Communication device and method in communication device

Abstract

Accurately construct content from latent vectors that include communication noise. [Solution] The first communication device implements machine learning. The first communication device receives data through the network, derives a latent vector including communication noise based on the data, and learns the latent vector.

Description

Technical Field

[0001] The present disclosure relates to a communication device and a method in the communication device, and particularly to a technique for communication of a communication device implementing machine learning.

Background Art

[0002] Techniques related to machine learning (ML) represented by artificial intelligence (AI) are widely used. In particular, a technique called generative AI is widely used. In generative AI, new content such as images and texts is generated based on a prompt.

[0003] In addition, a mechanism called the Internet of Things (IoT) is widely used. In IoT, all kinds of things such as home appliances, automobiles, and factory equipment are connected to the Internet and exchange information with each other. IoT is implemented by one or more IoT devices. It is expected that a mechanism for autonomous communication of IoT devices implementing AI will also be implemented.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

[0005] The first communication device according to the present disclosure includes a communication unit and a control unit implementing machine learning. The communication unit is configured to receive data through a network, and the control unit is configured to derive a latent vector including communication noise based on the data and learn the latent vector.

[0006] ]> A method performed by the first communication device relating to this disclosure includes receiving data over a network, deriving a latent vector including communication noise based on the data, and learning the latent vector, wherein the first communication device implements machine learning. [Brief explanation of the drawing]

[0007] [Figure 1] A communication system according to the first embodiment is shown. [Figure 2] A mobile communication network according to the first embodiment is shown. [Figure 3] A wireless LAN network according to the first embodiment is shown. [Figure 4] The physical configuration of the communication device according to the first embodiment is shown. [Figure 5] The logical configuration of the communication device according to the first embodiment is shown. [Figure 6] The forward diffusion process according to the latent diffusion model in the first embodiment is shown. [Figure 7] The first embodiment shows a reverse diffusion process using a latent diffusion model. [Figure 8] A learning model according to the first embodiment is shown. [Figure 9] The process for transmitting content according to the first embodiment is shown. [Figure 10] An AI communication model according to the second embodiment is shown. [Figure 11] An AI communication model according to the second embodiment is shown. [Figure 12] A semantic packet according to the second embodiment is shown. [Figure 13] The following describes the processing performed by a semantic entity according to the second embodiment. [Figure 14] The following describes the processing performed by a semantic entity according to the second embodiment. [Figure 15] The logical configuration of the communication device according to the third embodiment is shown. [Figure 16]Shows the semantic packet according to the third embodiment. [Figure 17] Shows the process of generating a semantic packet according to the third embodiment. [Figure 18] Shows the process of determining whether to construct content according to the third embodiment. [Figure 19] Shows the logical configuration of the communication device according to the fourth embodiment. [Figure 20] Shows the forward diffusion process by the potential diffusion model according to the fourth embodiment. [Figure 21] Shows the process of learning the latent vector according to the fourth embodiment. [Figure 22] Shows the logical configuration of the communication device according to the fourth embodiment. [Figure 23] Shows the process of learning the latent vector according to the fourth embodiment.

Modes for Carrying Out the Invention

[0008] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the present specification and drawings, for elements that can be similarly described, duplicate descriptions may be omitted by assigning the same reference numerals.

[0009] Each of the embodiments described below is merely an example of a realizable configuration of the present disclosure. Each of the following embodiments can be appropriately modified or changed according to the configuration of the device to which the present disclosure is applied and various conditions. Not all combinations of elements included in each of the following embodiments are essential for realizing the configuration of the present disclosure, and some of the elements can be appropriately omitted. Therefore, the scope of the present disclosure is not limited by the configurations described in each of the following embodiments. As long as there is no contradiction, a configuration combining a plurality of configurations described in the following embodiments can also be adopted.

[0010] Hereinafter, one or more embodiments will be described with reference to the accompanying drawings. In the present specification and the drawings, for elements that can be similarly described, duplicate description is omitted by assigning the same reference numerals.

[0011] The description will be made in the following order. <Prior Art> <This Embodiment> 1. First Embodiment 1-1 Configuration of Communication System 1-2 Determination of Communication Mode 1-2-1 Determination of Communication Mode Based on Importance 1-2-2 Determination of Communication Mode Based on Communication Quality 1-2-3 Determination of Communication Mode Based on Input Parameters 1-2-4 Determination of Communication Mode Based on Feedback from Receiving-Side Communication Device 1-2-5 Determination of Communication Mode Based on Request from Receiving-Side Communication Device 1-2-6 Determination of Communication Mode Based on Evaluation of Index 1-3 Content Transmission Processing 2. Second Embodiment 2-1 Communication in Hierarchical Structure 2-2 Hierarchical Structure Based on OSI Reference Model 2-3 Hierarchical Structure Based on TCP / IP Model 2-4 Processing in Semantic Layer of Transmitting-Side Communication Device 2-5 Processing in Semantic Layer of Receiving-Side Communication Device 3. Third Embodiment 3-1 Processing in Decodeless Mode 3-2 Processing in Transmitting-Side Communication Device According to Decodeless Mode 3-3 Processing in Receiving-Side Communication Device According to Decodeless Mode 4. Fourth Embodiment 4-1 Learning of Latent Vector with Communication Noise Added 4-2 Addition of Communication Noise in Learning Process 4-3 Detection of Communication Noise in Dialogue 5. Other Embodiments 6. Addendum

[0012] <Conventional Technology> Generative AI is widely used. Generative AI generates new content, including text, images, audio, and / or sensor sequences, based on prompts. Assuming that IoT devices implement generative AI, it is assumed that these IoT devices can communicate autonomously with each other. IoT devices implementing generative AI are expected to exchange content generated by the generative AI over a network. For example, in a medical setting, IoT devices implementing generative AI are expected to generate medical images from a patient's condition and send the generated medical images to a server. In the above scenario, it is anticipated that the amount of data transmitted over the network will increase significantly.

[0013] Patent Document 1 discloses a technology for devices implementing artificial neural networks (ANNs) to communicate with each other via a network. In the technology described in Patent Document 1, a common spreading model is implemented in both the transmitting and receiving devices. The transmitting device sends only lightweight conditional data and fine-tuning LoRA (low-rank adaptation) weights to the receiving device. The receiving device constructs video data based on the conditional data and LoRA weights.

[0014] In the technology described in Patent Document 1, the ANN outputs video data based on conditional data. Instead of transmitting video data, the transmitting device transmits conditional data and LoRA weights to the receiving device. The receiving device constructs video data based on the conditional data and LoRA weights. The conditional data and LoRA weights are smaller in size than the video data.

[0015] In the technology described in Patent Document 1, the receiving device implements a common ANN with the transmitting device. The receiving device constructs video data based on conditional data and LoRA weights using the ANN. In this way, the receiving device can construct its video data without the video data generated by the transmitting device being transmitted.

[0016] The technology described in Patent Document 1 can reduce the amount of data transmitted over a network by allowing the receiving device to construct video data with a smaller amount of data. On the other hand, the video data constructed by the receiving device does not necessarily match the video data generated by the transmitting device; in other words, the fidelity of reproducing the video data may be inferior. The technology described in Patent Document 1 cannot address cases where the receiving device needs to reproduce the video data with higher fidelity.

[0017] <This embodiment> 1. First Embodiment A first embodiment will be described. This embodiment may be implemented by a communication system. The communication system may include at least two or more communication devices. In the communication system, two or more communication devices may communicate with each other. In this embodiment, an example in which two communication devices communicate with each other will be described.

[0018] In this embodiment, one of the two communication devices may use a generative AI to output a latent vector and generate content based on the latent vector. Alternatively, one communication device may transmit the generated content or latent vector to the other communication device. In this way, the two communication devices may communicate (interact) with each other through the generative AI.

[0019] The latent vector may be information obtained by mapping the features of the input data to a vector space of a predetermined number of dimensions. For example, the latent vector may be data extracted from the input data via an encoder, or it may be a set of numbers expressed in a number of dimensions lower than the number of dimensions of the input data, while retaining the features of the input data. Furthermore, the latent vector may be expressed in a form that makes it impossible to recover the original input information; for example, the latent vector may be encrypted, anonymized, or disrupted.

[0020] In the following description, the communication device that transmits the generated content or latent vector may be referred to as the transmitting communication device. The communication device that receives the content or latent vector from the transmitting communication device may be referred to as the receiving communication device. The two communication devices may implement AI models with the same or different learning models. In other words, a symmetric AI model or an asymmetric AI model may be implemented between the two communication devices.

[0021] 1-1 Communication System Configuration Referring to Figure 1, the configuration of the communication system S according to this disclosure will be described. As shown in Figure 1, the communication system S according to this embodiment may include a first communication device 10 and a second communication device 20. In this embodiment, an example in which two communication devices communicate with each other will be described, but the number of communication devices is merely illustrative.

[0022] Both the first communication device 10 and the second communication device 20 may be implemented as IoT devices. For example, both the first communication device 10 and the second communication device 20 may be implemented in a robot operating in a factory. By communicating with each other, either the first communication device 10 or the second communication device 20 may control the other. Both the first communication device 10 and the second communication device 20 may be referred to as "IoT devices," "edge devices," or "AI devices." The terms "device," "equipment," "device," and "node" may be used interchangeably.

[0023] The first communication device 10 and the second communication device 20 may be connected via a network NW. The network NW may be implemented, for example, in accordance with the technical standards defined by the Third Generation Partnership Project (3GPP). The technical standards defined by 3GPP may include Long Term Evolution (LTE), 5G New Radio (NR), and Beyond 5G (6G). Hereinafter, a network conforming to such wireless specifications will be referred to as a "mobile communication network."

[0024] As shown in Figure 2, when the network NW is implemented by a mobile communication network, the first communication device 10 and the second communication device 20 may operate as user equipment (UE) as defined by 3GPP technical standards. The UE may be connected to a base station (BS) and a core network (CN). The base station BS and the core network CN manage the movement of the first communication device 10 and the second communication device 20, enabling them to communicate wirelessly with each other. The first communication device 10 and the second communication device 20 may also communicate wirelessly with each other via sidelink communication. Since the mobile communication network is implemented according to known technical specifications, a detailed explanation thereof is omitted in this specification.

[0025] In addition to, or instead of, mobile communication networks, networks may be implemented in accordance with technical standards specified by, for example, the Institute of Electrical and Electronics Engineers (IEEE). These technical standards may include IEEE 802.11. Hereinafter, networks conforming to such wireless communication technologies will be referred to as "wireless LAN networks."

[0026] As shown in Figure 3, when the network NW is implemented as a wireless LAN network, the first communication device 10 and the second communication device 20 may operate as stations (STAs) as defined by the IEEE 802.11 technical standard. The STAs may be connected to access points (APs). The first communication device 10 and the second communication device 20 can communicate wirelessly with each other by the access point AP detecting the wireless signal from the STA and performing authentication processing on both sides. Either or both of the first communication device 10 and the second communication device 20 may operate as APs. Note that since wireless LAN networks are implemented according to known technical specifications, a detailed explanation thereof is omitted in this specification.

[0027] In the configuration described in Figures 2 and 3, the first communication device 10 and the second communication device 20 communicate with each other wirelessly. However, this embodiment is not limited to this configuration. The first communication device 10 and the second communication device 20 may also communicate with each other using a wired connection.

[0028] Next, the physical configuration of the first communication device 10 will be described with reference to Figure 4. The physical configuration of the first communication device 10 may also be applied to the physical configuration of the second communication device 20. Therefore, the description of the physical configuration of the second communication device 20 will be omitted. As shown in Figure 4, the first communication device 10 may include a processor 101, memory 102, auxiliary storage device 103, transceiver 104, and antenna 105 as physical components. The above elements provided in the first communication device 10 may be interconnected by an internal bus. Note that the first communication device 10 may include physical components other than those shown in Figure 4.

[0029] The processor 101 may be a computing element that realizes various functions of the first communication device 10. The processor 101 may be implemented by a neural processing unit (NPU), a central processing unit (CPU), and a graphics processing unit (GPU).

[0030] Memory 102 may be implemented using a storage medium such as Random Access Memory (RAM). Memory 102 may also be an element for temporarily storing programs and data. Programs and data may be used to perform various processes in the first communication device 10. The program may include one or more instructions for the operation of the first communication device 10. The processor 101 may realize the functions of the first communication device 10 by unpacking the program stored in memory 102 into memory 102 and executing the unpacked program.

[0031] The auxiliary storage device 103 may be implemented using storage media such as ROM (Read Only Memory), HDD (Hard Disk Drive), and SSD (Solid State Drive). The auxiliary storage device 103 may also be an element that permanently stores programs and data. The programs and data may be used to perform various processes in the first communication device 10. For example, the auxiliary storage device 103 may store an AI model, which will be described later.

[0032] The transceiver 104 may be a circuit that performs various signal processing to realize wireless communication. The transceiver 104 may also include a baseband processor and an RF circuit. The transceiver 104 may transmit and receive data and control signals to and from the second communication device 20 via the antenna 105. If the network NW is implemented by a mobile communication network, the transceiver 104 may communicate wirelessly with the second communication device 20 via a base station STA. If the network NW is implemented by a wireless LAN network, the transceiver 104 may communicate wirelessly with the second communication device 20 via an access point AP.

