WO2022025303A1 - Method for performing joint reasoning in wireless communication system, and apparatus therefor - Google Patents

Abstract

A joint reasoning method is disclosed. A method according to one embodiment of the present specification may comprise the steps of: receiving particular values for joint reasoning from two or more terminals; and performing a classification operation by using the received particular values and a global AI model stored in a memory, and thus an optimal AI processing result can be derived by collecting AI processing result values of a plurality of devices or median values derived in the process thereof. A terminal, a server, and a system of the present specification can be linked to an artificial intelligence module, a drone (unmanned aerial vehicle (UAV)), a robot, an augmented reality (AR) device, a virtual reality (VR) device, a device related to a 5G or 6G service, and the like.

Description

Method and apparatus for performing federated inference in a wireless communication system

The present specification relates to a method and apparatus for performing federated inference in a wireless communication system.

Recently, with the development of deep learning (DL) technology, the number of products and services applied with artificial intelligence (AI)/machine learning (ML) technology is increasing. Currently, most products and services transmit data acquired from user equipment (UE) to the cloud, and perform learning and inference in the cloud.

A framework that sends data acquired from a terminal to the cloud as it is has various problems. First, power consumption is required for a terminal connected wirelessly to a communication network to transmit data to the cloud, but since many wirelessly connected terminals operate on batteries, frequent transmission of large amounts of data time can be greatly reduced. Second, in general, deep learning requires a large amount of data, but continuously transmitting data by a plurality of terminals may significantly increase the load on the network. Third, performing all learning and estimation in the cloud greatly increases the amount of computation in the cloud, increases power consumption, increases feedback latency to the terminal, and, as a result, increases service cost or quality become a degrading factor. Fourth, privacy and security issues are likely to arise.

SUMMARY OF THE INVENTION The present specification aims to solve the above-mentioned needs and/or problems.

In addition, the present specification provides a method for performing joint inference in a wireless communication system that obtains more accurate estimation results by collecting estimation results performed by a plurality of devices that have limitations in complicated calculation and estimation due to limitations in battery and hardware, and a method thereof It aims to implement the device.

In addition, the present specification, when performing an estimation operation in a server or a terminal performing a server function by using data obtained from a plurality of terminals, minimizes the data transmission amount and power consumption of each terminal, and increases the accuracy of the final estimation An object of the present invention is to implement a method and an apparatus for performing federated inference in a wireless communication system.

The method according to an embodiment of the present specification includes receiving specific values for joint inference from two or more terminals, and performing a classification operation using the received specific values and a global AI model stored in a memory.

The method may further include creating one or more terminal groups including at least some of the two or more terminals.

In addition, the terminal group may be grouped based on at least one of (i) identification information of a local AI model stored in the terminal, (ii) location information of the terminal, or (iii) CSI (Channel State Information). .

In addition, when two or more terminal groups exist, each terminal group may have a different local AI model.

In addition, the local AI model is a neural network model including one or more layers having one or more nodes and a non-linenar activatoin layer, and the specific value is an input value of the non-linear activation layer or It may be any one of the output values.

Also, the nonlinear activation layer may be a first softmax layer.

In addition, the method further comprises the step of selecting any one of the terminals included in the terminal group as a master UE, wherein the two or more terminals transmitting a specific value for the joint inference are configured only with the representative terminal. can

A method according to another embodiment of the present specification includes receiving specific values for joint inference from two or more external terminals, generating an intermediate value using the received specific values and a local AI model stored in a memory, and the and transmitting the generated intermediate value to a server for federated inference.

A terminal according to another embodiment of the present specification includes one or more transceivers, one or more processors, and one or more memories connected to the one or more processors and storing instructions, the instructions being sent to the one or more processors. When executed by one or more processors to support operations for federated inference, the operations include: the operation of receiving specific values for federated inference from two or more external terminals, the received specific values and the local AI model stored in the memory and generating an intermediate value by using the intermediate value and transmitting the generated intermediate value to a server for federated inference.

A method for performing joint inference in a wireless communication system according to an embodiment of the present specification and an effect of the apparatus will be described as follows.

In the present specification, more accurate estimation results may be obtained by collecting estimation results performed by a plurality of devices having limitations in complicated calculation and estimation due to limitations in battery and hardware.

In addition, the present specification can minimize the data transmission amount and power consumption of each terminal and increase the accuracy of the final estimation when performing an estimation operation in a server or a terminal performing a server function using data obtained from a plurality of terminals have.

The effects obtainable in the present specification are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those of ordinary skill in the art to which this specification belongs from the description below. .

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, which are included as a part of the detailed description to help the understanding of the present specification, provide embodiments of the present specification, and together with the detailed description, explain the technical features of the present specification.

1 illustrates physical channels and general signal transmission used in a 3GPP system.

2 is a diagram illustrating an example of a communication structure that can be provided in a 6G system.

3 illustrates a perceptron structure.

4 illustrates a multilayer perceptron structure.

5 illustrates a deep neural network structure.

6 illustrates a convolutional neural network structure.

7 illustrates a filter operation in a convolutional neural network.

8 illustrates a neural network structure in which a cyclic loop exists.

9 illustrates the operational structure of a recurrent neural network.

10 shows an example of an electromagnetic spectrum.

11 shows an example of a THz communication application.

12 shows an example of an electronic device-based THz wireless communication transceiver.

13 shows an example of a method for generating an optical device-based THz signal.

14 shows an example of an optical element-based THz wireless communication transceiver.

15 illustrates a photonic source based transmitter structure.

16 illustrates the structure of an optical modulator.

17 is a flowchart of an associative inference method according to various embodiments of the present specification.

18 to 20 are exemplary views of the associative reasoning method of FIG. 17 .

20 to 22 are exemplary diagrams of various classification algorithms of the associative inference method of FIG. 17 .

21 to 23 are exemplary diagrams of various classification algorithms of the associative inference method of FIG. 17 .

24 is a flowchart of a learning method of an AI model according to an embodiment of the present specification.

25 to 27 are exemplary diagrams of the learning method of FIG. 24 .

28 illustrates a communication system applied to this specification.

29 illustrates a wireless device applicable to this specification.

Hereinafter, the embodiments disclosed in the present specification will be described in detail with reference to the accompanying drawings, but the same or similar components are assigned the same reference numbers regardless of reference numerals, and redundant description thereof will be omitted. The suffixes "module" and "part" for components used in the following description are given or mixed in consideration of only the ease of writing the specification, and do not have distinct meanings or roles by themselves. In addition, in describing the embodiments disclosed in the present specification, if it is determined that detailed descriptions of related known technologies may obscure the gist of the embodiments disclosed in the present specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in the present specification, and the technical spirit disclosed in this specification is not limited by the accompanying drawings, and all changes included in the spirit and scope of the present specification , should be understood to include equivalents or substitutes.

Terms including an ordinal number such as 1st, 2nd, etc. may be used to describe various elements, but the elements are not limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

When an element is referred to as being “connected” or “connected” to another element, it is understood that it may be directly connected or connected to the other element, but other elements may exist in between. it should be On the other hand, when it is said that a certain element is "directly connected" or "directly connected" to another element, it should be understood that the other element does not exist in the middle.

The singular expression includes the plural expression unless the context clearly dictates otherwise.

In the present application, terms such as “comprises” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

The following techniques can be used in various radio access systems such as CDMA, FDMA, TDMA, OFDMA, SC-FDMA, and the like. CDMA may be implemented with a radio technology such as Universal Terrestrial Radio Access (UTRA) or CDMA2000. TDMA may be implemented with a radio technology such as Global System for Mobile communications (GSM)/General Packet Radio Service (GPRS)/Enhanced Data Rates for GSM Evolution (EDGE). OFDMA may be implemented with a radio technology such as IEEE 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802-20, Evolved UTRA (E-UTRA), and the like. UTRA is part of the Universal Mobile Telecommunications System (UMTS). 3GPP (3rd Generation Partnership Project) Long Term Evolution (LTE) is a part of Evolved UMTS (E-UMTS) using E-UTRA and LTE-A (Advanced)/LTE-A pro is an evolved version of 3GPP LTE. 3GPP NR (New Radio or New Radio Access Technology) is an evolved version of 3GPP LTE/LTE-A/LTE-A pro. 3GPP 6G may be an evolved version of 3GPP NR.

For clarity of explanation, description is based on a 3GPP communication system (eg, LTE, NR, etc.), but the technical spirit of the present specification is not limited thereto. LTE refers to technology after 3GPP TS 36.xxx Release 8. In detail, LTE technology after 3GPP TS 36.xxx Release 10 is referred to as LTE-A, and LTE technology after 3GPP TS 36.xxx Release 13 is referred to as LTE-A pro. 3GPP NR refers to technology after TS 38.xxx Release 15. 3GPP 6G may refer to technology after TS Release 17 and/or Release 18. "xxx" stands for standard document detail number. LTE/NR/6G may be collectively referred to as a 3GPP system. For background art, terms, abbreviations, etc. used in the description of this specification, reference may be made to matters described in standard documents published before this specification. For example, you can refer to the following documents:

3GPP LTE

- 36.211: Physical channels and modulation

- 36.212: Multiplexing and channel coding

- 36.213: Physical layer procedures

- 36.300: Overall description

- 36.331: Radio Resource Control (RRC)

3GPP NR

- 38.211: Physical channels and modulation

- 38.212: Multiplexing and channel coding

- 38.213: Physical layer procedures for control

- 38.214: Physical layer procedures for data

- 38.300: NR and NG-RAN Overall Description

- 38.331: Radio Resource Control (RRC) protocol specification

Physical Channels and Frame Structure

Physical channels and general signal transmission

1 illustrates physical channels and general signal transmission used in a 3GPP system. In a wireless communication system, a terminal receives information through a downlink (DL) from a base station, and the terminal transmits information through an uplink (UL) to the base station. Information transmitted and received between the base station and the terminal includes data and various control information, and various physical channels exist according to the type/use of the information they transmit and receive.

When the terminal is powered on or newly enters a cell, the terminal performs an initial cell search operation, such as synchronizing with the base station (S11). To this end, the terminal receives a primary synchronization signal (PSS) and a secondary synchronization signal (SSS) from the base station, synchronizes with the base station, and obtains information such as a cell ID. Thereafter, the terminal may receive a physical broadcast channel (PBCH) from the base station to obtain intra-cell broadcast information. On the other hand, the UE may receive a downlink reference signal (DL RS) in the initial cell search step to check the downlink channel state.

After the initial cell search, the UE receives a Physical Downlink Control Channel (PDCCH) and a Physical Downlink Control Channel (PDSCH) according to information carried on the PDCCH to obtain more specific system information. It can be done (S12).

