Encoding and decoding source information using distributed channel encoders and decoders

Abstract

A method and apparatus for encoding source information for transmission on a transmission channel is disclosed. The method includes causing a source encoder to generate a plurality of feature probability distributions representing aspects of the source information. The method also includes receiving the plurality of feature probability distributions at an input of a distributed channel encoder, where the distributed channel encoder is implemented using a polarized flow network. The method further includes causing the distributed channel encoder to convert the plurality of feature probability distributions into a plurality of distributed codewords of an increased dimension output for transmission over the transmission channel. Methods and apparatus for decoding the output plurality of distributed codewords to regenerate the source information are also disclosed.

Description

[0001] This application is a divisional application of the Chinese patent application No. 202180049625.5, filed on June 29, 2021, with the title of “Encoding and decoding source information using distributed channel encoders and decoders”, the whole content of the original application is incorporated herein by reference. BACKGROUND TECHNICAL FIELD

[0002] The present invention relates generally to wireless communications, and more specifically, to encoding source information using a distributed channel encoder for transmission over a communication channel, and decoding information received over the channel using a distributed channel decoder.

[0003] Description of Related Art

[0004] In a wireless communication system, information is encoded before transmission to prevent noise, distortion, and degradation over a wireless or wired transmission channel. A transmitter can encode k-bit source information into an n-bit codeword by adding redundant bits according to a predefined channel coding scheme. The codeword is then modulated onto a waveform and transmitted over the transmission channel. During transmission, the waveform is affected by channel noise. A decoder that receives the noisy codeword and is aware of the coding scheme is used to decode the source information from the noisy and distorted codeword.

[0005] Variational autoencoders (VAEs) provide a promising approach for encoding source information according to machine learning techniques. A VAE encoder outputs k features, where each feature is a Gaussian distribution defined by a mean and a variance. A probabilistic VAE decoder recovers the source information from a k-value vector, including k samples that have a mean and a variance associated with the Gaussian distribution. For VAEs, the assumption of independent and identically distributed (i.i.d.) for each information bit does not hold. According to the i.i.d. assumption, if one bit in an information block is decoded incorrectly, the entire information block must be discarded and retransmitted. Instead of ensuring that each bit is transmitted correctly, VAEs cause the k-value vector to draw samples from a Gaussian distribution that will conform to the desired mean and variance.

[0006] While VAEs can be implemented to improve performance, there is still a need for improvement to address some inefficiencies in traditional VAE encoding / decoding schemes. SUMMARY

[0007] According to one aspect disclosed, a method for encoding source information for transmission over a transmission channel is provided. The method includes causing a source encoder to generate a plurality of feature probability distributions representing aspects of the source information. The method also includes receiving the plurality of feature probability distributions at an input of a distribution channel encoder, wherein the distribution channel encoder is implemented using a polar stream network. The method further includes causing the distribution channel encoder to convert the plurality of feature probability distributions into a plurality of distribution codewords of an augmented output for transmission over the transmission channel.

[0008] The source encoder can include a source encoder neural network, and the method can further include training the source encoder neural network to generate the plurality of feature probability distributions in response to receiving a source information training set.

[0009] Training the source encoder neural network can include configuring a source decoder neural network to receive the plurality of feature probability distributions from the source encoder neural network at an input of the source decoder neural network, training the source encoder neural network and the source decoder neural network to reproduce the source information training set at an output of the source decoder neural network.

[0010] The training can include training the source encoder neural network and the source decoder neural network to objectives that maximize a likelihood between the output of the source decoder neural network and the source information training set, and that minimize a relative entropy between the output of the source encoder neural network and a target distribution.

[0011] The target distribution can be a Gaussian distribution.

[0012] The Gaussian distribution can include a multi-dimensional Gaussian distribution, and each feature probability distribution can include a vector of mean values and variance values.

[0013] The source encoder can be implemented in a transmitter, and the method can further include transmitting configuration information to a receiver, the configuration information defining the source decoder neural network and facilitating configuration of a source decoder at the receiver to decode subsequent transmissions from the transmitter.

[0014] Each feature probability distribution can include a mean value and a variance value.

[0015] The method can include training a neural network portion of the polar stream network to configure the polar stream network for performing the conversion between the feature probability distributions and the output plurality of distribution codewords.