[0033] Next, the logical configuration of the first communication device 10 will be described with reference to Figure 5. The logical configuration of the first communication device 10 may also be applied to the logical configuration of the second communication device 20. Therefore, the description of the logical configuration of the second communication device 20 will be omitted. As shown in Figure 5, the first communication device 10 may include a control unit 110, a communication unit 120, and a storage unit 130 as logical components.

[0034] The control unit 110 may be implemented by a processor 101 and a memory 102. The control unit 110 may perform various processes in the first communication device 10. The control unit 110 may also control components of the first communication device 10, including a communication unit 120 and a storage unit 130.

[0035] The communication unit 120 may be implemented by a transceiver 104 and an antenna 105. The communication unit 120 may communicate wirelessly with the second communication device 20 and send and receive data and control signals to and from the second communication device 20. The storage unit 130 may be implemented by an auxiliary storage device 103. The storage unit 130 may store data generated by the learning model described later.

[0036] The control unit 110 may implement a generative AI that generates content based on the input information. The generative AI may be implemented by a deep learning model. Specifically, the generative AI may be implemented by a variational autoencoder (VAE), contrastive language-image pre-training (CLIP), generative adversarial networks (GAN), diffusion models, and / or large language models (LLM). The content may include text, images, audio, and / or sensor sequences.

[0037] The content may be generated based on the input information. The input information may include conditional information. The conditional information may be referred to as a prompt. The input information may be input to the generating AI by text, images, audio, and / or a series of sensors. The input information may be entered by the user. Alternatively, the input information may be entered by a request message (request message) or response message (response message) from a receiving communication device (second communication device 20).

[0038] The control unit 110 may include a vector output unit 110a and a content generation unit 110b. The vector output unit 110a may output a latent vector based on input information. The vector output unit 110a may also output a second latent vector based on a first latent vector transmitted from, for example, a second communication device 20. In this case, the input information may include the first latent vector transmitted from the second communication device 20. The first communication device 10 and the second communication device 20 may communicate by exchanging information (messages) that include latent vectors. Details of such communication cases will be described later.

[0039] The content generation unit 110b may generate content based on latent vectors. The vector output unit 110a and the content generation unit 110b may function in the transmitting communication device. In this embodiment, the vector output unit 110a and the content generation unit 110b may be implemented by CLIP, VAE, and LDM or LLM.

[0040] CLIP may be a learning model that learns text-image pairs collected from the internet. CLIP may implement a CLIP text encoder and a CLIP image encoder. The CLIP text encoder may convert the linguistic meaning of input text data into vectors. The CLIP image encoder may convert the visual features of input image data into vectors.

[0041] A VAE may be a learning model that learns from images collected on the internet and generates new content in the form of variational derivatives of the learned data. A VAE may implement a VAE encoder and a VAE decoder. The VAE encoder may embed an image into a latent vector by compressing a high-dimensional image into a low-dimensional latent space based on the input image data. The latent space represents an abstract representation of the image features. The latent vector represents a specific point in the latent space and may be called a tensor. The VAE decoder may generate content from a latent vector from which noise has been removed by LDM.

[0042] LDM may be a learning model that repeatedly applies noise to latent vectors, performs learning, and then removes the noise from the noisy latent vectors. Such noise may be called diffusion noise. LDM may remove diffusion noise from latent vectors that have been given diffusion noise. In LDM, a layered U-NET may remove diffusion noise from latent vectors based on input information (conditional information). The U-NET is a convolutional neural network (CNN) used for image recognition. Diffusion noise may be applied by CLIP and / or VAE. The same applies to LLM.

[0043] Figure 6 shows the process of learning latent vectors using a latent diffusion model in LDM. The process shown in Figure 6 may also be called the forward diffusion process. As shown in Figure 6, the VAE encoder may output latent vectors by encoding the image into latent space. The latent vectors may be input to the LDM. In the LDM, a hierarchical structure may be used to add diffusion noise to the latent vectors.

[0044] Figure 7 shows the process of generating content using a latent spread model in LDM. The process shown in Figure 7 may also be called the despreading process. As shown in Figure 7, in LDM, a layered U-NET may remove spreading noise from the latent vector based on the input information. LDM may output a latent vector from which spreading noise has been removed. The input information may be provided by CLIP and / or VAE. The latent vector may be input to a VAE decoder. The VAE decoder may generate content by decoding the latent vector into the pixel-word space.

[0045] Figure 8 shows a learning model implemented by the transmitting and receiving communication devices. As shown in Figure 8, in the transmitting communication device (first communication device 10), the vector output unit 110a may be implemented by a CLIP text encoder and a VAE encoder. The content generation unit 110b may be implemented by a VAE decoder. Although Figure 8 shows an example where the generated content is an image, the generated content is not limited to images and may include text, audio, and / or sensor sequences.

[0046] The control unit 110 may also include a content construction unit 110c. The content construction unit 110c may construct content based on latent vectors and input information transmitted from another communication device. The content construction unit 110c may function in the receiving communication device. In this embodiment, the content construction unit 110c may be implemented by a VAE decoder and an LDM or LLM.

[0047] As described above, the LDM may generate content by removing spreading noise from a latent vector to which spreading noise has been added. The latent vector may be transmitted from another communication device and may be received by the communication unit 120. The LDM may similarly generate content from the latent vector.

[0048] As shown in Figure 8, in the receiving communication device (second communication device 20), the content construction unit 110c may be implemented by a VAE decoder. The receiving communication device may generate content using the content construction unit 110c based on the latent vector transmitted from the transmitting communication device. Alternatively, the receiving communication device may output the content transmitted from the transmitting communication device.

[0049] The control unit 110 may further include a communication mode determination unit 110d for determining the communication mode. In this embodiment, the first communication device 10 may communicate with the second communication device 20 according to a plurality of communication modes. The plurality of communication modes may include a first communication mode, a second communication mode, and a third communication mode.

[0050] In the first communication mode, the first communication device 10 may transmit content generated by the content generation unit 110b to the second communication device 20. The first communication mode can accommodate cases where the receiving communication device (second communication device 20) is required to reproduce the content with higher fidelity. The first communication mode may also be called reproduction accuracy-focused communication or bit-level mode.

[0051] In the second communication mode, the first communication device 10 may transmit to the second communication device 20 the latent vector output by the vector output unit 110a and the input information input to the vector output unit 110a. The second communication device 20 may construct content based on the latent vector and input information transmitted from the first communication device 10. The content may be constructed by the content construction unit 110c. In the second communication mode, only the latent vector may be transmitted, or the content may be constructed based only on the latent vector. The second communication mode can reduce the amount of data transmitted over the network, for example. The second communication mode may be called the latent generation communication mode or semantic mode.

[0052] In the third communication mode, the first communication device 10 may transmit to the second communication device 20 the content generated by the content generation unit 110b, the latent vector output by the vector output unit 110a, and the input information input to the vector output unit 110a. The third communication mode is a combination of the first and second communication modes. In the third communication mode, a portion of the content generated by the content generation unit 110b, along with the latent vector and input information corresponding to the other portion of the content, may be transmitted. The third communication mode may be called Hybrid Mode.

[0053] In addition to the communication modes described above, the multiple communication modes may further include a fourth communication mode. In the fourth communication mode, the first communication device 10 may transmit only conditional information to the second communication device 20. The second communication device 20 may construct content based on the conditional information transmitted from the first communication device 10. The content may be constructed by the content construction unit 110c. The fourth communication mode can further reduce the amount of data transmitted over the network compared to, for example, the second communication mode. On the other hand, in the fourth communication mode, for example, the fidelity of content reproduction at the receiving communication device may be even lower compared to the second communication mode.

[0054] The third communication mode may be a combination of the first communication mode and the fourth communication mode. In this case, the first communication device 10 may transmit to the second communication device 20 the content generated by the content generation unit 110b and the input information input to the vector output unit 110a. Alternatively, a portion of the content generated by the content generation unit 110b and the input information corresponding to the other portion of the content may be transmitted.

[0055] The communication mode determination unit 110d may decide whether to adopt one of the above-mentioned communication modes based on communication quality, etc. As shown in Figure 8, in the receiving communication device (second communication device 20), the communication mode determination unit 110d may decide on one of the communication modes. Details of determining the communication mode will be described later.

[0056] As described above, in this embodiment, depending on the communication mode, the transmitting communication device may transmit the generated content to the receiving communication device, or it may transmit a latent vector and / or input information. When transmitting a latent vector and / or input information, the receiving communication device may construct content based on the latent vector and / or input information. In this case, although the content generated by the transmitting communication device is not transmitted, it can be said that content is being transmitted in the sense that the content is conveyed to the receiving communication device in a different form. Therefore, in the following, transmitting a latent vector and input information along with the content may be included in the definition of content transmission.

[0057] The learning model for implementing the generative AI shown in Figure 8 is merely illustrative. Other learning models may be used to implement the generative AI. Furthermore, while this embodiment describes an example of transmitting content generated by the generative AI between communication devices, it is not limited to such an example. For example, it may also include an example of transmitting content generated by an AI for analyzing and predicting text and images, i.e., machine learning, between communication devices.

[0058] The AI ​​used to analyze and predict the text and images described above may be called an analytical AI to distinguish it from a generative AI that generates new content. The analytical AI may output a latent vector based on the input information such as text and images. The analytical AI may also acquire content, including images, based on the latent vector. In other words, in this embodiment, a communication device implementing machine learning may output a latent vector and acquire or generate content based on the latent vector. The following embodiment describes an example in which a communication device implementing machine learning generates content, but the embodiment may also include an example in which content is acquired.

[0059] 1-2 Determining the Communication Mode The following describes an example in which the first communication device 10 corresponds to the transmitting communication device and the second communication device 20 corresponds to the receiving communication device. In the second communication mode, the first communication device 10 may transmit latent vectors and input information to the second communication device 20 instead of transmitting the generated content. In the fourth communication mode, the first communication device 10 may transmit only input information to the second communication device 20. The second communication device 20 may construct content based on the latent vectors and input information. On the other hand, in the first communication mode, the first communication device 10 may transmit the generated content to the second communication device 20.

[0060] In the second communication mode, compared to the first communication mode, the first communication device 10 can transmit content to the second communication device 20 with a smaller amount of data. On the other hand, in the second communication mode, the content constructed by the second communication device 20 does not necessarily match the content generated by the first communication device 10, and the fidelity of content reproduction may be inferior. In the fourth communication mode, since content is constructed based only on input information, the fidelity of content reproduction may be even inferior.

[0061] Communication using the first communication mode can increase the fidelity with which content is reproduced in the second communication device 20. On the other hand, communication using the first communication mode cannot reduce the amount of data transmitted over the network. Communication using the second communication mode can reduce the amount of data transmitted over the network. On the other hand, communication using the second communication mode may result in lower fidelity with which content is reproduced in the second communication device 20. The level of content reproduction fidelity and data reduction in communication using the third communication mode is intermediate between that of communication using the first communication mode and communication using the second communication mode.

[0062] The above-mentioned communication modes can be summarized in Table 1 regarding the fidelity of content reproduction and the level of data reduction. [Table 1]

[0063] For example, consider a case where the first communication device 10 generates a medical image of a patient in a medical setting and transmits the medical image to the second communication device 20. In such a case, it is conceivable that the medical image will be used for medical purposes in the second communication device 20. Since the medical image may affect the patient's life and health, it is necessary for the second communication device 20 to reproduce the medical image with higher fidelity. For this reason, in the above case, communication using the first communication mode or the third communication mode is preferable to communication using the second communication mode or the fourth communication mode.

[0064] Furthermore, let's consider a case where the quality of communication between the first communication device 10 and the second communication device 20 is significantly low. This deterioration in communication quality is often caused by an increase in network traffic. In such a case, the first communication device 10 may need to reduce the amount of data it transmits. For this reason, in the above case, communication using the second or fourth communication mode is preferable to communication using the first or third communication mode.

[0065] In this embodiment, the communication mode determination unit 110d may determine the communication mode based on several indicators. In this embodiment, the indicators may include the importance of the content, the communication quality, the input parameters, and the feedback from the receiving communication device. The communication mode determination unit 110d may determine the communication mode based on an evaluation of the above-mentioned indicators.

[0066] 1-2-1 Determining the communication mode based on importance The communication mode determination unit 110d may determine the communication mode based on the importance of the content. Importance may be determined from the perspective of safety, accuracy, and / or priority (urgency), for example. Importance may also be determined based on input information. For example, the communication mode determination unit 110d may determine that the content is of high importance based on the input information indicating that it generates medical images. Alternatively, the communication mode determination unit 110d may determine that the content is of high importance based on the input information indicating that it generates content to maintain safety.

[0067] Furthermore, importance may be determined based on the latent vector. Alternatively, importance may be determined based on the content generated from the latent vector. In other words, importance may be determined during the process of generating content from the latent vector. The communication mode determination unit 110d may, for example, determine that the content is highly important based on the latent vector, on the fact that it is related to a medical image.

[0068] The importance level may be determined by a deep learning model implemented in the control unit 110 learning content (text, images, audio, and / or sensor sequences) collected from the internet. Alternatively, the importance level may be determined by a rule-based method. For example, the importance level may be determined based on whether or not the input information contains predetermined keywords.

