training images, and if the loss function is still large (even if it has been reduced), back propagation may be applied again to identify the gradient of the loss function with respect to each weight. The weights may be updated again, and the process may continue to be repeated until the difference between the predicted and actual values ​​is minimized.

[0082] A network node such as a UE (e.g., 104 / 350) or a base station (e.g., 102, 180, 310) can use federated learning (FL) to train a neural network (e.g., neural network 500). Compared to centralized machine learning techniques that typically upload all local data sets to one server, FL allows the generation of high-quality ML models without aggregating distributed data. Therefore, FL facilitates parallel processing, significantly reduces the costs associated with message exchange, and protects data privacy.

[0083] In the FL framework, nodes such as UEs can learn a global ML model by passing messages between nodes through a central coordinator. For example, a node can provide weights, biases, gradients, or other ML information to a central coordinator (e.g., a base station, RSU, edge server, etc.), which aggregates the weights, biases, gradients, or other ML information to generate a global ML model.

[0084] Further, each node in the FL environment utilizes a dataset to locally train and update a coordinated global ML model. The dataset may be a local dataset that a node or device can obtain for a certain ML task (e.g., object detection, etc.). The data in a given dataset may be preloaded or may be accumulated throughout the life of a device (e.g., UE) or otherwise accumulated during the use of the device. For example, an accumulated dataset may include recorded data observed by a node from an onboard sensor such as a camera and stored locally at the node.

[0085] A global ML model can be defined by its model architecture and model weights. An example of a model architecture is a neural network, such as Figure 5 A CNN or other neural network architecture is described, which may include multiple hidden layers, multiple neurons in each layer, and synapses connecting the neurons together. Model weights are applied to data passed through the various layers of the ML model for processing by the various neurons.

[0086] refer to Figures 6 to 12In some non-limiting aspects, according to some aspects of the present disclosure, system 600 is configured to implement a procedure for federated domain generalization.

[0087] Figure 6 An example of a FDG architecture is illustrated. The federated server device 602 initializes a global ML model (WOG) 604 (1) (e.g., a classifier in a pre-configured ML architecture such as Figure 6 The federated server device 602 broadcasts the global ML model 604 to participating devices or federated client devices 606 (1)-(n) (e.g., UEs) (collectively, "federated client devices 606"). Some non-limiting examples of the federated server device 602 include a central coordinator (e.g., a base station, an RSU, an edge server, a virtualized network element, etc.). Some non-limiting examples of the federated client devices 606 include UEs, base stations, network nodes, smart phones and computing devices, Internet of Things (IoT) devices, drones, robots, process automation equipment, sensors, control devices, vehicles, transportation equipment, tactile interaction devices, virtual and augmented reality (VR and AR) devices, industrial machines, etc.

[0088] like Figure 6 As illustrated, the federated client device 606 may receive federated information 608 from the federated server device 602. In some aspects, the federated information 608 includes the global ML model 604. After receiving the global ML model 604, the federated client device 606 may iteratively perform the federated learning process. For example, at a given iteration t, the global ML model 604 is received (W t G), the federated client devices 606 (1)-(n) use a local dataset specific to the domain of the corresponding local ML model 610 to locally train the global ML model 604 into their respective local ML models 610, which local dataset can be a preconfigured dataset, such as a training set and / or sensory data from the environment. After training, each federated client device 606 can update the weights of the local ML model 610 of the federated client device 606 until the loss function or cost function of the model is minimized or optimized. Accordingly, due to having different weights, each federated client device 606 can have a different local ML model 610.

[0089] In addition, as described in detail herein, each federated client device 606 can identify a client-shared portion 612 of a local ML model 610 that is separate from a client-specific portion associated with a specific domain of the federated client device 606. As described in detail herein, in some aspects, a federated client device 606 can identify its corresponding client-shared portion 612 based on a batch normalization (BN) layer of the local ML model 610 of the federated client device 606. In particular, the BN layer is employed to ensure that the local ML model 610 learns a client-invariant representation. In addition, by isolating the client-shared portion 612, the federated client device 606 avoids negatively affecting the performance of the global ML model 604 of a federated client device 614 associated with a domain unknown to the federated server device 602 by preventing domain-specific features from being included in the global ML model 604.

[0090] In addition, in some aspects, the federated client devices 606(1)-606(n) may send client shared portion (CSP) information 616(1)-616(n) including multiple client shared portions 612(1)-612(n) to the federated server device 602. In some aspects, to reduce transmission traffic, the CSP information 616 for a particular federated client device 606 may be, for example, one or more modified weights of the client shared portions 612, one or more weights of the client shared portions 612 that have changed beyond a predefined threshold, and / or a multiplicative or additive increment or percentage of the modified weights of the client shared portions 612. Upon receiving the CSP information 616(1)-616(n), the federated server device 602 may aggregate the corresponding weights within the client shared portions 612 (e.g., by averaging the weights or performing other calculations on the weights). In addition, the federated server device 602 may update the first global ML model 604(1) to generate a second global ML model (W) including the aggregated weights. t+1 The federated server device 602 may receive a second global ML model 604(2) and send federated domain (FD) information 608 including the second global ML model 604(2) to another federated client device 614 (e.g., a local ML model 610 in an unseen domain). Upon receiving the second global ML model 604(2), the federated client device 614 may adopt the global model 604 as a local model 618 to perform ML tasks (e.g., generate inference information) in a domain unknown to the federated server device 602.