On the other hand, when first accessing the base station or there is no radio resource for signal transmission, the terminal may perform a random access procedure (RACH) for the base station (S13 to S16). To this end, the UE transmits a specific sequence as a preamble through a Physical Random Access Channel (PRACH) (S13 and S15), and a response message to the preamble through the PDCCH and the corresponding PDSCH ((Random Access (RAR)) Response) message) In the case of contention-based RACH, a contention resolution procedure may be additionally performed (S16).

After performing the procedure as described above, the UE performs PDCCH/PDSCH reception (S17) and Physical Uplink Shared Channel (PUSCH)/Physical Uplink Control Channel (Physical Uplink) as a general uplink/downlink signal transmission procedure. Control Channel (PUCCH) transmission (S18) may be performed. In particular, the UE may receive downlink control information (DCI) through the PDCCH. Here, the DCI includes control information such as resource allocation information for the terminal, and different formats may be applied according to the purpose of use.

On the other hand, the control information that the terminal transmits to the base station through the uplink or the terminal receives from the base station includes a downlink/uplink ACK/NACK signal, a channel quality indicator (CQI), a precoding matrix index (PMI), and a rank indicator (RI). ) and the like. The UE may transmit the above-described control information such as CQI/PMI/RI through PUSCH and/or PUCCH.

Structures of uplink and downlink channels

Downlink Channel Structure

The base station transmits a related signal to the terminal through a downlink channel to be described later, and the terminal receives the related signal from the base station through a downlink channel to be described later.

(1) Physical Downlink Shared Channel (PDSCH)

PDSCH carries downlink data (eg, DL-shared channel transport block, DL-SCH TB), and modulation methods such as Quadrature Phase Shift Keying (QPSK), 16 Quadrature Amplitude Modulation (QAM), 64 QAM, and 256 QAM are available. applies. A codeword is generated by encoding the TB. A PDSCH can carry multiple codewords. Scrambling and modulation mapping are performed for each codeword, and modulation symbols generated from each codeword are mapped to one or more layers (Layer mapping). Each layer is mapped to a resource together with a demodulation reference signal (DMRS), is generated as an OFDM symbol signal, and is transmitted through a corresponding antenna port.

(2) Physical Downlink Control Channel (PDCCH)

The PDCCH carries downlink control information (DCI) and a QPSK modulation method is applied. One PDCCH is composed of 1, 2, 4, 8, 16 CCEs (Control Channel Elements) according to an Aggregation Level (AL). One CCE consists of six REGs (Resource Element Groups). One REG is defined as one OFDM symbol and one (P)RB.

The UE obtains DCI transmitted through the PDCCH by performing decoding (aka, blind decoding) on the set of PDCCH candidates. A set of PDCCH candidates decoded by the UE is defined as a PDCCH search space set. The search space set may be a common search space or a UE-specific search space. The UE may acquire DCI by monitoring PDCCH candidates in one or more search space sets configured by MIB or higher layer signaling.

Uplink Channel Structure

The terminal transmits a related signal to the base station through an uplink channel to be described later, and the base station receives the related signal from the terminal through an uplink channel to be described later.

(1) Physical Uplink Shared Channel (PUSCH)

PUSCH carries uplink data (eg, UL-shared channel transport block, UL-SCH TB) and/or uplink control information (UCI), and CP-OFDM (Cyclic Prefix - Orthogonal Frequency Division Multiplexing) waveform (waveform) , DFT-s-OFDM (Discrete Fourier Transform - spread - Orthogonal Frequency Division Multiplexing) is transmitted based on the waveform. When the PUSCH is transmitted based on the DFT-s-OFDM waveform, the UE transmits the PUSCH by applying transform precoding. For example, when transform precoding is not possible (eg, transform precoding is disabled), the UE transmits a PUSCH based on the CP-OFDM waveform, and when transform precoding is possible (eg, transform precoding is enabled), the UE transmits the CP-OFDM PUSCH may be transmitted based on a waveform or a DFT-s-OFDM waveform. PUSCH transmission is dynamically scheduled by a UL grant in DCI, or based on higher layer (eg, RRC) signaling (and/or Layer 1 (L1) signaling (eg, PDCCH)) semi-statically. Can be scheduled (configured grant). PUSCH transmission may be performed on a codebook-based or non-codebook-based basis.

(2) Physical Uplink Control Channel (PUCCH)

The PUCCH carries uplink control information, HARQ-ACK and/or a scheduling request (SR), and may be divided into a plurality of PUCCHs according to the PUCCH transmission length.

6G system general

6G (wireless) systems have (i) very high data rates per device, (ii) very large number of connected devices, (iii) global connectivity, (iv) very low latency, (v) battery- It aims to reduce energy consumption of battery-free IoT devices, (vi) ultra-reliable connections, and (vii) connected intelligence with machine learning capabilities. The vision of the 6G system can be in four aspects: intelligent connectivity, deep connectivity, holographic connectivity, and ubiquitous connectivity, and the 6G system can satisfy the requirements shown in Table 1 below. That is, Table 1 is a table showing an example of the requirements of the 6G system.

6G systems include Enhanced mobile broadband (eMBB), Ultra-reliable low latency communications (URLLC), massive machine-type communication (mMTC), AI integrated communication, Tactile internet, High throughput, High network capacity, High energy efficiency, Low backhaul and It may have key factors such as access network congestion and enhanced data security. FIG. 2 is a diagram showing an example of a communication structure that can be provided in a 6G system.

6G systems are expected to have 50 times higher simultaneous wireless connectivity than 5G wireless communication systems. URLLC, a key feature of 5G, will become an even more important technology by providing an end-to-end delay of less than 1ms in 6G communication. 6G systems will have much better volumetric spectral efficiencies as opposed to frequently used areal spectral efficiencies. The 6G system can provide very long battery life and advanced battery technology for energy harvesting, so mobile devices will not need to be charged separately in the 6G system. New network characteristics in 6G may be as follows.

- Satellites integrated network: 6G is expected to be integrated with satellites to provide a global mobile population. The integration of terrestrial, satellite and public networks into one wireless communication system is very important for 6G.

- Connected intelligence: Unlike previous generations of wireless communication systems, 6G is revolutionary and will update the evolution of wireless from “connected things” to “connected intelligence”. AI may be applied in each step of a communication procedure (or each procedure of signal processing to be described later).

- Seamless integration wireless information and energy transfer: The 6G wireless network will deliver power to charge the batteries of devices such as smartphones and sensors. Therefore, wireless information and energy transfer (WIET) will be integrated.

- Ubiquitous super 3D connectivity: access to networks and core network functions of drones and very low Earth orbiting satellites will create super 3D connectivity in 6G ubiquitous.

In the above new network characteristics of 6G, some general requirements may be as follows.

- Small cell networks: The idea of small cell networks was introduced to improve the received signal quality as a result of improved throughput, energy efficiency and spectral efficiency in cellular systems. As a result, small cell networks are essential characteristics for communication systems beyond 5G and Beyond 5G (5GB). Accordingly, the 6G communication system also adopts the characteristics of the small cell network.

- Ultra-dense heterogeneous network: Ultra-dense heterogeneous networks will be another important characteristic of 6G communication systems. A multi-tier network composed of heterogeneous networks improves overall QoS and reduces costs.

- High-capacity backhaul: A backhaul connection is characterized as a high-capacity backhaul network to support high-capacity traffic. High-speed fiber optics and free-space optics (FSO) systems may be possible solutions to this problem.

- Radar technology integrated with mobile technology: High-precision localization (or location-based service) through communication is one of the functions of the 6G wireless communication system. Therefore, the radar system will be integrated with the 6G network.

- Softwarization and virtualization: Softening and virtualization are two important features that underlie the design process in 5GB networks to ensure flexibility, reconfigurability and programmability. In addition, billions of devices can be shared in a shared physical infrastructure.

Core implementation technology of 6G system

Artificial Intelligence

The most important and newly introduced technology for 6G systems is AI. AI was not involved in the 4G system. 5G systems will support partial or very limited AI. However, the 6G system will be AI-enabled for full automation. Advances in machine learning will create more intelligent networks for real-time communication in 6G. Incorporating AI into communications can simplify and enhance real-time data transmission. AI can use numerous analytics to determine how complex target tasks are performed. In other words, AI can increase efficiency and reduce processing delays.

Time-consuming tasks such as handovers, network selection, and resource scheduling can be performed instantly by using AI. AI can also play an important role in M2M, machine-to-human and human-to-machine communication. In addition, AI can be a rapid communication in BCI (Brain Computer Interface). AI-based communication systems can be supported by metamaterials, intelligent structures, intelligent networks, intelligent devices, intelligent cognitive radios, self-sustaining wireless networks, and machine learning.

Recently, attempts have been made to integrate AI with wireless communication systems, but these have been focused on the application layer and network layer, especially deep learning, in the field of wireless resource management and allocation. However, these studies are gradually developing into the MAC layer and the physical layer, and in particular, attempts to combine deep learning with wireless transmission in the physical layer are appearing. AI-based physical layer transmission means applying a signal processing and communication mechanism based on an AI driver rather than a traditional communication framework in a fundamental signal processing and communication mechanism. For example, deep learning-based channel coding and decoding, deep learning-based signal estimation and detection, deep learning-based MIMO mechanism, AI-based resource scheduling and It may include an allocation (allocation) and the like.

Machine learning may be used for channel estimation and channel tracking, and may be used for power allocation, interference cancellation, and the like in a physical layer of a downlink (DL). In addition, machine learning may be used for antenna selection, power control, symbol detection, and the like in a MIMO system.

However, the application of DNN for transmission in the physical layer may have the following problems.

Deep learning-based AI algorithms require large amounts of training data to optimize training parameters. However, due to a limitation in acquiring data in a specific channel environment as training data, a lot of training data is used offline. This is because static training on training data in a specific channel environment may cause a contradiction between dynamic characteristics and diversity of a wireless channel.

In addition, current deep learning mainly targets real signals. However, signals of the physical layer of wireless communication are complex signals. In order to match the characteristics of a wireless communication signal, further research on a neural network for detecting a complex domain signal is needed.

Hereinafter, machine learning will be described in more detail.

Machine learning refers to a set of actions that trains a machine to create a machine that can perform tasks that humans can or cannot do. Machine learning requires data and a learning model. In machine learning, data learning methods can be roughly divided into three types: supervised learning, unsupervised learning, and reinforcement learning.

Neural network learning is to minimize output errors. Neural network learning repeatedly inputs learning data into the neural network, calculates the output and target errors of the neural network for the training data, and backpropagates the neural network error from the output layer of the neural network to the input layer in the direction to reduce the error. ) to update the weight of each node in the neural network.