[0069] 1-2-2 Determination of communication mode based on communication quality The communication mode determination unit 110d may determine the communication mode based on communication quality. Communication quality may be determined from the perspective of, for example, the signal-to-noise ratio (SNR), round trip time (RTT), block error rate (BLER), communication channel, bandwidth, and / or power. Communication quality may also be determined based on measurement results. For example, the communication mode determination unit 110d may measure the reference signal received power (RSRP) based on the reference signal (RS) and determine whether the SNR is high or low based on the measurement results. Furthermore, if the communication mode determination unit 110d determines that the SNR is high, it may also determine that the communication quality is high.

[0070] Furthermore, the communication mode determination unit 110d may, for example, measure the difference between the time RS was transmitted and the time RS was received, and determine whether the RTT is high or low based on the measurement result. Also, if the communication mode determination unit 110d determines that the RTT is low, it may determine that the communication quality is high.

[0071] 1-2-3 Determining the communication mode based on input parameters The communication mode determination unit 110d may determine the communication mode based on the input parameters. The input parameters may be input to the communication mode determination unit 110d. For example, a mode that reproduces content with high fidelity may be defined, and parameters indicating that mode may be input to the communication mode determination unit 110d. Based on the input parameters, the communication mode determination unit 110d may determine that the content to be transmitted needs to be reproduced with high fidelity.

[0072] Furthermore, a mode for transmitting a low amount of data may be defined, and parameters indicating that mode may be input to the communication mode determination unit 110d. Based on the input parameters, the communication mode determination unit 110d may determine that it is necessary to reduce the amount of data to be transmitted.

[0073] Furthermore, for example, a device type for the first communication device 10 may be defined, and parameters indicating the device type may be input to the communication mode determination unit 110d. For example, an Ultra-Reliable and Low Latency Communications (URLLC) device may be defined as the device type. Based on the input parameters indicating a URLLC device, the communication mode determination unit 110d may determine that it is necessary to reduce the amount of data to be transmitted in order to achieve low latency.

[0074] 1-2-4 Determination of communication mode based on feedback from the receiving communication device The second communication device 20 may send feedback to the first communication device 10. The feedback may be sent as a response message (response telegram). For example, the feedback may indicate that the content needs to be reproduced with high fidelity. Alternatively, the feedback may indicate that the amount of data to be transmitted needs to be reduced. The communication mode determination unit 110d may determine the communication mode based on the feedback.

[0075] The communication mode determination unit 110d may determine that the content to be transmitted needs to be reproduced with high fidelity, based on feedback indicating that the content needs to be reproduced with high fidelity. Alternatively, the communication mode determination unit 110d may determine that the amount of data to be transmitted needs to be reduced, based on feedback indicating that the amount of data to be transmitted needs to be reduced.

[0076] 1-2-5 Determination of communication mode based on requests from the receiving communication device The second communication device 20 may send a request to the first communication device 10. The request may be sent as a request message (request telegram). For example, the request may indicate that the content needs to be reproduced with high fidelity. The communication mode determination unit 110d may determine the communication mode based on the request.

[0077] The communication mode determination unit 110d may determine that the content to be transmitted needs to be reproduced with high fidelity, based on the requirement indicating that the content needs to be reproduced with high fidelity. Alternatively, the communication mode determination unit 110d may determine that the amount of data to be transmitted needs to be reduced, based on the requirement indicating that the amount of data to be transmitted needs to be reduced.

[0078] 1-2-6 Determination of communication mode based on evaluation of indicators The communication mode determination unit 110d may determine the importance of the content and set a content fidelity score based on that importance. If the communication mode determination unit 110d determines that the content is of high importance, it may set a high content fidelity score. On the other hand, if the communication mode determination unit 110d determines that the content is of low importance, it may set a low content fidelity score. The content fidelity score may indicate the level of fidelity required for the second communication device 20 to reproduce the content.

[0079] The communication mode determination unit 110d may determine the communication quality and set a data volume reduction score based on the communication quality. If the communication mode determination unit 110d determines that the communication quality is high, it may set a low data volume reduction score. On the other hand, if the communication mode determination unit 110d determines that the communication quality is low, it may set a low data volume reduction score. The data volume reduction score may indicate the level at which the amount of data transmitted by the first communication device 10 through the network needs to be reduced.

[0080] Furthermore, the communication mode determination unit 110d may set a content fidelity score and / or a data volume reduction score based on the input parameters. In addition, the communication mode determination unit 110d may set a content fidelity score and / or a data volume reduction score based on feedback / requests from the second communication device 20.

[0081] In this way, the communication mode determination unit 110d may set the content fidelity score and the data volume reduction score based on the evaluation of the above-mentioned indicators. Finally, the communication mode determination unit 110d may determine the communication mode based on the content fidelity score and the data volume reduction score. In order to set the content fidelity score and the data volume reduction score, one or more of the above-mentioned indicators may be evaluated. Alternatively, only one of the content fidelity score and the data volume reduction score may be set.

[0082] The communication mode determination unit 110d may, for example, decide to adopt the first communication mode or the third communication mode if it determines that the content fidelity score is high. Alternatively, the communication mode determination unit 110d may decide to adopt the second communication mode or the fourth communication mode if it determines that the content fidelity score is low and the data volume reduction score is high. The content fidelity score and / or data volume reduction score may be set by a deep learning model implemented by the control unit 110 learning evaluations of the above-mentioned indicators.

[0083] The communication mode determination unit 110d may, for example, set a score separately for each part of the content. For example, let's assume the content is an image. For example, the communication mode determination unit 110d may identify a region of interest (ROI) and other areas (e.g., background area) within the generated image, determine the importance of the ROI and the background area separately, and set a content fidelity score based on the importance. The content parts are not limited to areas of an image, but may also include words, sentences, paragraphs of text, time intervals of audio, or intervals of sensor data. In other words, the content parts may include partial units of content that include text, images, audio, and / or sensor sequences.

[0084] The communication mode determination unit 110d may decide to adopt a third communication mode, for example, if it sets a high content fidelity score for the ROI and a low content fidelity score for the background area. In this case, content may be generated only for the ROI by the content generation unit 110b, and latent vectors may be output only for the background area by the vector output unit 110a. The first communication device 10 may transmit to the second communication device 20 the content generated for the ROI and the latent vectors output for the background area using the third communication mode.

[0085] In the third communication mode, input information for the background area may be transmitted to the second communication device 20. Regarding the input information, the input information input to the vector output unit 110a may be transmitted. Alternatively, regarding the input information, the input information input to the vector output unit 110a may be updated, and the updated input information may be transmitted. For example, if a low content fidelity score is set for the background area, the input information may be updated to indicate that the background area should be painted white.

[0086] Furthermore, if the content fidelity score is set low for the background region, the third communication mode may transmit only the input information for the background region to the second communication device 20 along with the content generated for the ROI. In this case, the latent vector does not need to be transmitted. In other words, the third communication mode may be a combination of the first communication mode and the fourth communication mode.

[0087] 1-3 Content Delivery Processing Next, with reference to Figure 9, the process by which the first communication device 10 transmits content to the second communication device 20 will be described. In the example shown in Figure 9, the first communication device 10 and the second communication device 20 may interact by sending and receiving content generated by the generation AI.

[0088] In step S901, the control unit 110 in the first communication device 10 may receive the input information. The input information may be entered by the user, or it may be entered as a response from the second communication device 20.

[0089] Next, the control unit 110 may output a latent vector based on the input information (step S902). The latent vector may be output by the vector output unit 110a. The vector output unit 110a may implement a CLIP text encoder and / or a VAE encoder.

[0090] Next, the control unit 110 may generate content based on the input information received in step S901 and the latent vector output in step S902 (step S903). The content may be generated by the content generation unit 110b. The content may include text, images, audio, and / or a series of sensors.

[0091] Next, the control unit 110 may set a content fidelity score (step S904). The content fidelity score may be set by the communication mode determination unit 110d. The content fidelity score may be set based on the importance of the content, input parameters, and / or feedback / requests from the second communication device 20.

[0092] Next, the control unit 110 may set a data volume reduction score (step S905). The data volume reduction score may be set by the communication mode determination unit 110d. The data volume reduction score may be set based on communication quality, input parameters, and / or feedback / requests from the second communication device 20.

[0093] Next, the control unit 110 may determine the communication mode based on either or both of the content reproduction fidelity score set in step S904 and the data volume reduction score set in step S905 (step S906). The communication mode may also be determined by the communication mode determination unit 110d.

[0094] In step S906, the content fidelity score may be set separately for each part of the content. Based on the score set for each part of the content, it may be decided to adopt a third communication mode.

[0095] If it is decided in step S906 to adopt the first communication mode, the communication unit 120 may transmit the content generated in step S903 to the second communication device 20 (step S907). The second communication device 20 may output the received content.

[0096] If it is decided in step S906 to adopt the second communication mode, the communication unit 120 may transmit to the second communication device 20 the input information input in step S901 and the latent vector output in step S902 (step S908). In step S908, the input information may not be transmitted, and only the latent vector may be transmitted.

[0097] Although not shown in Figure 9, in step S906, it may be decided to adopt a fourth communication mode instead of a second communication mode based on the content fidelity score. If it is decided to adopt a fourth communication mode, in step S908, the communication unit 120 may transmit only the input information entered in step S901 to the second communication device 20. The second communication device 20 may construct content based on the received latent vector and / or input information and output the constructed content.

[0098] If it is decided in step S906 to adopt the third communication mode, the communication unit 120 may transmit the content generated in step S903 to the second communication device 20, as well as the input information input in step S901 and the latent vector output in step S902 (step S909). In step S909, the input information may not be transmitted, and only the latent vector may be transmitted.

[0099] The content transmitted in step S909 may correspond to portions of the content where a high content fidelity score is set. The latent vector and input information transmitted in step S909 may correspond to portions of the content where a low content fidelity score is set.

[0100] The third communication mode adopted in step S909 may represent a combination of the first communication mode and the second communication mode. Alternatively, the third communication mode may represent a combination of the first communication mode and the fourth communication mode. In this case, the communication unit 120 may transmit the content generated in step S903 to the second communication device 20, as well as transmit only the input information entered in step S901. The second communication device 20 may construct content based on the received latent vector and / or input information, and combine the content received from the first communication device 10 with the constructed content. The second communication device 20 may also output the combined content.

[0101] Based on the processing described in Figure 9, an appropriate communication mode may be adopted based on the content fidelity score and / or data volume reduction score. According to the adopted communication mode, the first communication device 10 may transmit the generated content or latent vector / input information.

[0102] The first embodiment has been described above. According to the first embodiment, it is possible to appropriately address both cases in which it is necessary to reproduce content with higher fidelity and cases in which it is necessary to reduce the amount of data transmitted over the network.

[0103] 2. Second Embodiment 2-1 Communication in a Layered Structure Next, a second embodiment will be described. In the second embodiment, in addition to or instead of the features described in the first embodiment, an example is included in which the communication devices communicate in a hierarchical structure. In the second embodiment as well, an example will be described in which the first communication device 10 corresponds to the transmitting communication device and the second communication device 20 corresponds to the receiving communication device.

[0104] In this embodiment, a communication model may be defined for mutually transmitting content generated by machine learning between communication devices. Hereinafter, the communication model described in this embodiment may be referred to as an AI communication model or an ML communication model. The AI ​​communication model may be implemented using a hierarchical structure. In the hierarchical structure of the AI ​​communication model, each layer may function independently for the communication devices to communicate with each other. Each layer may implement an independent communication protocol. The AI ​​communication model may be implemented in both the first communication device 10 and the second communication device 20 that communicate with each other.

[0105] The AI ​​communication model may include a layer for providing the functions described in the first embodiment. In this embodiment, such a layer may be referred to as a semantic layer. The semantic layer may perform roles such as determining the communication mode and / or determining the importance of the content.

[0106] 2-2 Hierarchical structure based on the OSI reference model AI communication models may be based on the Open Systems Interconnect (OSI) Reference Model. The OSI Reference Model is a conceptual model developed by the International Organization for Standardization (ISO). In the OSI Reference Model, seven layers function independently to realize communication functionality. These seven layers, from top to bottom, are the Application Layer, Presentation Layer, Session Layer, Transport Layer, Network Layer, Data Link Layer, and Physical Layer.

[0107] In this embodiment, the AI ​​communication model may be defined based on the OSI reference model. Figure 10 shows an AI communication model defined based on the OSI reference model. As shown in Figure 10, the AI ​​communication model may include, from the top layer, the application layer, semantic layer, transport layer, network layer, data link layer, and physical layer. In other words, in the AI ​​communication model shown in Figure 10, the semantic layer may be defined instead of the presentation layer and session layer in the OSI reference model.

[0108] The application layer may play a role defined in the OSI reference model. Alternatively, the application layer may interpret latent vectors according to the format in learning models such as CLIP and VAE. As described above, the first communication device 10 and the second communication device 20 may communicate by exchanging information (messages) that include latent vectors. In this case, the application layer in the first communication device 10 may interpret the latent vectors transmitted from the second communication device 20. The application layer may pass the interpreted latent vectors to lower layers.

[0109] The semantic layer may play a role such as determining the importance of content and / or determining the communication quality, as described in the first embodiment. The semantic layer may also play a role in determining the communication mode. Furthermore, the semantic layer may play a role in controlling the coordination between the learning models (LDM, VAE, and CLIP). In other words, the semantic layer may output latent vectors and / or generate content based on these latent vectors. The semantic layer may also output latent vectors and / or generate content based on these latent vectors depending on the communication mode. The semantic layer may implement the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d described in the first embodiment.