[0091] Figure 7 is a diagram illustrating example communications and components of a federated server device and a federated client device. Figure 7As illustrated, the federated system 700 may include one or more federated server devices 702(1)-702(n) (e.g., federated server device 602), one or more federated client devices 704(1)-704(n) in a domain known to the one or more federated server devices 702 (e.g., UE 104, federated client devices 606(1)-606(n)), and one or more federated client devices 706(1)-706(n) in a domain unknown to the one or more federated server devices 702 (e.g., UE 104, federated client devices 606(1)-606(n)). In addition, the federated system 700 may include one or more communication networks 707(1)-707(n). In some implementations, the communication network 707 may include one or more of a wired and / or wireless private network, a personal area network, a local area network, a wide area network, and / or the Internet. Furthermore, in some aspects, federated server device 702, federated client devices 704(1)-704(n), and / or federated client devices 706(1)-706(n) may be configured to communicate via communication networks 107(1)-107(n).

[0092] like Figure 7 As illustrated, federated client devices (e.g., federated client devices 704(1)-704(n) and federated client devices 706(1)-706(n)) may include one or more local models 708(1)-708(n), a federated learning component 140 (e.g., a training component 710, a CSP detector component 712, and an alpha generator 714), and one or more applications 716(1)-716(n). As described herein, in some aspects, federated client devices (e.g., federated client device 704(1) and federated client device 706(1)) may employ local model 708(1) to perform tasks. Further, in some aspects, one or more applications 716(1)-716(n) may employ local model 708(1)-708(n) to perform application tasks. For example, application 716 may be an image processing application and employ local model 708 to perform image processing tasks (e.g., object recognition). Further, as Figure 7 As illustrated, each federated server device 702 may include a federated domain generalization management (FDGM) component 718 for managing federated domain generalization within the federated system 700 via a global model generation component 719 .

[0093] As described herein, in some aspects, the federated system 700 can implement federated domain generalization. Specifically, the federated server device 702 can use data collected from the federated client devices 702(1)-(n) in the domain known to the federated server device 704(1)-704(n) to create a generalized global model 720 via a global model generation component 719, and deploy the generalized global model 720 to the federated client devices 706(1)-706(n). Specifically, the FDGM component 718 can initialize the global model 720 and send the global model 720 to the federated client devices 704(1)-704(n) in the domain known to the federated server device 702. In addition, the training component 710 of the federated client device 704 can train the global model 720 to generate a local model 708. In some aspects, the local model 708 will include a client variant layer and a client invariant layer, such as Figure 6 Additionally, in some aspects, the training component 710 can be configured to retrain the local model 708.

[0094] In some aspects, the client-side invariant layer of the federated client devices 704(1)-704(n) may be overfitted to the domain of the federated client devices 704(1)-704(n). Therefore, the federated client device 704 may augment the local data with the data distribution of other federated client devices to mitigate the overfitting problem without any privacy leakage. Specifically, the training component 710 may interpolate instance statistics of the sample data and global statistics associated with multiple other federated client devices 704. Specifically, the training component 710 may interpolate the mean and variance of the instance sample with the global BN statistics (e.g., running mean and variance) of the other federated client devices 704 received from the federated server device 702. By randomly interpolating these two statistics, the local data is augmented with the data distribution of other federated client devices without any privacy leakage. As relative to Figure 8 Described in more detail, the training component 710 can employ a local training loss function, a local client-independent feature learning loss function, and a client-independent classifier learning loss function generated by the aforementioned interpolation when training the local model 708 via batch normalization enhancement.

[0095] In response to the training or retraining of the local model 708, the CSP detector component 712 can detect the client-side invariant layers of the local model 708. In some aspects, the CSP detector component 712 can identify the client-side invariant layers of the local model 708 based on the batch normalization information of the local model 708. For example, the CSP detector component 712 can identify the batch normalization layers of the local model 708 as client-side variant layers and the remaining layers as client-side invariant information, and send the client-side invariant information layers to the federated server device 702 as the CSP information 722.

[0096] In addition, the CSP detector component 712 may send CSP information 722 to the federated server device 702. In some aspects, the CSP information 722 may include inputs, parameters, weights, gradients, and / or scaling factors associated with the client-side invariant layer. Additionally, in some aspects, the scaling factor may be a multiplicand or an addend. Upon receiving the CSP information 722 from the federated client device 704, the federated server device 702 may update the global model 720. For example, the FDGM component 718 may incorporate the CSP information 722 received from the federated client device 704 into the global model 720. Additionally, the FDGM component 718 may deploy the updated global model 720 to the federated client device 706 associated with the unknown domain. By employing the client-side invariant layer of the local model 708 to collaboratively train the global model 720, the federated server may avoid domain shifts, which may significantly reduce the performance of the global model 720, because the client-side variant layer is trained in view of the specific domain of the federated client device 704.

[0097] After deploying the global model 720 to a federated client device 706 associated with a domain unknown to the federated server device 702, the federated client device 706 can employ the global model 720 to perform ML tasks to generate inference information. In some aspects, given that the domain is unknown to the federated server device 702, the global model 720 may not generalize to the domain of the federated client device 706. In addition, the federated client device 706 cannot access the local batch normalization statistics of the federated client device 704. Therefore, in some aspects, such as Fig. 9 As illustrated, during inference generation, the federated client device 706 can interpolate the global batch normalization statistics 724 using the instance test statistics 726. In addition, the federated client device 706 can employ an alpha generator (e.g., a zero-shot adapter) to interpolate the global batch normalization statistics 724 using the instance test statistics 726. In some aspects, the federated client device 706 can employ an alpha generator (e.g., a zero-shot adapter) to interpolate the global batch normalization statistics 724 using the instance test statistics 726. Fig. 9 Federated learning is illustrated to train a zero-shot adapter.