[0016] The output plurality of distribution codewords can include a plurality of non-Gaussian distributions.

[0017] The source encoder can be configured to generate a k-dimensional feature probability distribution, and the polar stream network can be configured to convert the k-dimensional feature probability distribution received at an input of the polar stream network to an n-dimensional output of a plurality of distribution codewords, where k is less than n; and training the neural network portion of the polar stream network can include keeping excess input dimensions of the polar stream network constant to generate an augmented n-dimensional output of a plurality of distribution codewords.

[0018] The distribution channel encoder can be implemented in a transmitter, and the method can further include transmitting configuration information to a receiver, the configuration information including an identification of the excess input dimensions and information facilitating configuration of a reverse polar stream network to implement a decoder at the receiver.

[0019] Training the neural network portion of the polar stream network can include causing a receiver to train a plurality of polar stream network configurations, and the method can further include receiving one of the plurality of polar stream network configurations at the transmitter for performing the conversion between the feature probability distribution and the output plurality of distribution codewords.

[0020] The method can include receiving the plurality of distributions representing source information of the transmission at an input of a distribution channel decoder of a receiver, wherein the distribution channel decoder is implemented using a polar stream network; causing the distribution channel decoder to convert the plurality of distributions to a plurality of feature distributions of a reduced dimension output; causing a source decoder to regenerate the source information from the output plurality of feature probability distributions.

[0021] According to another aspect disclosed, a method for encoding source information for transmission over a transmission channel is provided. The method includes causing a first source encoder to generate a first plurality of feature probability distributions representing aspects of the source information. The method further includes rearranging the source information to generate converted source information; causing a second source encoder to generate a second plurality of feature probability distributions representing aspects of the converted source information. The method further includes causing a first distribution channel encoder to convert the first plurality of feature probability distributions to a first augmented dimension output of a plurality of distribution codewords for transmission over the transmission channel. The method further includes causing a second distribution channel encoder to convert the second plurality of feature probability distributions to a second augmented dimension output of a plurality of distribution codewords for transmission over the transmission channel.

[0022] The method can include receiving the first plurality of distributions representing the source information at an input of a first distribution channel decoder, wherein the first distribution channel decoder is implemented using a polar stream network; causing the first distribution channel decoder to convert the plurality of distributions to a plurality of feature distributions of a reduced dimension output; causing a first source decoder to regenerate the source information from the output plurality of feature probability distributions. The method can further include receiving a second plurality of distributions representing converted source information at an input of a second distribution channel decoder, wherein the second distribution channel decoder is implemented using a polar stream network; causing the second distribution channel decoder to convert the second plurality of distributions to a plurality of feature distributions of a reduced dimension output; causing a second source decoder to regenerate the converted source information from the output plurality of feature probability distributions. The method can further include performing iterative cancellation from the regenerated source information and the converted source information to generate a refined regenerated result of the source information.

[0023] According to another aspect disclosed, there is provided an encoder apparatus for encoding source information for transmission over a transmission channel. The apparatus includes a source encoder for generating a plurality of feature probability distributions representing aspects of the source information; a distribution channel encoder having an input for receiving the plurality of feature probability distributions, implemented using a polar stream network. The distribution channel encoder is for converting the plurality of feature probability distributions to a plurality of distribution code words of an increased dimension output for transmission over the transmission channel.

[0024] The encoder apparatus can be implemented within a transmitter for sending configuration information to a receiver, the configuration information facilitating configuring a polar stream network of the receiver to decode subsequent transmissions from the transmitter.

[0025] The transmitter can be associated with a sensor generating the source information, the encoder apparatus being operable to encode the source information for transmission to the receiver.

[0026] Other aspects and features of the present application will become apparent to those ordinarily skilled in the art upon review of the following detailed description in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0027] In the drawings illustrating the disclosed embodiments:

[0028] Figure 1A is a block diagram of an exemplary communication system provided by one disclosed embodiment;

[0029] Figure 1B is a block diagram of an exemplary electronic device provided by one disclosed embodiment;

[0030] Figure 1C This is a block diagram of an example base station provided in one of the disclosed embodiments;

[0031] Figure 2 This is a block diagram of a transmitter provided in one of the disclosed embodiments;

[0032] Figure 3 This is a block diagram of a receiver used to decode source information transmitted from the transmitter shown in Figure 1.