[0110] The transport layer may perform a role defined in the OSI reference model. The transport layer may also perform priority control over communication in a first, second, and third communication mode, along with quality of service (QoS) control. For example, communication in a second communication mode may have higher priority than communication in a first communication mode.

[0111] The network layer may perform a role defined in the OSI reference model. Alternatively, the network layer may identify the content transmitted by a first communication mode and the latent vectors and input information transmitted by a second communication mode, and route the packets. For such identification, for example, an ID indicating the communication mode (e.g., a communication mode ID) or a flag indicating that latent vectors are being transmitted (e.g., a semantic flag) may be used.

[0112] The data link layer may play a role defined in the OSI reference model. The data link layer may also provide the semantic layer with information such as SNR, BLER, and bandwidth. As described in the first embodiment, SNR, BLER, and bandwidth may be used as indicators for determining communication quality.

[0113] The physical layer may perform a role defined in the OSI reference model. It may also perform the role of encoding a mixture of content and latent vectors transmitted by the third communication mode. Furthermore, the physical layer may prioritize encoding latent vectors when communication quality is low.

[0114] In Figure 10, a semantic layer is defined instead of a presentation layer and a session layer, but the configuration is not limited to this. An AI communication model may include a presentation layer and a session layer along with the semantic layer. In this case, the semantic layer may be located at any layer between the application layer and the transport layer.

[0115] 2-3 Layered structure based on the TCP / IP model Instead of the example above, the AI ​​communication model may be based on the Transmission Control Protocol (TCP) / Internet Protocol (IP) model. The TCP / IP model is a communication model used on IP networks. In the TCP / IP model, four layers function independently to realize communication functionality. These four layers, from top to bottom, are the application layer, transport layer, internet layer, and network interface layer.

[0116] In this embodiment, the AI ​​communication model may be defined based on the TCP / IP model. Hereinafter, the AI ​​communication model defined based on the OSI reference model may be referred to as the first AI communication model. The AI ​​communication model defined based on the TCP / IP model may be referred to as the second AI communication model.

[0117] Figure 11 shows a second AI communication model. As shown in Figure 11, the second AI communication model may include, from the top layer, an application layer, a semantic layer, a transport layer, an internet layer, and a network interface layer. In other words, in the second AI communication model shown in Figure 11, a semantic layer may be additionally defined in the TCP / IP model.

[0118] The application layer may correspond to the application layer in the first AI communication model. The semantic layer may correspond to the semantic layer in the first AI communication model. The transport layer may correspond to the transport layer in the first AI communication model. The internet layer may correspond to the network layer in the first AI communication model. The network interface layer may correspond to the data link layer and physical layer in the first AI communication model.

[0119] The hierarchical structures shown in Figures 10 and 11 are illustrative examples only. A new AI communication model separate from the OSI reference model and the TCP / IP model may be defined. Such a communication model may have a hierarchical structure and may include at least a layer that performs the role provided by the semantic layer described above.

[0120] The AI ​​communication model may be implemented in both the first communication device 10 and the second communication device 20. The semantic layer in the first communication device 10 may communicate with the semantic layer in the second communication device 20. Each of the other layers in the first communication device 10 may also communicate with the corresponding layer in the second communication device 20.

[0121] 2-4 Processing at the semantic layer of the transmitting communication device In the semantic layer of the transmitting communication device (first communication device 10), independent entities may operate under the control of the control unit 110. Such entities may be referred to as semantic entities. Semantic entities may operate in cooperation with upper and lower layers.

[0122] A semantic entity may receive data from a higher layer and generate packets based on the data received from the higher layer. Packets generated by a semantic entity may be called semantic packets. Semantic packets may be generated by adding header information to data received from a higher layer. A semantic packet may include a header area and a data area. Control information may be set in the header area by the semantic entity. Data may be set in the data area.

[0123] The header area of ​​the semantic packet may be referenced by the semantic entity in both the first communication device 10 and the second communication device 20. Therefore, the header area of ​​the semantic packet may contain information necessary for the operation of the semantic entity in each of the first communication device 10 and the second communication device 20.

[0124] A semantic packet will be described with reference to Figure 12. As shown in Figure 12, a semantic packet may include a header area and a data area. The header area may include control information set by the semantic entity. The control information may indicate information necessary for the first communication device 10 to transmit content to the second communication device 20. The data area may include data set by the semantic entity. The data area may include content, latent vectors, and / or input information.

[0125] A semantic entity may determine a communication mode. The communication mode may be determined as described in the first embodiment. The semantic entity may set information indicating the determined communication mode in the header area. As shown in Figure 12, the header area of ​​the semantic packet may include a communication mode ID. The communication mode ID may indicate the communication mode determined by the semantic entity in the first communication device 10.

[0126] By setting a communication mode ID in the header area of ​​the semantic packet, the second communication device 20 can decide whether to output the content transmitted from the first communication device 10 or to construct content based on the transmitted latent vector. In this way, information indicating which communication mode is adopted can be sent and received between the first communication device 10 and the second communication device 20. Note that the communication mode ID does not necessarily have to be sent and received between the first communication device 10 and the second communication device 20. For example, the communication mode may be pre-set in either or both of the first communication device 10 and the second communication device 20.

[0127] A semantic entity may determine the importance of the content. The importance may be determined as described in the first embodiment. The semantic entity may set information indicating the determined importance in the header area. As shown in Figure 12, the header area of ​​the semantic packet may include the importance level. The importance level may indicate the importance level determined by the semantic entity in the first communication device 10.

[0128] By setting an importance level in the header area of ​​the semantic packet, the second communication device 20 can provide feedback on the importance level indicated by the first communication device 10. For example, the second communication device 20 may provide feedback to the first communication device 10 indicating that the content transmitted from the first communication device 10 is of higher importance. In this way, information indicating the importance of the content can be sent and received between the first communication device 10 and the second communication device 20.

[0129] The semantic entity in the second communication device 20 may set information in the header area that indicates the communication mode requested by the second communication device 20, which is feedback to the communication mode ID indicated by the first communication device 10. As shown in Figure 12, the header area of ​​the semantic packet may include a preferred communication mode ID. The preferred communication mode ID may indicate the communication mode requested by the semantic entity in the second communication device 20.

[0130] The semantic entity in the second communication device 20 may set information in the header area that indicates the importance level determined by the second communication device 20, which is feedback to the importance level indicated by the first communication device 10. As shown in Figure 12, the header area of ​​the semantic packet may include the reception side importance level. The reception side importance level may indicate the importance level determined by the semantic entity in the second communication device 20.

[0131] A semantic entity may output a latent vector based on the input information. The latent vector may be output as described in the first embodiment. The semantic entity may also generate content based on the latent vector. The content may be generated as described in the first embodiment. Depending on the determined communication mode, the semantic entity may set the content and / or the latent vector / input information in the data area.

[0132] If a semantic entity decides to adopt the first communication mode, it may set content in the data area. If a semantic entity decides to adopt the second or fourth communication mode, it may set latent vectors and / or input information in the data area. If a semantic entity decides to adopt the third communication mode, it may set content and latent vectors and / or input information in the data area.

[0133] The information included in the header area shown in Figure 12 is illustrative only. For example, in addition to, or instead of, the information shown in Figure 12, a flag indicating that a latent vector is being transmitted (e.g., a semantic flag) may be set in the header area. Also, if the data area includes both a latent vector and content, information indicating the start position of the latent vector and / or the start position of the content may be set in the header area.

[0134] Once a semantic entity generates a semantic packet as described above, it may pass the semantic packet to a lower layer. The semantic entity may also notify the lower layer of the communication mode ID. Furthermore, the semantic entity may notify the lower layer of a flag indicating that a latent vector is being transmitted. This information may be used in the network layer.

[0135] Semantic packets generated in the semantic layer may be transmitted by the communication unit 120 to the receiving communication device (second communication device 20).

[0136] Next, with reference to Figure 13, the processing performed by the semantic entity of the first communication device 10 will be described. In the example shown in Figure 13, semantic packets may be generated in the semantic layer of the first communication device 10. The semantic entity may implement a vector output unit 110a, a content generation unit 110b, a content construction unit 110c, and a communication mode determination unit 110d.

[0137] In step S1301, the semantic entity may receive data from a higher layer. Next, the semantic entity may receive input information and output a latent vector based on the input information (step S1302). The latent vector may be output as described in the first embodiment.

[0138] Next, the semantic entity may generate content based on the latent vector (step S1303). The content may be generated as described in the first embodiment. Next, the semantic entity may determine the communication mode (step S1304). The communication mode may be determined as described in the first embodiment.

[0139] Next, the semantic entity may set information in the header area (step S1305). The information set in the header area may include the communication mode ID and / or severity level mentioned above.

[0140] Next, the semantic entity may set the content generated in step S1303 and / or the latent vector / input information output in step S1302 in the data area, according to the communication mode determined in step S1304 (step S1306). In this way, a semantic packet including a header area and a data area may be generated.

[0141] Next, the semantic entity may pass the semantic packet generated in step S1306 to the lower layer (step S1307). At this time, the semantic entity may notify the lower layer of the communication mode ID, etc.

[0142] As described in Figure 13, the first communication device 10 may operate at the semantic layer. Alternatively, the first communication device 10 may generate semantic packets at the semantic layer and transmit them to the second communication device 20.

[0143] 2-5 Processing at the semantic layer of the receiving communication device Semantic entities may also operate in the semantic layer of the receiving communication device (second communication device 20) under the control of the control unit 110. Semantic entities may also operate in cooperation with upper and lower layers.

[0144] A semantic entity may receive data from a lower layer and extract a semantic packet from the data received from the lower layer. The semantic entity may perform processing based on the information set in the header area of ​​the semantic packet.

[0145] A semantic entity may determine the communication mode adopted to transmit the semantic packet based on the communication mode ID set in the header area. Depending on the determined communication mode, the semantic entity may output content or construct content based on latent vectors.

[0146] If the semantic entity determines that the first communication mode is being adopted, it may output the content contained in the data area. If the semantic entity determines that the second or fourth communication mode is being adopted, it may construct content based on the latent vectors and / or input information contained in the data area. If the semantic entity determines that the third communication mode is being adopted, it may construct content based on the latent vectors and / or input information contained in the data area and combine the content contained in the data area with the constructed content.

[0147] Next, with reference to Figure 14, the processing performed by the semantic entity of the second communication device 20 will be described. In the example shown in Figure 14, semantic packets may be processed in the semantic layer of the second communication device 20. The semantic entity may implement a vector output unit 110a, a content generation unit 110b, a content construction unit 110c, and a communication mode determination unit 110d.

[0148] In step S1401, the semantic entity may receive data from a lower layer. The data received from the lower layer may include a semantic packet transmitted from the first communication device 10. Next, the semantic entity may extract a semantic packet from the data received in step S1401 (step S1402). The semantic packet may include a header area and a data area. The header area may include information set by the first communication device 10. The data area may include content and / or latent vector / input information set by the first communication device 10.

[0149] Next, the semantic entity may determine the communication mode adopted to transmit the semantic packet based on the communication mode ID included in the header area (step S1403). If it is determined in step S1403 that the first communication mode has been adopted, the semantic entity may output the contents included in the data area (step S1404).

[0150] If it is determined in step S1403 that a second communication mode is being adopted, the semantic entity may construct content based on the latent vector / input information contained in the data area and output the constructed content (step S1405).

[0151] If it is determined in step S1403 that a third communication mode is being adopted, the semantic entity may construct content based on the latent vector / input information contained in the data area and combine the constructed content with the content contained in the data area (step S1406). The semantic entity may output the combined content.

[0152] As described in Figure 14, the second communication device 20 may operate at the semantic layer. Alternatively, the second communication device 20 may receive semantic packets at the semantic layer and output content from the semantic packets.

[0153] The semantic entity of the second communication device 20 may send a response packet to the first communication device 10 in response to receiving a semantic packet from the first communication device 10. The response packet may include feedback in response to receiving the semantic packet from the first communication device 10. The response packet may be generated at the semantic layer. The second communication device 20 may set the requested communication mode ID and / or the receiving importance level as feedback in the header area of ​​the response packet.

[0154] The response packet may have the same layout as the semantic packet described with respect to Figure 12. Alternatively, the response packet may have a different layout from the semantic packet described with respect to Figure 12.

[0155] The semantic entity of the second communication device 20 may send a request packet to the first communication device 10 before receiving a semantic packet from the first communication device 10. The request packet may indicate a request to the first communication device 10 to transmit content. The request packet may be generated at the semantic layer. The second communication device 20 may set a request communication mode ID and / or a receiver importance level in the header area of ​​the request packet as a request to the first communication device 10.

[0156] The request packet may have the same layout as the semantic packet described with respect to Figure 12. Alternatively, the request packet may have a different layout from the semantic packet described with respect to Figure 12.

[0157] The semantic entity of the first communication device 10 may, in response to receiving a request packet from the second communication device 20, transmit a semantic packet to the second communication device 20. The semantic entity may also determine the communication mode based on the request communication mode ID and / or recipient importance level indicated in the request packet.

[0158] The second embodiment has been described above. According to the second embodiment, in communication between two communication devices, processing including determining the communication mode can be made independent of processing in other layers.

