Sending device and method, receiving device and method, and corresponding network for semantic communication

The distributed network with task-oriented mapping and semantic hashing functions addresses inefficiencies in detecting transmission errors in semantic communication, enabling efficient error detection and reducing re-transmissions by utilizing a K-level hierarchy of features and adaptive thresholds.

WO2025185819A1PCT designated stage Publication Date: 2025-09-11HUAWEI TECH CO LTD +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2024/055804
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-03-06
Publication Date
2025-09-11

AI Technical Summary

Technical Problem

Existing methods for detecting transmission errors in semantic communication are inefficient, leading to high rejection rates and increased re-transmissions due to the inability to classify errors as relevant or irrelevant to a specific task, and are computationally intensive, making them impractical for real-time processing and low-energy sensors.

Method used

A distributed network comprising a sending device and a receiving device that utilizes task-oriented mapping, specifically semantic hashing functions, to determine a semantic key based on a K-level hierarchy of features, enabling efficient detection of transmission errors by comparing the semantic key with a check semantic key, and applying an adaptive threshold for acknowledgement responses.

Benefits of technology

The solution allows for efficient detection of relevant transmission errors, reducing re-transmissions and computational overhead, while maintaining communication efficiency comparable to conventional CRC-based error detection, and is applicable to various classes of input data and tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2024055804_12092025_PF_FP_ABST
    Figure EP2024055804_12092025_PF_FP_ABST
Patent Text Reader

Abstract

A distributed network including a sending device and a receiving device, where the sending device configured to obtain data to be transmitted, where the data to be transmitted is of an input class and relates to a task to be performed. The sending device is configured to determine a semantic key by applying a task-oriented mapping to the data, and transmit the data and the semantic key to the receiving device. The receiving device configured to receive the data and the semantic key, determine a check semantic key by applying the same task-oriented mapping to the received data, and compare the check semantic key to the semantic key in order to determine if there is a transmission error. The distributed network enables an efficient detection of the relevant transmission errors in the semantic communications by virtue of using the task-oriented mapping.
Need to check novelty before this filing date? Find Prior Art