Supervised learning uses training data in which the correct answer is labeled in the training data, and in unsupervised learning, the correct answer may not be labeled in the training data. That is, for example, learning data in the case of supervised learning related to data classification may be data in which categories are labeled for each of the training data. The labeled training data is input to the neural network, and an error can be calculated by comparing the output (category) of the neural network with the label of the training data. The calculated error is back propagated in the reverse direction (ie, from the output layer to the input layer) in the neural network, and the connection weight of each node of each layer of the neural network may be updated according to the back propagation. A change amount of the connection weight of each node to be updated may be determined according to a learning rate. The computation of the neural network on the input data and the backpropagation of errors can constitute a learning cycle (epoch). The learning rate may be applied differently depending on the number of repetitions of the learning cycle of the neural network. For example, in the early stage of learning a neural network, a high learning rate can be used to increase the efficiency by allowing the neural network to quickly obtain a certain level of performance, and in the late learning period, a low learning rate can be used to increase the accuracy.

The learning method may vary depending on the characteristics of the data. For example, when the purpose of accurately predicting data transmitted from a transmitter in a communication system is at a receiver, it is preferable to perform learning using supervised learning rather than unsupervised learning or reinforcement learning.

The learning model corresponds to the human brain, and the most basic linear model can be considered. ) is called

The neural network cord used as a learning method is largely divided into deep neural networks (DNN), convolutional deep neural networks (CNN), and recurrent boltzmann machine (RNN) methods. have.

An artificial neural network is an example of connecting several perceptrons.

Referring to FIG. 3, when an input vector x=(x1,x2,...,xd) is input, each component is multiplied by a weight (W1,W2,...,Wd), and after summing all the results, The whole process of applying the activation function is called a perceptron. The huge artificial neural network structure may extend the simplified perceptron structure shown in FIG. 3 to apply input vectors to different multidimensional perceptrons. For convenience of description, an input value or an output value is referred to as a node.

Meanwhile, the perceptron structure shown in FIG. 3 can be described as being composed of a total of three layers based on an input value and an output value. An artificial neural network in which H (d+1)-dimensional perceptrons exist between the 1st layer and the 2nd layer and K (H+1)-dimensional perceptrons exist between the 2nd layer and the 3rd layer can be expressed as shown in FIG. 4 .

The layer where the input vector is located is called the input layer, the layer where the final output value is located is called the output layer, and all the layers located between the input layer and the output layer are called hidden layers. In the example of FIG. 4 , three layers are disclosed, but when counting the actual number of artificial neural network layers, the input layer is counted except for the input layer, so it can be viewed as a total of two layers. The artificial neural network is constructed by connecting the perceptrons of the basic blocks in two dimensions.

The aforementioned input layer, hidden layer, and output layer can be jointly applied in various artificial neural network structures such as CNN and RNN to be described later as well as multi-layer perceptron. As the number of hidden layers increases, the artificial neural network becomes deeper, and a machine learning paradigm that uses a sufficiently deep artificial neural network as a learning model is called deep learning. Also, an artificial neural network used for deep learning is called a deep neural network (DNN).

The deep neural network shown in FIG. 5 is a multilayer perceptron composed of eight hidden layers + output layers. The multi-layered perceptron structure is referred to as a fully-connected neural network. In a fully connected neural network, a connection relationship does not exist between nodes located in the same layer, and a connection relationship exists only between nodes located in adjacent layers. DNN has a fully connected neural network structure and is composed of a combination of a number of hidden layers and activation functions, so it can be usefully applied to figure out the correlation between input and output. Here, the correlation characteristic may mean a joint probability of input/output.

‘On the other hand, depending on how a plurality of perceptrons are connected to each other, various artificial neural network structures different from the aforementioned DNN can be formed.

In DNN, nodes located inside one layer are arranged in a one-dimensional vertical direction. However, in FIG. 6 , it may be assumed that the nodes are two-dimensionally arranged with w horizontally and h vertical nodes (convolutional neural network structure of FIG. 6 ). In this case, since a weight is added per connection in the connection process from one input node to the hidden layer, a total of hХw weights must be considered. Since there are hХw nodes in the input layer, a total of h ² w ² weights are needed between two adjacent layers.

The convolutional neural network of FIG. 6 has a problem in that the number of weights increases exponentially according to the number of connections, so instead of considering the connection of all modes between adjacent layers, it is assumed that a filter with a small size exists in FIG. 7 As in Fig., the weighted sum and activation function calculations are performed on the overlapping filters.

One filter has a weight corresponding to the number corresponding to its size, and weight learning can be performed so that a specific feature on an image can be extracted and output as a factor. In FIG. 7 , a filter with a size of 3Х3 is applied to the upper left 3Х3 region of the input layer, and an output value obtained by performing weighted sum and activation function operations on the corresponding node is stored in z22.

The filter performs weight sum and activation function calculations while moving horizontally and vertically at regular intervals while scanning the input layer, and places the output value at the current filter position. Such a calculation method is similar to a convolution operation on an image in the field of computer vision, so a deep neural network with such a structure is called a convolutional neural network (CNN), and a hidden layer generated as a result of a convolution operation is called a convolutional layer. Also, a neural network having a plurality of convolutional layers is called a deep convolutional neural network (DCNN).

In the convolution layer, the number of weights can be reduced by calculating the weighted sum by including only nodes located in the region covered by the filter in the node where the filter is currently located. Due to this, one filter can be used to focus on features for a local area. Accordingly, CNN can be effectively applied to image data processing in which physical distance in a two-dimensional domain is an important criterion. Meanwhile, in CNN, a plurality of filters may be applied immediately before the convolution layer, and a plurality of output results may be generated through the convolution operation of each filter.

Meanwhile, there may be data whose sequence characteristics are important according to data properties. Considering the length variability and precedence relationship of the sequence data, one element in the data sequence is input at each timestep, and the output vector (hidden vector) of the hidden layer output at a specific time is input together with the next element in the sequence. A structure in which this method is applied to an artificial neural network is called a recurrent neural network structure.

Referring to FIG. 8 , a recurrent neural network (RNN) connects elements (x1(t), x2(t), ,..., xd(t)) of a certain gaze t on a data sequence to a fully connected neural network. In the process of inputting, the weighted sum and activation function are calculated by inputting the hidden vectors (z1(t-1), z2(t-1),..., zH(t-1)) for the immediately preceding time point t-1 during the input process. structure to be applied. The reason why the hidden vector is transferred to the next time point in this way is that information in the input vector at previous time points is considered to be accumulated in the hidden vector of the current time point.

Referring to FIG. 8 , the recurrent neural network operates in a predetermined time sequence with respect to an input data sequence.

The hidden vector (z1(1),z2(1),.. .,zH(1)) is input together with the input vector (x1(2),x2(2),...,xd(2)) of time 2, and then the vector of the hidden layer (z1( 2),z2(2) ,...,zH(2)) are determined. This process is repeatedly performed until time point 2, time point 3, ,, and time point T.

On the other hand, when a plurality of hidden layers are arranged in a recurrent neural network, this is called a deep recurrent neural network (DRNN). The recurrent neural network is designed to be usefully applied to sequence data (eg, natural language processing).

As a neural network core used as a learning method, in addition to DNN, CNN, and RNN, Restricted Boltzmann Machine (RBM), deep belief networks (DBN), Deep Q-Network and It includes various deep learning techniques such as, and can be applied to fields such as computer vision, voice recognition, natural language processing, and voice/signal processing.

Recently, attempts have been made to integrate AI with wireless communication systems, but these have been focused on the application layer and network layer, especially deep learning, in the field of wireless resource management and allocation. However, these studies are gradually developing into the MAC layer and the physical layer, and in particular, attempts to combine deep learning with wireless transmission in the physical layer are appearing. AI-based physical layer transmission means applying a signal processing and communication mechanism based on an AI driver rather than a traditional communication framework in a fundamental signal processing and communication mechanism. For example, deep learning-based channel coding and decoding, deep learning-based signal estimation and detection, deep learning-based MIMO mechanism, AI-based resource scheduling and It may include an allocation (allocation) and the like.

THz (Terahertz) communication

The data rate can be increased by increasing the bandwidth. This can be accomplished by using sub-THz communication with a wide bandwidth and applying advanced large-scale MIMO technology. THz waves, also known as sub-millimeter radiation, typically exhibit a frequency band between 0.1 THz and 10 THz with corresponding wavelengths in the range of 0.03 mm-3 mm. The 100GHz-300GHz band range (Sub THz band) is considered a major part of the THz band for cellular communication. Sub-THz band Addition to mmWave band increases 6G cellular communication capacity. Among the defined THz bands, 300GHz-3THz is in the far-infrared (IR) frequency band. The 300GHz-3THz band is part of the broadband, but at the edge of the wideband, just behind the RF band. Thus, this 300 GHz-3 THz band shows similarities to RF. 10 shows an example of an electromagnetic spectrum.

The main characteristics of THz communication include (i) widely available bandwidth to support very high data rates, and (ii) high path loss occurring at high frequencies (high directional antennas are indispensable). The narrow beamwidth produced by the highly directional antenna reduces interference. The small wavelength of the THz signal allows a much larger number of antenna elements to be integrated into devices and BSs operating in this band. This allows the use of advanced adaptive nesting techniques that can overcome range limitations.

Optical wireless technology

OWC technology is envisioned for 6G communications in addition to RF-based communications for all possible device-to-access networks. These networks connect to network-to-backhaul/fronthaul network connections. OWC technology has already been used since the 4G communication system, but will be used more widely to meet the needs of the 6G communication system. OWC technologies such as light fidelity, visible light communication, optical camera communication, and FSO communication based on a light band are well known technologies. Communication based on optical radio technology can provide very high data rates, low latency and secure communication. LiDAR can also be used for ultra-high-resolution 3D mapping in 6G communication based on wide bands.

FSO backhaul network

The transmitter and receiver characteristics of an FSO system are similar to those of a fiber optic network. Thus, data transmission in an FSO system is similar to that of a fiber optic system. Therefore, FSO can be a good technology to provide backhaul connectivity in 6G systems along with fiber optic networks. Using FSO, very long-distance communication is possible even at distances of 10,000 km or more. FSO supports high-capacity backhaul connections for remote and non-remote areas such as sea, space, underwater, and isolated islands. FSO also supports cellular BS connectivity.