[0033] Figure 4 This is a block diagram of the training configuration, which is used to train the source encoder of the transmitter shown in Figure 1 and... Figure 2 The source decoder of the receiver shown;

[0034] Figure 5 This is a block diagram of the forward polarization stream network for the distributed channel encoder used in the transmitter shown in Figure 1, and a block diagram of the network used in... Figure 2 The diagram shows a block diagram of the inverse polarization stream network of the distributed channel decoder for the receiver.

[0035] Figure 6 This is a schematic diagram of a neural network, which is used to implement... Figure 5 The scaling and offset functions used in the polarization flow network shown;

[0036] Figure 7 This is a block diagram of the training configuration used to train the polarization stream network of the distributed channel encoder;

[0037] Figure 8 yes Figure 3 The diagram shows the configuration of the distributed channel decoder and source decoder of the receiver.

[0038] Figure 9 This is a simplified block diagram of an alternative encoder embodiment;

[0039] Figure 10 This is a block diagram of an alternative configuration for training polarimetric flow networks;

[0040] Figure 11 This is a block diagram of another disclosed embodiment, which implements a feedback loop for decoding encoded transmissions from an encoder. Detailed Implementation

[0041] Figure 1Ais an exemplary communications system 100 in which embodiments of the application can be implemented. Generally, the system 100 enables multiple wireless or wired elements to communicate data and other content. The system 100 can be designed to provide content (voice, data, video, text) through broadcast, narrowcast, user equipment to user equipment, etc. The system 100 can achieve efficient operation through sharing of resources, such as bandwidth.

[0042] In this example, the communications system 100 includes wireless communication networks 102 (including electronic devices (EDs) 110-114) and radio access networks (RANs) 120 and 122. The system 100 also includes a core network 130, a public switched telephone network (PSTN) 132, the Internet 134, and other networks 136. Although a certain number of these components or elements are shown in FIG. 1, the system 100 can include any reasonable number of these components or elements.

[0043] The EDs 110-114 are configured to operate and / or communicate in the system 100. For example, the EDs 110-114 are configured to transmit and / or receive signals through wireless communication channels. Each of the EDs 110-114 represents any suitable end-user device for wireless operation, and can include (or can be referred to as) a user equipment (UE), a wireless transmit / receive unit (WTRU), a mobile station, a mobile user unit, a cellular telephone, a station (STA), a machine type communication (MTC) device, a personal digital assistant (PDA), a smartphone, a laptop, a computer, a touchpad, a wireless sensor, a consumer electronics device, etc.

[0044] In Figure 1AIn particular embodiments, the RANs 120 and 122 include base stations 140 and 142, respectively. The base stations 140 and 142 are used to wirelessly interface a one or more of the EDs 110-114 to enable access to any of the other base stations, the core network 130, the PSTN 132, the Internet 134, and / or the other networks 136. For example, the base stations 140-142 can include (or be) one or more of a base transceiver station (BTS), a Node-B, an evolved NodeB (eNodeB), a Home eNodeB, a gNodeB (sometimes referred to as a “giga” NodeB), a transmission point (TP), a transmit / receive point (TRP), a site controller, an access point (AP), or a wireless router, among several other possibilities. Any of the EDs 110-114 can alternatively, or additionally, be used to connect, access, and / or communicate with any of the other base stations 140-142, the Internet 134, the core network 130, the PSTN 132, the other networks 136, or any combination or subset thereof. Optionally, the system can include a RAN, such as the RAN 120; wherein the respective base stations 140 access the core network 130 through the Internet 134.