[0159] 3. Third Embodiment 3-1 Processing in Decodeless Mode Next, a third embodiment will be described. In addition to the features described in the first and second embodiments, the third embodiment includes an example in which the receiving communication device does not construct (decode) the content. In the third embodiment, an example will be described in which the first communication device 10 corresponds to the transmitting communication device and the second communication device 20 corresponds to the receiving communication device.

[0160] IoT devices that implement machine learning are expected to be able to communicate with each other autonomously. For example, consider the case of controlling robots operating in a factory. In this case, it is conceivable that the server equipment controlling the robot and the IoT devices embedded in the robot could communicate with each other autonomously.

[0161] In the above case, it is assumed that the server device and the robot-side IoT device can communicate with each other by sending and receiving content / latent vectors. In other words, no human intervention is required in their interaction. Human intervention in the interaction may include the human operating one or both of the communication devices to enable communication between them. Human intervention in the interaction may also include the human referring to content generated and output by the communication devices through machine learning.

[0162] The content output by the communication device through machine learning may include text, images, audio, and / or sensor sequences. Such content is perceived by humans. When humans are not involved in the interaction, it is not necessary to output the above content. In this embodiment, in such cases, the content does not need to be output (decoded). Instead, the communication devices may interact with each other using latent vectors.

[0163] Communication devices interacting with each other using latent vectors may include, for example, the following cases. For example, a robot-side IoT device may capture an image of its surroundings and output a latent vector using the captured image as input information. The robot-side IoT device may be one of the communication devices. The robot-side IoT device may also send information (messages) indicating the latent vector to a server device. The server device may recognize objects present near the robot based on the latent vector and send commands to the robot-side IoT device based on the recognized objects. The server device may be the other of the communication devices. In the above case, the server device can send commands to the robot-side IoT device without constructing content based on the latent vector.

[0164] In this embodiment, in the case described above, the content construction unit 110c of the receiving communication device (second communication device 20) does not need to construct content based on the latent vector and / or input information received from the transmitting communication device (first communication device 10). Instead, the second communication device 20 may output a latent vector based on the latent vector and / or input information received from the first communication device 10, and may transmit the latent vector and / or input information to the first communication device 10.

[0165] In this embodiment, a mode in which the second communication device 20 does not construct content may be defined as a decodeless mode. In the decodeless mode, the latent vector may be communicated without being decoded into human-perceptible content. On the other hand, a mode in which the second communication device 20 constructs content may be defined as a decoded mode. In the decoded mode, the latent vector may be decoded into human-perceptible content. When the first communication device 10 and the second communication device 20 operate in the decodeless mode, the second communication device 20 does not need to decode the content. Whether or not to operate in the decodeless mode may be determined by either the first communication device 10 or the second communication device 20.

[0166] When the first communication device 10 and the second communication device 20 communicate in decodeless mode, they may communicate by exchanging information (messages) that indicate latent vectors. In this case, the application layer of the second communication device 20 may interpret the latent vectors contained in the message and pass the interpreted latent vectors to the semantic layer. Alternatively, the semantic layer of the second communication device 20 may interpret the latent vectors contained in the message.

[0167] Referring to Figure 15, the logical configuration of the first communication device 10 according to the third embodiment will be described. The logical configuration of the first communication device 10 shown in Figure 15 may also be applied to the logical configuration of the second communication device 20. Therefore, the description of the logical configuration of the second communication device 20 will be omitted. As shown in Figure 15, the first communication device 10 may include a control unit 110, a communication unit 120, and a storage unit 130 as logical components.

[0168] In the configuration shown in Figure 15, the communication unit 120 and the storage unit 130 may be the same as the communication unit 120 and storage unit 130 shown in Figure 5. Also, the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d included in the control unit 110 may be the same as the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d shown in Figure 5.

[0169] As shown in Figure 15, the control unit 110 according to the third embodiment may further include a decode execution decision unit 110e. The decode execution decision unit 110e may decide whether or not to construct content in the second communication device 20, that is, whether or not to interact in decodeless mode. The decode execution decision unit 110e may be implemented in the semantic layer, as described in the second embodiment.

[0170] The decode execution decision unit 110e may, by default, decide to interact in decodeless mode. In this case, the decode execution decision unit 110e may, in principle, decide to interact in decodeless mode. Alternatively, the decode execution decision unit 110e may decide to interact in decodeless mode based on pre-set conditions. For example, each communication device may be set to interact in decodeless mode, and the unit may decide to interact in decodeless mode based on this setting.

[0171] If the decode execution decision unit 110e decides to communicate in decodeless mode, the communication mode decision unit 110d does not need to decide on a communication mode. If the decode execution decision unit 110e decides to communicate in decodeless mode, it may decide to adopt the second communication mode or the fourth communication mode. If the decode execution decision unit 110e decides to communicate in decodeless mode, the latent vector and / or input information output by the vector output unit 110a may be transmitted to the second communication device 20.

[0172] The decode execution decision unit 110e may decide whether or not to engage in dialogue in decodeless mode based on whether or not a human intervenes in the dialogue. Hereinafter, dialogue (communication) between communication devices without human intervention may be referred to as MLtoML dialogue (communication) (ML2ML dialogue (communication)) or AItoAI dialogue (communication) (AI2AI dialogue (communication)). On the other hand, dialogue (communication) between communication devices with human intervention may be referred to as HumantoML communication (H2ML dialogue (communication)) or HumantoAI dialogue (communication) (H2AI dialogue (communication)).

[0173] The decoding execution decision unit 110e may decide to communicate in decoding mode if the interaction with the second communication device 20 is an ML2ML interaction. The decoding execution decision unit 110e may decide to communicate in decoding mode if the interaction with the second communication device 20 is an H2ML interaction. If the decoding execution decision unit 110e decides to communicate in decoding mode, the communication mode may be determined by the communication mode decision unit 110d, as described in the first embodiment.

[0174] The decode execution decision unit 110e may decide to interact in decode mode, for example, based on the fact that the first communication device 10 has been operated by a human. For example, the first communication device 10 may be provided with a touch sensor, and the touch sensor may detect human operation.

[0175] The decode execution decision unit 110e may decide to interact in decode mode based on the fact that information, such as input information, has been input to the first communication device 10. For example, if text, images, audio, and / or a series of sensors are input to the first communication device 10, the decode execution decision unit 110e may consider this to be input by a human and decide to interact in decode mode.

[0176] The decode execution decision unit 110e may decide whether or not to interact in decodeless mode based on a request or response from the second communication device 20. The second communication device 20 may also have the decode execution decision unit 110e implemented. In the second communication device 20, the decode execution decision unit 110e may also decide whether or not to interact in decodeless mode. The second communication device 20 may also decide whether or not to interact in decodeless mode, similar to the first communication device 10.

[0177] The example of the decode execution decision unit 110e deciding whether or not to interact in decodeless mode is merely illustrative. In addition to, or instead of, whether or not a human intervenes in the interaction, the decision of whether or not to interact in decodeless mode may be based on other conditions.

[0178] A semantic packet according to a third embodiment will be described with reference to Figure 16. As shown in Figure 16, a semantic packet may include a header area and a data area. In the configuration shown in Figure 16, the data area may be the same as the data area shown in Figure 12. Also, the communication mode ID, importance level, requested communication mode ID, and receiving side importance level included in the header area may be the same as the communication mode ID, importance level, requested communication mode ID, and receiving side importance level shown in Figure 12.

[0179] As shown in Figure 16, the header area of ​​the semantic packet according to the third embodiment may further include a decodeless flag. The decodeless flag may indicate whether or not the second communication device 20 constructs content, that is, whether or not to interact in decodeless mode.

[0180] When the first communication device 10 decides whether or not to interact in decodeless mode, the first communication device 10 may set information in the decodeless flag in the header area. For example, if the first communication device 10 decides to interact in decodeless mode, the first communication device 10 may set the decodeless flag in the header area to 1 (interaction in decodeless mode). In this case, the first communication device 10 may set latent vectors and / or input information in the data area. Furthermore, the first communication device 10 does not have to generate content based on latent vectors.

[0181] If the first communication device 10 decides to interact in decode mode, it may set the decodeless flag in the header area to 0 (interaction in decode mode). In this case, the first communication device 10 may set content, or latent vectors and / or input information, in the data area. The first communication device 10 may also generate content based on latent vectors.

[0182] The second communication device 20 may also implement a decode execution decision unit 110e. The decode execution decision unit 110e of the second communication device 20 may decide whether or not to construct content based on a decodeless flag set in the header area of ​​the semantic packet transmitted from the first communication device 10.

[0183] When the second communication device 20 decides whether or not to communicate in decodeless mode, the second communication device 20 may set information in the decodeless flag in the header area. In this case, the second communication device 20 may send a semantic packet containing the decodeless flag to the first communication device 10 as a response packet or request packet. The first communication device 10 may decide whether or not to communicate in decodeless mode based on the response packet or request packet.

[0184] If the second communication device 20 decides to communicate in decodeless mode, it may interpret the latent vector contained in the semantic packet transmitted from the first communication device 10, and may output a second latent vector based on the interpreted first latent vector. The latent vector may be interpreted by the application layer or the semantic layer. The second communication device 20 may also transmit a semantic packet containing the second latent vector to the first communication device 10. In this way, the first communication device 10 and the second communication device 20 may communicate using latent vectors without transmitting or constructing content.

[0185] 3-2 Processing in the transmitting communication device according to the decodeless mode Next, referring to Figure 17, the process by which the first communication device 10 generates semantic packets depending on whether or not it interacts in decodeless mode will be described. In the example shown in Figure 17, the semantic layer of the first communication device 10 may decide whether or not to interact in decodeless mode. The semantic entity may implement a vector output unit 110a, a content generation unit 110b, a content construction unit 110c, a communication mode determination unit 110d, and a decode execution determination unit 110e.

[0186] In step S1701, the semantic entity may receive data from a higher layer. Next, the semantic entity may receive input information and output a latent vector based on the input information (step S1702). The latent vector may be output as described in the first embodiment.

[0187] Next, the semantic entity may decide whether or not to interact using the decode mode (step S1703). Whether or not to interact using the decode mode may be determined as described above.

[0188] If it is decided in step S1703 to interact in decodeless mode, the process may proceed to step S1706. If it is decided in step S1703 to interact in decode mode, the semantic entity may generate content based on the latent vector (step S1704). The content may be generated as described in the first embodiment. Next, the semantic entity may determine the communication mode (step S1705). The communication mode may be determined as described in the first embodiment.

[0189] Next, the semantic entity may set information in the header area (step S1706). The information set in the header area may include the decodeless flag, communication mode ID and / or severity level mentioned above.

[0190] Next, the semantic entity may set the content generated in step S1704 and / or the latent vector / input information output in step S1702 in the data area, depending on whether or not it interacts using the decodeless mode determined in step S1703, and depending on the communication mode determined in step S1705 (step S1707). In this way, a semantic packet including a header area and a data area may be generated.

[0191] Next, the semantic entity may pass the semantic packet generated in step S1707 to the lower layer (step S1708). At this time, the semantic entity may notify the lower layer of the decodeless flag and / or the communication mode ID, etc.

[0192] As described in Figure 17, the first communication device 10 may decide whether or not the first communication device 10 and the second communication device 20 will communicate in decodeless mode. The first communication device 10 may also generate semantic packets depending on whether or not the first communication device 10 and the second communication device 20 will communicate in decodeless mode.

[0193] In the process shown in Figure 17, the first communication device 10 decides whether to communicate in decodeless mode depending on whether the dialogue between the first communication device 10 and the second communication device 20 is an MLtoML dialogue. However, the second communication device 20 may decide whether to communicate in decodeless mode using a similar method. In this case, the second communication device 20 may set a decodeless flag in the header area of ​​the response packet or request packet.

[0194] 3-3 Processing in the receiving communication device according to the decodeless mode Next, referring to Figure 18, the process by which the first communication device 10 determines whether or not to construct content depending on whether or not to interact in decodeless mode will be described. In the example shown in Figure 18, the decision of whether or not to interact in decodeless mode may be made in the semantic layer of the second communication device 20. The semantic entity may implement a vector output unit 110a, a content generation unit 110b, a content construction unit 110c, a communication mode determination unit 110d, and a decode execution determination unit 110e.

[0195] In step S1801, the semantic entity may receive data from a lower layer. The data received from the lower layer may include a semantic packet transmitted from the first communication device 10. Next, the semantic entity may extract a semantic packet from the data received in step S1801 (step S1802). The semantic packet may include a header area and a data area. The header area may include information set by the first communication device 10. The data area may include content and / or latent vector / input information set by the first communication device 10.

[0196] Next, the semantic entity may decide whether or not to interact in decode mode based on the decodeless flag included in the header area (step S1803). If it decides to interact in decodeless mode in step S1803, the semantic entity may output the latent vector / input information included in the data area (step S1804). In this case, processing ends.

[0197] If it is decided in step S1803 to interact using the decode mode, the semantic entity may determine the communication mode adopted to send the semantic packet based on the communication mode ID included in the header area (step S1805).

[0198] If it is determined in step S1805 that the first communication mode is adopted, the semantic entity may output the content contained in the data area (step S1806). If it is determined in step S1805 that the second communication mode is adopted, the semantic entity may construct content based on the latent vector / input information contained in the data area and output the constructed content (step S1807).