Massive MIMO technology

One of the key technologies to improve spectral efficiency is to apply MIMO technology. As MIMO technology improves, so does the spectral efficiency. Therefore, large-scale MIMO technology will be important in 6G systems. Since the MIMO technology uses multiple paths, a multiplexing technique and a beam generation and operation technique suitable for the THz band should also be considered important so that a data signal can be transmitted through one or more paths.

blockchain

Blockchain will become an important technology for managing large amounts of data in future communication systems. Blockchain is a form of distributed ledger technology, which is a database distributed across numerous nodes or computing devices. Each node replicates and stores an identical copy of the ledger. The blockchain is managed as a peer-to-peer network. It can exist without being managed by a centralized authority or server. Data on the blockchain is collected together and organized into blocks. Blocks are linked together and protected using encryption. Blockchain in nature perfectly complements IoT at scale with improved interoperability, security, privacy, reliability and scalability. Therefore, blockchain technology provides several features such as interoperability between devices, traceability of large amounts of data, autonomous interaction of different IoT systems, and large-scale connection stability of 6G communication systems.

3D Networking

The 6G system integrates terrestrial and public networks to support vertical expansion of user communications. 3D BS will be provided via low orbit satellites and UAVs. Adding a new dimension in terms of elevation and associated degrees of freedom makes 3D connections significantly different from traditional 2D networks.

quantum communication

In the context of 6G networks, unsupervised reinforcement learning of networks is promising. Supervised learning methods cannot label the massive amounts of data generated by 6G. Unsupervised learning does not require labeling. Thus, this technique can be used to autonomously build representations of complex networks. Combining reinforcement learning and unsupervised learning allows networks to operate in a truly autonomous way.

drone

Unmanned Aerial Vehicles (UAVs) or drones will become an important element in 6G wireless communication. In most cases, high-speed data wireless connections are provided using UAV technology. A BS entity is installed in the UAV to provide cellular connectivity. UAVs have certain features not found in fixed BS infrastructure, such as easy deployment, strong line-of-sight links, and degrees of freedom with controlled mobility. During emergencies such as natural disasters, the deployment of terrestrial communications infrastructure is not economically feasible and sometimes cannot provide services in volatile environments. A UAV can easily handle this situation. UAV will become a new paradigm in the field of wireless communication. This technology facilitates the three basic requirements of wireless networks: eMBB, URLLC and mMTC. UAVs can also serve several purposes, such as improving network connectivity, fire detection, disaster emergency services, security and surveillance, pollution monitoring, parking monitoring, incident monitoring, and more. Therefore, UAV technology is recognized as one of the most important technologies for 6G communication.

Cell-free Communication

Tight integration of multiple frequencies and heterogeneous communication technologies is very important in 6G systems. As a result, users can seamlessly move from one network to another without having to make any manual configuration on the device. The best network is automatically selected from the available communication technologies. This will break the limitations of the cell concept in wireless communication. Currently, user movement from one cell to another causes too many handovers in high-density networks, causing handover failures, handover delays, data loss and ping-pong effects. 6G cell-free communication will overcome all of this and provide better QoS. Cell-free communication will be achieved through multi-connectivity and multi-tier hybrid technologies and different heterogeneous radios of devices.

Integration of wireless information and energy transmission

WIET uses the same fields and waves as wireless communication systems. In particular, the sensor and smartphone will be charged using wireless power transfer during communication. WIET is a promising technology for extending the life of battery-charging wireless systems. Therefore, devices without batteries will be supported in 6G communication.

Integration of sensing and communication

An autonomous wireless network is a function that can continuously detect dynamically changing environmental conditions and exchange information between different nodes. In 6G, sensing will be tightly integrated with communications to support autonomous systems.

Consolidation of access backhaul networks

The density of access networks in 6G will be enormous. Each access network is connected by backhaul connections such as fiber optic and FSO networks. To cope with a very large number of access networks, there will be tight integration between the access and backhaul networks.

Holographic Beamforming

Beamforming is a signal processing procedure that adjusts an antenna array to transmit a radio signal in a specific direction. A smart antenna or a subset of an advanced antenna system. Beamforming technology has several advantages such as high call-to-noise ratio, interference prevention and rejection, and high network efficiency. Hologram beamforming (HBF) is a new beamforming method that is significantly different from MIMO systems because it uses a software-defined antenna. HBF will be a very effective approach for efficient and flexible transmission and reception of signals in multi-antenna communication devices in 6G.

Big Data Analytics

Big data analytics is a complex process for analyzing various large data sets or big data. This process ensures complete data management by finding information such as hidden data, unknown correlations and customer propensity. Big data is gathered from a variety of sources such as videos, social networks, images and sensors. This technology is widely used to process massive amounts of data in 6G systems.

Large Intelligent Surface (LIS)

In the case of the THz band signal, the linearity is strong, so there may be many shaded areas due to obstructions. By installing the LIS near these shaded areas, the LIS technology that expands the communication area, strengthens communication stability and enables additional additional services becomes important. The LIS is an artificial surface made of electromagnetic materials, and can change the propagation of incoming and outgoing radio waves. LIS can be seen as an extension of massive MIMO, but the array structure and operation mechanism are different from those of massive MIMO. In addition, LIS has low power consumption in that it operates as a reconfigurable reflector with passive elements, that is, only passively reflects the signal without using an active RF chain. There are advantages to having Also, since each of the passive reflectors of the LIS must independently adjust the phase shift of the incoming signal, it can be advantageous for a wireless communication channel. By properly adjusting the phase shift via the LIS controller, the reflected signal can be gathered at the target receiver to boost the received signal power.

Terahertz (THz) wireless communication general

THz wireless communication uses wireless communication using a THz wave having a frequency of approximately 0.1 to 10THz (1THz=1012Hz), and can mean terahertz (THz) band wireless communication using a very high carrier frequency of 100GHz or more. . THz wave is located between RF (Radio Frequency)/millimeter (mm) and infrared band, (i) It transmits non-metal/non-polar material better than visible light/infrared light, and has a shorter wavelength than RF/millimeter wave, so it has high straightness. Beam focusing may be possible. In addition, since the photon energy of the THz wave is only a few meV, it is harmless to the human body. The frequency band expected to be used for THz wireless communication may be a D-band (110 GHz to 170 GHz) or H-band (220 GHz to 325 GHz) band with low propagation loss due to absorption of molecules in the air. The standardization discussion on THz wireless communication is being discussed centered on the IEEE 802.15 THz working group in addition to 3GPP, and the standard documents issued by the IEEE 802.15 Task Group (TG3d, TG3e) may specify or supplement the content described in this specification. have. THz wireless communication may be applied to wireless recognition, sensing, imaging, wireless communication, THz navigation, and the like.

11 is a diagram showing an example of THz communication application.

As shown in FIG. 11 , a THz wireless communication scenario may be classified into a macro network, a micro network, and a nanoscale network. In the macro network, THz wireless communication can be applied to vehicle-to-vehicle connection and backhaul/fronthaul connection. THz wireless communication in micro networks is applied to indoor small cells, fixed point-to-point or multi-point connections such as wireless connections in data centers, and near-field communication such as kiosk downloading. can be

Table 2 below is a table showing an example of a technique that can be used in the THz wave.

THz wireless communication can be classified based on a method for generating and receiving THz. The THz generation method can be classified into an optical device or an electronic device-based technology. 12 is a diagram illustrating an example of an electronic device-based THz wireless communication transceiver.

A method of generating THz using an electronic device includes a method using a semiconductor device such as a Resonant Tunneling Diode (RTD), a method using a local oscillator and a multiplier, and an integrated circuit based on a compound semiconductor HEMT (High Electron Mobility Transistor). MMIC (Monolithic Microwave Integrated Circuits) method using In the case of FIG. 12 , a doubler, tripler, or multiplier is applied to increase the frequency, and it is radiated by the antenna through the subharmonic mixer. Since the THz band forms a high frequency, a multiplier is essential. Here, the multiplier is a circuit that has an output frequency that is N times that of the input, matches the desired harmonic frequency, and filters out all other frequencies. Also, beamforming may be implemented by applying an array antenna or the like to the antenna of FIG. 12 . In FIG. 12 , IF denotes an intermediate frequency, tripler, multipler denote a multiplier, PA Power Amplifier denotes, LNA denotes a low noise amplifier, and PLL denotes a phase lock circuit (Phase). -Locked Loop).

13 is a diagram illustrating an example of a method of generating an optical device-based THz signal, and FIG. 14 is a diagram illustrating an example of an optical device-based THz wireless communication transceiver.

Optical device-based THz wireless communication technology refers to a method of generating and modulating a THz signal using an optical device. The optical element-based THz signal generation technology is a technology that generates a high-speed optical signal using a laser and an optical modulator, and converts it into a THz signal using an ultra-high-speed photodetector. In this technology, it is easier to increase the frequency compared to the technology using only electronic devices, it is possible to generate a high-power signal, and it is possible to obtain a flat response characteristic in a wide frequency band. 13, a laser diode, a broadband optical modulator, and a high-speed photodetector are required to generate an optical device-based THz signal. In the case of FIG. 13 , a THz signal corresponding to a difference in wavelength between the lasers is generated by multiplexing the light signals of two lasers having different wavelengths. In FIG. 13 , an optical coupler refers to a semiconductor device that uses light waves to transmit electrical signals to provide coupling with electrical insulation between circuits or systems, and UTC-PD (Uni-Traveling Carrier Photo-) Detector) is one of the photodetectors, which uses electrons as active carriers and reduces the movement time of electrons by bandgap grading. UTC-PD is capable of photodetection above 150GHz. 14, EDFA (Erbium-Doped Fiber Amplifier) represents an erbium-doped optical fiber amplifier, PD (Photo Detector) represents a semiconductor device capable of converting an optical signal into an electrical signal, and OSA represents various optical communication functions (photoelectric It represents an optical module (Optical Sub Aassembly) in which conversion, electro-optical conversion, etc.) are modularized into one component, and DSO represents a digital storage oscilloscope.

The structure of the photoelectric converter (or photoelectric converter) will be described with reference to FIGS. 15 and 16 . 15 illustrates a structure of a photoinc source-based transmitter, and FIG. 16 illustrates a structure of an optical modulator.

In general, a phase of a signal may be changed by passing an optical source of a laser through an optical wave guide. At this time, data is loaded by changing electrical characteristics through a microwave contact or the like. Accordingly, an optical modulator output is formed as a modulated waveform. The photoelectric modulator (O/E converter) is an optical rectification operation by a nonlinear crystal (nonlinear crystal), photoelectric conversion (O / E conversion) by a photoconductive antenna (photoconductive antenna), a bunch of electrons in the light beam (bunch of) THz pulses can be generated by, for example, emission from relativistic electrons. A terahertz pulse (THz pulse) generated in the above manner may have a length in units of femtoseconds to picoseconds. An O/E converter performs down conversion by using non-linearity of a device.