Description

[0001]DISTRIBUTED NETWORK, SENDING DEVICE, RECEIVING DEVICE, AND METHODS FOR USE INDISTRIBUTED NETWORK TECHNICAL FIELDThe present disclosure relates generally to the field of wireless communication; and more specifically, to a distributed networkcomprising a sending device and a receiving device, a sending device, a receiving device, and methods for use in the distributed network comprising the sending device and the receiving device, the sending device, and the receiving device, respectively. BACKGROUND In contemporary technological landscapes, numerous applications and services, spanning a wide range of domains, such as robotics, autonomous driving, traffic management, and smart factories, are increasingly reliant on advanced techniques like object recognition and computer vision. In such applications and services, a central computing device is used to collect sensed data from diverse distributed sensors and perform complex tasks, such as security monitoring, navigation, or traffic prediction. However, this process encounters significant technical challenges due to the burgeoning volume and intricacy of sensor data alongside privacy constraints, and renders direct transmission of all sensor data to the central computing device. To address these challenges, a promising solution involves processing the sensor data at a semantic level, emphasizing the intended meaning and utility of the sensor data over the exact representation of the sensor data. The semantic communication, where semantic interpretations of sensor data are transmitted instead of raw sensor data, holds the potential to substantially reduce thecommunication costs and latency. However, the semantic communication introduces novel challenges (e.g., transmissionerrors), which are not encountered in data-driven communication, such as defining semantic meaning and ensuring accurate interpretation of a received semantic message amidst channel noise-induced “semantic error”, or “semantic transmission error”. Unlike data-driven communication, where transmission errors are considered as distortions in the data’s representation, the semantic transmission errors can alter the underlying meaning of the received semantic message, and therefore, generate transmission error detection challenges. Currently, certain attempts have been made in order to detect transmission errors in the semantic communication, for example, in an approach, a semantic message is considered akin to a conventional binary message and traditional error-detection techniques, like cyclic redundancy check (CRC) parity bits or error-correction codes are employed. This approach is effective at detecting transmission errors, however lacks the ability to classify the transmission errors as relevant or irrelevant to a specifictask, resulting in high rejection rate and increased re-transmissions. In another approach, a semantic analysis is used, in whicha received semantic message is checked for coherence with background knowledge and past messages. This approach is capableof detecting inconsistencies however, susceptible to error misdetection when an erroneous message is consistent with thebackground knowledge or with the past messages. Furthermore, the semantic analysis is computationally intensive hence, impractical for applications involving real-time processing and low-energy sensors. In a yet another approach, a combination of aforementioned two approaches is used, in which a large semantic message is split into little portions and sent progressively to a receiver. Upon reception, the receiver evaluates semantically the received data and acknowledges each portion of data. Such progressive message transmission offers some advantages, for example, inference on a partial message, partial message re-transmission, but still remains highly impractical due to several transmission rounds for transmitting a single message.Thus, there exists a technical problem of inefficient detection of relevant transmission errors in semantic communications.Therefore, in light of the foregoing discussion, there exists a need to overcome the aforementioned drawbacks associated withthe conventional methods of detecting transmission errors in semantic communications. SUMMARYThe present disclosure provides a distributed network comprising a sending device and a receiving device, a method for thedistributed network comprising the sending device and the receiving device, a sending device, a method for the sending device,a receiving device, and a method for the receiving device. The present disclosure provides a solution to the existing problemof inefficient detection of relevant transmission errors in semantic communications. An aim of the present disclosure is toprovide a solution that overcomes at least partially the problems encountered in prior art, and provide an improved distributednetwork comprising a sending device and a receiving device, an improved method for the distributed network comprising thesending device and the receiving device, an improved sending device, an improved method for the sending device, an improvedreceiving device, and an improved method for the receiving device. The object of the present disclosure is achieved by the solutions provided in the enclosed independent claims. Advantageous implementations of the present disclosure are further defined in the dependent claims.In one aspect, the present disclosure provides a distributed network comprising a sending device and a receiving device, wherethe sending device comprising a sending controller configured to obtain data to be transmitted, where the data to be transmittedis of an input class and relates to a task to be performed, where the input class and the task to be performed are associated witha K-level hierarchy of features. The sending controller is further configured to determine a semantic key by applying a task-oriented mapping to the data, where the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmit the data to be transmitted and the semantic key to the receiving device. The receiving devicecomprises a receiving controller configured to receive the data and the semantic key, determine a check semantic key byapplying the same task-oriented mapping to the received data, and compare the check semantic key to the semantic key in orderto determine if there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data. The receiving controller 110 is further configured to determine the level of thetransmission error based on a K-level of the feature for which the mismatch is detected, and then determine that the data issuccessfully received if the level of the transmission error is below a relevance threshold level and send an acknowledgementresponse, ACK, in response thereto, and determine that the data is not successfully received if the level of the transmissionerror is above the relevance threshold level and send a negative acknowledgement response, NACK, in response thereto. The disclosed distributed network comprising the sending device and the receiving device enables an efficient detection of the relevant transmission errors in the semantic communications by virtue of using the task-oriented mapping that is semantic hashing functions. Moreover, the distributed network manifests computational and communication efficiency proportional to conventional CRC-based error detection approach. The semantic hashes are much smaller than initial data representation andhence, the semantic hashes introduce little communication overhead. Concurrently, the distributed network enables thedetection of relevant transmission errors for the given task and also, applies an adaptive threshold for accepting partially correct messages hence, reducing re-transmissions. Additionally, the distributed network utilizing the semantic hashing functions does not require any complex semantic analysis.In an implementation form, the receiving controller is further configured to compute a weighted L1 distance between thereceived semantic hash and the check semantic key as the level of the transmission error and determine that the level of the transmission error is above the relevance threshold level if the weighted L1 distance is larger than the relevance threshold level. The computation of the L1 distance between the received semantic hash and the check semantic key does not depend on the semantic task anymore and hence, the semantic hash is applicable on any class of input data and any specific task.In a further implementation form, the K-level hierarchy of features is based on a training of a neural network utilized to learnthe most relevant features of the K-level hierarchy of features. This is advantageous to learn the most relevant features based on the training of the neural network in terms of computational efficiency because in some scenarios, the specific task to be solved is quite difficult to model analytically or the underlying model is unknown.In a further implementation form, the K-level hierarchy of features is based on a training of a neural network, where the neuralnetwork is a randomly initialized neural network having an encoder / decoder architecture, wherein a role of the encoder is to map an input vector into a latent feature vector of dimension K where each dimension corresponds to one feature, and a role ofthe decoder is to predict the true label from a latent vector, and where the encoder is configured to receive an input vectorprovide a K-dimensional feature vector to which a random mask is applied which randomly selects k < K and masks all lowerk features of the feature vector, where the decoder is configured to receive the resulted vector and provide a prediction of thetrue label, whereby an error vector is back-propagated to the neural network, whereby weights are updated. The learning of the K-level hierarchy of features based on the training of the neural network is advantageous to reduce the dataredundancy and to obtain the relevant information.In a further implementation form, the random mask is configured to erase information from lowest features and preservinginformation in top features which encourages the encoder to put more relevant information in top features.The preservation of relevant information in top features and erasure of the information in lowest features results in a reductionof the communication overhead. In a further implementation form, the task-oriented mapping is based on receiving input data and outputting the semantic key being a short code of a predefined structure, wherein the short code depends on the input data’s relevant features from the K- level hierarchy of features. The utilization of the task-oriented mapping is efficient in terms of computational and communication overhead.In a further implementation form, the task-oriented mapping is based on a learning using a neural network, utilizingdiscriminative learning, where a discriminative L1 loss is utilized to encourage the neural network to assign close codes tosimilar features while assigning distant codes to dissimilar features in a following way: vectors with same level-1 featuresshould be close to each other (and remaining far apart), and among these vectors, the vectors with same level-2 features shouldbe close to each other.This is advantageous in terms of providing huge flexibility for learning the task-oriented mapping using the neural networkutilizing the discriminative learning. In another aspect, the present disclosure provides a method for a distributed network comprising a sending device and areceiving device, the method comprising: the sending device obtaining data to be transmitted, wherein the data to be transmittedis of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associatedwith a K-level hierarchy of features, determining a semantic key by applying a task-oriented mapping to the data, wherein thetask-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmitting thedata to be transmitted and the semantic key to the receiving device. The method further comprises the receiving device receivingthe data and the semantic key, determining a check semantic key by applying the same task-oriented mapping to the receiveddata, and comparing the check semantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data. The methodfurther comprising the receiving device determining the level of the transmission error based on a K-level of the feature forwhich the mismatch is detected, and then determining that the data is successfully received if the level of the transmission error is below a relevance threshold level and sending an acknowledgement response, ACK, in response thereto, and determining that the data is not successfully received if the level of the transmission error is above the relevance threshold level and sending a negative acknowledgement response, NACK, in response thereto. The use of the disclosed method enables an efficient detection of the relevant transmission errors in the semantic communications by virtue of using the task-oriented mapping that is semantic hashing functions. Moreover, the computationaland communication efficiency proportional to the conventional CRC-based error detection approach is obtained by utilizingthe method. Concurrently, the detection of relevant transmission errors for the given task is obtained on using the disclosedmethod. Moreover, an adaptive threshold is applied for accepting partially correct messages hence, reducing re-transmissions.Additionally, by virtue of utilizing the semantic hashing functions, the method does not include any complex semantic analysis.In a yet another aspect, the present disclosure provides a sending device comprising a sending controller configured to obtaindata to be transmitted, where the data to be transmitted is of an input class and relates to a task to be performed, wherein theinput class and the task to be performed are associated with a K-level hierarchy of features, determine a semantic key byapplying a task-oriented mapping to the data, wherein the task-oriented mapping provides a mapping of the data to the semantickey based on K-level hierarchy of, and transmit the data to be transmitted and the semantic key to the receiving device.This is advantageous to apply the task-oriented mapping to the data to reduce the communication and computational overheadat the sending device. Moreover, the use of the task-oriented mapping (i.e., semantic hashing function) at the sending device eliminates the requirement of semantic analysis.In a yet another aspect, the present disclosure provides a method for use in a sending device, the method comprising obtainingdata to be transmitted, where the data to be transmitted is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, determining a semantic key by applying a task-oriented mapping to the data, wherein the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmitting the data to be transmitted and the semantic key to the receiving device. The method achieves all the advantages and technical features of the sending device after execution.In a yet another aspect, the present disclosure provides receiving device comprising a receiving controller configured to receivedata and a semantic key, where the data received is of an input class and relates to a task to be performed, where the input class and the task to be performed are associated with a K-level