[0199] If it is determined in step S1805 that a third communication mode is being adopted, the semantic entity may construct content based on the latent vector / input information contained in the data area and combine the constructed content with the content contained in the data area (step S1808). The semantic entity may output the combined content.

[0200] As described in Figure 18, the second communication device 20 may decide whether or not the first communication device 10 and the second communication device 20 will communicate in decodeless mode based on the semantic packets transmitted from the first communication device 10. The second communication device 20 may also decide whether or not to construct content depending on whether or not the first communication device 10 and the second communication device 20 will communicate in decodeless mode.

[0201] In the process shown in Figure 18, the first communication device 10 decides whether to communicate in decodeless mode depending on whether the dialogue between the first communication device 10 and the second communication device 20 is an MLtoML dialogue. However, the second communication device 20 may decide whether to communicate in decodeless mode using a similar method. In this case, the second communication device 20 may set a decodeless flag in the header area of ​​the response packet or request packet.

[0202] The third embodiment has been described above. According to the third embodiment, when it is not necessary to construct content, the two communication devices can communicate using latent vectors without decoding the content.

[0203] 4. Fourth Embodiment 4-1 Learning latent vectors with added communication noise Next, a fourth embodiment will be described. In addition to the features described in the first to third embodiments, or instead, the fourth embodiment includes an example in which the transmitting and / or receiving communication devices learn latent vectors containing communication noise. In the fourth embodiment, an example will be described in which the first communication device 10 corresponds to the receiving communication device and the second communication device 20 corresponds to the transmitting communication device.

[0204] The first communication device 10 and the second communication device 20 may send and receive content over the network. For example, if the content includes images, the image data may be transmitted according to the Real-time Transport Protocol (RTP). Because image data has a large data size, the transmitting communication device often does not retransmit the image data even if a transmission error occurs, and the receiving communication device often tolerates the loss of images. RTP is a UDP-based protocol that tolerates the loss of images.

[0205] As illustrated above, content transmitted over a network may contain errors or omissions. These omissions and omissions can occur in the content during communication and can act as noise. Such noise may be referred to as communication noise. Communication noise may include additive white Gaussian noise (AWGN), Rayleigh fading, packet loss, quantization errors, frequency offset, and phase noise.

[0206] In this embodiment, a latent vector including communication noise that may occur during communication may be learned. By learning a latent vector including communication noise, for example, even if communication noise occurs in data transmitted and received between two communication devices under poor communication conditions, content can be accurately constructed from a latent representation including communication noise.

[0207] 4-2 Adding communication noise during the learning process In this embodiment, the first communication device 10 may add communication noise to the latent vector and learn the latent vector to which the communication noise has been added. As described above, the first communication device 10 may implement a learning model using LDM. The LDM may add spreading noise to the latent vector and learn the latent vector to which the spreading noise has been added. In this embodiment, the first communication device 10 may learn the latent vector to which the communication noise has been added during the learning process.

[0208] Referring to Figure 19, the logical configuration of the first communication device 10 according to the fourth embodiment will be described. The logical configuration of the first communication device 10 shown in Figure 19 may also be applied to the logical configuration of the second communication device 20. Therefore, the description of the logical configuration of the second communication device 20 will be omitted. As shown in Figure 19, the first communication device 10 may include a control unit 110, a communication unit 120, and a storage unit 130 as logical components.

[0209] In the configuration shown in Figure 19, the communication unit 120 and the storage unit 130 may be the same as the communication unit 120 and storage unit 130 shown in Figure 5. Also, the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d included in the control unit 110 may be the same as the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d shown in Figure 5. Furthermore, the decode execution determination unit 110e included in the control unit 110 may be the same as the decode execution determination unit 110e shown in Figure 15.

[0210] As shown in Figure 19, the control unit 110 according to the fourth embodiment may further include a communication noise application unit 110f. The communication noise application unit 110f may apply communication noise to the latent vector. The communication noise application unit 110f may be implemented in the semantic layer, as described in the second embodiment.

[0211] The communication noise addition unit 110f may add communication noise to the latent vector by discarding, for example, a portion of a packet containing content such as an image. According to the RTP described above, the transmitting side generates an RTP packet containing an image, and a sequence number is assigned to the RTP packet. When the receiving side receives the RTP packet, it reconstructs the image based on the sequence number. When the communication noise addition unit 110f collects content including an image by RTP, it may add communication noise to the latent vector by discarding a portion of the packet based on the sequence number. In other words, the communication noise addition unit 110f may add communication noise to the latent vector by discarding a portion of the data.

[0212] As mentioned above, discarding data based on the sequence number assigned to the RTP packet is merely an example. According to other protocols, and / or at other layers, some data may be discarded based on the sequence number or other factors. Furthermore, if a layer other than the semantic layer determines the sequence number, that layer may notify the semantic layer of the sequence number.

[0213] The methods for adding communication noise described above are merely illustrative. Communication noise may be added to latent vectors by other methods. Communication noise may also be added to data sampled using a mixture distribution. Sampling from a mixture distribution may be performed by a selection step from a categorical distribution and an extraction step from individual distributions.

[0214] Figure 20 illustrates the process of learning latent vectors using a latent spread model in LDM. As shown in Figure 20, the VAE encoder may output latent vectors by encoding an image into latent space. The latent vectors may be input to the LDM. In the LDM, a layered structure may be used to add communication noise and spreading noise to the latent vectors. Communication noise may be added in the first layer of the layered structure.

[0215] In the learning process shown in Figure 20, content collected from the internet may be used for learning. In this case, the first communication device 10 may receive content from the internet and output a latent vector based on the received content. In other words, the first communication device 10 may derive the latent vector through the network.

[0216] The content to be learned is not limited to content collected on the internet. For example, when the first communication device 10 interacts with the second communication device 20, it may receive content from the second communication device 20 and learn the received content. In this case, the first communication device 10 may output a latent vector from the content and add communication noise to the latent vector. Alternatively, the first communication device 10 may receive a latent vector from the second communication device 20 and add communication noise to the received latent vector. In either case, the first communication device 10 may derive a latent vector from the second communication device 20 via the network.

[0217] As described above, communication noise can occur during communication. In other words, the latent vector transmitted from the transmitting communication device may contain communication noise. When the receiving communication device constructs content based on such a latent vector, the constructed content will be affected by the communication noise. In this embodiment, the first communication device 10 may learn a latent vector to which communication noise has been added.

[0218] As described above, in LDM, the layered U-NET may remove spreading noise from the latent vector. As described above, LDM can remove communication noise in the despreading process by learning latent vectors to which communication noise has been added in addition to spreading noise. In this way, LDM can identify communication noise when constructing content. Therefore, the constructed content becomes less susceptible to the effects of communication noise.

[0219] Next, with reference to Figure 21, the process by which the first communication device 10 learns latent vectors will be described. In the example shown in Figure 21, the first communication device 10 may derive latent vectors from the second communication device 20 in the semantic layer, or it may learn the derived latent vectors. The semantic entity may implement a vector output unit 110a, a content generation unit 110b, a content construction unit 110c, a communication mode determination unit 110d, a decode execution determination unit 110e, and a communication noise addition unit 110f.

[0220] In step S2101, the semantic entity may receive data from a lower layer. The data received from the lower layer may include a semantic packet transmitted from the second communication device 20. Next, the semantic entity may extract a semantic packet from the data received in step S2101 (step S2102). The semantic packet may include a header area and a data area. The header area may include information set by the first communication device 10. The data area may include content and / or latent vectors set by the second communication device 20.

[0221] Next, the semantic entity may determine whether the content or latent vector is contained in the data area (step S2103). Whether the content or latent vector is contained in the data area may be determined based on information set in the header area of ​​the semantic packet. Whether the content or latent vector is contained in the data area may also be determined by whether any of the first to third communication modes are employed.

[0222] If it is determined in step S2103 that the latent vector is included in the data area, the latent vector may be output and the process may proceed to step S2105. If it is determined in step S2103 that the content is included in the data area, the semantic entity may output a latent vector from the content (step S2104). The latent vector may be output as described in the first embodiment.

[0223] Next, the semantic entity may add communication noise to the latent vector output in step S2104 (step S2105). Communication noise may be added by discarding a portion of the packet (a portion of the data). Next, the semantic entity may learn the latent vector to which communication noise was added in step S2105 (step S2106). Learning may be performed by adding spreading noise to the latent vector.

[0224] As described in Figure 21, the first communication device 10 may add communication noise to the latent vector based on the content / latent vector transmitted from the second communication device 20. The first communication device 10 may also learn the latent vector to which the communication noise has been added.

[0225] 4-3 Learning about communication noise in dialogue In the example described above, the first communication device 10 introduces communication noise during the learning process. Below, we will describe an example in which the first communication device 10 determines or predicts that communication noise has occurred when it interacts with the second communication device 20.

[0226] Referring to Figure 22, another logical configuration of the first communication device 10 according to the fourth embodiment will be described. The logical configuration of the first communication device 10 shown in Figure 22 may also be applied to the logical configuration of the second communication device 20. Therefore, the description of the logical configuration of the second communication device 20 will be omitted. As shown in Figure 22, the first communication device 10 may include a control unit 110, a communication unit 120, and a storage unit 130 as logical components.

[0227] In the configuration shown in Figure 22, the communication unit 120 and the storage unit 130 may be the same as the communication unit 120 and storage unit 130 shown in Figure 5. Also, the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d included in the control unit 110 may be the same as the vector output unit 110a, content generation unit 110b, content construction unit 110c, and communication mode determination unit 110d shown in Figure 5. Furthermore, the decode execution determination unit 110e included in the control unit 110 may be the same as the decode execution determination unit 110e shown in Figure 15.

[0228] As shown in Figure 22, the control unit 110 according to the fourth embodiment may further include a communication noise prediction unit 110g. The communication noise prediction unit 110g may predict that the content or latent vector transmitted from the second communication device 20 contains communication noise. The communication noise prediction unit 110g may be implemented in the semantic layer, as described in the second embodiment.

[0229] In this embodiment, an error correction code (ECC) or an error detection code (EDC) may be incorporated into the data transmitted and received between the first communication device 10 and the second communication device 20. The communication noise prediction unit 110g may detect, based on the ECC or EDC, whether or not an error has occurred in the data transmitted and received between the first communication device 10 and the second communication device 20. If the communication noise prediction unit 110g detects an error, it does not have to perform error correction. By not performing error correction, communication noise will be included in the content / latent vector transmitted from the second communication device 20.

[0230] Furthermore, the communication noise prediction unit 110g may detect whether or not a portion of the data transmitted from the second communication device 20 is missing. Data loss may be detected, for example, based on the sequence number assigned to the RTP packet when the data is transmitted and received according to RTP. Alternatively, data loss may be detected based on the sequence number assigned to the data according to a protocol other than RTP. If the communication noise prediction unit 110g detects data loss, it does not need to request retransmission of the data. By not requesting retransmission, communication noise will be included in the content / latent vector transmitted from the second communication device 20.

[0231] The error detection and / or data loss detection described above may be performed at layers other than the semantic layer. If an error / data loss is detected at a layer other than the semantic layer, that layer may notify the semantic layer of the error / data loss. The method for predicting communication noise described above is merely illustrative. Communication noise may be added to the latent vector by other methods.

[0232] If the first communication device 10 detects that communication noise has occurred in the content transmitted from the second communication device 20, it may output a latent vector from the content containing the communication noise, or it may learn the latent vector. Alternatively, if the first communication device 10 detects that communication noise has occurred in the latent vector transmitted from the second communication device 20, it may learn the latent vector containing the communication noise.

[0233] In the learning process described above, content collected from the internet may be used for learning. In this case, the first communication device 10 may receive content from the internet and may detect when communication noise occurs in the received content. When the first communication device 10 detects that communication noise has occurred in the content collected from the internet, it may output a latent vector based on the content containing the communication noise.

[0234] Next, with reference to Figure 23, the process by which the first communication device 10 learns latent vectors will be described. In the example shown in Figure 23, the first communication device 10 may derive latent vectors from the second communication device 20 in the semantic layer, or it may learn the derived latent vectors. The semantic entity may implement a vector output unit 110a, a content generation unit 110b, a content construction unit 110c, a communication mode determination unit 110d, a decode execution determination unit 110e, and a communication noise prediction unit 110g.

[0235] In step S2301, the semantic entity may receive data from a lower layer. The data received from the lower layer may include a semantic packet transmitted from the second communication device 20. Next, the semantic entity may extract a semantic packet from the data received in step S2301 (step S2302). The semantic packet may include a header area and a data area. The header area may include information set by the first communication device 10. The data area may include content and / or latent vectors set by the second communication device 20.

[0236] Next, the semantic entity may determine whether the content or latent vector is contained in the data area (step S2303). Whether the content or latent vector is contained in the data area may be determined based on information set in the header area of ​​the semantic packet. Whether the content or latent vector is contained in the data area may also be determined by whether any of the first to third communication modes are employed.

[0237] If it is determined in step S2303 that the latent vector is included in the data area, the latent vector may be output and the process may proceed to step S2305. If it is determined in step S2303 that the content is included in the data area, the semantic entity may output a latent vector from the content (step S2304). The latent vector may be output as described in the first embodiment.

[0238] Next, the semantic entity may detect that communication noise has occurred in the latent vector output in step S2304 (step S2305). Communication noise may be detected through error detection / packet loss (data loss) detection. Next, the semantic entity may learn a latent vector containing communication noise in step S2305 (step S2306). Learning may be performed by adding spreading noise to the latent vector.