Considering the THz spectrum usage, a number of contiguous GHz bands for fixed or mobile service use for the terahertz system are used. likely to use According to the outdoor scenario standard, available bandwidth may be classified based on oxygen attenuation of 10^2 dB/km in a spectrum up to 1 THz. Accordingly, a framework in which the available bandwidth is composed of several band chunks may be considered. As an example of the framework, if the length of a terahertz pulse (THz pulse) for one carrier is set to 50 ps, the bandwidth (BW) becomes about 20 GHz.

Effective down conversion from the IR band to the THz band depends on how the nonlinearity of the O/E converter is utilized. That is, in order to down-convert to a desired terahertz band (THz band), the O/E converter having the most ideal non-linearity for transfer to the terahertz band (THz band) is design is required. If an O/E converter that does not fit the target frequency band is used, there is a high possibility that an error may occur with respect to the amplitude and phase of the corresponding pulse.

In a single carrier system, a terahertz transmission/reception system may be implemented using one photoelectric converter. Although it depends on the channel environment, as many photoelectric converters as the number of carriers may be required in a far-carrier system. In particular, in the case of a multi-carrier system using several broadbands according to the above-described spectrum usage-related scheme, the phenomenon will become conspicuous. In this regard, a frame structure for the multi-carrier system may be considered. The down-frequency-converted signal based on the photoelectric converter may be transmitted in a specific resource region (eg, a specific frame). The frequency domain of the specific resource region may include a plurality of chunks. Each chunk may be composed of at least one component carrier (CC).

The above salpin 6G communication technology may be applied in combination with the method or apparatus proposed in the present specification to be described later, or may be supplemented to specify or clarify the technical characteristics of the methods proposed in the present specification. On the other hand, the joint inference method proposed in the present specification may be applied in combination with communication services by 3G, 4G, and 5G technologies as well as the 6G communication technology described above.

17 is a flowchart of an associative inference method according to various embodiments of the present specification.

The associative inference method according to various embodiments of the present specification may be implemented by one or more processors included in a terminal or a server as a wireless communication system. In various embodiments, a base station or terminal includes one or more memories storing instructions, one or more transceivers, and one or more processors, wherein the one or more processors support operations for associative inference. The following specification describes operations by one or more processors of the terminal.

Referring to FIG. 17 , the server receives specific values for federated inference from terminals included in the federated inference system (S1710).

The server receives specific values for federated inference from two or more terminals. In an embodiment, the two or more terminals may be a master UE of a terminal group including a plurality of terminals. That is, assuming a communication environment in which a plurality of terminals exist, the server may not receive specific values from all of the plurality of terminals, but may receive specific values only from a representative terminal that is a part of the plurality of terminals. The server or the terminal may create a terminal group through a control signal before receiving specific values, and select any one of terminals included in the created terminal group as a representative terminal. When the representative terminal is selected, since the server does not receive the specific values from other terminals than the representative terminal, participation of other factors impairing communication performance such as interference can be minimized. On the other hand, the generation of the terminal group will be described in detail in S1730 below.

The server or terminal may create one or more terminal groups including at least some of the two or more terminals (S1720).

The terminal group is configured based on a predetermined parameter. In an embodiment, the terminal group may be grouped based on at least one of identification information of a local AI model stored in the terminal, location information of the terminal, and CSI. The local AI model refers to a model for generating the specific value by first receiving raw data. In an embodiment, the local AI model refers to a neural network (NN) model including one or more layers including one or more nodes and a nonlinear activation layer.

In an embodiment, the server may transmit a control signal for generation of a terminal group to terminals located within a cell of the base station. Transmission of the control signal by the server may be performed in response to receiving a message requesting association inference from the requesting terminal in the cell. That is, upon receiving a request for federated inference from the requesting terminal, the server performs grouping for all terminals in the cell, and the grouped terminal groups transmit a specific value to the server to support the federated inference.

In another embodiment, terminals may perform a terminal group using device to device (D2D) communication. That is, the terminal group is created by transmitting and receiving predetermined parameters for generating the aforementioned terminal group according to sidelink communication.

In this case, two or more terminal groups created by the server or terminal may exist. For example, if it is assumed that some of the plurality of terminals have a first local AI model, and some of the remaining terminals have a second local AI model, in this case, two or more terminals have a first local AI model with a first local AI model. group and may be grouped into a second group with a second local AI model.

Meanwhile, in an embodiment, the specific value refers to any one of an input value or an output value of the nonlinear activation layer, and in another embodiment, the specific value refers to an input value or an output value of a layer other than the nonlinear activation layer You may.

In an embodiment, the deactivation layer may be a softmax layer. The softmax layer may calculate a probability value for data applied to the input layer of the local AI model. The probability value means a probability value for each class that has been previously learned for a classification operation. The highest value among these probability values may be subsequently selected through the argmax function, and the classification result will be determined as an entity or class corresponding to the highest probability value.

The server performs a classification operation using the received specific values and the global AI model stored in the memory (S1730).

The global AI model refers to the AI model stored in the server's memory. The global AI model may be a result of federated inference and other federated learning. Unlike the local AI model, the global AI model is not personalized according to users.

In an embodiment, the global AI model may be implemented as a weighted sum model or a deep neural network model.

The weighted sum model refers to a model for assigning a preset weight to specific values received from the server, and adding the weighted specific values. In this case, the preset weight may be set differently according to the type of terminal group that transmits a specific value to the server. For example, a first weight may be assigned to a specific value received from the first group, and a second weight may be assigned to a specific value received from the second group. The result value calculated using the weighted sum model is then converted into probability values for a plurality of classes by the softmax layer. Any one of the probability values is selected based on the argmax function, and a classification result is generated as information corresponding to the selected probability value.

The deep neural network model, unlike the weighted sum model, is an artificial neural network model composed of a plurality of layers. When specific values received from a server are provided to an input layer, it passes through one or more hidden layers and outputs a result value from an output layer. The result value corresponds to a kind of weighted sum, and a subsequent process is the same as the above-described weighted sum model, and thus a description thereof will be omitted.

Conventionally, one device calculates a probability value for each class using an AI model stored in a memory, and a classification operation is performed based on the calculated probability value. The method according to various embodiments of the present specification, unlike this operation, allows one device to transmit an intermediate value generated in the process of calculating a probability value using an AI model to another device or the probability value to another device. . At this time, since different intermediate values or probability values can be received from a plurality of devices before final inference is performed at the last stage, more accurate inference results can be calculated as well as the computing performance of a single device. It has the advantage of being able to overcome limitations.

In the following specification, the local AI model may be used interchangeably with the local model, and the global AI model may be used interchangeably with the global model.

18 to 20 are exemplary views of the associative reasoning method of FIG. 17 . 18 exemplifies a method of inferring by collecting the intermediate or final values of the AI model in the server, and FIGS. 19 and 20 illustrate a method of inferring by collecting the intermediate or final values in the server terminal (Server UE) do.

Referring to FIG. 18 , a plurality of terminals 1821a, 1821b, 1823, 1831a, 1831b, 1831c, 1833 connected to a server 1850 and a wireless or wired communication network cooperate for the same purpose artificial intelligence-based inference (Inference) can do. Each of the terminals 1821a, 1821b, 1823, 1831a, 1831b, 1831c, and 1833 has AI models 1829 and 1839 for artificial intelligence-based inference in one or more memories. Each terminal stores different AI models 1829 and 1839 in memory based on the performance (eg, computing power) of the processor provided in each terminal. For example, if the computing capability of the first terminal is better than that of the second terminal, the first AI model stored in the first terminal may be more complex than the second AI model of the second terminal. The first and second AI models have structures that are distinct from each other, and as a result, even when the same input data 1810 is applied, different outputs may be output.

Here, since inference results of one or more terminals having the same AI model have similar characteristics, one or more terminals having the same AI model may form one group.

That is, the plurality of terminals 1821a, 1821b, 1823, 1831a, 1831b, 1831c, and 1833 illustrated in FIG. 18 have a first group 1820 and a second group 1830 according to the types of AI models 1829 and 1839. ) as an example. Here, as described above, the first and second groups 1820 and 1830 are classified based on the types of the AI models 1829 and 1839 . The structures of the AI models 1829 and 1839 stored in each terminal are stored differently based on the performance of the processor.

Referring back to Figure 18, a plurality of terminals (1821a, 1821b, 1823, 1831a, 1831b, 1831c, 1833) obtain the same target data (Target Data, 1810), the target data stored in each memory AI model ( 1829, 1839) as input. Thereafter, the plurality of terminals 1821a, 1821b, 1823, 1831a, 1831b, 1831c, 1833, by one or more processors, using the AI models 1829 and 1839 data in response to the input of the target data can be transmitted The data is data for final reasoning. The target data includes data for multiple viewpoints on the same target. As such, as a result of using data from multiple viewpoints for the same target, even if the AI model used by each terminal for AI processing is the same, the inference result for the target may be different depending on the viewpoint of each terminal.

In the AI models 1829 and 1839, the layer where the input vector is located is the input layer, the layer where the final output value is located is the output layer, and all hidden layers located between the input layer and the output layer. ) is included. Here, the data for final inference includes an intermediate value input to the output layer or the final output value.

The server 1850 collects a plurality of data received from a plurality of terminals 1821a, 1821b, 1823, 1831a, 1831b, 1831c, 1833 to generate a final output related to the target data 1810, and the final inference result is It is determined based on the final output. In this case, the final output is performed based on a technique using a class-wise sum or a technique using a class-based average and a deep neural network. In addition, the final reasoning result refers to a result corresponding to the highest value by comparing the final output. Specific details on this will be described with reference to FIGS. 21 to 23 .

On the other hand, the embodiments of the present specification are not limited to connection with the server 1850 . Even if there is no server 1850 that collects data for final reasoning, some embodiments of the present specification may generate the same final reasoning result using a terminal functioning as the server 1850 . In this case, the terminal performing the function of the server 1850 is defined as a server terminal (Server UE).

Referring to FIG. 19 , even in the absence of the server 1850 of FIG. 18 , the server terminal 1933 performing the same function performs inference by adding data for final inference. In this case, the server terminal 1933 may be selected based on a preset priority. The priority is determined based on the performance of the processor, the battery performance, and the remaining amount of the plurality of terminals (1921a, 1921b, 1921c, 1931a, 1931b, 1931c, 1933). On the other hand, since there may be cases where there are two or more terminals with the best performance by comparing the performance, the priority may be set to a terminal having excellent communication channel quality according to the distance between the terminals and the base station. .