[0098] Additionally, federated client device 704 and / or federated client device 706 may include a receiver component 728 and a transmitter component 730. Receiver component 728 may include, for example, a radio frequency (RF) receiver for receiving signals as described herein. Transmitter component 730 may include, for example, an RF transmitter for transmitting signals as described herein. In one aspect, receiver component 728 and transmitter component 730 may be co-located on a transceiver (e.g., Fig.10 In the transceiver 1010 shown.

[0099] Additionally, federated server device 706 may include a receiver component 732 and a transmitter component 734. Receiver component 732 may include, for example, a radio frequency (RF) receiver for receiving signals as described herein. Transmitter component 734 may include, for example, an RF transmitter for transmitting signals as described herein. In one aspect, receiver component 732 and transmitter component 734 may be co-located on a transceiver (e.g., Fig.13 In the transceiver 1310 shown.

[0100] Figure 8 800 is a diagram illustrating an example of client-independent learning with data augmentation according to some aspects of the present disclosure. As described herein, because each local model is trained in a different domain, each local model is fitted to its own specific domain, which may lead to unsatisfactory performance of the global model generated by aggregating CSP information associated with different local models. Accordingly, as Figure 8 As illustrated, in some aspects, a training component (e.g., training component 710) can configure a local model 802 (e.g., local model 708) to augment local data with data distributions from other federated client devices without any privacy leakage. Specifically, the local model 802 can interpolate local BN statistics 804 (i.e., statistics of local samples) using mixed instance statistics and global statistics (mixed statistics) 806 during training.

[0101] For example, let X and Y refer to the input and label space respectively. The kth client in the federation has single-domain data where n k is the number of samples. The set of distributed source domain data from K federated clients is represented as {D1,...,D k In federated domain generalization, there is domain shift across clients, where each client data D k From a different domain-specific distribution than other clients (X k ,Y) sampling. The target test domain data from the unseen environment is denoted as D t , its distribution (X t ,Y) is offset from the training data. The feature extractor parameterized by θ is denoted as F θ (in is the feature extractor for the kth client), and the classifier parameterized by φ is denoted as C φ (in is the classifier for the kth client). In some aspects, a federated system (e.g., federated system 700) is designed to aggregate K distributed federated client models trained on To learn a generalized global model Make the global model generalize to the unseen target domain D t .

[0102] In local training, the k-th local model is trained on the k-th dataset using the cross entropy loss function 808 as follows:

[0103]

[0104] Before starting local training, the client receives the global model parameters {θ G ,φ G} and use the global parameters to initialize the local model {θ k ,φ k}.Then, and Train on local data for long epochs. Although {θ k ,φ k}Use the global model to initialize, but it may not rely solely on learning client-invariant representations using the cross entropy loss of single-domain data, because directly using data from other federated clients raises privacy issues. To address this issue, in some aspects, the federated system uses statistics in the BN layer of the global model to generate diverse domains. For example, the training component can set all batch normalization layers in the backbone network to local BN layers:

[0105]

[0106] in and are the BN statistics of the l-th BN layer of the k-th client. These BN statistics are calculated as running means and variances, indicates the input tensor for the kth client, and and are learnable affine parameters.

[0107] As described in this paper, the local feature f i,k It is obtained by using batch normalization at the local client level, while enhancing the feature f i,Δ Obtained by a hybrid instance and global statistics approach that interpolates instance statistics and global statistics in a randomized manner, which results in an intermediate feature that can be a variety of features Standardization, such as Figure 8 As illustrated. This process can be repeated at each BN layer. Using these features, the local model is able to learn client-side invariant representations using Equations 5 and 6.

[0108] The training component compares the mean and standard deviation of each sample with the global statistic {μ G ,σ G} Mix, that is, running mean and standard deviation, as follows:

[0109]

[0110] in and Indicates the instance mean and standard deviation of the intermediate features of the i-th input of the l-th BN layer along the channel axis, respectively. is the interpolation weight vector, where each element is independently sampled from a uniform distribution U(0,1), and C l is the feature size in the lth BN layer.

[0111] In some aspects, features normalized by instance statistics contain local representative characteristics, while features normalized using global statistics consist of global representations. By randomly interpolating these two statistics in all BN layers, the federated system can obtain more diverse data by leveraging characteristics of both local and global domains.

[0112] In addition, the enhanced feature f Δ The statistics are interpolated and the training component can use the client-independent form of f Δ To train the local model 802. Additionally, Figure 8 As illustrated, the local model 802 may generate a client-independent feature loss 810. The client-independent feature loss 810 may be determined as follows:

[0113]

[0114] By adopting the loss function of the client-irrelevant feature loss 810, the local model 802 can extract the client-irrelevant features by minimizing the distance between the original features and the enhanced features.