[0239] As described in Figure 23, the first communication device 10 may detect that communication noise has occurred in the content / latent vector transmitted from the second communication device 20. Alternatively, the first communication device 10 may learn the latent vector containing the communication noise.

[0240] The fourth embodiment has been described above. According to the fourth embodiment, even if communication noise occurs in the data transmitted and received between two communication devices under poor communication conditions, content can be accurately constructed from latent representations that include the communication noise.

[0241] 5. Other Embodiments While embodiments of the disclosure have been described above, this disclosure is not limited to these embodiments. It will be understood by those skilled in the art that these embodiments are merely illustrative and that various modifications are possible without departing from the scope and spirit of this disclosure.

[0242] For example, the steps in the process described herein do not necessarily have to be executed chronologically in the order shown in the sequence diagram. For example, the steps in the process may be executed in a different order than that shown in the sequence diagram, or they may be executed in parallel. Also, some of the steps in the process may be deleted, or additional steps may be added to the process.

[0243] Furthermore, methods including processing the above-mentioned components may be provided, and programs for causing a processor to perform the processing of the above-mentioned components may be provided. A non-transitory computer-readable medium on which such a program is recorded may also be provided. Naturally, such devices, modules, methods, programs, and computer-readable non-transitory media are also included in this disclosure.

[0244] As described above, this embodiment can contribute to achieving the Sustainable Development Goals (SDGs) by achieving the effects described above. Specifically, this embodiment can contribute to achieving SDG Goal 9, "Build resilient infrastructure, promote inclusive and sustainable industrialization and foster technological innovation."

[0245] 6. Addendum Some or all of the above embodiments and modifications may also be described as follows, but are not limited to the contents of the following appendix. Hereinafter, a relationship is expressed in which appendixes that are subordinate to multiple appendixes are subordinate to appendixes that are subordinate to multiple appendixes. All of the appendix dependency relationships expressed below are included in the above embodiments.

[0246] Appendix Group A (Note A1) A first communication device, Communications Department and, It comprises a control unit that implements machine learning, The control unit, Based on the input information, output a latent vector. Based on the aforementioned latent vector, content is generated or retrieved. It is configured to determine whether to transmit the latent vector and / or the input information, or to transmit the content, The communication unit is configured to transmit the latent vector and / or the input information or content to the second communication device based on the decision. The first communication device.

[0247] (Appendix A2) The control unit, Determine the importance of the aforementioned content, Based on the aforementioned importance, it is determined whether to transmit the latent vector and / or the input information, or to transmit the content. The first communication device described in Appendix A1 is further configured as follows.

[0248] (Note A3) The first communication device as described in Appendix A2, wherein the control unit is further configured to determine the importance based on the content.

[0249] (Note A4) The first communication device according to Appendix A2, wherein the control unit is further configured to determine the importance based on the latent vector.

[0250] (Note A5) The first communication device as described in Appendix A2, wherein the control unit is further configured to determine the importance based on the input information.

[0251] (Note A6) The control unit, Based on the measurements, the communication quality is determined. Based on the aforementioned communication quality, it is determined whether to transmit the latent vector and / or the input information, or to transmit the content. The first communication device according to any one of Appendices A1 to A5, further configured as such.

[0252] (Appendix A7) The control unit is further configured to determine whether to transmit the potential vector and / or the input information or to transmit the content based on the input parameter, the first communication device according to any one of Appendices A1 to A6.

[0253] (Appendix A8) The input parameter indicates whether it is necessary to reproduce the content with high fidelity in the second communication device, the first communication device according to Appendix A7.

[0254] (Appendix A9) The input parameter indicates whether it is necessary to reduce the amount of data to be transmitted, the first communication device according to Appendix A7 or A8.

[0255] (Appendix A10) The input parameter is set for each device type, the first communication device according to any one of Appendices A7 or A9.

[0256] (Appendix A11) The communication unit is further configured to receive a response message from the second communication device, The control unit is further configured to determine whether to transmit the potential vector and / or the input information or to transmit the content based on the response message, The first communication device according to any one of Appendices A1 to A10.

[0257] (Appendix A12) The response message indicates whether it is necessary to reproduce the content with high fidelity in the second communication device, the first communication device according to Appendix A11.

[0258] (Appendix A13) The response message is the first communication device as described in Appendix A11 or A12, indicating whether or not it is necessary to reduce the amount of data to be transmitted.

[0259] (Note A14) The communication unit is further configured to receive a request message from the second communication device. The control unit is further configured to determine, based on the request message, whether to transmit the latent vector and / or the input information, or to transmit the content. The first communication device described in any one of the appendices A1 to A13.

[0260] (Note A15) The request message indicates whether or not the second communication device needs to reproduce the content with high fidelity, as described in Appendix A14, for the first communication device.

[0261] (Note A16) The request message is the first communication device as described in Appendix A14 or A15, indicating whether or not it is necessary to reduce the amount of data to be transmitted.

[0262] (Note A17) The control unit, Identify the first part and the second part of the aforementioned content, It is decided to transmit the content relating to the first part and the latent vector and / or input information relating to the second part. The first communication device as described in any one of the appendices A1 to A16, further configured as follows:

[0263] (Note A18) The control unit, The importance of the first part and the second part mentioned above is determined separately. Based on the aforementioned importance, it is decided to transmit the content relating to the first part and the latent vector and / or input information relating to the second part. The first communication device according to Appendix A17, which is further configured as described above.

[0264] (Appendix A19) The first communication device according to any one of Appendices A1 to A18, wherein the input information indicates condition information.

[0265] (Appendix A20) The input information indicates the first potential vector transmitted from the second communication device, and the control unit is further configured to output a second potential vector based on the first potential vector. The first communication device according to any one of Appendices A1 to A18.

[0266] (Appendix A21) A method executed by a first communication device, wherein the first communication device implements machine learning, outputs a potential vector based on input information, generates or acquires content based on the potential vector, determines whether to transmit the potential vector and / or the input information, or to transmit the content, transmits the potential vector and / or the input information, or the content, to a second communication device based on the determination, and includes a method.

[0267] (Appendix A22) A program in a first communication device, wherein the first communication device implements machine learning, and when the program is executed, it causes a processor in the first communication device to output a potential vector based on input information, generate or acquire content based on the potential vector, determine whether to transmit the potential vector and / or the input information, or to transmit the content, Based on the above decision, the latent vector and / or the input information, or the content, is transmitted to the second communication device. A program that executes something.

[0268] (Note A23) A computer-readable non-temporary recording medium storing a program in a first communication device, wherein the first communication device implements machine learning, and the program, when executed, is processed by the processor in the first communication device. Based on the input information, output a latent vector, Based on the aforementioned latent vector, generate or retrieve content, To determine whether to transmit the latent vector and / or the input information, or to transmit the content, Based on the above decision, the latent vector and / or the input information, or the content, is transmitted to the second communication device. A non-temporary recording medium that enables execution.

[0269] (Note A24) A second communication device, Communications Department and, It comprises a control unit that implements machine learning, The communication unit is configured to transmit latent vectors and / or information or content input to the first communication device from the first communication device. The control unit, When the latent vector and / or the input information is received, content is constructed based on the latent vector and / or the input information, and the constructed content is output. The system is configured to output the received content when it receives the aforementioned content. The second communication device.

[0270] (Note A25) A method performed by a second communication device, wherein the second communication device implements machine learning. Transmitting latent vectors and / or information or content input to the first communication device from the first communication device, When the latent vector and / or the input information is received, content is constructed based on the latent vector and / or the input information, and the constructed content is output. When the aforementioned content is received, the received content is output, Methods that include...

[0271] (Note A26) A program for a second communication device, wherein the second communication device implements machine learning, and when the program is executed, the processor in the second communication device, Transmitting latent vectors and / or information or content input to the first communication device from the first communication device, When the latent vector and / or the input information is received, content is constructed based on the latent vector and / or the input information, and the constructed content is output. When the aforementioned content is received, the received content is output, A program that executes something.

[0272] (Note A27) A computer-readable non-temporary recording medium storing a program in a second communication device, wherein the second communication device implements machine learning, and the program, when executed, is transmitted to the processor in the second communication device. Transmitting latent vectors and / or information or content input to the first communication device from the first communication device, When the latent vector and / or the input information is received, content is constructed based on the latent vector and / or the input information, and the constructed content is output. When the aforementioned content is received, the received content is output, A non-temporary recording medium that enables execution.

[0273] Appendix B group (Note B1) A first communication device, Communications Department and, It comprises a control unit that implements machine learning, In order to communicate with a second communication device in a hierarchical structure, in a predetermined first layer within the hierarchical structure, The control unit, It is configured to receive data from higher layers. Based on the information entered, generate or retrieve content. A packet is generated that includes control information for transmitting the content to the second communication device and the data. The communication unit is configured to transmit the packet to the second layer of the second communication device, which corresponds to the first layer. The first communication device.

[0274] (Note B2) The first communication device as described in Appendix B1, wherein the control unit is further configured to pass the packets to a lower layer.

[0275] (Note B3) The control unit, Based on the input information, output a latent vector. The latent vector and / or the input information are set in the packet. The first communication device as described in Appendix B1 or B2, further configured as follows.

[0276] (Note B4) The control unit, Based on the information entered, generate or retrieve content. The content is set in the packet. The first communication device as described in Appendix B1 or B2, further configured as follows.

[0277] (Note B5) The control unit, Based on the input information, output a latent vector. Based on the aforementioned latent vector, content is generated or retrieved. The latent vector and / or the input information are set in the packet, or the content is set in the packet. The first communication device as described in either Appendix B1 or B4, further configured as follows:

[0278] (Note B6) The control unit sets the second layer as the control information, wherein the control unit sets the latent vector and / or the input information, or information indicating whether to transmit the content, as the control information. This is the first communication device as described in Appendix B5.

[0279] (Note B7) The control unit, Determine the importance of the aforementioned content, Based on the importance, the latent vector and / or the input information is set in the packet, or the content is set in the packet. The first communication device described in Appendix B6 is further configured as follows.

[0280] (Note B8) The first communication device according to Appendix B7, wherein the control unit is further configured to determine the importance based on the content.

[0281] (Note B9) The first communication device according to Appendix B7, wherein the control unit is further configured to determine the importance based on the latent vector.

[0282] (Note B10) The first communication device as described in Appendix B7, wherein the control unit is further configured to determine the importance based on the input information.

[0283] (Note B11) The control unit sets the importance as the control information, and is the first communication device as described in any one of the appendices B7 to B10.

[0284] (Note B12) The control unit, Based on the measurements, the communication quality is determined. Based on the aforementioned communication quality, the latent vector and / or the input information is set in the packet, or the content is set in the packet. The first communication device as described in any one of the appendices B6 to B11, further configured as follows:

[0285] (Note B13) The first communication device according to any one of the appendices B6 to B12, wherein the control unit is further configured to set the latent vector and / or the input information into the packet, or to set the content into the packet, based on the input parameters.

[0286] (Note B14) The input parameters indicate whether or not the second communication device needs to reproduce the content with high fidelity, as described in Appendix B13 for the first communication device.

[0287] (Note B15) The input parameter indicates whether or not it is necessary to reduce the amount of data to be transmitted, as described in Appendix B13 or B14 of the first communication device.

[0288] (Note B16) The input parameters are set for each device type, and the first communication device is as described in either Appendix B13 or B15.

[0289] (Note B17) The communication unit is further configured to receive response packets from the corresponding layer. The control unit is further configured to determine, based on the response packet, whether to transmit the latent vector and / or the input information or the content to the second layer. The first communication device described in Appendix B5.

[0290] (Note B18) The communication unit is further configured to receive request packets from the corresponding layer. The control unit is further configured to determine, based on the request packet, whether to transmit the latent vector and / or the input information or the content to the second layer. The first communication device described in Appendix B5.

[0291] (Note B19) The input information is a first communication device described in any one of the appendices B1 to B18, indicating conditional information.

[0292] (Note B20) The input information indicates the first latent vector transmitted from the second communication device. The first communication device according to any one of the appendices B1 to B18, wherein the control unit is further configured to output a second latent vector based on the first latent vector.

[0293] (Note B21) A method performed by a first communication device, wherein the first communication device implements machine learning. In order to communicate with a second communication device in a hierarchical structure, in a predetermined first layer within the hierarchical structure, Receiving data from higher layers, To generate or retrieve content based on the input information, To generate a packet containing control information for transmitting the content to the second communication device and the data, The packet is transmitted to the second layer of the second communication device, which corresponds to the first layer. Methods that include...

[0294] (Note B22) A program for a first communication device, wherein the first communication device implements machine learning, and when the program is executed, it is performed on the processor of the first communication device. In order to communicate with a second communication device in a hierarchical structure, in a predetermined first layer within the hierarchical structure, Receiving data from higher layers, To generate or retrieve content based on the input information, To generate a packet containing control information for transmitting the content to the second communication device and the data, The packet is transmitted to the second layer of the second communication device, which corresponds to the first layer. A program that executes something.

[0295] (Note B23) A computer-readable non-temporary recording medium storing a program in a first communication device, wherein the first communication device implements machine learning, and the program, when executed, is processed by the processor in the first communication device. In order to communicate with a second communication device in a hierarchical structure, in a predetermined first layer within the hierarchical structure, Receiving data from higher layers, To generate or retrieve content based on the input information, To generate a packet containing control information for transmitting the content to the second communication device and the data, The packet is transmitted to the second layer of the second communication device, which corresponds to the first layer. A non-temporary recording medium that enables execution.