In addition, referring to FIG. 20 , the system of FIG. 20 further includes one or more representative terminals (Master UEs, 2023 and 2033 ) selected for each group as compared with FIG. 19 . The representative terminals 2023 and 2033 refer to terminals that collect data for final inference from groups 2020 and 2030 consisting of two or more terminals having the same AI model 2029 and 2039. The representative terminals 2023 and 2033 may increase or decrease according to the number of groups. Accordingly, in a system including two or more groups, the representative terminal may include two or more representative terminals corresponding to the number of the groups. In an embodiment, any one of the two or more representative terminals may be selected as the server terminal 2033 .

The representative terminal 2023 or the server terminal 2033 may be determined based on a preset priority. The priority is determined based on at least one of processor information (eg, processor performance information) of the terminal, battery performance and residual amount, and communication channel quality according to a distance between the terminal and the base station. The representative terminal 2023 is determined by comparing priorities between two or more terminals included in a group. The server terminal 1933 is determined by comparing the priorities between two or more representative terminals 2023 and 2033 included in the system.

Meanwhile, the terminal applied to the federated inference method according to various embodiments of the present specification may selectively perform any one function of the representative terminal or the server terminal.

In the wireless communication system, the first terminal may check the connection state with the server for federated inference based on information related to the communication environment. If the server is not identified, the first terminal may request transmission of specific values for federated inference to two or more second terminals. That is, when there is no terminal serving as a server, the first terminal serves to collect information in order to reduce the load on the terminal that will serve as a server.

Thereafter, the first terminal performs a classification operation using the specific values received in response to the request and a pre-trained local AI model, or jointly infers an aggregated value calculated based on the specific values. may be transmitted to a third terminal for At this time, if the classification operation is performed, the first terminal can be regarded as performing a function as a server terminal, and if the operation of transmitting to the third terminal is performed, the first terminal performs a function as a representative terminal rather than a server terminal it can be seen that

In addition, the first and second terminals may constitute one terminal group, and the third terminal may belong to another terminal group other than the terminal group. Here, the terminal group may be grouped based on at least one of (i) identification information of a local AI model stored in the terminal, (ii) location information of the terminal, or (iii) CSI (Channel State Information).

If there are two or more terminal groups, each terminal group may have a different local AI model, and the local AI model includes one or more layers including one or more nodes and a non-linear activation layer. It is a neural network model that includes The specific value is either an input value or an output value of the nonlinear activation layer, in order to more effectively implement the advantage of the above-described joint inference method. In this case, the nonlinear activation layer may be a softmax layer.

On the other hand, any one of the first operation of performing the above-described classification operation or the second operation of transmitting the aggregated value to the third terminal includes processor performance, battery performance, and remaining battery capacity provided in the first to third terminals. It may be selectively determined based on at least a part of That is, the first terminal may determine whether it is the most suitable terminal as a server terminal based on at least a part of processor performance, battery performance, and residual amount of terminals belonging to a terminal group different from that of the terminal group to which the first terminal belongs. If the first terminal has the best performance among one or more terminals to be checked, the first terminal will perform the classification operation, otherwise, other terminals may support the classification operation.

Meanwhile, the server terminal of FIGS. 19 and 20 is not limited to performing the function of the server, and may transmit data for final inference received from other terminals to the server.

The associative inference method according to various embodiments of the present specification uses various techniques for final inference. Various techniques for the final inference will be described with reference to FIGS. 21 to 23 below.

21 to 23 are exemplary diagrams of various classification algorithms of the associative inference method of FIG. 17 .

Here, the examples of FIGS. 21 to 23 may be combined with the above-described various embodiments of FIGS. 18 to 20 . For example, the devices of the groups shown in FIGS. 21 to 23 correspond to the groups of devices shown in FIGS. 18 to 20 . In addition, the server of FIGS. 21 to 23 corresponds to the server of FIG. 18 or the server terminal of FIGS. 19 and 20 .

Referring to FIG. 21 , a metric transmitted to the server 2150 or the representative terminal may be an output value of a softmax layer.

The metric indicates the degree of similarity between two or more matrices or vectors. Specifically, the metric is a distance between two or more matrices or vectors in a vector space through an operation between matrices or vectors, and indicates the degree of similarity between vectors.

In the field of artificial intelligence technology, the metric is generated as an output when input data is provided to the input layer of a neural network for machine learning. That is, the metric may be different according to the weight and bias of the neural network trained to output a target result. In the following specification, it is exemplified that one or more terminals transmit the metric to the server 2150, but is not limited thereto. For example, one or more terminals may transmit a vector value or a matrix value that is an input value or an output value of the softmax layers 2129s and 2139s of the AI models 2129 and 2139 in addition to the above metrics.

Descriptions related to the above-mentioned metrics are common in the following specifications, and overlapping descriptions are omitted.

Referring back to FIG. 21 , two or more terminals form one or more groups based on the types of AI models 2129 and 2139 stored in the memory. For example, terminals of Group A 2120 have a first AI model 2129 , and terminals of Group B 2130 have a second AI model 2139 . The first model 2129 and the second model 2139 have different neural network structures.

The terminals of the group A 2120 and the terminals of the group B 2130 transmit the metric Pi generated using the first AI model 2129 and the second AI model 2139, respectively, to the server 2150. . At this time, the transmitted metric is a value that has passed through the softmax layers 2129s and 2139s of each AI model 2129 and 2139 and has a real value between 0 and 1. As for the transmitted metric, the values of the metric for each class are summed in the server 2150 through a class-wise sum, and preset weights are given to the sum. Here, the preset weights may have different values for each group of the aforementioned terminals. For example, a weight A(α _A ) is given to the class-specific sum of metrics received from Group A 2120 , and a weight B(α _B ) is given to the class-specific sum of metrics received from Group B 2130 . can be given.

Thereafter, the server 2150 may calculate the final result by summing the respective weighted result values. In this case, the server 2150 may calculate a classification result corresponding to any one of a plurality of classes through the argmax function 2252 .

On the other hand, various embodiments of the present specification are not limited to the embodiment of FIG. 21 , and for example, the metric transmitted to the server 2150 may be an input value rather than an output value of the softmax layers 2129s and 2139s. . Hereinafter, it will be described with reference to FIG.

Referring to FIG. 22 , the metric may be an input value of the softmax layers 2229s and 2239s. The embodiment of FIG. 22 has some differences in that an operation of removing the offset of the metric transmitted to the server 2250 and the metric received from the server 2250 is additionally performed, and the rest of the common parts are those of FIG. Since it overlaps with the embodiment, it is omitted.

In this case, the server 2250 additionally performs an operation of removing the offset before the class-by-class sum, unlike the exemplary embodiment of FIG. 21 . This is because the values that do not pass through the softmax layers 2229s and 2239s do not have a value within the range of 0 to 1 and may have various distributions according to the neural network model.

Accordingly, the server 2250 selects the largest metric value for each group as an offset so that the range of the metric received for each group is within the same range, and additionally performs an operation of subtracting it.

Meanwhile, in some embodiments, even when the output values of the softmax layers 2229s and 2239s are transmitted, input values of the softmax layers 2229s and 2239s may be calculated as shown in FIG. 22 . For example, by taking the logarithm of the output values of the softmax layers 2229s and 2239s, an input value related to the output value may be obtained. In this case, the server 2250 does not perform an offset removal operation, and a logarithmic operation is required instead. If all of the weights corresponding to each group are fixed to 1, the sum of the input values of the sptmax layer is expressed as a product of the output values associated with the input values.

Meanwhile, FIGS. 21 and 22 exemplify a case in which the server or the representative terminal generates an inference result from a value calculated by applying a preset weight to a class-wise sum, but FIG. 23 shows the server or A case in which the representative terminal generates an inference result using a deep neural network (DNN, 2355) is exemplified.

Unlike FIGS. 21 and 22 , FIG. 23 stores a deep neural network 2355 . The stored deep neural network 2355 receives parameters received from two or more terminal groups 2320 and 2330 as input, and is used to generate an output for joint inference.

In this case, the metric received by the server 2350 is either an input value or an output value of the softmax layers 2329s and 2339s illustrated in FIGS. 21 and 22 . When the server 2350 receives metrics from two or more terminal groups 2320 and 2330, the server 2350 calculates a class-wise average for each group.

The deep neural network 2355 stored in the server 2350 has three kinds of factors as inputs. The factors include, for example, (1) a metric for which an average value is calculated for each group, (2) reliability for each group, and/or (3) the number of terminals. The metric to which the class-based average is applied is an essential factor, but the reliability for each group and/or the number of terminals is an optional factor. That is, the reliability of each group or the number of terminals may be omitted or added as needed, but when added, the reliability of the federated inference may be further improved.

The group-specific reliability refers to an index or factor indicating the performance of an AI model stored or used for each group, and the number of terminals refers to the number of terminals belonging to the group.

On the other hand, the intermediate value calculated through the deep neural network 2355 is passed through the softmax layer 2356 and the argmax function 2352 located at the end of the deep neural network by the processor provided in the server 2350, the final result of used for inference.

In summary, when the server 2350 receives metrics from two or more terminal groups, the server 2350 classifies the metrics for each group and calculates a class-based average. Thereafter, the server 2350 provides the average value of the calculated metrics, the reliability of each group, and/or the number of terminals as an input to the deep neural network 2355, and the output of the deep neural network 2355 is a softmax layer ( 2356) to a value between 0 and 1. Finally, the server 2350 compares the output values of the softmax layer 2356 , selects the largest value, and generates information on the result of federated inference as a classification result corresponding to the selected value.

On the other hand, in some embodiments of the present specification, the third AI model may be defined as including all of the deep neural network stored in the server 2350 , the softmax layer 2356 and the argmax function 2352 .

As such, when the third AI model is used, a new technical problem can be solved unlike FIGS. 21 and 22 . The new description point will be described below in terms of group reliability, the number of terminals belonging to the group, and the number of terminal groups.

First, the performance of estimation may be different depending on the hardware specification and learning algorithm of each terminal, and if the same weight is given to the results of all terminal groups, there may be a problem in that the overall performance is lowered by the results of the group with low performance. . For these problems, the server can evaluate the performance of the AI model for each group and use the index (eg, accuracy, etc.) as reliability. Accordingly, group-specific reliability may be included as one of the input factors of the deep neural network.

Next, when the number of terminals is applied as an input to the deep neural network, if there is a difference in the number of terminals between groups, a group in which many terminals participate may have a greater influence on the final result. In this case, when the input values of the deep neural network are too large or too small, or when there is a large difference in range from other input values, a learning problem may occur. In contrast, the server may set a reference number of terminals for normalization and limit the maximum number of terminals. For example, if the reference number of terminals is 100 and the maximum number of terminals is limited to 200, the number of 1 to 200 terminals may be expressed as a real number between 0.01 and 2.0, respectively.