[0115] In addition, the training component can train the local classifier to classify features from other domains, so that the local classifier can be a client-independent classifier. Trained using client-independent classification loss function 812, as follows:

[0116]

[0117] Where L CE is the cross entropy loss 808, F θk，c is the kth client and the classifier is updated using this loss function. In some aspects, the client-independent learning performed by the training component can be considered as a regularization method that forces the local model not to deviate from the global model. In addition, the overall loss of the local optimization is as follows:

[0118] L total=(1-λ1)·L CE +λ1·L CACL +λ2·L CAFL (Equation 7)

[0119] Fig. 9 900 is a diagram illustrating an example of an example alpha generator 902 (e.g., a zero-shot adapter) in accordance with some aspects of the present disclosure. As described herein, a global model generated using aggregated CSP information (e.g., CSP information 616) may not generalize satisfactorily to domains (e.g., federated client devices 706) that are unknown to a federated server (e.g., federated server device 702). Accordingly, in some aspects, when the domain of the federated client device is unknown to the federated server device, a federated client device employing a global model may dynamically use instance statistics during inference as well as global BN statistics that fully reflect the test distribution to generalize the global model received from the federated server device. For example, as Fig. 9 As illustrated, the federated client device may utilize the statistics 906 of the test input to interpolate the running mean and variance 904 in the BN layer as follows:

[0120]

[0121] in and Indicates the instance mean and variance of the input tensor in the lth BN layer, and is used to normalize the test input tensor, and α l is an interpolation parameter that adjusts the contribution of the instance statistics of the test sample. In some aspects, α is selected to test the domain or the test input, for example, a fixed value for a specific test domain or a dynamic value for the test input. However, in some aspects, the federated client may not be able to select an appropriate α, and the search space of α is very large, where α is in [0, 1] and each α l May be different from α in other layers. Accordingly, the federated client device may employ an α generator 902 configured to generate α for each input in both known and unknown domains.

[0122] For example, Fig. 9 As illustrated, the alpha generator 902 can be added to the batch normalization layer of the global model. In addition, the input 908 to the alpha generator 902 can be the channel-by-channel distance between the instance statistics and the global statistics. Output 910 can be a l In some respects, a l Set to a scalar value to interpolate instance statistics and global statistics with the same weights along the channel axis in the lth BN layer, which reduces the network size of the zero-shot adapter and alleviates overfitting on local data.

[0123] In some aspects, a training component (e.g., training component 710) can simulate training samples in a federated learning phase. and The test scenario forwarded to the main network is used to train the α generator 902, where the BN statistics are replaced with the interpolated BN statistics in equations 6 and 7. In addition, the α generator 902 is updated using the cross entropy loss to generate the optimal interpolation parameter a based on the difference between the instance statistics and the global statistics. l . In some aspects, the training procedure of the α generator 902 may not affect the performance of the main network, because its purpose is only to learn how to interpolate instance statistics and global statistics to minimize the cross entropy loss on the trained main network. Therefore, the main network and the α generator 902 are trained alternately, thereby freezing the other model. For example, when the α generator 902 is being trained, the FC layer and ReLU layer of the α generator 902 are being updated, while the convolution layer, ReLU layer, BN layer, and FC layer of the main network are frozen. In order to prevent the adapter from overfitting to each training domain, the training component may apply a reparameterization method. For example, the interpolation parameters can be generated by sampling from a Gaussian distribution, which is reparameterized by the adapter as follows: a l =T(δ l z l +∈ l ),in And z l Sampling from N(0,1). In some aspects, T(·) is to ensure that a l Clamp function in the range [0,1].

[0124] During the inference phase, a l is parameterized as ∈ l , which is δ l z l +∈ l In addition, the normalization of the test input can be performed at each BN layer by using Equations 6 and 7 This is performed using interpolated statistics, which is done via a single forward pass.

[0125] Fig.101000 is a diagram illustrating an example of a hardware implementation of a federated client device 1002 (e.g., UE 104, federated client device 606, federated client device 704, federated client device 706) employing a processing system 1014. The processing system 1014 may be implemented using a bus architecture represented generally by a bus 1024. The bus 1024 may include any number of interconnecting buses and / or bridges, depending on the specific application of the processing system 1014 and the overall design constraints. The bus 1024 links together various circuits including one or more processors and / or hardware components represented by the processor 1004, models 708(1)-708(n), the federated learning component 140, the training component 710, the CSP detector component 712, the alpha generator 714, one or more applications 716(1)-716(n), and a computer-readable medium (e.g., a non-transitory computer-readable medium) / memory 1006. The bus 1024 may also link various other circuits, such as timing sources, peripherals, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further.

[0126] The processing system 1014 may be coupled to the transceiver 1010. The transceiver 1010 may be coupled to one or more antennas 1020. The transceiver 1010 provides components for communicating with various other devices via a transmission medium. The transceiver 1010 receives signals from the one or more antennas, extracts information from the received signals, and provides the extracted information to the processing system 1014 (specifically, the receiver component 728). The receiver component 728 may receive the federation information 608 and the global model 720. In addition, the transceiver 1010 receives information from the processing system 1014 (specifically, the transmitter component 730) and generates a signal to be applied to the one or more antennas based on the received information. In addition, the transmitter component 730 may transmit the CSP information 616 and / or the CSP information 722.

[0127] The processing system 1014 includes a processor 1004 coupled to a computer-readable medium / memory 1006 (e.g., a non-transitory computer-readable medium). The processor 1004 is responsible for general processing, including executing software stored on the computer-readable medium / memory 1006. The software, when executed by the processor 1004, causes the processing system 1014 to perform the various functions described above for any particular device. The computer-readable medium / memory 1006 may also be used to store data manipulated by the processor 1004 when executing the software. The processing system 1014 also includes one or more local models 708(1)-708(n), a federated learning component 140, a training component 710, a CSP detector component 712, an alpha generator 714, and one or more applications 716(1)-716(n). The foregoing components may be software components running in the processor 1004, resident / stored in the computer-readable medium / memory 1006, one or more hardware components coupled to the processor 1004, or some combination thereof. Processing system 1014 may be a component of federated client device 1002 and may include memory 360 and / or at least one of TX processor 368, RX processor 356, and controller / processor 359. Alternatively, processing system 1014 may be the entire federated client device (e.g., see Figure 3 350 Figure 6 The federated client device 606, Figure 7 Federal client device 704, Figure 7 of the federated client device 706).