hierarchy of features, and where the semantic key is determined by applying a task-oriented mapping to the prior to transmission data, where the task-oriented mapping provides a mapping of thedata to the semantic key based on K-level hierarchy of features. The receiving controller is further configured to determine acheck semantic key by applying the same task-oriented mapping to the received data, and compare the check semantic key tothe semantic key in order to determine if there is a transmission error by determining that there is a mismatch between thesemantic key and the check semantic key for a feature of the data, and if so. The receiving controller is further configured to determine the level of the transmission error based on a K-level of the feature for which the mismatch is detected, and then determine that the data is successfully received if the level of the transmission error is below a relevance threshold level and send an acknowledgement response, ACK, in response thereto, and determine that the data is not successfully received if the level of the transmission error is above the relevance threshold level and send a negative acknowledgement response, NACK, in response thereto. This is advantageous to apply the same task-oriented mapping to the received data to detect the relevant transmission errors in the received data. Moreover, the use of the task-oriented mapping (i.e., semantic hashing function) enables the receiving deviceto decide which portion of the received data is analysed for future use and which portion of the received data is required to be re-transmitted. The use of the task-oriented mapping reduces the communication and computational overhead at the receiving device.In a yet another aspect, the present disclosure provides method for use in a receiving device, the method comprising receivingdata and a semantic key, where the data received is of an input class and relates to a task to be performed, where the input class and the task to be performed are associated with a K-level hierarchy of features, and where the semantic key is determined by applying a task-oriented mapping to the prior to transmission data, where the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of features. The method further comprises determining a check semantic key by applying the same task-oriented mapping to the received data and comparing the check semantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so. The method further comprises determining the level of the transmission error based on a K-level of the feature for which the mismatch is detected, and then determining that the data is successfully received if the level of the transmission error is below a relevance threshold level and sending an acknowledgement response, ACK, in response thereto, and determining that the data is not successfully received if the level of the transmission error is above the relevance threshold level and sending a negative acknowledgement response, NACK, in response thereto.The method achieves all the advantages and technical features of the receiving device after execution.It is to be appreciated that all the aforementioned implementation forms can be combined. It has to be noted that all devices, elements, circuitry, units and means described in the present application could be implemented in the software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionalities described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Even if, in the following description of specific embodiments, a specific functionality or step to be performed by external entities is not reflected in the description of a specific detailed element of that entity which performs that specific step or functionality, it should be clear for a skilled person that these methods and functionalities can be implemented in respective software or hardware elements, or any kind of combination thereof. It will be appreciated that features of the present disclosure are susceptible to being combined in various combinations without departing from the scope of the present disclosure as defined by the appended claims. Additional aspects, advantages, features and objects of the present disclosure would be made apparent from the drawings and the detailed description of the illustrative implementations construed in conjunction with the appended claims that follow. BRIEF DESCRIPTION OF THE DRAWINGSThe summary above, as well as the following detailed description of illustrative embodiments, is better understood when readin conjunction with the appended drawings. For the purpose of illustrating the present disclosure, exemplary constructions ofthe disclosure are shown in the drawings. However, the present disclosure is not limited to specific methods andinstrumentalities disclosed herein. Moreover, those skilled in the art will understand that the drawings are not to scale. Whereverpossible, like elements have been indicated by identical numbers. Embodiments of the present disclosure will now be described, by way of example only, with reference to the following diagrams wherein: FIG.1 illustrates a distributed network comprising a sending device and a receiving device, in accordance with an embodiment of the present disclosure; FIG.2 illustrates detection of relevant transmission errors in a distributed network comprising a sending device and a receiving device, in accordance with an embodiment of the present disclosure; FIG. 3 illustrates a space of an input data and a space of semantic hashes, in accordance with an embodiment of the present disclosure; FIG. 4 illustrates learning a hierarchy of features using neural networks, in accordance with an embodiment of the present disclosure;FIG. 5 illustrates learning a semantic hashing function using a neural network, in accordance with an embodiment of the presentdisclosure; FIG.6 illustrates a joint training of semantic hashing function using a neural network, in accordance with an embodiment of the present disclosure; FIGs. 7A-7B collectively, is a flowchart of a method for a distributed network comprising a sending device and a receiving device, in accordance with an embodiment of the present disclosure; FIG.8 is a block diagram that illustrates various exemplary components of a sending device, in accordance with an embodiment of the present disclosure;FIG. 9 is a flowchart of a method for a sending device, in accordance with an embodiment of the present disclosure;FIG. 10 is a block diagram that illustrates various exemplary components of a receiving device, in accordance with an embodiment of the present disclosure;FIG. 11 is a flowchart of a method for a receiving device, in accordance with an embodiment of the present disclosure; andFIG.12 illustrates an exemplary implementation scenario of neural networks based distributed Joint Inference-Source-Channel- Coding (JISCC) in a distributed network, in accordance with an embodiment of the present disclosure. In the accompanying drawings, an underlined number is employed to represent an item over which the underlined number is positioned or an item to which the underlined number is adjacent. A non-underlined number relates to an item identified by a line linking the non-underlined number to the item. When a number is non-underlined and accompanied by an associated arrow, the non-underlined number is used to identify a general item at which the arrow is pointing. DETAILED DESCRIPTION OF EMBODIMENTS The following detailed description illustrates embodiments of the present disclosure and ways in which they can beimplemented. Although some modes of carrying out the present disclosure have been disclosed, those skilled in the art wouldrecognize that other embodiments for carrying out or practicing the present disclosure are also possible.FIG. 1 illustrates a distributed network comprising a sending device and a receiving device, in accordance with an embodimentof the present disclosure. With reference to FIG. 1, there is shown a distributed network 100 comprising a sending device 102and a receiving device 104. There is further shown a communication network 106 through which the sending device 102 andthe receiving device 104 communicates with each other. The sending device 102 comprises a sending controller 108 and the receiving device 104 comprises a receiving controller 110.The sending device 102 and the receiving device 104 are configured for semantic communication (i.e., task-orientedcommunication) in the distributed network 100. In the distributed network 100, the receiving device 104 is configured to detect relevant transmission errors in received messages in the semantic communication. When implementing the Automatic Repeat- Request (ARQ) in the semantic communication, the understanding of the receiving device 104 about the received messages and the transmission errors plays a significant role. The detection of transmission errors in the semantic communication is different from the traditional data-driven communications. For example, assuming an image of a cat, the background pixels of the image of the cat can be entirely replaced without changing the image’s core meaning (i.e., “a cat”). On the other hand, assuming a text sentence, the sentence’s meaning can be entirely changed just by adding a word “no”. These examples show that there is no direct correspondence between the data representation (e.g., pixels of an image, letters in text) and its semanticmeaning. In contrast to the data-driven communication, the relevant information in semantic communications is inherentlylinked to a specific task to solve (e.g., navigation, accident detection) and to a class of processed data (e.g., images recorded by an autonomous vehicle, legal documents). This is shown that the distributed network 100 includes merely one sending device (i.e., the sending device 102) and one receiving device (i.e., the receiving device 104) for sake of simplicity. In another implementation scenario, the distributed network 100 may include hundreds to thousands of sending devices and two or morereceiving devices to process the sensor data.The sending device 102 may include suitable logic, circuitry, interfaces and / or code that is configured to communicate withthe receiving device 104 via the communication network 106 (e.g., a wireless channel). Examples of the sending device 102may include, but are not limited to, an Internet-of-Things (IoT) device, a smart phone, a machine type communication (MTC) device, a computing device, an evolved universal mobile telecommunications system (UMTS) terrestrial radio access (E- UTRAN) NR-dual connectivity (EN-DC) device, a drone, a customized hardware for wireless telecommunication, a transmitter, or any other portable or non-portable electronic device.The receiving device 104 may include suitable logic, circuitry, interfaces and / or code that is configured to receive data fromthe sending device 102, via the communication network 106. Examples of the receiving device 104 may include, but are notlimited to, an Internet-of-Things (IoT) controller, a base station, a server, a smart phone, a customized hardware for wireless telecommunication, a receiver, or any other portable or non-portable electronic device.The communication network 106 includes a medium (e.g., a wireless channel) through which the sending device 102,potentially communicates with the receiving device 104. Examples of the communication network 106 may include, but are not limited to, a cellular network (e.g., a 2G, a 3G, long-term evolution (LTE) 4G, a 5G, or 5G NR network, such as sub 6GHz, cmWave, or mmWave communication network), a wireless sensor network (WSN), a cloud network, a Local AreaNetwork (LAN), a vehicle-to-network (V2N) network, a Metropolitan Area Network (MAN), and / or the Internet. The sendingdevice 102 in the distributed network 100 is configured to connect to the receiving device 104, in accordance with variouswireless communication protocols. Examples of such wireless communication protocols, communication standards, and technologies may include, but are not limited to, IEEE 802.11, 802.11p, 802.15, 802.16, 1609, Worldwide Interoperability for Microwave Access (Wi-MAX), Transmission Control Protocol and Internet Protocol (TCP / IP), User Datagram Protocol (UDP), Hypertext Transfer Protocol (HTTP), Long-term Evolution (LTE), File Transfer Protocol (FTP), Enhanced Data GSM Environment (EDGE), Voice over Internet Protocol (VoIP), a protocol for email, instant messaging, and / or Short Message Service (SMS), and / or other cellular or IoT communication protocols.The sending controller 108 may include suitable logic, circuitry, interfaces and / or code that is configured to obtain data to betransmitted to the receiving controller of the receiving device 104. Examples of the sending controller 108 may include, but arenot limited to, an integrated circuit, a processor, a co-processor, a microprocessor, a microcontroller, a complex instruction setcomputing (CISC) processor, an application-specific integrated circuit (ASIC) processor, a reduced instruction set (RISC) processor, a very long instruction word (VLIW) processor, a central processing unit (CPU), a state machine, a data processing unit, and other processors or circuits. Moreover, the sending controller 108 may refer to one or more individual controllers,controlling devices, a controlling unit that is part of a machine.The receiving controller 110 may include suitable logic, circuitry, interfaces and / or code that is configured to receive data fromthe sending controller 108 of the sending device 102. Examples of the receiving controller 110 may include, but are not limitedto, an integrated circuit, a processor, a co-processor, a microprocessor, a microcontroller, a complex instruction set computing(CISC) processor, an application-specific integrated circuit (ASIC) processor, a reduced instruction set (RISC) processor, a very long instruction word (VLIW) processor, a central processing unit (CPU), a state machine, a data processing unit, andother processors or circuits. Moreover, the receiving controller 110 may refer to one or more individual controllers, controllingdevices, a controlling unit that is part of a machine.FIG. 2 illustrates detection of relevant transmission errors in a distributed network comprising a sending device and a receivingdevice, in accordance with an embodiment of the present disclosure. FIG. 2 is described in conjunction with elements fromFIG. 