Finally, in the AI model of the server that can support the number of various groups, the number of inputs to the deep neural network may vary if the number of participating terminal groups is changed in some cases. For example, in a specific situation, two groups of 10 terminals may participate, and in other circumstances, 4 groups of 5 terminals may participate. As such, various combinations may occur according to the number of terminals for each group and the number of groups, and the server may have different deep neural network models for each of the various combinations. That is, the server must have a number of deep heart networks corresponding to the number of combinations. In contrast, in various embodiments of the present specification, the server may repeatedly apply one AI model having two inputs. For example, if it is assumed that there are Group A, Group B, and Group C, the result is inferred by first collecting the Group A and Group B, and the final result is obtained by using the inferred result value from the value received from Group C. It is to perform a way of inferring a value. In this case, the order of inference for each group may be based on a preset priority, but is not limited thereto.

How AI models are trained

24 is a flowchart of a learning method of an AI model according to an embodiment of the present specification.

Referring to FIG. 24 , the server may check the learning state of the local AI model ( S2410 ). At this time, the server determines whether new learning is required based on the learning state of the local AI model. The learning state refers to data representing the performance of the local AI model. The performance is determined based on the magnitude of the training error and the difference between the training error and the generalization error. For example, if the learning error is not sufficiently small or the difference between the learning error and the generalization error is not sufficiently small, the performance is evaluated as insufficient.

A case in which learning of the local AI model is required will be described (see FIG. 25 ).

If the server needs to learn the local AI model, it learns the global AI model and the local AI model using the training dataset (S2410: Yes, S2420).

The learning method at this time is the same as a normal supervised learning method. That is, parameters related to both the local AI model and the global AI model are updated by a general supervised learning method. This update may be performed based on a difference between a value calculated according to a loss function or a cost function and a labeled correct answer value.

When the learning is completed, the server transmits the learned local AI model to each individual terminal that has transmitted the local AI model before learning (S2430).

On the other hand, there may be a case in which learning of the local AI model is not required and only the global AI model needs to be learned. At this time, a method of learning only the global AI model will be described (see FIG. 26 ).

If learning of the local AI model is not required, each local AI model may be received from a plurality of terminals associated with the federated inference system (S2410: No, S2440).

Here, the reason for receiving the local AI model is to provide the values calculated using the local AI models as input data of the global AI model for joint inference with the global AI model, and in this step, Learning is not performed.

Thereafter, the server may generate a training dataset including a value (eg, a metric) generated based on the received local AI models ( S2450 ). On the other hand, the training dataset refers to a dataset for training only the global AI model, and is distinguished from other training datasets in this specification. The training dataset in S2450 consists of metrics generated using local AI models received from multiple terminals, and other training datasets in the specification refer to raw data or data processed from the raw data. The other training data sets will be described in the following learning data acquisition method.

The server may train the global AI model using the generated training dataset (S2460).

Here, in the training of the global AI model, unlike S2410, back-propagation is processed only in the global AI model, and parameters of the pre-trained local AI model are not affected (see FIG. 27 ). For example, referring to FIG. 27 , one or more terminals each obtain at least a portion of a previously provided data set 2710 , respectively, and transmit the at least a portion of the data 2715-1, ,,, , 2715-5 respectively. It is provided to the input layer of the AI model (2729, 2739). 27 illustrates a case in which each AI model 2729, 2739 is provided in individual terminals 2720, 2730, but each AI model 2729, 2739 is a server ( 2750) and the same operation may be performed only within the server 2750.

Thereafter, the server 2750 performs an inference operation using the collected metrics 2752 and 2753 and the AI models 2759 and 2759s provided in the server, and the correct answer value labeled on the initially provided data 2710 (label ) to calculate the gradient for backpropagation. The remaining description of supervised learning using the calculated gradient is omitted because it overlaps with the above description of S2410.

How to acquire training data

In order to improve the performance of AI models, it is important to obtain high-quality training data. The performance of the system in the actual field can be improved by further learning the AI model using data from the field where the system is applied.

Meanwhile, since it is inefficient for each terminal to transmit all raw data (eg, photos, sensor data, etc.) to the server, the server can efficiently obtain learning data based on the following operations.

The server sets parameters necessary for securing learning data related to each terminal. For example, a parameter necessary for securing the data is a threshold probability (P _threshold ), a minimum difference (P _{min-difference} ), a data retention time (T _retain ), or a flag (flag_newestdat_prior) for checking whether the latest data is prioritized at least one of In this way, the parameters for securing the set learning data may be transmitted to the terminals.

The P _threshold represents a threshold of the highest probability value generated through the softmax layer. The P _{min-difference} represents the threshold of the minimum difference between the highest probability value and the next highest value. The T _retain represents the time the data is maintained in the storage space. The flag_newestdat_prior, if the value is 1, indicates that the latest data has priority, and if the value is 0, the latest data does not have priority.

If the highest probability value of the final inference result is less than P _threshold or the difference between the highest probability value and the next highest probability value is less than P _{min-difference} , the terminal stores the training data in a storage space (eg memory) and maintains it during T _retain do. If an event to store new data occurs in a state where there is no available storage space in the storage space, the terminal checks the flag.

If the value of the flag is 1 and the storage space is not all filled with the preserved data, the terminal deletes the longest stored data among the data not designated as the preserved data and then stores new data in the free space. Preserved data refers to data that is not deleted even after T _retain , and is not deleted even if new data is collected to secure space to store it. When the value of the flag is 0 or the storage space is all filled with preserved data, the terminal discards new data without storing it. That is, the terminal can manage the storage space by determining whether to store or discard the new data based on the flag for determining whether the data is prioritized.

Meanwhile, when one or more conditions are satisfied, the server requests that each terminal transmit raw data among data already used for learning, or requests that each terminal preserves the raw data. The condition includes a first condition in which the number of terminals that have transmitted an inference result that does not match the final reasoning result of the server is equal to or greater than a set number and a second condition in which a deviation of the inference result between the terminals is equal to or greater than a set value. The server may perform any one of the requested operations when one or more of the first and second conditions are satisfied.

The request to transmit the raw data is effective when controlled by the server, and the request to preserve the raw data is effective when controlled by the server terminal serving as the server, but only in that case does not limit

Thereafter, when receiving a request for data transmission from the server, each terminal, if there is the data in the storage space, transmits the data to the server and then deletes the data from the storage space. On the other hand, if there is no data in the storage space, it notifies the server that there is no data. In this case, when the server receives data from each terminal, the server stores the received data.

If data preservation is requested from the server, each terminal designates the data as preservation data if there is the data in the storage space. At this time, the server receives the stored data stored in each terminal through a predetermined procedure. For example, the stored data may be received by the terminal at a time when the terminal can access the server, or may be manually moved to the server by the user.

As such, the secured data may be used as training data through a labeling operation.

Devices used in wireless communication systems

Although not limited thereto, the various proposals of the present specification may be applied to various fields requiring wireless communication/connection (eg, 6G) between devices.

Hereinafter, it will be exemplified in more detail with reference to the drawings. In the following drawings/descriptions, the same reference numerals may represent the same or corresponding hardware blocks, software blocks, or functional blocks, unless otherwise indicated.

28 illustrates a communication system applied to this specification.

Referring to FIG. 28 , the communication system 1 applied to the present specification includes a wireless device, a base station, and a network. Here, the wireless device means a device that performs communication using a wireless access technology (eg, 6G, 5G NR (New RAT), LTE (Long Term Evolution)), and may be referred to as a communication/wireless/5G device. . Although not limited thereto, the wireless device may include a robot 100a, a vehicle 100b-1, 100b-2, an eXtended Reality (XR) device 100c, a hand-held device 100d, and a home appliance 100e. ), an Internet of Thing (IoT) device 100f, and an AI device/server 400 . For example, the vehicle may include a vehicle equipped with a wireless communication function, an autonomous driving vehicle, a vehicle capable of performing inter-vehicle communication, and the like. Here, the vehicle may include an Unmanned Aerial Vehicle (UAV) (eg, a drone). XR devices include AR (Augmented Reality)/VR (Virtual Reality)/MR (Mixed Reality) devices, and include a Head-Mounted Device (HMD), a Head-Up Display (HUD) provided in a vehicle, a television, a smartphone, It may be implemented in the form of a computer, a wearable device, a home appliance, a digital signage, a vehicle, a robot, and the like. The portable device may include a smart phone, a smart pad, a wearable device (eg, a smart watch, smart glasses), a computer (eg, a laptop computer), and the like. Home appliances may include a TV, a refrigerator, a washing machine, and the like. The IoT device may include a sensor, a smart meter, and the like. For example, the base station and the network may be implemented as a wireless device, and the specific wireless device 200a may operate as a base station/network node to other wireless devices.

The wireless devices 100a to 100f may be connected to the network 300 through the base station 200 . AI (Artificial Intelligence) technology may be applied to the wireless devices 100a to 100f , and the wireless devices 100a to 100f may be connected to the AI server 400 through the network 300 . The network 300 may be configured using a 3G network, a 4G (eg, LTE) network, or a 5G (eg, NR) network. The wireless devices 100a to 100f may communicate with each other through the base station 200/network 300, but may also communicate directly (e.g. sidelink communication) without passing through the base station/network. For example, the vehicles 100b-1 and 100b-2 may perform direct communication (e.g. Vehicle to Vehicle (V2V)/Vehicle to everything (V2X) communication). Also, the IoT device (eg, sensor) may communicate directly with other IoT devices (eg, sensor) or other wireless devices 100a to 100f.

Wireless communication/ connection 150a and 150b may be performed between the wireless devices 100a to 100f/base station 200 - the base station 200/wireless devices 100a to 100f. Here, the wireless communication/connection may be performed through various wireless access technologies (eg, 6G, NR) for uplink/downlink communication 150a and sidelink communication 150b (or D2D communication). Through the wireless communication/ connection 150a and 150b, the wireless device and the base station/wireless device may transmit/receive wireless signals to each other. For example, the wireless communication/ connection 150a and 150b may transmit/receive signals through various physical channels based on all/part of the process of FIG. A1 . To this end, based on the various proposals of the present specification, various configuration information setting processes for wireless signal transmission/reception, various signal processing processes (eg, channel encoding/decoding, modulation/demodulation, resource mapping/demapping, etc.) , at least a part of a resource allocation process may be performed.

29 illustrates a wireless device applicable to this specification.