[0128] The aforementioned components may be one or more of the aforementioned components of the federated client device 1002 and / or the processing system 1014 of the federated client device 1002 configured to perform the functions recited by the aforementioned components. As described above, the processing system 1014 may include the TX processor 368, the RX processor 356, and the controller / processor 359. Thus, in one configuration, the aforementioned components may be the TX processor 368, the RX processor 356, and the controller / processor 359 configured to perform the functions recited by the aforementioned components.

[0129] See also Figures 1 to 11 In operation, the federated learning client device 1002 may perform a method of wireless communication 1100. The method may be performed by the entire federated client device or a component of the federated client device (e.g., parameter set management component 140 / 510, processor 1004, computer readable medium / memory 1006, TX processor 368, RX processor 356, and / or controller / processor 359).

[0130] At block 1102, method 1100 includes identifying client-side invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server. For example, in some aspects, at least one of federated client device 606, federated client device 704, federated client device 706, local model 708, local model 802, processor 1004, computer-readable medium / memory 1006, CSP detector component 712 may be configured to or may include means for identifying client-side invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server. For example, CSP detector component 712 may detect a client-side invariant layer of local model 708. In some aspects, CSP detector component 712 may identify the client-side invariant layer of local model 708 based on batch normalization information of local model 708. For example, CSP detector component 712 may identify a batch normalization layer of local model 708 as a client-side variant layer and identify the remaining layers as client-side invariant information.

[0131] At box 1104, method 1100 includes: sending the client invariant information to the federated server. The federated server can be configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client invariant information and other client invariant information of another neural network for performing the ML task in a second domain known to the federated server. For example, in some aspects, at least one of the federated client device 1002, the federated client device 606, the federated client device 704, the federated client device 706, the CSP detector component, the processor 1004, the computer-readable medium / memory 1006, and / or the transmitter component 534 can be configured to or may include a component for sending the client invariant information to the federated server, and the federated server is configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client invariant information and other client invariant information of another neural network for performing the ML task in a second domain known to the federated server. For example, the federated client device 1002 can send the client invariant information layer to the federated server device 702 as the CSP information 722.

[0132] See also Figures 1 to 12 In operation, the federated learning client device 1002 may perform the method 1200 of wireless communication. The method may be performed by the entire federated client device or a component of the federated client device (e.g., parameter set management component 140 / 510, processor 1004, computer readable medium / memory 1006, TX processor 368, RX processor 356, and / or controller / processor 359).

[0133] At block 1202, method 1200 includes receiving a neural network for performing a machine learning task from a federated server. The neural network may have been generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server. For example, in some aspects, at least one of the federated client device 606, the federated client device 704, the federated client device 706, the processor 1004, the computer-readable medium / memory 1006, the training component 710, and / or the receiver component 728 may be configured to or may include a component for receiving a neural network for performing a machine learning task from a federated server, the neural network generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server. For example, the federated client device 706 may receive a global model 720 from the federated server device 702. As described herein, the federated server device 702 may generate the global model 720 based on CSP information 722 received from a plurality of federated client devices 704 associated with domains known to the federated server device 702.

[0134] At block 1204, method 1200 includes generating inference information in a domain unknown to the federated server via the neural network based on an alpha generator added to a batch normalization layer of the neural network. For example, in some aspects, at least one of federated client device 1002, federated client device 606, federated client device 704, federated client device 706, local model 708, global model 720, alpha generator 714, processor 1004, and / or computer-readable medium / memory 1006 may be configured to or may include components for generating inference information in a domain unknown to the federated server via the neural network based on an alpha generator added to a batch normalization layer of the neural network. For example, after deploying global model 720 to federated client device 706 associated with a domain unknown to federated server device 702, federated client device 706 may employ global model 720 to perform an ML task to determine to generate inference information. Additionally, the federated client device 706 can employ the alpha generator 902 (eg, a zero-shot adapter) to interpolate the global batch normalization statistics 724 using the instance test statistics 726 during generation of inference information.

[0135] Fig.131300 is a diagram illustrating an example of a hardware implementation of a federated server device 1302 employing a processing system 1314. The processing system 1314 may be implemented using a bus architecture generally represented by a bus 1324. The bus 1324 may include any number of interconnecting buses and bridges, depending on the specific application of the processing system 1314 and the overall design constraints. The bus 1324 links together various circuits including one or more processors and / or hardware components represented by the processor 1304, the federated domain generalization management component 718, the global model generation component 719, and the computer-readable medium / memory 1306. The bus 1324 may also link various other circuits such as timing sources, peripherals, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further.

[0136] The processing system 1314 may be coupled to the transceiver 1310. The transceiver 1310 is coupled to one or more antennas 1320. The transceiver 1310 provides components for communicating with various other devices over a transmission medium. The transceiver 1310 receives signals from the one or more antennas 1320, extracts information from the received signals, and provides the extracted information to the processing system 1314 (specifically, the receiver component 732). The receiver component 732 may receive the CSP information 616 and the CSP information 722 (1)-722 (n). In addition, the transceiver 1310 receives information from the processing system 1314 (specifically, the transmitter component 734) and generates a signal to be applied to the one or more antennas 1320 based on the received information. In addition, the transmitter component 734 may transmit the federation information 608 and the global model 720.