1. With reference to FIG. 2, there is shown the distributed network 100 comprising the sending device 102 and thereceiving device 104. There is further shown an input data 202, a wireless channel 208, a first sequence of operations 204 and206 executed by the sending controller 108 of the sending device 102 and a second sequence of operations 210 to 224 executedby the receiving controller 110 of the receiving device 104. In operation, the distributed network 100 comprising the sending device 102 and the receiving device 104, where the sending device 102 comprises the sending controller 108 configured to obtain data to be transmitted, where the data to be transmitted is of an input class and relates to a task to be performed, where the input class and the task to be performed are associated witha K-level hierarchy of features. In the FIG. 2, there is shown the input data 202 (may be represented as ^^) which the sendingcontroller 108 is configured to transmit to the receiving device 104. In the semantic communication, the input data 202 (i.e.,^^) which is required to be transmitted to the receiving device 104 is related to the input class or a class of processed data (e.g.,images recorded by an autonomous vehicle, legal documents, and the like) and to the specific task required to be solved (e.g.,navigation, accident detection, and the like). The input class and the specific task to be solved are related to various inputfeatures of the input data 202 (i.e., ^^), although some input features are more relevant than other features for a given task(e.g., a species vs a breed of an animal for object classification, wind speed vs temperature in storm prediction, etc.). Therefore,for each task and the class of the input data 202 (i.e., ^^), there is defined the K-level hierarchy of relevant input’s features.Each feature in the K-level hierarchy of features can be a category or a number within a range of values. The features are ordered from the most relevant (e.g., level 1 features) to the least relevant (e.g., level K features). Examples of such hierarchies can be (assuming K=3): •Level 1: “animal / machine”, Level 2: “species / type”, Level 3: “breed / model”,• Level 1: “rain level”, Level 2: “wind speed”, Level 3: “pressure”,• Level 1: “hot / cold”, Level 2: ‘upper / lower range’, Level 3: ‘upper-upper / upper-lower / lower-upper / … ‘, •Level 1: “NN feature 1”, Level 2: “NN feature 2”, Level 3: “NN feature 3”.The K-level hierarchy of features can be designed using some domain-knowledge (e.g., weather models) or learned using neural networks, shown and described, for example, in FIG.4. The K-level hierarchy of features does not have to represent all possible relevant features related to the given task (e.g., just K selected).In accordance with an embodiment, the data to be transmitted is the result of an analysis of raw data and wherein the analysisis based on a task to be performed and provides one or more features of the data to be transmitted. The input data 202 (i.e., ^^) which is to be transmitted to the receiving device 104 is obtained by analyzing the raw sensor data. The analysis of the raw sensor data is performed according to the specific task required to be solved. By performing the analysis of the raw sensor data, the one or more relevant features of the input data 202 (i.e., ^^) can be obtained. In accordance with an embodiment, the sending controller 108 is configured to obtain the data by receiving the raw data. The input data 202 (i.e., ^^) is obtained by receiving the raw sensor data. The raw sensor data corresponds to exact representation of the data sensed by various sensing devices used in a distributed network.In accordance with an embodiment, the sending controller 108 is configured to obtain the data by analyzing the raw data. In animplementation scenario, the input data 202 (i.e., ^^) may be a pre-processed data obtained by applying pre-processing on the raw sensor data.In accordance with an embodiment, the raw data represents an image. In an implementation scenario, the raw sensor data canbe pixels of the image and the input data 202 (i.e., ^^) can be an encoded image using JPG format or encoded pixels of the image.In accordance with an embodiment, the input class and the task to be performed are associated with the K-level hierarchy offeatures based on a Hierarchical Classification utilizing K-means. The Hierarchical Classification is a type of classificationtask which involves a hierarchical tree and each label which is not at the root of the hierarchical tree is associated with aprevious label at an upper level. This means that every tree node, is a possible classification label and the tree leaves are themost specific labels which we can classify a sample with. Also, referred to as hierarchical multiclass classification using K-means or K-level clustering. The K-level hierarchy of features is based on the given class of the input data 202 (i.e., ^^) and the specific task to be resolved. The general rule for the K-level hierarchy of features is rather simple, for example, level-1 features are the most relevantfeatures, level-2 features are less relevant and level-K features are the least relevant features. However, the K-level hierarchyof features can be learnt from the input data 202 using K-means or other unsupervised learning technique. Typically, thehierarchical classification utilizing K-means refers to a method of organizing data into a hierarchical structure based onsimilarity. In hierarchical classification, the data is grouped into a hierarchy of clusters, with each cluster representing a groupof similar data points. The process begins by initially partitioning data into K clusters using the K-means algorithm, which aimsto minimize the within cluster-variance. Once the initial clusters are formed, a hierarchical structure is built by iteratively merging or splitting clusters based on their similarity. In another implementation scenario, a supervised learning technique (e.g., training a neural network) may also be used for defining the K-level hierarchy of features. In a yet another implementationscenario, the K-level hierarchy of features can be simply hard-coded by a designer who has a good understanding of the specifictask to be resolved and the class of the input data 202 (i.e., ^^).Furthermore, the low-level features do not have to be sub-groups of high-level features (e.g., a dog is a sub-group of an animal). Another example is weather prediction based on multiple sensing data (e.g., average temperature, pressure, wind speed, etc.) and based on studies it is known that average temperature is more significant than pressure, and pressure is more significant than wind speed. Thus, the hierarchy of features can be defined as: level-1 feature is an average temperature, level-2 feature is pressure, and level-3 feature is wind speed.Moreover, in a yet another implementation scenario, the K-level hierarchy of features can be obtained using domain knowledgeand various models. In execution of such technique, it is assumed that a user has a good knowledge of the input data 202 (i.e., ^^) and the specific task to be solved. For example, weather models can be used to determine the K-level hierarchy of features (e.g., pressure, wind speed, rain intensity) for predicting a storm. Another example, may be hierarchy of objects that an autonomous vehicle may be aware of during driving (e.g., information about people on the crossroad and the color of the trafficlights has more significance over the information about the distant buildings).In accordance with an embodiment, the K-level hierarchy of features is based on a training of a neural network utilized to learnthe most relevant features of the K-level hierarchy of features. In an implementation scenario, a training dataset and the neuralnetwork can be utilized to learn the most relevant features of the K-level hierarchy of features, shown and described, for example, in FIG.4.The sending controller 108 is further configured to determine a semantic key by applying a task-oriented mapping to the data,where the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmitthe data to be transmitted and the semantic key to the receiving device 104. At operation 204, the sending controller 108 isconfigured to apply the task-oriented mapping, for example, semantic hashing function (i.e., ℎ^^ℎ()) to the input data 202(i.e., ^^). The semantic hashing function can be designed for any class of the input data 202 and a task to be resolved. The semantic hashing function is utilized for detection of relevant transmission errors in the semantic communication. The semantic hashing function is designed offline for the given class of the input data 202 and the given K-level hierarchy of features. The semantic hashing function may either be any algorithm or a trained neural network, which efficiently maps the input data 202to the semantic key (i.e., ℎ^ = ℎ (^^)). The sending controller 108 is further configured to transmit the semantic key (i.e., ℎ^)and the input data 202 (i.e., ^^) to the receiving device 104. At operation 206, the sending controller 108 is configured to applyencoding and signal modulation on the input data 202 (i.e., ^^) and the semantic key (i.e., ℎ^) and transmit a triple [ID, ^^,ℎ^] that is, identity of the sending device 102, (ID), encoded input data (i.e., ^^) and the encoded semantic key (i.e., ℎ^) tothe receiving device 104 through a wireless channel 208. The wireless channel 208 corresponds to a noisy channel in thecommunication network 106 (of FIG.1).In accordance with an embodiment, the task-oriented mapping is based on receiving input data and outputting the semantic keybeing a short code of a predefined structure, where the short code depends on the input data’s relevant features from the K-level hierarchy of features. The task-oriented mapping that is, semantic hashing function, ℎ^^ℎ (), associates pieces of theinput data 202 with short codes (i.e., ℎ^) of a predefined structure. These short codes have a special property, shown and described, for example, in FIG.3. The value of the semantic hashing function depends on the relevant features from the K-level features of the input data 202.In accordance with an embodiment, the task-oriented mapping is based on a natural hash format [level 1 feature value, level 2 feature value, …, level K feature value]. In an implementation scenario, the task-oriented mapping (i.e., the semantic hashing function) can be determined based on the natural hash format, i.e., [level 1 feature value, level 2 feature value, …, level Kfeature value]. The natural hash format is easy to compute and enables an easy identification of features which are different forboth hashes. However, the use of the natural hash format is not preferrable for privacy reasons. The codes are more prone to distortion with uniform channel noise because noise added to the level 1 features introduces more distortion than noise introduced to level K features.In accordance with an embodiment, the task-oriented mapping is based on K-level nested lattices. In another implementationscenario, the task-oriented mapping (i.e., the semantic hashing function) can be determined based on the K-level nested lattices,where each lattice corresponds to a unique feature level. In the K-level nested lattices, the features are assigned to codes having a nested “tree-like” structure. With such design, the codes which differ at the top features are likely to be far apart from eachother, while the codes which differ at the low features are likely to be similar. The K-level nested lattices have a simple design,and are easy to compute and robust to channel distortion. However, except for some sophisticated designs of the K-level nested lattices, the number of distinctive values per feature should be relatively similar for all feature levels.In accordance with an embodiment, the task-oriented mapping is based on a learning using a neural network, utilizingdiscriminative learning, where a discriminative L1 loss is utilized to encourage the neural network to assign close codes tosimilar features while assigning distant codes to dissimilar features in a following way: vectors with same level-1 features should be close to each other (and remaining far apart), and among these vectors, the vectors with same level-2 features should be close to each other. The learning of the semantic hashing function using the neural network relies on the discriminative learning, as shown and described, for example, in FIG. 5. The discriminative learning uses the discriminative L1 loss which encourages the neural network to assign the codes having a small difference to similar features and assign the codes having a large difference to dissimilar features. The codes to similar and dissimilar features can be assigned in the following way: the vectors with the same level-1 features are close to each other and remaining far apart. Also, among these vectors, the vectorswith same level-2 features should be close to each other, etc. In this way, the vectors with the same level-K features should beclose to each other.The receiving device 104 comprising the receiving controller 110 is configured to receive the data and the semantic key anddetermine a check semantic key by applying the same task-oriented mapping to the received data. The receiving controller 110 is configured to receive the triple [ID, ^^, ℎ^] that is, identity of the sending device 102, (ID), encoded input data (i.e., ^^) and the encoded semantic key (i.e., ℎ^) from the sending controller 108. At operation 210, the receiving controller 110 is configured to apply the signal demodulation and decoding on the received identity of the sending device 102 (ID), the encoded input data (i.e., ^^) and the encoded semantic key (i.e., ℎ^) and obtain the received data (i.e., ^^) and the semantic key (i.e.,ℎ^^). At operation 212, the receiving controller 110 is further configured to determine the check semantic key by applying thesame task-oriented mapping (i.e., the semantic hashing function, ℎ) to the received data (i.e., ^^). The same task-oriented mapping (i.e., the semantic hashing function, ℎ) is used at each of the sending device 102 and the receiving device 104.The receiving controller 110 is further configured to compare the check semantic key to the semantic key in order to determineif there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so, determine the level of the transmission error based on a K-level of the feature for which themismatch is detected. After computing the check semantic key, the receiving controller 110 is configured to compute adifference (^^) between the check semantic key (i.e., ℎ(^^)) and the semantic key (i.e., ℎ^^) at operation 214 in order todetermine the transmission error. The transmission error may be induced due to noise addition during transmission of datathrough the wireless channel 208. If the semantic key (i.e., ℎ^^) and the check semantic key (i.e., ℎ(^^)) are different fromeach other for a specific feature of the input data 202 then, it is determined that the transmission error is present in the receiveddata. After determining the presence of the transmission error, at operation 216, the receiving controller 110 is configured todetermine the level of the transmission error based on the K-level feature of the input data 202 for which the mismatch is detected.In accordance with an embodiment, the receiving controller 110 is further configured to compute a weighted L1 distancebetween the received semantic hash and the check semantic key as the level of the transmission error and determine that the level of the transmission error is above the relevance threshold level if the weighted L1 distance is larger than the relevancethreshold level. The level of the transmission error is determined by computing the weighted L1 distance between the receivesemantic key (i.e., ℎ^^) and the check semantic key (i.e., ℎ(^^)). If the weighted L1 distance is larger than the relevancethreshold level it means that the wireless channel 208 introduced intolerable distortion to the input data (i.e., ^^). In accordance with an embodiment, the weighted L1 distance is the distance between the semantic key and the check semantickey. The weighted L1 distance