[0045] The EDs 110-114 and the base stations 140-142 are examples of communication devices that can be used to implement some or all of the functionality and / or embodiments described herein. In particular embodiments, the EDs 110-114 and the base stations 140-142 can include one or more processors and / or other circuitry configured to perform various operations described herein. Figure 1AIn the illustrated embodiment, base station 140 forms part of RAN 120, which may include other base stations, one or more base station controllers (BSCs), one or more radio network controllers (RNCs), relay nodes, components, and / or devices. Any base station 140 or 142 may be a single element (as shown) or multiple elements distributed within a corresponding RAN or other device. Furthermore, base station 140 forms part of RAN 120, which may include other base stations, components, and / or devices. Each of base stations 140 to 142 may be used to operate within a specific geographic area or region to transmit and / or receive radio signals, where the geographic area or region is sometimes referred to as a "coverage area." Cells may be further divided into cell sectors, and base stations 140 to 142 may, for example, use multiple transceivers to provide services to multiple sectors. In some embodiments, base stations 140 to 142 may be implemented as piconodes or femtonodes, where radio access technologies support such nodes. In some embodiments, multiple transceivers can be deployed for each coverage area using multiple-input multiple-output (MIMO) technology. (See Appendix) Figure 1A The number of RANs 120 and 122 shown is merely exemplary. Any number of RANs can be considered when designing the system 100.

[0046] The base stations 140 to 142 communicate with one or more of the EDs 110 to 114 over one or more air interfaces 150 and 152 using wireless communication links (e.g., RF, μWave, IR, etc.). The air interfaces 150 and 152 can utilize any suitable wireless access technology. For example, the system 100 can implement one or more channel access methods in the air interfaces 150 and 152, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal frequency division multiple access (OFDMA), or single-carrier frequency division multiple access (SC-FDMA).

[0047] The base stations 140 to 142 can implement Universal Mobile Telecommunication System (UMTS) Terrestrial Radio Access (UTRA) to establish air interface 150 using wideband CDMA (WCDMA). In this regard, the base stations 140 to 142 can implement HSPA, HSPA+, and / or the like, optionally including HSDPA and / or HSUPA. Alternatively, the base stations 140 to 142 can implement LTE, LTE-A, and / or LTE-B to establish air interface 150 with Evolved UMTS Terrestrial Radio Access (E-UTRA). It is contemplated that the system 100 can use multiple access technologies including schemes as described above. Other wireless technologies for implementing the air interface include IEEE 802.11, 802.15, 802.16, CDMA1800, CDMA1800 1X, CDMA1800 EV-DO, IS-1800, IS-95, IS-856, GSM, EDGE, and GERAN. Of course, other multiple access schemes and wireless protocols can be utilized.

[0048] The RANs 120 and 122 are in communication with the core network 130 to provide the EDs 110 to 114 with access to various services, such as voice, data, and other services. It will be appreciated that the RANs 120 and 122 and / or the core network 130 can be in direct or indirect communication with one or more other RANs (not shown) that can or can not be served by the core network 130 and can or can not utilize the same radio access technology as the RANs 120 and / or 122. The core network 130 can also serve as a gateway for the RANs 120 and 122 and / or the EDs 110 to 114 to access other networks (such as the PSTN 132, the Internet 134, and the other networks 136). In addition, some or all of the EDs 110 to 114 can include functionality for communicating with different wireless networks on different wireless links using different wireless technologies and / or protocols. The PSTN 132 can include a circuit- switched telephone network that provides plain old telephone service (POTS). The Internet 134 can include computer networks and / or sub-networks (intranets) that contain protocols such as IP, TCP, and UDP. The EDs 110 to 114 can be multi-mode devices capable of operating according to multiple wireless access technologies and contain multiple transceivers necessary to support such technologies.

[0049] The RANs 120 and 122, the base stations 140 and 142, and the core network 130 can be collectively referred to as "network equipment". The network equipment components can be physically distributed within the coverage area. The core network 130 typically includes computer processor hardware connected between the PSTN 132, the Internet 134, the other network 136, and the RANs 120 and 122 to provide services to the EDs 110 to 114.

[0050] Figure 1B and Figure 1C It shows the implementation Figure 1A The network 102 shown is an exemplary device that can be used. Specifically, Figure 1B An example of ED 160 is shown. Figure 1C An exemplary base station 180 is shown. These components can be used in communication system 100 or any other suitable system.