[0137] The processing system 1314 includes a processor 1304 coupled to a computer-readable medium / memory 1306 (e.g., a non-transitory computer-readable medium). The processor 1304 is responsible for general processing, including executing software stored on the computer-readable medium / memory 1306. The software, when executed by the processor 1304, causes the processing system 1314 to perform the various functions described above for any particular device. The computer-readable medium / memory 1306 can also be used to store data manipulated by the processor 1304 when executing the software. The processing system 1314 also includes a federated domain generalization management component 718 and a global model generation component 719. The aforementioned components can be software components running in the processor 1304, resident / stored in the computer-readable medium / memory 1306, one or more hardware components coupled to the processor 1304, or some combination thereof. The processing system 1314 can be a component of the base station 313 and can include a memory 376 and / or at least one of the TX processor 316, the RX processor 370, and the controller / processor 375. Alternatively, the processing system 1314 may be an entire base station (see, for example, Figure 3 310 Figure 6 The federal server device 602, Figure 7 Federal server device 702).

[0138] The aforementioned components may be one or more of the aforementioned components of the federated server device 1302 and / or the processing system 1314 of the federated server device 1302 configured to perform the functions recited by the aforementioned components. As described above, the processing system 1314 may include the TX processor 316, the RX processor 370, and the controller / processor 375. Thus, in one configuration, the aforementioned components may be the TX processor 316, the RX processor 370, and the controller / processor 375 configured to perform the functions recited by the aforementioned components.

[0139] Sample Clauses

[0140] Clause 1. A federated client device, comprising: one or more memories storing computer-executable instructions; one or more processors coupled to the one or more memories and configured to execute the computer-executable instructions, alone or in combination, to cause the federated client device to: identify client-invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server; and send the client-invariant information to the federated server, the federated server being configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client-invariant information and other client-invariant information of another neural network for performing the ML task in a second domain known to the federated server.

[0141] Clause 2. A federated client device according to clause 1, wherein the one or more processors are further configured, individually or in combination, to cause the federated client device to: receive federated ML information from the federated server; and generate the neural network based on the federated ML information.

[0142] Clause 3. A federated client device as described in any of clauses 1 to 3, wherein the client-invariant information includes at least one of a weight, a gradient, or a scaling factor, the scaling factor being a multiplicand or a summand.

[0143] Clause 4. A federated client device according to any one of clauses 1 to 4, wherein in order to identify the client-side invariant information of the neural network, the one or more processors are configured, individually or in combination, to: select an invariant layer of the neural network that is not a batch normalization layer of the neural network as the client-side invariant information.

[0144] Clause 5. A federated client device according to any one of clauses 1 to 5, wherein the one or more processors are further configured, individually or in combination, to enable the federated client device to: train the neural network by interpolating instance statistics of sample data and global statistics associated with multiple other federated client devices.

[0145] Clause 6. The federated client device of clause 5, wherein the global statistics comprise batch-normalized statistics collected from the plurality of other federated client devices.

[0146] Clause 7. A federated client device according to Clause 6, wherein, in order to train the neural network, the one or more processors are configured, individually or in combination, to: train the neural network by interpolating the mean and variance of the instance statistics using the mean and variance of the batch normalization statistics.

[0147] Clause 8. A federated client device according to any one of clauses 1 to 7, wherein the federated server is a first federated server, the ML task is a first ML task, the neural network is a first neural network, and the one or more processors are configured, individually or in combination, to cause the federated client device to: receive federated ML information from a second federated server, the federated ML information comprising a second neural network, the second neural network generated by the second federated server for performing a second ML task and generated by the second federated server using multiple client-invariant information of multiple other neural networks for performing the second ML task in multiple domains known to the second federated server; and generate inference information in a domain unknown to the second federated server via the second neural network based on an alpha generator added to a batch normalization layer of the second neural network.

[0148] Clause 9. A federated client device according to clause 8, wherein the alpha generator is a zero-shot adapter, and in order to determine the inference information, the one or more processors are configured individually or in combination to cause the federated client device to: interpolate global batch normalization statistics using instance test statistics at the batch normalization layer via the zero-shot adapter to generate learning parameters based on the channel-by-channel distance between the global batch normalization statistics and the instance test statistics.

[0149] Clause 10. The federated client device of Clause 9, wherein the at least one processor is further configured to train the zero-shot adapter via federated learning.

[0150] Clause 11. A federated client device according to clause 9, wherein the plurality of client-invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

[0151] Clause 12. A method comprising: identifying client-invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server; and sending the client-invariant information to the federated server, the federated server being configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client-invariant information and other client-invariant information of another neural network for performing the ML task in a second domain known to the federated server.

[0152] Clause 13. The method of clause 12, further comprising: receiving federated ML information from the federated server; and generating the neural network based on the federated ML information.

[0153] Clause 14. The method of any one of clauses 12 to 13, wherein the client-invariant information comprises at least one of a weight, a gradient, or a scaling factor, the scaling factor being a multiplicand or a summand.

[0154] Clause 15. A method according to any one of clauses 12 to 14, wherein identifying the client-side invariant information of the neural network includes: selecting an invariant layer of the neural network that is not a batch normalization layer of the neural network as the client-side invariant information.

[0155] Clause 16. The method of any one of clauses 12 to 15, further comprising: training the neural network by interpolating instance statistics of the sample data and global statistics associated with a plurality of other federated client devices.

[0156] Clause 17. The method of clause 16, wherein the global statistics comprise batch-normalized statistics collected from the plurality of other federated client devices.

[0157] Clause 18. The method of clause 17, wherein training the neural network comprises training the neural network by interpolating the mean and variance of the instance statistics using the mean and variance of the batch normalization statistics.