is measured as the distance between the semantic key (i.e., ℎ^^) and the check semantic key(i.e., ℎ(^^)).In accordance with an embodiment, the receiving controller 110 is further configured to compute the weighted L1 metric overfinite fields to measure differences between the semantic key and the check semantic key, wherein there exist values d1 > d2> … dK > 0 such that L1_distance (semantic key (i.e. mapping of data to be transmitted), check semantic key (i.e. mapping of data received)) >= d1 implies a difference at the level-1 feature between data to be transmitted and data received, d1 > L1_distance (semantic key, check semantic key) >= d2 implies a difference at the level-2 feature between data to be transmitted and data received, dK-1 > L1_distance (semantic key, check semantic key) >= dK implies a difference at the level-K feature between data to be transmitted and data received, and L1_distance (semantic key, check semantic key) = 0 implies full agreement of data to be transmitted and data received with respect to the K-level hierarchy. Code: [2, -12, 1013] corresponding features: [‘heavy rain’, ‘-12 degrees C’, ‘1013hPa’].The receiving controller 110 is further configured to determine that the data is successfully received if the level of thetransmission error is below a relevance threshold level and send an acknowledgement response, ACK, in response thereto. Atoperation 218, the receiving controller 110 is configured to determine the level of the transmission error below the relevancethreshold level. The determining the level of the transmission error below the relevance threshold level means that the wirelesschannel 208 introduced the tolerable distortion during transmission and the receiving controller 110 is configured to accept the received data (i.e., ^^). Thereafter, at operation 220, the receiving controller 110 is configured to send a positive acknowledgement (ACK) to the sending device 102 as [ID, ‘ACK’, prompt]. The receiving controller 110 is configured to use the relevance threshold level for either accepting or rejecting the received data (i.e., ^^). The relevance threshold levelcorresponds to the highest feature level in the K-level hierarchy of features at which the transmission error is still tolerable bythe receiving controller 110. Moreover, the relevance threshold level can be adaptive depending on the context, for example, a high threshold level is used for a first message and a low threshold level is used for a re-transmitted message. In accordance with an embodiment, the acknowledgement response comprises a prompt indicating a distance between thesemantic key and the check semantic key. The acknowledgement response sent to the sending device 102 comprises the promptwhich provides the information about the understanding of the receiving device 104 related to the transmitted data or detected errors (e.g., a distance between the hashes). The prompt can be useful for the sending device 102 who may efficiently adapt thereceived information and get ready for future data transmission, and the like.In accordance with an embodiment, the prompt indicates an action for a future transmission of data. After accepting the receiveddata (i.e., ^^), the receiving controller 110 is configured to send the positive acknowledgement to the sending device 102. The positive acknowledgement comprises the prompt, which provides the information to the sending device 102 for transmitting afuture data packet.The receiving controller 110 is further configured to determine that the data is not successfully received if the level of thetransmission error is above the relevance threshold level and send a negative acknowledgement response, NACK, in responsethereto. At operation 222, the receiving controller 110 is configured to determine the level of the transmission error above therelevance threshold level. The determining the level of the transmission error above the relevance threshold level means thatthe wireless channel 208 introduced the intolerable distortion during transmission and the receiving controller 110 is configured to reject the received data (i.e., ^^). Thereafter, at operation 224, the receiving controller 110 is configured to send a negative acknowledgement (NACK) to the sending device 102 as [ID, ‘NACK’, prompt]. In accordance with an embodiment, the negative acknowledgement response comprises a prompt indicating a distance betweenthe semantic key and the check semantic key. The negative acknowledgement response sent to the sending device 102comprises the prompt which provides the information about the understanding of the receiving device 104 related to the detected transmission errors, for example, the distance between the semantic key and the check semantic key. The prompt can be useful for the sending device 102 who may efficiently adapt the received information and get ready for partial data re- transmission, data correction, and the like. In accordance with an embodiment, the prompt indicates an action for a retransmission of the data. After rejecting the receiveddata (i.e., ^^), the receiving controller 110 is configured to send the negative acknowledgement to the sending device 102.The negative acknowledgement comprises the prompt, which provides the information to the sending device 102 for data re- transmission, data correction, and the like.Thus, the distributed network 100 comprising the sending device 102 and the receiving device 104 enables an efficient detectionof the relevant transmission errors in the semantic communications by virtue of using the task-oriented mapping that is semantic hashing functions. Moreover, the distributed network 100 manifests computational and communication efficiency proportional to conventional CRC-based error detection approach. The semantic hashes are much smaller than initial data representation and hence, the semantic hash functions introduce little communication overhead. The computation of the semantic hashes can be highly efficient even when implemented using neural networks. The semantic hashing functions can be designed for any class of the input data 202 and any kind of specific task to be resolved. Moreover, the computing a difference between the two semantic hash functions is performed using the L1 metric which does not depend on the semantic task anymore. The semantic hash functions allow one to detect by how much the input data U1, U2 differ from each other which is advantageous over the conventional CRC approach. Concurrently, the distributed network 100 enables the detection of relevant transmission errors for the given task and also, applies an adaptive threshold for accepting partially correct messages hence, reducing re- transmissions. Additionally, the distributed network 100 utilizing the semantic hashing functions does not require any complexsemantic analysis.Moreover, the distributed network 100 may be useful for real-time wireless communication with robotic devices. In suchimplementation scenarios, either a robotic device (or a robot) or an autonomous vehicle (or a semi-autonomous vehicle) isrequired to solve some complex tasks based on information received wirelessly from a remote sensing device. In order toperform the task accurately and reliably, the robotic device or the autonomous vehicle (or the semi-autonomous vehicle) isrequired to receive relevant information from the remote sensing device with a minimum delay. By using the task-orientedmapping (i.e., the semantic hashes), the robotic device or the autonomous vehicle (or the semi-autonomous vehicle) may havethe capability to validate the usefulness of received data despite of the presence of the transmission errors, resulting in asignificant reduction of costly re-transmissions. Furthermore, even if the received data is significantly distorted, the correctportion of the received data validated by the semantic hashes can still be useful for the robotic device or the autonomous vehicle(or the semi-autonomous vehicle). FIG. 3 illustrates a space of an input data and a space of semantic hashes, in accordance with an embodiment of the present disclosure. FIG. 3 is described in conjunction with elements from FIGs. 1 and 2. With reference to FIG. 3, there is shown a space of an input data 302 and a space of semantic hashes 304. In an implementation scenario, the space of input data 302 may have images represented as pixels, which have little correspondence with the relevance of the data (e.g., completely different images of dogs). The space of input data 302 corresponds to a high-dimensional data space which have similar data 302A as well as dissimilar data 302B. The similar data 302A (may be represented as ^) on which the task-oriented mapping (i.e., the semantic hashing function) is applied todetermine the semantic key (i.e., ℎ). The space of semantic hashes 304 corresponds to a low-dimensional space of semantichashes, in which similar data 304A (e.g., images of dogs) is close to each other, whereas dissimilar data 304B are far apart from each other. A small difference between the semantic keys (or short codes) implies high semantic similarity of the associated inputs (and vice-versa). The semantic hashing functions are initially designed for data mining and data retrieval (e.g., for searching images similar to a reference image). In the distributed network 100, the semantic hashes are used in the context of semantic communications and error detection. FIG. 4 illustrates learning a hierarchy of features using neural networks, in accordance with an embodiment of the presentdisclosure. FIG. 4 is described in conjunction with elements from FIGs. 1, 2 and 3. With reference to FIG. 4, there is shown aneural network 400 comprising an encoder 402 and a decoder 404. There is further shown an input vector 406, latent featurevector of dimension K 408, a random mask 410, a uniform noise 412 a predicted label 414 and an error vector 416.In accordance with an embodiment, the K-level hierarchy of features is based on a training of the neural network 400, wherethe neural network 400 is a randomly initialized neural network having an encoder / decoder architecture, wherein a role of the encoder 402 is to map the input vector 406 into the latent feature vector of dimension K 408 where each dimension correspondsto one feature, and a role of the decoder 404 is to predict a true label from the latent vector. In an implementation scenario, thelearning of hierarchy of features can be achieved by training of the neural network 400 comprising the encoder 402 and thedecoder 404. The encoder 402 is configured to map the input vector 406 (may also be represented as ^^∗) into the latent feature vector of dimension K 408. The latent feature vector of dimension K 408 may also be referred to as a K-dimensional feature vector. The decoder 404 is configured to predict the true label (may also be referred to as the predicted label 414) using the latent feature vector of dimension K 408.The encoder 402 is configured to receive the input vector 406 and provide a K-dimensional feature vector to which the randommask 410 is applied which randomly selects k < K and masks all lower k features of the feature vector. The input vector 406is passed through the encoder 402 to obtain the latent feature vector of dimension K 408 (i.e., the K-dimensional feature vector). The random mask 410 is applied on the latent feature vector of dimension K 408 (i.e., the K-dimensional feature vector) which randomly selects k<K features and masks all lower k features of the latent feature vector of dimension K 408 (i.e., the K-dimensional feature vector) and a masked vector is generated. Thereafter, the uniform noise 412 is added to themasked vector and a resulted vector is generated having the added uniform noise 412.In accordance with an embodiment, the random mask 410 is configured to erase information from lowest features and preserving information in top features which encourages the encoder 402 to put more relevant information in top features. Consequently, a hierarchical representation is obtained in which the top features are the most relevant and the lowest featuresare the least relevant.In accordance with an embodiment, wherein the training of the neural network 400 further comprises applying the uniformnoise 412 to the input vector 406, whereby the encoder 402 is encouraged to maintain a smooth scaling of features. The uniformnoise 412 is added to the masked vector, which is obtained after masking of all lower k<K features of the latent feature vector of dimension K 408. The masked vector enables the encoder 402 to maintain the smooth scaling of features, which is beneficial if the feature vectors are required to be quantized.The decoder 404 is configured to receive the resulted vector and provide a prediction of the true label, whereby an error vectoris back-propagated the neural network 400, whereby weights are updated. The resulted vector is passed to the decoder 404 andthe decoder is configured to predict the true label, i.e., the predicted label 414 (may also be represented as ^^^∗). The differencebetween a true label and the predicted label 414 is computed and a loss vector is computed. Moreover, the error vector 416 isback propagated from the decoder 404 to the encoder 402 and accordingly the weights (i.e., weights of convolution layers) inthe encoder 402 and the decoder 404 are updated.FIG. 5 illustrates learning a semantic hashing function using a neural network, in accordance with an embodiment of the presentdisclosure. FIG.5 is described in conjunction with elements from FIGs.1, 2, 3 and 4. With reference to FIG.5, there is showna neural network 500 comprising a first encoder 502 and a second encoder 504. There is further shown a group of input vectors506, a group of K-dimensional feature vectors 508, and a group of coded vectors 510.The neural network 500 is used for learning a semantic hash function (i.e., ℎ^^ℎ()) from K-level hierarchy of features. In the neural network 400 (of FIG.4), mere one input sample (i.e., ^^∗) is used for training of the neural network 400. However, inthe FIG. 5, a group of inputs is used simultaneously, for learning of the neural network 500. In an implementation scenario, fortraining of the neural network 500, the group of input vectors 506 (may also be represented as ^∗ ∗ ∗^ , ^^, , … , ^^ ) may have agroup of “similar” data inputs. By use of the group of input vectors 506, a group of either all similar or dissimilar feature vectorscan be obtained. As shown in the FIG. 5, the group of K-dimensional feature vectors 508 (may also be represented as^∗^, ^ ∗^, , … , ^^∗) is obtained using the group of input vectors 506 (i.e., ^∗^ , ^ ∗^, , … , ^^∗). The group of K-dimensional feature vectors 508 (i.e., ^∗^, ^ ∗^, , … , ^^∗) can be obtained, for example, using the first encoder 502. The first encoder 502 corresponds to theencoder 402 of the neural network 400. Moreover, the group of K-dimensional feature vectors 508 (i.e., ^∗ ∗^, ^^, , … , ^^∗) can be generated artificially. The first encoder 502 is a trained encoder (whose training is done as shown and described, for example, in FIG.4) with frozen weights. The group of K-dimensional feature vectors 508 (i.e., ^∗^, ^ ∗^, , … , ^^∗) is provided as an input tothe second encoder 504, which is configured to assign codes to every vector of the group of K-dimensional feature vectors 