Referring to FIG. 29 , the first wireless device 100 and the second wireless device 200 may transmit/receive wireless signals through various wireless access technologies (eg, 6G, LTE, NR). Here, {first wireless device 100, second wireless device 200} is {wireless device 100x, base station 200} of FIG. 26 and/or {wireless device 100x, wireless device 100x) } can be matched.

The first wireless device 100 includes one or more processors 102 and one or more memories 104 , and may further include one or more transceivers 106 and/or one or more antennas 108 . The processor 102 controls the memory 104 and/or the transceiver 106 and may be configured to implement the functions, procedures and/or methods described/suggested above. For example, the processor 102 may process information in the memory 104 to generate first information/signal, and then transmit a wireless signal including the first information/signal through the transceiver 106 . In addition, the processor 102 may receive the radio signal including the second information/signal through the transceiver 106 , and then store information obtained from signal processing of the second information/signal in the memory 104 . The memory 104 may be connected to the processor 102 and may store various information related to the operation of the processor 102 . For example, the memory 104 may store software code including instructions for performing some or all of the processes controlled by the processor 102 , or for performing the procedures and/or methods described/suggested above. . Here, the processor 102 and the memory 104 may be part of a communication modem/circuit/chip designed to implement a wireless communication technology (eg, 6G, LTE, NR). The transceiver 106 may be coupled to the processor 102 , and may transmit and/or receive wireless signals via one or more antennas 108 . The transceiver 106 may include a transmitter and/or a receiver. The transceiver 106 may be used interchangeably with a radio frequency (RF) unit. In this specification, a wireless device may refer to a communication modem/circuit/chip.

The second wireless device 200 includes one or more processors 202 , one or more memories 204 , and may further include one or more transceivers 206 and/or one or more antennas 208 . The processor 202 controls the memory 204 and/or the transceiver 206 and may be configured to implement the functions, procedures, and/or methods described/suggested above. For example, the processor 202 may process the information in the memory 204 to generate third information/signal, and then transmit a wireless signal including the third information/signal through the transceiver 206 . In addition, the processor 202 may receive the radio signal including the fourth information/signal through the transceiver 206 , and then store information obtained from signal processing of the fourth information/signal in the memory 204 . The memory 204 may be connected to the processor 202 and may store various information related to the operation of the processor 202 . For example, the memory 204 may store software code including instructions for performing some or all of the processes controlled by the processor 202 , or for performing the procedures and/or methods described/suggested above. . Here, the processor 202 and the memory 204 may be part of a communication modem/circuit/chip designed to implement a wireless communication technology (eg, LTE, NR). The transceiver 206 may be coupled to the processor 202 and may transmit and/or receive wireless signals via one or more antennas 208 . The transceiver 206 may include a transmitter and/or a receiver. The transceiver 206 may be used interchangeably with an RF unit. In this specification, a wireless device may refer to a communication modem/circuit/chip.

Hereinafter, hardware elements of the wireless devices 100 and 200 will be described in more detail. Although not limited thereto, one or more protocol layers may be implemented by one or more processors 102 , 202 . For example, one or more processors 102 , 202 may implement one or more layers (eg, functional layers such as PHY, MAC, RLC, PDCP, RRC, SDAP). The one or more processors 102 and 202 may generate one or more Protocol Data Units (PDUs) and/or one or more Service Data Units (SDUs) according to the functions, procedures, proposals and/or methods disclosed herein. One or more processors 102 , 202 may generate messages, control information, data, or information according to the functions, procedures, proposals and/or methods disclosed herein. The one or more processors 102 and 202 generate a signal (eg, a baseband signal) including PDUs, SDUs, messages, control information, data or information according to the functions, procedures, proposals and/or methods disclosed herein. , to one or more transceivers 106 and 206 . One or more processors 102 , 202 may receive signals (eg, baseband signals) from one or more transceivers 106 , 206 , PDUs, SDUs, and/or SDUs according to the functions, procedures, proposals and/or methods disclosed herein. , a message, control information, data or information can be obtained.

One or more processors 102, 202 may be referred to as a controller, microcontroller, microprocessor, or microcomputer. One or more processors 102 , 202 may be implemented by hardware, firmware, software, or a combination thereof. For example, one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), one or more Digital Signal Processing Devices (DSPDs), one or more Programmable Logic Devices (PLDs), or one or more Field Programmable Gate Arrays (FPGAs) may be included in one or more processors 102 , 202 . The functions, procedures, proposals and/or methods disclosed in this document may be implemented using firmware or software, and the firmware or software may be implemented to include modules, procedures, functions, and the like. Firmware or software configured to perform the functions, procedures, proposals, and/or methods disclosed herein is included in one or more processors 102, 202, or stored in one or more memories 104, 204, to one or more processors 102, 202) can be driven. The functions, procedures, proposals and/or methods disclosed in this document may be implemented using firmware or software in the form of code, instructions, and/or a set of instructions.

One or more memories 104 , 204 may be coupled with one or more processors 102 , 202 , and may store various forms of data, signals, messages, information, programs, code, instructions, and/or instructions. The one or more memories 104 and 204 may be comprised of ROM, RAM, EPROM, flash memory, hard drives, registers, cache memory, computer readable storage media, and/or combinations thereof. One or more memories 104 , 204 may be located inside and/or external to one or more processors 102 , 202 . Additionally, one or more memories 104 , 204 may be coupled to one or more processors 102 , 202 through various technologies, such as wired or wireless connections.

One or more transceivers 106 , 206 may transmit user data, control information, radio signals/channels, etc. referred to in the methods and/or operational flowcharts of this document to one or more other devices. The one or more transceivers 106 and 206 may receive user data, control information, radio signals/channels, etc. referred to in the functions, procedures, proposals, methods, and/or flowcharts of operations disclosed herein, and the like, from one or more other devices. For example, one or more transceivers 106 , 206 may be coupled to one or more processors 102 , 202 and may transmit and receive wireless signals. For example, one or more processors 102 , 202 may control one or more transceivers 106 , 206 to transmit user data, control information, or wireless signals to one or more other devices. In addition, one or more processors 102 , 202 may control one or more transceivers 106 , 206 to receive user data, control information, or wireless signals from one or more other devices. Further, one or more transceivers 106, 206 may be coupled to one or more antennas 108, 208, and the one or more transceivers 106, 206 may be coupled to one or more of the transceivers 106, 206 via the one or more antennas 108, 208 for the functions, procedures, and procedures disclosed herein. , may be set to transmit and receive user data, control information, radio signals/channels, etc. mentioned in a proposal, a method and/or an operation flowchart. In this document, one or more antennas may be a plurality of physical antennas or a plurality of logical antennas (eg, antenna ports). The one or more transceivers 106, 206 convert the received radio signal/channel, etc. from the RF band signal to process the received user data, control information, radio signal/channel, etc. using the one or more processors 102, 202. It can be converted into a baseband signal. One or more transceivers 106 and 206 may convert user data, control information, radio signals/channels, etc. processed using one or more processors 102 and 202 from baseband signals to RF band signals. To this end, one or more transceivers 106 , 206 may include (analog) oscillators and/or filters.

The above-described specification can be implemented as computer-readable code on a medium in which a program is recorded. The computer-readable medium includes all types of recording devices in which data readable by a computer system is stored. Examples of computer-readable media include Hard Disk Drive (HDD), Solid State Disk (SSD), Silicon Disk Drive (SDD), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc. There is also a carrier wave (eg, transmission over the Internet) that is implemented in the form of. Accordingly, the above detailed description should not be construed as restrictive in all respects but as exemplary. The scope of this specification should be determined by a reasonable interpretation of the appended claims, and all modifications within the scope of equivalents of this specification are included in the scope of this specification.

Claims (15)

1. A method of performing Federated Inferencing (FI) of a server in a wireless communication system, comprising:

   Receiving specific values for federated inference from two or more terminals;

   performing a classification operation using the received specific values and a pre-trained global AI model;

   A method comprising
2. According to claim 1,

   creating one or more terminal groups including at least some of the two or more terminals;

   A method further comprising:
3. 3. The method of claim 2,

   The terminal group is

   (i) the identification information of the local AI model stored in the terminal, (ii) the location information of the terminal, or (iii) CSI (Channel State Information) will be grouped based on at least one, the method.
4. 3. The method of claim 2,

   When there are two or more terminal groups, each terminal group has a different local AI model.
5. 5. The method of claim 4,

   The local AI model is

   It is a neural network model including one or more layers having one or more nodes and a non-linenar activatoin layer,

   The specific value is either an input value or an output value of the non-linear activation layer.
6. 6. The method of claim 5,

   wherein the non-linear activation layer is a first softmax layer.
7. 3. The method of claim 2,

   selecting any one of the terminals included in the terminal group as a master UE;

   further comprising,

   The method, wherein the two or more terminals transmitting the specific value for the joint inference are configured only with the representative terminal.
8. In a wireless communication system, a method for performing joint inference of a first terminal, comprising:

   checking a connection state with a server for federated inference based on information related to a communication environment;

   requesting transmission of specific values for federated inference to two or more second terminals based on the server not being confirmed; and

   A classification operation is performed using the specific values received in response to the request and a pre-trained local AI model, or an aggregated value calculated based on the specific values is transmitted to a third terminal for joint inference to do;

   A method comprising
9. 9. The method of claim 8,

   The first and second terminals constitute one terminal group,

   The third terminal belongs to a terminal group other than the terminal group, the method.
10. 10. The method of claim 9,

    The terminal group is

    (i) identification information of the local AI model stored in the terminal, (ii) the location information of the terminal, or (iii) CSI (Channel State Information) will be grouped based on at least one of, the method.
11. 10. The method of claim 9,

    When there are two or more terminal groups, each terminal group has a different local AI model.
12. 12. The method of claim 11,

    The local AI model is

    It is a neural network model including one or more layers having one or more nodes and a non-linenar activatoin layer,

    The specific value is either an input value or an output value of the non-linear activation layer.
13. 13. The method of claim 12,

    wherein the non-linear activation layer is a first softmax layer.
14. 9. The method of claim 8,

    Any one of the first operation of performing the classification operation or the second operation of transmitting the aggregated value to the third terminal is at least a portion of processor performance, battery performance, and remaining battery capacity provided in the first to third terminals Selectively determined based on the method.
15. one or more transceivers;

    one or more processors; and

    one or more memories coupled to the one or more processors for storing instructions;

    The instructions, when executed by the one or more processors, cause one or more processors to support operations for federated inference, the operations comprising:

    Receiving specific values for federated inference from two or more external terminals;

    generating an intermediate value using the received specific values and a local AI model stored in a memory; and

    transmitting the generated intermediate value to a server for federated inference;

    Including, the terminal.

Images