[0158] Clause 19. A method according to any one of clauses 12 to 18, wherein the federated server is a first federated server, the ML task is a first ML task, the neural network is a first neural network, and the method further includes: receiving federated ML information from a second federated server, the federated ML information including a second neural network, the second neural network generated by the second federated server for performing a second ML task and generated by the second federated server using multiple client-invariant information of multiple other neural networks for performing the second ML task in multiple domains known to the second federated server; and generating inference information in a domain unknown to the second federated server via the second neural network based on an alpha generator added to a batch normalization layer of the second neural network.

[0159] Clause 20. A method according to clause 19, wherein the alpha generator is a zero-shot adapter, and determining the inference information comprises: interpolating global batch normalization statistics using instance test statistics at the batch normalization layer via the zero-shot adapter to generate learning parameters based on a channel-by-channel distance between the global batch normalization statistics and the instance test statistics.

[0160] Clause 21. The method of clause 20, further comprising: training the zero-shot adapter via federated learning.

[0161] Clause 22. The method of clause 20, wherein the plurality of client-side invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

[0162] Clause 23. A federated client device, comprising: one or more memories storing computer-executable instructions; one or more processors coupled to the one or more memories and configured to execute the computer-executable instructions, alone or in combination, to cause the federated client device to: receive a neural network for performing a machine learning task from a federated server, the neural network generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server; and generate inferred information in a domain unknown to the federated server via the neural network based on an alpha generator added to a batch normalization layer of the neural network.

[0163] Clause 24. A federated client device according to clause 23, wherein the alpha generator is a zero-shot adapter, and in order to determine the inference information, the one or more processors are configured individually or in combination to: interpolate global batch normalization statistics using instance test statistics at the batch normalization layer via the zero-shot adapter to generate learning parameters based on the channel-by-channel distance between the global batch normalization statistics and the instance test statistics.

[0164] Clause 25. The federated client device of Clause 24, wherein the one or more processors, individually or in combination, are further configured to cause the federated client device to train the zero-shot adapter via federated learning.

[0165] Clause 26. A federated client device according to any one of clauses 23 to 25, wherein the plurality of client-invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

[0166] Clause 27. A method comprising: receiving a neural network for performing a machine learning task from a federated server, the neural network being generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server; and generating inferred information in a domain unknown to the federated server via the neural network based on an alpha generator added to a batch normalization layer of the neural network.

[0167] Clause 28. A method according to clause 27, wherein the alpha generator is a zero-shot adapter, and determining the inference information includes: interpolating global batch normalization statistics using instance test statistics at the batch normalization layer via the zero-shot adapter to generate learning parameters based on the channel-by-channel distance between the global batch normalization statistics and the instance test statistics.

[0168] Clause 29. The method of clause 28, further comprising: training the zero-shot adapter via federated learning.

[0169] Clause 30. A method according to any one of clauses 27 to 29, wherein the plurality of client-side invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

[0170] The previous description is provided to enable any person of ordinary skill in the art to practice the various aspects described herein. Various modifications to these aspects will be easily understood by those of ordinary skill in the art, and the universal principles defined in this article can be applied to other aspects. The claims are not intended to be limited to the aspects shown in this article, but should be granted the full scope consistent with the claims in language, wherein the singular reference of the elements is not intended to represent "there is and only one", but "one or more", unless otherwise stated. Unless otherwise specified, the term "some" refers to one or more. Combinations such as "at least one of A, B or C", "one or more of A, B or C", "at least one of A, B and C", "one or more of A, B and C" and "A, B, C or any combination thereof" include any combination of A, B or C, and may include multiples of A, multiples of B or multiples of C. Specifically, combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B and C," "one or more of A, B, and C," and "A, B, C, or any combination thereof" may be only A, only B, only C, A and B, A and C, B and C, or A and B and C, wherein any such combination may include one or more members of A, B, or C. All structural and functional equivalents of the elements of the various aspects described throughout the present disclosure that are or later will be known to one of ordinary skill in the art are expressly incorporated herein by reference and are intended to be covered by the claims. In addition, nothing disclosed herein is intended to be dedicated to the public, regardless of whether such disclosure is explicitly recited in the claims. Words such as "module," "mechanism," "element," "device," etc. cannot replace the word "component." Therefore, no claim element will be understood to be a part-plus-function unless the element is explicitly recited using the phrase "component for..."

Claims

1. A federated client device, comprising: one or more memories storing computer-executable instructions; and one or more processors coupled to the one or more memories and configured to, alone or in combination, execute the computer-executable instructions to cause the federated client device to: client-side invariant information identifying a neural network for performing a machine learning (ML) task in a first domain known to the federated server; and The client-invariant information is sent to the federated server, and the federated server is configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client-invariant information and other client-invariant information of another neural network for performing the ML task in a second domain known to the federated server.

2. The federated client device of claim 1 , wherein the one or more processors, alone or in combination, are further configured to cause the federated client device to: receiving federated ML information from the federated server; and The neural network is generated based on the federated ML information.

3. The federated client device of claim 1, wherein the client-invariant information comprises at least one of a weight, a gradient, or a scaling factor, the scaling factor being a multiplicand or a summand.

4. The federated client device of claim 1 , wherein to identify the client-invariant information of the neural network, the one or more processors are configured, individually or in combination, to: An invariant layer of the neural network that is not a batch normalization layer of the neural network is selected as the client-side invariant information.

5. The federated client device of claim 1 , wherein the one or more processors, alone or in combination, are further configured to cause the federated client device to: The neural network is trained by interpolating instance statistics of the sample data and global statistics associated with a plurality of other federated client devices.