508(i.e., ^∗^, ^ ∗^, , … , ^^∗ ) and generate the group of coded vectors 510 (may also be represented as ^∗ ∗^̂, ^^̂, , … , ^^̂∗). The group of coded vectors 510 (i.e., ^̂∗, ^^̂ ∗, , … , ) corresponds to a group of “similar” codes. Depending on whether the selected group feature vectors (i.e., the group of K-dimensional feature vectors 508) is similar or dissimilar, a discriminative loss (or acontractive loss) function is used, which supports the second encoder 504 to keep the codes of similar features close to eachother and the codes of dissimilar features far apart from each other. The training of the second encoder 504 is continued untilconvergence. During inference, the output (i.e., the group of coded vectors 510) of the second encoder 504 is additionallyquantized at operation 512. The training of the neural network 500 is advantageous by virtue of the huge flexibility. The neural network 500 can be configured to assign the codes to the K-level features of mixed types (e.g., categories and numerical values) and of variable resolution. However, the neural network 500 (particularly, the second encoder 504) is required to be fine-tuned to ensure that the second encoder 504 generates balanced codes with a distribution corresponding to the intended similarity of features.FIG. 6 illustrates a joint training of semantic hashing function using a neural network, in accordance with an embodiment ofthe present disclosure. FIG. 6 is described in conjunction with elements from FIGs. 1, 2, 3, 4 and 5. With reference to FIG. 6,there is shown a neural network 600 comprising an encoder 602 and a decoder 604. There is further shown a group of inputvectors 606, a group of coded vectors 608, and a group of predicted vectors 610. There is further shown a space of an inputdata 612 and a space of semantic hashes 614. The space of input data 612 corresponds to a high-dimensional data space which have similar data as well as dissimilar data. The similar data (may be represented as ^^∗) on which the task-oriented mapping (i.e., the semantic hashing function) is appliedto determine the semantic key (i.e., ℎ). The space of semantic hashes 614 corresponds to a low-dimensional space of semantichashes, in which similar data (may be represented as ^∗^) is close to each other, whereas dissimilar data are far apart from eachother. The training of the semantic hashing function can be done without discovering or learning the K-level hierarchy of features. In FIG. 6, there is shown training of the neural network 600 for which it is assumed that K-level hierarchy of features already exists. For training of the neural network 600, a combination of two approaches as shown and described, for example, in FIGs. 4 and 5, is used. The encoder / decoder architecture (i.e., the encoder 402 and the decoder 404) used for learning the K-level hierarchy of features (as shown in FIG.4) is utilized. Also, the way by which the group of input vectors 506 is processed by use of the first encoder 502, as shown and described in FIG.5, is utilized. By properly grouping similar input vectors (i.e., thegroup of input vectors 606) and by using a following loss function according to Equation (1) The encoder 602 is configured to learn to directly assign the codes the group of input vectors 606 (may also be represented as ^∗^ , ^ ∗^, , … , ^^∗) by taking into consideration the inputs’ semantic similarity. The group of input vectors 606 (i.e., ^∗^ , ^ ∗^, , … , ^^∗) corresponds to a group of “similar” data inputs. The group of input vectors 606 (i.e., ^∗^ , ^ ∗^, , … , ^^∗) is provided as an input to the encoder 602 and the encoder 602 is configured to generate the group of coded vectors 608 (may also be represented as ^∗^, ^ ∗^, , … , ^^∗ ) by applying the semantic hash (i.e., ℎ(. )) on the group of input vectors 606 (i.e., ^∗ ∗^ , ^^, , … The group of coded vectors 608 (i.e., ^∗^, ^ ∗^, , … , ^^∗) corresponds to a group of “similar” semantic hashes and provided as an input to the decoder 604. The decoder 604 is configured to generate the group of predicted vectors 610 (may also be represented as ^^∗^, ^^ ∗^, , … , ^^^∗ ) by applying a function (i.e., ^(. )) on the group of coded vectors 608 (i.e., ^∗^, ^ ∗^, , … , ^^∗). When the training of the neural network 600 is finished, the decoder 604 is discarded and the encoder 602 becomes a semantic hashing function.The training of the semantic hash function (i.e., ℎ^(. )) is described in Algorithm 1.Algorithm 1: Training the semantic hash function ℎ^(. )Input: Training dataset ^^^^^^ = {(^^ , ^^)}^^^^. Output: Optimized ^. 1: Initialization: 2: Initialize ^, ^.3: Repeat: 4: Randomly select a sample (^∗^ , ^∗^) from ^^^^^^.5: Select ^ ≪ ^ ‘similar’ samples 6: Encode samples 7: Decode samples 8: Compute gradients of ∇^^^^,^ℒ^ (^, ^) and update (^, ^).9: Until: convergence of (^, ^).FIGs. 7A-7B collectively, is a flowchart of a method for a distributed network comprising a sending device and a receivingdevice, in accordance with an embodiment of the present disclosure. FIGs. 7A-7B are described in conjunction with elementsfrom FIGs. 1, 2, 3, 4, 5 and 6. With reference to FIGs. 7A-7B, there is shown a method 700 for the distributed network 100comprising the sending device 102 and the receiving device 104 (of FIG. 1). The method 700 includes steps 702 to 718 (thesteps 702 to 710 are shown in FIG. 7A, and the steps 712 to 718 are shown in FIG. 7B). The sending device 102 and the receiving device 104 of the distributed network 100 are configured to execute the method 700. There is provided the method 700 for the distributed network 100 comprising the sending device 102 and the receiving device104. The use of the method 700 enables an efficient detection of the relevant transmission errors in the semanticcommunications by virtue of using the task-oriented mapping that is semantic hashing functions. Moreover, the computationaland communication efficiency proportional to the conventional CRC-based error detection approach is obtained by utilizing the method 700. Concurrently, the detection of relevant transmission errors for the given task is obtained on using the method 700. Moreover, an adaptive threshold is applied for accepting partially correct messages hence, reducing re-transmissions. Additionally, by virtue of utilizing the semantic hashing functions, the method 700 does not include any complex semantic analysis.Referring to FIG. 7A, at step 702, at the sending device 102, the method 700 comprises obtaining data to be transmitted, wherethe data to be transmitted is of an input class and relates to a task to be performed, where the input class and the task to beperformed are associated with a K-level hierarchy of features. The data to be transmitted to the receiving device 104 is relatedto the input class (e.g., images recorded by an autonomous vehicle, legal documents, and the like) and to the specific task required to be solved (e.g., navigation, accident detection, and the like). The input class and the specific task to be solved are related to various input features of the data, arranged as the K-level hierarchy of features.At step 704, at the sending device 102, the method 700 further comprises determining a semantic key by applying a task-oriented mapping to the data, where the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of features. The task-oriented mapping (i.e., the semantic hashing function) is applied on the data to obtain thesemantic key, which is further transmitted along with the data to the receiving device 104. At step 706, at the sending device 102, the method 700 further comprises transmitting the data to be transmitted and the semantickey to the receiving device 104. The data and the semantic key are transmitted to the receiving device 104.At step 708, at the receiving device 104, the method 700 further comprises receiving the data and the semantic key.At step 710, at the receiving device 104, the method 700 further comprises determining a check semantic key by applying thesame task-oriented mapping to the received data. The same task-oriented mapping (i.e., the semantic hashing function) isapplied on the received data to obtain the check semantic key.Now referring to FIG. 7B, at step 712, at the receiving device 104, the method 700 further comprises comparing the checksemantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatchbetween the semantic key and the check semantic key for a feature of the data. The difference between the check semantic keyand the semantic key (i.e., the received semantic key) is computed to determine the presence of the transmission error. At step 714, at the receiving device 104, the method 700 further comprises determining the level of the transmission error based on a K-level of the feature for which the mismatch is detected. In accordance with an embodiment, the method 700 further comprises the receiving device 104 computing a weighted L1distance between the received semantic hash and the check semantic key as the level of the transmission error and determinethat the level of the transmission error is above the relevance threshold level if the weighted L1 distance is larger than therelevance threshold level. The level of the transmission error is determined by computing the weighted L1 distance betweenthe received semantic key and the check semantic key. If the weighted L1 distance is larger than the relevance threshold level it means that a significant distortion gets added to the data during transmission over a wireless channel (e.g., the wireless channel 208, of FIG.2). In accordance with an embodiment, the method 700 further comprises the receiving device 104 computing the weighted L1metric over finite fields to measure differences between the semantic key and the check semantic key, where there exist valuesd1 > d2 > … dK > 0 such that L1_distance (semantic key (i.e. mapping of data to be transmitted), check semantic key (i.e. mapping of data received)) >= d1 implies a difference at the level-1 feature between data to be transmitted and data received, d1 > L1_distance (semantic key, check semantic key) >= d2 implies a difference at the level-2 feature between data to be transmitted and data received, dK-1 > L1_distance (semantic key, check semantic key) >= dK implies a difference at the level-K feature betweendata to be transmitted and data received, and L1_distance (semantic key, check semantic key) = 0 implies full agreement of data to be transmitted and data receivedwith respect to the K-level hierarchy. The use of the semantic key and check semantic key enables the detection of relevant differences between various inputs (d1, d2, ..., dK) based on the weighted L1 metric.At step 716, at the receiving device 104, the method 700 further comprises determining that the data is successfully received ifthe level of the transmission error is below a relevance threshold level and sending an acknowledgement response, ACK, inresponse thereto. In response of determining the level of the transmission error below the relevance threshold level, a positiveacknowledgement response (i.e., ACK) is sent to the sending device 102.At step 718, at the receiving device 104, the method 700 further comprises determining that the data is not successfully receivedif the level of the transmission error is above the relevance threshold level and sending a negative acknowledgement response,NACK, in response thereto. In response of determining the level of the transmission error above the relevance threshold level,the negative acknowledgement response (i.e., NACK) is sent to the sending device 102.In accordance with an embodiment, the K-level hierarchy of features is based on a training of a neural network (e.g., the neuralnetwork 400) utilized to learn the most relevant features of the K-level hierarchy of features and where the K-level hierarchyof features is based on a training of a neural network (i.e., the neural network 400), where the neural network (i.e., the neuralnetwork 400) is a randomly initialized neural network having an encoder / decoder architecture (e.g., the encoder 402 and thedecoder 404), wherein a role of the encoder (i.e., the encoder 402) is to map an input vector (e.g., the input vector 406) into alatent feature vector of dimension K (e.g., the latent feature vector of dimension K 408) wherein each dimension corresponds to one feature, and a role of the decoder (i.e., the decoder 404) is to predict the true label from a latent vector, and wherein the encoder (i.e., the encoder 402) is configured to receive an input vector provide a K-dimensional feature vector to which arandom mask (e.g., the random mask 410) is applied which randomly selects k < K and masks all lower k features of the featurevector, wherein the decoder (i.e., the decoder 404) is configured to receive the resulted vector and provide a prediction of thetrue label, whereby an error vector is back-propagated to the neural network, whereby weights are updated. The training of the neural network to learn the K-level hierarchy of features is shown and described in detail, for example, in FIG.4. The steps 702 to 718 are only illustrative and other alternatives can also be provided where one or more steps are added, one or more steps are removed, or one or more steps are provided in a different sequence without departing from the scope of the claims herein. In one aspect, there is provided a computer program product comprising program instructions for performing the method 700 when executed by one or more processors (e.g., the sending controller 108 of the sending device 102 and the receiving controller 110 of the receiving device 104) in the distributed network 100. In another aspect, the present disclosure provides a non- transitory computer-readable medium having stored thereon, computer-implemented instructions that, when executed by a computer, causes the computer to execute operations of the method 700. FIG.8 is a block diagram that illustrates various exemplary components of a sending device, in accordance with an embodiment of the present disclosure. FIG.8 is described in conjunction with elements from FIGs.1, 2, 3, 4, 5, 6, and 7A-7B. With referenceto FIG. 8, there is shown a block diagram 800 of the sending device 102 comprising the sending controller 108, a memory 802and a communication interface 804.The memory 802 may include suitable logic, circuitry, and / or interfaces that is configured to store machine code and / orinstructions executable by the sending controller 108. Examples of implementation of the memory 802 may include, but arenot limited to, an Electrically Erasable Programmable Read-Only Memory (EEPROM), Random Access Memory (RAM),Read Only Memory (ROM), Hard Disk Drive (HDD), Flash memory, a Secure Digital (SD) card, Solid-State Drive (SSD), acomputer readable storage medium, and / or CPU cache memory. The memory 802 may store an operating system and / or acomputer program product to operate the sending device 102. A computer readable storage medium for providing a non- transient memory may include, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.The communication interface 804 may include suitable logic, circuitry, and / or interfaces that is configured to transmit data tobe transmitted and a semantic key to a receiving device (e.g., the receiving device 104). Examples of the communication interface 804 may