[0296] (Note B24) A second communication device, Communications Department and, It comprises a control unit that implements machine learning, In order to communicate with the first communication device in a hierarchical structure, in a predetermined second layer within the hierarchical structure that corresponds to the first layer in the first communication device, The communication unit is configured to receive packets from the first layer, The control unit is configured to output a latent vector or content based on the control information and data contained in the packet. The second communication device.

[0297] (Note B25) A method performed by a second communication device, wherein the second communication device implements machine learning. In order to communicate with the first communication device in a hierarchical structure, in a predetermined second layer within the hierarchical structure that corresponds to the first layer in the first communication device, Receiving packets from the aforementioned first layer, The control unit outputs a latent vector or outputs content based on the control information and data contained in the packet. Methods that include...

[0298] (Note B26) A program for a second communication device, wherein the second communication device implements machine learning, and when the program is executed, the processor in the second communication device, In order to communicate with the first communication device in a hierarchical structure, in a predetermined second layer within the hierarchical structure that corresponds to the first layer in the first communication device, Receiving packets from the aforementioned first layer, The control unit outputs a latent vector or outputs content based on the control information and data contained in the packet. A program that executes something.

[0299] (Note B27) A computer-readable non-temporary recording medium storing a program in a second communication device, wherein the second communication device implements machine learning, and the program, when executed, is transmitted to the processor in the second communication device. In order to communicate with a second communication device in a hierarchical structure, in a predetermined first layer within the hierarchical structure, Receiving data from higher layers, To generate or retrieve content based on the input information, To generate a packet containing control information for transmitting the content to the second communication device and the data, The packet is transmitted to the second layer of the second communication device, which corresponds to the first layer. A non-temporary recording medium that enables execution.

[0300] Appendix C group (Note C1) A first communication device, Communications Department and, It comprises a control unit that implements machine learning, The control unit, Based on the input information, output a latent vector. In the second communication device, it is determined whether or not to construct content based on the latent vector. The communication unit is configured to transmit the latent vector and / or the input information or content to the second communication device based on the decision. The first communication device.

[0301] (Note C2) The first communication device as described in Appendix C1, wherein the communication unit is further configured to transmit to the second communication device a packet containing information indicating whether or not the second communication device constructs the content.

[0302] (Note C3) If the control unit decides not to construct the content in the second communication device, the communication unit is further configured to transmit the latent vector and / or the input information to the second communication device, as described in Appendix C1 or C2.

[0303] (Note C4) The first communication device according to any one of the appendices C1 to C3, wherein the control unit is further configured to determine whether or not to construct the content in the second communication device based on whether or not a human is involved in the dialogue between the first communication device and the second communication device.

[0304] (Note C5) The first communication device according to any one of the appendices C1 to C3, wherein the control unit is further configured by default to decide not to construct the content in the second communication device.

[0305] (Appendix C6) The first communication device according to any one of the appendices C1 to C3, wherein the control unit is further configured to determine whether or not to construct the content in the second communication device based on pre-set conditions.

[0306] (Note C7) The first communication device according to any one of the appendices C1 to C3, wherein the control unit is further configured to determine whether or not to construct the content in the second communication device based on the input information.

[0307] (Note C8) The communication unit is further configured to receive response packets from the second communication device. The control unit is further configured to determine whether or not to construct the content in the second communication device based on the response packet. The first communication device described in any one of the appendices C1 to C4.

[0308] (Note C9) The communication unit is further configured to receive request packets from the second communication device. The control unit is further configured to determine whether or not to construct the content in the second communication device based on the request packet. The first communication device described in any one of the appendices C1 to C4.

[0309] (Note C10) The first communication device according to any one of the appendices C1 to C9, wherein the control unit is further configured to determine whether to transmit the latent vector and / or the input information, or the content, when the second communication device decides to construct the content.

[0310] (Note C11) The first communication device according to any one of the appendices C1 to C10, wherein the control unit is further configured to operate in a decodeless mode, which communicates the latent vector and / or the input information without decoding it into human-perceptible content, when the second communication device decides not to construct the content.

[0311] (Note C12) A method performed by a first communication device, wherein the first communication device implements machine learning. Based on the input information, output a latent vector, In the second communication device, it is determined whether or not to construct content based on the latent vector, Based on the above decision, the latent vector and / or the input information, or the content, is transmitted to the second communication device. Methods that include...

[0312] (Note C13) A program for a first communication device, wherein the first communication device implements machine learning, and when the program is executed, it is performed on the processor of the first communication device. Based on the input information, output a latent vector, In the second communication device, it is determined whether or not to construct content based on the latent vector, Based on the above decision, the latent vector and / or the input information, or the content, is transmitted to the second communication device. A program that executes something.

[0313] (Note C14) A computer-readable non-temporary recording medium storing a program in a first communication device, wherein the first communication device implements machine learning, and the program, when executed, is processed by the processor in the first communication device. Based on the input information, output a latent vector, In the second communication device, it is determined whether or not to construct content based on the latent vector, Based on the above decision, the latent vector and / or the input information, or the content, is transmitted to the second communication device. A non-temporary recording medium that enables execution.

[0314] (Note C15) A second communication device, Communications Department and, It comprises a control unit that implements machine learning, The communication unit is configured to receive a latent vector from the first communication device. The control unit is configured to determine whether or not to construct content based on the latent vector. The second communication device.

[0315] (Note C16) The communication unit is further configured to receive information from the first communication device indicating whether or not to construct the content. The control unit is further configured to determine whether or not to construct the content based on the latent vector, based on the information. The second communication device described in Appendix C15.

[0316] (Note C17) The aforementioned latent vector is the first latent vector, The second communication device according to either Appendix C15 or C16, wherein if the control unit decides not to construct the content, the communication unit is further configured to output a second latent vector based on the first latent vector.

[0317] (Note C18) The second communication device according to any one of the appendices C15 to C17, wherein the control unit is further configured to determine whether or not to construct the content based on whether or not a human is involved in the dialogue between the first communication device and the second communication device.

[0318] (Note C19) The second communication device as described in any one of the appendices C15 to C17, wherein the control unit is further configured by default to decide not to build the content.

[0319] (Note C20) The second communication device according to any one of the appendices C15 to C17, wherein the control unit is further configured to determine whether or not to construct the content based on pre-set conditions.

[0320] (Note C21) The second communication device according to any one of the appendices C15 to C17, wherein the control unit is further configured to determine whether or not to construct the content based on the input information.

[0321] (Note C22) The second communication device according to any one of the appendices C15 to C17, wherein the communication unit is further configured to transmit a response packet to the first communication device indicating whether or not to construct the content.

[0322] (Note C23) The second communication device according to any one of the appendices C15 to C17, wherein the communication unit is further configured to transmit a request packet to the first communication device indicating whether or not to construct the content.

[0323] (Note C24) The second communication device according to any one of the appendices C15 to C23, wherein the control unit is further configured to determine whether or not to construct the content based on the latent vector when it has decided to construct the content.

[0324] (Note C25) The second communication device according to any one of Appendix C15 to C24, wherein the control unit is further configured to operate in a decodeless mode, which communicates the latent vector without decoding it into human-perceptible content, if it is decided not to construct the content.

[0325] (Note C26) A method performed by a second communication device, wherein the second communication device implements machine learning. Receiving a latent vector from the first communication device, The decision of whether or not to construct content based on the aforementioned latent vector, Methods that include...

[0326] (Note C27) A program for a second communication device, wherein the second communication device implements machine learning, and when the program is executed, the processor in the second communication device, Receiving a latent vector from the first communication device, The decision of whether or not to construct content based on the aforementioned latent vector, A program that executes something.

[0327] (Note C28) A computer-readable non-temporary recording medium storing a program in a second communication device, wherein the second communication device implements machine learning, and the program, when executed, is transmitted to the processor in the second communication device. Receiving a latent vector from the first communication device, The decision of whether or not to construct content based on the aforementioned latent vector, A non-temporary recording medium that enables execution.

[0328] Appendix Group D (Note D1) A first communication device, Communications Department and, It comprises a control unit that implements machine learning, The aforementioned communication unit is configured to receive data via the network. The control unit, Based on the above data, a latent vector including communication noise is derived. The system is configured to learn the aforementioned latent vector. The first communication device.

[0329] (Note D2) The first communication device according to Appendix D1, wherein the control unit is further configured to apply the communication noise to the latent vector.

[0330] (Note D3) The first communication device according to Appendix D2, wherein the control unit is further configured to add the communication noise to the latent vector by discarding a portion of the data.

[0331] (Note D4) The first communication device according to Appendix D1, wherein the control unit is further configured to detect when communication noise is present in the data.

[0332] (Note D5) The first communication device as described in Appendix D4, wherein the control unit is further configured to detect, by error detection, that communication noise is occurring in the data.

[0333] (Note D6) The first communication device according to Appendix D5, wherein the control unit is further configured not to perform error detection on the data in response to the error detection.

[0334] (Note D7) The first communication device as described in Appendix D4, wherein the control unit is further configured to detect when communication noise is present in the data by detecting data loss.

[0335] (Note D8) The first communication device according to Appendix D7, wherein the control unit is further configured not to request retransmission of the data in response to the detection of the data loss.

[0336] (Note D9) The first communication device according to any one of the appendices D1 to D8, wherein the control unit is further configured to derive the latent vector by having the communication unit receive content through a network and the control unit outputting the latent vector based on the content.

[0337] (Note D10) The communication unit is further configured to receive the data from the second communication device, The first communication device is configured to interact with the second communication device. The first communication device described in any one of the appendices D1 to D8.

[0338] (Note D11) The control unit, Determine whether the aforementioned data contains content, If it is determined that the data contains the content, the latent vector is output based on the content. The first communication device described in Appendix D10, further configured to derive the aforementioned latent vector by means of the above.

[0339] (Note D12) The control unit, Determine whether the latent vector is included in the aforementioned data. If it is determined that the packet contains the latent vector, the latent vector is output. The first communication device described in Appendix D10, further configured to derive the aforementioned latent vector by means of the above.

[0340] (Note D13) The first communication device according to any one of the appendices D1 to D12, wherein the control unit is further configured to generate content by removing the communication noise from the latent vector.

[0341] (Note D14) The communication noise includes one or more of the following: additive white Gaussian noise, Rayleigh fading, packet loss, quantization error, frequency offset, and phase noise, as described in any one of the first communication device described in any one of the appendices D1 to D13.

[0342] (Note D15) A method performed by a first communication device, wherein the first communication device implements machine learning. Receiving data over a network, Based on the aforementioned data, a latent vector including communication noise is derived, Learning the aforementioned latent vector, Methods that include...

[0343] (Note D16) A program for a first communication device, wherein the first communication device implements machine learning, and when the program is executed, it is performed on the processor of the first communication device. Receiving data over a network, Based on the aforementioned data, a latent vector including communication noise is derived, Learning the aforementioned latent vector, A program that executes something.

[0344] (Note D17) A computer-readable non-temporary recording medium storing a program in a first communication device, wherein the first communication device implements machine learning, and the program, when executed, is processed by the processor in the first communication device. Receiving data over a network, Based on the aforementioned data, a latent vector including communication noise is derived, Learning the aforementioned latent vector, A non-temporary recording medium that enables execution. [Explanation of Symbols]

[0345] 10 First communication device 20 Second communication device 101 Processors 102 memory 104 Transmitter / Receiver 105 Antenna 110 Control Unit 120 Communications Department 201 Processor 202 memory 204 Transmitter / Receiver 205 Antenna 210 Control Unit 220 Communications Department

Claims

1. A first communication device, Communications Department and, It comprises a control unit that implements a machine learning model, The aforementioned communication unit is configured to receive data via the network. The control unit, The system detects that a first noise is present in the data, and that this first noise is noise generated during the communication of the data. By detecting the first noise, a latent vector containing the first noise is derived based on the data. The machine learning model is configured to learn to add a second noise to the latent vector. The first communication device.

2. The first communication device according to claim 1, wherein the control unit is further configured to detect, by error detection, that the first noise is occurring in the data.

3. The first communication device according to claim 2, wherein the control unit is further configured not to perform error correction on the data in response to the error detection.

4. The first communication device according to claim 1, wherein the control unit is further configured to detect that the first noise is occurring in the data by detecting data loss.

5. The first communication device according to claim 4, wherein the control unit is further configured not to request retransmission of the data in response to the detection of the data loss.

6. A method performed by a first communication device, wherein the first communication device implements a machine learning model. Receiving data over a network, The method involves detecting that a first noise is present in the data, wherein the first noise is noise generated during the communication of the data. By detecting the first noise, a latent vector containing the first noise is derived based on the data, The machine learning model is trained to add a second noise to the latent vector, Methods that include...

7. A program for a first communication device, wherein the first communication device implements a machine learning model, and when the program is executed, it is sent to the processor in the first communication device. Receiving data over a network, The method involves detecting that a first noise is present in the data, wherein the first noise is noise generated during the communication of the data. By detecting the first noise, a latent vector containing the first noise is derived based on the data, The machine learning model is trained to add a second noise to the latent vector, A program that executes something.

Images