6. The federated client device of claim 5, wherein the global statistics comprise batch-normalized statistics collected from the plurality of other federated client devices.

7. The federated client device of claim 6, wherein to train the neural network, the one or more processors are configured, individually or in combination, to: The neural network is trained by interpolating the mean and variance of the instance statistics using the mean and variance of the batch normalization statistics.

8. The federated client device of claim 1 , wherein the federated server is a first federated server, the ML task is a first ML task, the neural network is a first neural network, and the one or more processors are configured, alone or in combination, to cause the federated client device to: receiving federated ML information from a second federated server, the federated ML information comprising a second neural network generated by the second federated server for performing a second ML task and generated by the second federated server using a plurality of client-invariant information of a plurality of other neural networks for performing the second ML task in a plurality of domains known to the second federated server; and Generate inference information in a domain unknown to the second federated server via the second neural network based on an alpha generator added to a batch normalization layer of the second neural network.

9. The federated client device of claim 8, wherein the alpha generator is a zero sample adapter, and to determine the inferred information, the one or more processors are configured, alone or in combination, to cause the federated client device to: Global batch normalization statistics are interpolated at the batch normalization layer via the zero-shot adapter using instance test statistics to generate learning parameters based on a channel-wise distance between the global batch normalization statistics and the instance test statistics.

10. The federated client device of claim 9, wherein the at least one processor is further configured to train the zero-shot adapter via federated learning.

11. The federated client device of claim 9, wherein the plurality of client-invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

12. A method comprising: client-side invariant information identifying a neural network for performing a machine learning (ML) task in a first domain known to the federated server; as well as The client-invariant information is sent to the federated server, and the federated server is configured to generate an ML model for performing the ML task in a domain unknown to the federated server based on the client-invariant information and other client-invariant information of another neural network for performing the ML task in a second domain known to the federated server.

13. The method according to claim 12, further comprising: receiving federated ML information from the federated server; as well as The neural network is generated based on the federated ML information.

14. The method of claim 12, wherein the client-invariant information comprises at least one of a weight, a gradient, or a scaling factor, the scaling factor being a multiplicand or a summand.

15. The method of claim 12, wherein the client-side invariant information identifying the neural network comprises: An invariant layer of the neural network that is not a batch normalization layer of the neural network is selected as the client-side invariant information.

16. The method according to claim 12, further comprising: The neural network is trained by interpolating instance statistics of the sample data and global statistics associated with a plurality of other federated client devices.

17. The method of claim 16, wherein the global statistics comprise batch-normalized statistics collected from the plurality of other federated client devices.

18. The method of claim 17, wherein training the neural network comprises: The neural network is trained by interpolating the mean and variance of the instance statistics using the mean and variance of the batch normalization statistics.

19. The method of claim 12, wherein the federated server is a first federated server, the ML task is a first ML task, the neural network is a first neural network, and the method further comprises: receiving federated ML information from a second federated server, the federated ML information comprising a second neural network generated by the second federated server for performing a second ML task and generated by the second federated server using a plurality of client-invariant information of a plurality of other neural networks for performing the second ML task in a plurality of domains known to the second federated server; as well as Generate inference information in a domain unknown to the second federated server via the second neural network based on an alpha generator added to a batch normalization layer of the second neural network.

20. The method of claim 19, wherein the alpha generator is a zero-sample adapter, and determining the inferred information comprises: Global batch normalization statistics are interpolated at the batch normalization layer via the zero-shot adapter using instance test statistics to generate learning parameters based on a channel-wise distance between the global batch normalization statistics and the instance test statistics.

21. The method according to claim 20, further comprising: The zero-shot adapter is trained via federated learning.

22. The method of claim 20, wherein the plurality of client-side invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

23. A federated client device comprising: one or more memories storing computer-executable instructions; and one or more processors coupled to the one or more memories and configured to, alone or in combination, execute the computer-executable instructions to cause the federated client device to: receiving, from a federated server, a neural network for performing a machine learning task, the neural network generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server; as well as Inferred information is generated via the neural network in a domain unknown to the federated server based on an alpha generator added to a batch normalization layer of the neural network.

24. The federated client device of claim 23, wherein the alpha generator is a zero sample adapter, and to determine the inferred information, the one or more processors are configured, individually or in combination, to: Global batch normalization statistics are interpolated at the batch normalization layer via the zero-shot adapter using instance test statistics to generate learning parameters based on a channel-wise distance between the global batch normalization statistics and the instance test statistics.

25. The federated client device of claim 24, wherein the one or more processors, alone or in combination, are further configured to cause the federated client device to train the zero-shot adapter via federated learning.

26. The federated client device of claim 23, wherein the plurality of client-invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

27. A method comprising: receiving, from a federated server, a neural network for performing a machine learning task, the neural network generated by the federated server using a plurality of client-invariant information received from a plurality of other neural networks associated with one or more domains known to the federated server; as well as Inferred information is generated via the neural network in a domain unknown to the federated server based on an alpha generator added to a batch normalization layer of the neural network.

28. The method of claim 27, wherein the alpha generator is a zero sample adapter, and determining the inferred information comprises: Global batch normalization statistics are interpolated at the batch normalization layer via the zero-shot adapter using instance test statistics to generate learning parameters based on a channel-wise distance between the global batch normalization statistics and the instance test statistics.

29. The method according to claim 28, further comprising: The zero-shot adapter is trained via federated learning.

30. The method of claim 27, wherein the plurality of client-side invariant information corresponds to layers of the plurality of other neural networks that are not batch normalization layers.

Images