include, but are not limited to, an antenna, a radio frequency transceiver, a network interface, a telematics unit, or any antenna suitable for use in an Internet-of-Things (IoT) device, a smart phone, a machine type communication (MTC) device, a computing device, an evolved universal mobile telecommunications system (UMTS) terrestrial radio access (E-UTRAN) NR-dual connectivity (EN-DC) device, a drone, a customized hardware for wireless telecommunication, a transmitter, or any other portable or non-portable electronic device. The communication interface 804 may wirelessly communicate by use of various wireless communication protocols. In operation, the sending device 102 comprising the sending controller 108 is configured to obtain data to be transmitted, wherethe data to be transmitted is of an input class and relates to a task to be performed, where the input class and the task to beperformed are associated with a K-level hierarchy of features. The sending controller 108 is configured to transmit the datarelated to the input class and the specific task to be solved. The input class and the specific task to be solved are related to various input features of the data, arranged as the K-level hierarchy of features. The sending controller 108 is further configured to determine a semantic key by applying a task-oriented mapping to the data, where the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmitthe data to be transmitted and the semantic key to the receiving device 104. By virtue of applying the task-oriented mapping tothe data, the sending controller 108 supports the detection of relevant transmission errors in the semantic communication.FIG. 9 is a flowchart of a method for a sending device, in accordance with an embodiment of the present disclosure. FIG. 9 isdescribed in conjunction with elements from FIGs. 1, 2, 3, 4, 5, 6, 7A-7B and 8. With reference to FIG. 9, there is shown amethod 900 for use in the sending device 102. The method 900 includes steps 902 to 906. The sending device 102 is configuredto execute the method 900.At step 902, the method 900 comprises obtaining data to be transmitted, where the data to be transmitted is of an input classand relates to a task to be performed, where the input class and the task to be performed are associated with a K-level hierarchyof features. At step 904, the method 900 further comprises determining a semantic key by applying a task-oriented mapping to the data, where the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy offeatures. At step 906, the method 900 further comprises transmitting the data to be transmitted and the semantic key to thereceiving device 104.The steps 902 to 906 are only illustrative and other alternatives can also be provided where one or more steps are added, one or more steps are removed, or one or more steps are provided in a different sequence without departing from the scope of the claims herein. In one aspect, there is provided a computer program product comprising program instructions for performing the method 900 when executed by one or more processors (e.g., the sending controller 108 of the sending device 102) in the distributed network 100. In another aspect, the present disclosure provides a non-transitory computer-readable medium having stored thereon,computer-implemented instructions that, when executed by a computer, causes the computer to execute operations of themethod 900.FIG. 10 is a block diagram that illustrates various exemplary components of a receiving device, in accordance with anembodiment of the present disclosure. FIG.10 is described in conjunction with elements from FIGs.1, 2, 3, 4, 5, 6, 7A-7B, 8 and 9. With reference to FIG.10, there is shown a block diagram 1000 of the receiving device 104 comprising the receiving controller 110, a memory 1002 and a communication interface 1004. The memory 1002 may include suitable logic, circuitry, and / or interfaces that is configured to store machine code and / orinstructions executable by the receiving controller 110. Examples of implementation of the memory 1002 are similar to that ofthe memory 802 (of FIG. 8). The memory 1002 may store an operating system and / or a computer program product to operatethe receiving device 104. A computer readable storage medium for providing a non-transient memory may include, but is notlimited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.The communication interface 1004 may include suitable logic, circuitry, and / or interfaces that is configured to receive data anda semantic key from a sending device (e.g., the sending device 102). Examples of the communication interface 1004 mayinclude, but are not limited to, an antenna, a radio frequency transceiver, a network interface, a telematics unit, or any antennasuitable for use in an Internet-of-Things (IoT) controller, a base station, a server, a smart phone, a customized hardware orother portable or non-portable communication devices. The communication interface 1004 may wirelessly communicate by use of various wireless communication protocols.In operation, the receiving device 104 comprising the receiving controller 110 is configured to receive data and a semantic key,where the data received is of an input class and relates to a task to be performed, where the input class and the task to be performed are associated with a K-level hierarchy of features, and where the semantic key is determined by applying a task- oriented mapping to the prior to transmission data, where the task-oriented mapping provides a mapping of the data to thesemantic key based on K-level hierarchy of features. The receiving controller 110 is configured to receive the data and thesemantic key transmitted by the sending controller 108 of the sending device 102 over a wireless channel, have been described in detail, for example, in FIG.2. The receiving controller 110 is further configured to determine a check semantic key by applying the same task-orientedmapping to the received data and compare the check semantic key to the semantic key in order to determine if there is atransmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so, determine the level of the transmission error based on a K-level of the feature for which the mismatch is detected. The receiving controller 110 is configured to compute a difference between the check semantic key and the semantic key to detect the presence of the transmission error. After determining the presence of the transmission error, the receivingcontroller 110 is further configured to determine the level of the transmission error using the hierarchy of the K-level features.The receiving controller 110 is further configured to determine that the data is successfully received if the level of thetransmission error is below a relevance threshold level and send an acknowledgement response, ACK, in response thereto, anddetermine that the data is not successfully received if the level of the transmission error is above the relevance threshold leveland send a negative acknowledgement response, NACK, in response thereto. In response of determining the transmission errorbelow the relevance threshold level, the receiving controller 110 is configured to transmit the positive acknowledgement to the sending device 102 to indicate the successful reception of the data. In response of determining the transmission error above the relevance threshold level, the receiving controller 110 is configured to transmit the negative acknowledgement to the sendingdevice 102 to indicate the unsuccessful reception of the data.FIG. 11 is a flowchart of a method for a receiving device, in accordance with an embodiment of the present disclosure. FIG.11 is described in conjunction with elements from FIGs.1, 2, 3, 4, 5, 6, 7A-7B, 8, 9 and 10. With reference to FIG.11, there is shown a method 1100 for use in the receiving device 104. The method 1100 includes steps 1102 to 1112. The receiving device 104 is configured to execute the method 1100. At step 1102, the method 1100 comprises receiving data and a semantic key, where the data received is of an input class and relates to a task to be performed, where the input class and the task to be performed are associated with a K-level hierarchy of features, and where the semantic key is determined by applying a task-oriented mapping to the prior to transmission data, wherethe task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of features. At step1104, the method 1100 further comprises determining a check semantic key by applying the same task-oriented mapping to thereceived data. At step 1106, the method 1100 further comprises comparing the check semantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatch between the semantic key and the checksemantic key for a feature of the data, and if so. At step 1108, the method 1100 further comprises determining the level of thetransmission error based on a K-level of the feature for which the mismatch is detected. At step 1110, the method 1100 furthercomprises determining that the data is successfully received if the level of the transmission error is below a relevance thresholdlevel and sending an acknowledgement response, ACK, in response thereto. At step 1112, the method 1100 further comprisesdetermining that the data is not successfully received if the level of the transmission error is above the relevance threshold level and sending a negative acknowledgement response, NACK, in response thereto. The steps 1102 to 1112 are only illustrative and other alternatives can also be provided where one or more steps are added, one or more steps are removed, or one or more steps are provided in a different sequence without departing from the scope of the claims herein. In one aspect, there is provided a computer program product comprising program instructions for performing the method 1100 when executed by one or more processors (e.g., the receiving controller 110 of the receiving device 104) in the distributed network 100. In another aspect, the present disclosure provides a non-transitory computer-readable medium having stored thereon, computer-implemented instructions that, when executed by a computer, causes the computer to execute operations of the method 1100.FIG. 12 illustrates an exemplary implementation scenario of neural networks based distributed Joint Inference-Source-Channel-Coding (JISCC) in a distributed network, in accordance with an embodiment of the present disclosure. FIG. 12 is described inconjunction with elements from FIGs.1, 2, 3, 4, 5, 6, 7A-7B, 8, 9, 10 and 11. With reference to FIG.12, there is shown an exemplary implementation scenario of a distributed network 1200 comprising a plurality of distributed sensing devices, for example, a first sensing device 1202A up to a Kth sensing device 1202K, connected to a fusion center 1204 through a wirelesschannel 1206. There is further shown a neural network at each of the plurality of distributed sensing devices, for example, afirst neural network 1208A for the first sensing device 1202A up to a Kth neural network 1208K for the Kth sensing device1202K and a neural network 1210 for the fusion center 1204. There is further shown a group of input vectors 1212 (may alsobe represented as ^) through which a plurality of input vectors (may be represented as ^^ , … , ^^) are obtained for each of theplurality of distributed sensing devices (i.e., 1202A, …, 1202K). There is further shown a sequence of operations 1214 to 1226. Each input vector is processed by the neural network comprised by each of the plurality of distributed sensing devices. The input to the semantic hashing function is not required to be raw sensing data (e.g., image pixels or text symbols). However, the semantic hashes are equally useful when applied on pre-processed data that is obtained, for example, by use of the neural networks at each sensing device. Furthermore, each neural network at each of the plurality of distributed sensing devices (i.e., 1202A, …, 1202K) and the neural network 1210 at the fusion center 1204 are jointly optimized to maximize the performance of predictions obtained at the fusion center 1204. This involves extracting relevant complementary features ^^at each sensing device (source coding), making the features robust against channel noise (channel coding), and performing inference ondistorted / partial information by the fusion center 1204 (inference coding).At operation 1214, the semantic hashing function is applied on the pre-processed data (i.e., ^^ , … , ^^) at each of the pluralityof distributed sensing device, that is the first sensing device 1202A up to the Kth sensing device 1202K. At operation 1218, encoding and signal modulation is applied on the pre-processed data and transmitted to the fusion center 1204 through the wireless channel 1206.At operation 1220, signal demodulation and decoding is applied on the received data at the fusion center 1204 for each of theplurality of distributed sensing device. At operation 1224, the same semantic hashing function is applied on the received data (i.e., ^^,…, ^^) to determine the check semantic key for each of the plurality of distributed sensing device. At operation 1226, a difference between the received semantic key and the check semantic key is computed to determine thepresence of the transmission errors in the received data for each of the plurality of distributed sensing device. The received data(i.e., ^^,…, ^^) is analyzed by the neural network 1210 at the fusion center 1204 for each of the plurality of distributedsensing device to determine a group of estimated input vectors 1228 (may also be represented as ^^). In the exemplary implementation scenario of the distributed network 1200, the semantic hashes enable the fusion center 1204 to decide whether the received data from each of the plurality of distributed sensing device (i.e., 1202A, …, 1202K) is either too noisy or to be useful. In case of determining the received data as too much noisy, the fusion center 1204 may ask to particular sensing devices for data re-transmission. In another case of the determining the received data to be useful, the fusion center 1204 may perform inference on a subset of received data. Modifications to embodiments of the present disclosure described in the foregoing are possible without departing from the scope of the present disclosure as defined by the accompanying claims. Expressions such as "including", "comprising", "incorporating", "have", "is" used to describe and claim the present disclosure are intended to be construed in a non-exclusive manner, namely allowing for items, components or elements not explicitly described also to be present. Reference to the singular is also to be construed to relate to the plural. The word "exemplary" is used herein to mean "serving as an example, instance or illustration". Any embodiment described as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments and / or to exclude the incorporation of features from other embodiments. The word "optionally" is used herein to mean "is provided in some embodiments and not provided in other embodiments". It is appreciated that certain features of the present disclosure, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the present disclosure, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable combination or as suitable in any other described embodiment of the disclosure.

Claims

CLAIMS1. A distributed network (100) comprising a sending device (102) and a receiving device (104), wherein the sending device(102) comprises a sending controller (108) configured toobtain data to be transmitted, wherein the data to be transmitted is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, determine a semantic key by applying a task-oriented mapping to the data, wherein the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmit the data to be transmitted and the semantic key to the receiving device (104), wherein the receiving device(104) comprises a receiving controller (110) configured to receive the data and the semantic key, determine a check semantic key by applying the same task-oriented mapping to the received data, compare the check semantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so, determine the level of the transmission error based on a K-level of the feature for which the mismatch is detected,and thendetermine that the data is successfully received if the level of the transmission error is below a relevance threshold level and send an acknowledgement response, ACK, in response thereto, and determine that the data is not successfully received if the level of the transmission error is above the relevance threshold level and send a negative acknowledgement response, NACK, in response thereto.

2. The distributed network (100) according to claim 1, wherein the receiving controller (110) is further configured to computea weighted L1 distance between the received semantic hash and the check semantic key as the level of the transmission error and determine that the level of the transmission error is above the relevance threshold level if the weighted L1 distance is larger than the relevance threshold level.

3. The distributed network (100) according to claim 2, wherein the weighted L1 distance is the distance between the semantic key and the check semantic key.

4. The distributed network (100) according to claim 2 or 3, wherein the receiving controller (110) is further configured tocompute the weighted L1 metric over finite fields to measure differences between the semantic key and the check semantickey, wherein there exist values d1 > d2 > … dK > 0 such that L1_distance (semantic key (i.e. mapping of data to be transmitted), check semantic key (i.e. mapping ofdata received)) >= d1 implies a difference at the level-1 feature between data to be transmitted and data received, d1 > L1_distance (semantic key, check semantic key) >= d2 implies a difference at the level-2 feature between data to be transmitted and data received, dK-1 > L1_distance (semantic key, check semantic key) >= dK implies a difference at the level-K featurebetween data to be transmitted and data received, and L1_distance (semantic key, check semantic key) = 0 implies full agreement of data to be transmitted anddata received with respect to the K-level hierarchy.

5. The distributed network (100) according to any preceding claim, wherein the negative acknowledgement response comprises a prompt indicating a distance between the semantic key and the check semantic key.

6. The distributed network (100) according to any preceding claim, wherein the acknowledgement response comprises a prompt indicating a distance between the semantic key and the check semantic key.

7. The distributed network (100) according to claim 5, wherein the prompt indicates an action for a retransmission of the data.

8. The distributed network (100) according to claim 6, wherein the prompt indicates an action for a future transmission of data.

9. The distributed network (100) according to any preceding claim, wherein the input class and the task to be performed are associated with the K-level hierarchy of features based on a Hierarchical Classification utilizing K-means.

10. The distributed network (100) according to any preceding claim, wherein the K-level hierarchy of features is based on atraining of a neural network utilized to learn the most relevant features of the K-level hierarchy of features,11. The distributed network (100) according to claim 10, wherein the K-level hierarchy of features is based on a training of aneural network (400), wherein the neural network (400) isa randomly initialized neural network having an encoder / decoder architecture, wherein a role of the encoder (402) isto map an input vector (406) into a latent feature vector of dimension K (408) wherein each dimension corresponds to one feature, and a role of the decoder (404) is to predict the true label from a latent vector, and wherein the encoder (402) is configured to receive an input vector (406) provide a K-dimensional feature vector to which arandom mask (410) is applied which randomly selects k < K and masks all lower k features of the feature vector, wherein thedecoder (404) is configured toreceive the resulted vector and provide a prediction of the true label, whereby an error vector is back-propagated tothe neural network (400), whereby weights are updated.

12. The distributed network (100) according to claim 11, wherein the random mask (410) is configured to erase informationfrom lowest features and preserving information in top features which encourages the encoder (402) to put more relevantinformation in top features.

13. The distributed network (100) according to claim 11 or 12, wherein the training of a neural network (400) further comprisesapplying uniform noise (412) to the input vector (406), whereby the encoder (402) is encouraged to maintain a smooth scalingof features.

14. The distributed network (100) according to any preceding claim, wherein the task-oriented mapping is based on receiving input data and outputting the semantic key being a short code of a predefined structure, wherein the short code depends on theinput data’s relevant features from the K-level hierarchy of features.

15. The distributed network (100) according to claim 14, wherein the task-oriented mapping is based on a natural hash format [level 1 feature value, level 2 feature value, …, level K feature value] 16. The distributed network (100) according to claim 14, wherein the task-oriented mapping is based on K-level nested lattices.

17. The distributed network (100) according to claim 14, wherein the task-oriented mapping is based on a learning using aneural network (500), utilizing discriminative learning, wherein a discriminative L1 loss is utilized to encourage the neuralnetwork (500) to assign close codes to similar features while assigning distant codes to dissimilar features in a following way:vectors with same level-1 features should be close to each other (and remaining far apart), and among these vectors, the vectors with same level-2 features should be close to each other.

18. The distributed network (100) according to any preceding claim, wherein the data to be transmitted is the result of ananalysis of raw data and wherein the analysis is based on a task to be performed and provides one or more features of the data to be transmitted.

19. The distributed network (100) according to any preceding claim, wherein the sending controller (108) is configured to obtain the data by receiving the data.

20. The distributed network (100) according to any preceding claim, wherein the sending controller (108) is configured toobtain the data by analyzing the raw data.

21. The distributed network (100) according to claim 20, wherein the raw data represents an image.

22. A method (700) for a distributed network (100) comprising a sending device (102) and a receiving device (104), the method(700) comprising: the sending device (102)obtaining data to be transmitted, wherein the data to be transmitted is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, determining a semantic key by applying a task-oriented mapping to the data, wherein the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmitting the data to be transmitted and the semantic key to the receiving device (104), wherein the method (700)further comprises the receiving device (104)receiving the data and the semantic key, determining a check semantic key by applying the same task-oriented mapping to the received data, comparing the check semantic key to the semantic key in order to determine if there is a transmission error bydetermining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so, determining the level of the transmission error based on a K-level of the feature for which the mismatch is detected, and then determining that the data is successfully received if the level of the transmission error is below a relevance threshold level and sending an acknowledgement response, ACK, in response thereto, and determining that the data is not successfully received if the level of the transmission error is above the relevance threshold level and sending a negative acknowledgement response, NACK, in response thereto.

23. The method (700) according to claim 22, wherein the method (700) further comprises the receiving device (108) computing a weighted L1 distance between the received semantic hash and the check semantic key as the level of the transmission error and determine that the level of the transmission error is above the relevance threshold level if the weighted L1 distance is larger than the relevance threshold level.

24. The method (700) according to claim 22 or 23, wherein the method (700) further comprises the receiving controller (110)computing the weighted L1 metric over finite fields to measure differences between the semantic key and the check semantic key, wherein there exist values d1 > d2 > … dK > 0 such that L1_distance (semantic key (i.e. mapping of data to be transmitted), check semantic key (i.e. mapping of data received)) >= d1 implies a difference at the level-1 feature between data to be transmitted and data received, d1 > L1_distance (semantic key, check semantic key) >= d2 implies a difference at the level-2 featurebetween data to be transmitted and data received, dK-1 > L1_distance (semantic key, check semantic key) >= dK implies a difference at the level-K featurebetween data to be transmitted and data received, and L1_distance (semantic key, check semantic key) = 0 implies full agreement of data to be transmitted anddata received with respect to the K-level hierarchy.

25. The method (700) according to claim 22, wherein the K-level hierarchy of features is based on a training of a neural network(400) utilized to learn the most relevant features of the K-level hierarchy of features and wherein the K-level hierarchy offeatures is based on a training of a neural network (400), wherein the neural network (400) isa randomly initialized neural network having an encoder / decoder architecture, wherein a role of the encoder (402) isto map an input vector (406) into a latent feature vector of dimension K (408) wherein each dimension corresponds to onefeature, and a role of the decoder (404) is to predict the true label from a latent vector, and wherein the encoder (402) is configured to receive an input vector (406) and provide a K-dimensional feature vector to whicha random mask (410) is applied which randomly selects k < K and masks all lower k features of the feature vector, wherein thedecoder (404) is configured toreceive the resulted vector and provide a prediction of the true label, whereby an error vector is back-propagated to the neural network (400), whereby weights are updated.

26. A sending device (102) comprising a sending controller (108) configured to obtain data to be transmitted, wherein the data to be transmitted is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, determine a semantic key by applying a task-oriented mapping to the data, wherein the task-oriented mappingprovides a mapping of the data to the semantic key based on K-level hierarchy of, andtransmit the data to be transmitted and the semantic key to the receiving device (104).

27. A method (900) for use in a sending device (102), the method (900) comprisingobtaining data to be transmitted, wherein the data to be transmitted is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, determining a semantic key by applying a task-oriented mapping to the data, wherein the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, and transmitting the data to be transmitted and the semantic key to the receiving device (104).

28. A receiving device (104) comprising a receiving controller (110) configured to receive data and a semantic key, wherein the data received is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, and wherein the semantic key is determined by applying a task-oriented mapping to the prior to transmission data, wherein the task-orientedmapping provides a mapping of the data to the semantic key based on K-level hierarchy of,determine a check semantic key by applying the same task-oriented mapping to the received data,compare the check semantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so, determine the level of the transmission error based on a K-level of the feature for which the mismatch is detected, and then determine that the data is successfully received if the level of the transmission error is below a relevance threshold level and send an acknowledgement response, ACK, in response thereto, and determine that the data is not successfully received if the level of the transmission error is above the relevance threshold level and send a negative acknowledgement response, NACK, in response thereto.

29. A method (1100) for use in a receiving device (104), the method (1100) comprising receiving data and a semantic key, wherein the data received is of an input class and relates to a task to be performed, wherein the input class and the task to be performed are associated with a K-level hierarchy of features, and wherein the semantic key is determined by applying a task-oriented mapping to the prior to transmission data, wherein the task-oriented mapping provides a mapping of the data to the semantic key based on K-level hierarchy of, determining a check semantic key by applying the same task-oriented mapping to the received data,comparing the check semantic key to the semantic key in order to determine if there is a transmission error by determining that there is a mismatch between the semantic key and the check semantic key for a feature of the data, and if so, determining the level of the transmission error based on a K-level of the feature for which the mismatch is detected, and then determining that the data is successfully received if the level of the transmission error is below a relevance threshold level and sending an acknowledgement response, ACK, in response thereto, and determining that the data is not successfully received if the level of the transmission error is above the relevance threshold level and sending a negative acknowledgement response, NACK, in response thereto.

30. A computer program product comprising program instructions for performing the method (700, 900, 1100) according toclaim 22, 27 or 29, when executed by one or more processors in a distributed network (100).

Citation Information

Patent Citations

  • Data processing method and device, server and storage medium

    CN112329933A

  • Semantic network data correction

    US20220224442A1

  • Semantic Communication and Programmable Protocol Stack: Media Access Control

    US20230413267A1