[0051] like Figure 1B As shown, the ED 160 includes at least one processing unit 162. The processing unit 162 implements various processing operations of the ED 160. For example, the processing unit 162 may perform signal encoding, data processing, power control, input / output processing, or any other function that enables the ED 160 to operate within the communication system 100. The processing unit 162 may also be used to implement some or all of the functions and / or embodiments described in more detail elsewhere herein. Each processing unit 162 includes any suitable processing or computing device for performing one or more operations. For example, each processing unit 162 may include a microprocessor, microcontroller, digital signal processor, field-programmable gate array, or application-specific integrated circuit, etc.

[0052] The ED 160 also includes at least one transceiver 164. The transceiver 164 is used to modulate data or other content for transmission via at least one antenna or Network Interface Controller (NIC) 166. The transceiver 164 is also used to demodulate data or other content received by the at least one antenna 166. Each transceiver 164 includes any suitable structure for generating signals for wireless or wired transmission and / or for processing signals received wirelessly or wiredly at the antenna 166. Each antenna 166 includes any suitable structure for transmitting and / or receiving wireless or wired signals 172. One or more transceivers 164 may be used in the ED 160. One or more antennas 166 may be used in the ED 160. Although the transceiver 164 is shown as a single functional unit, it may also be implemented using at least one transmitter and at least one separate receiver.

[0053] The ED 160 also includes one or more input / output devices 168 or interfaces (e.g., to...). Figure 1A (The wired interface of Internet 134 is shown). Input / output device 168 can interact with users or other devices in the network. Each input / output device 168 includes any suitable structure for providing or receiving information from a user, such as a speaker, microphone, keypad, keyboard, display, or touchscreen, including network interface communication.

[0054] Furthermore, the ED 160 includes at least one memory 170. The memory 170 stores instructions and data used, generated, or collected by the ED 160. For example, the memory 170 may store software instructions or modules for implementing some or all of the functions and / or embodiments described herein and executed by one or more processing units 162. Each memory 170 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices. Any suitable type of memory can be used, such as random access memory (RAM), read-only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, etc.

[0055] like Figure 1C As shown, the base station 180 includes at least one processing unit 182, at least one transmitter 184, at least one receiver 186, one or more antennas 188, at least one memory 190, and one or more input / output devices or interfaces 192. Transceivers (not shown) may be used in place of the transmitter 184 and the receiver 186. A scheduler 194 may be coupled to the processing unit 182. The scheduler 194 may be included within the base station 180 or may operate independently of the base station 180. The processing unit 182 implements various processing operations of the base station 180, such as signal encoding, data processing, power control, input / output processing, or any other functions. The processing unit 182 may also be used to implement some or all of the functions and / or embodiments detailed herein. Each processing unit 182 includes any suitable processing or computing device for performing one or more operations. For example, each processing unit 182 may include a microprocessor, microcontroller, digital signal processor, field-programmable gate array, or application-specific integrated circuit, etc.

[0056] Each transmitter 184 includes any suitable structure for generating signals for wireless or wired transmission to one or more EDs or other devices. Each receiver 186 includes any suitable structure for processing signals received from one or more EDs or other devices via wireless or wired transmission. Although illustrated separately, at least one transmitter 184 and at least one receiver 186 can be combined into a transceiver. Each antenna 188 includes any suitable structure for transmitting and / or receiving wireless or wired signals 172. Although a common antenna 188 is shown here as being coupled to both the transmitter 184 and the receiver 186, one or more antennas 188 can be coupled to the one or more transmitters 184, and one or more separate antennas 188 can be coupled to the one or more receivers 186. Each memory 190 includes any suitable volatile and / or non-volatile storage and retrieval devices, such as those described above in connection with the ED 160 shown in FIG. 1. The memory 190 stores instructions and data used, generated, or collected by the base station 180. For example, the memory 190 could store software Figure 1B instructions or modules for implementing part or all of the functionality and / or embodiments described herein and for execution by the one or more processing units 182.

[0057] Each input / output device 192 can interact with a user or other devices in the network. Each input / output device 192 includes any suitable structure for providing information to or receiving / providing information from a user, including network interface communications.

[0058] It should be understood that the functions of the various elements shown or described in connection with the embodiments herein can be provided within Figures 1A-1COne or more steps of the example methods provided herein can be performed by corresponding units or modules. For example, a signal can be transmitted by a transmitting unit or module. A signal can be received by a receiving unit or module. A signal can be processed by a processing unit or module. Other steps can be performed by machine learning (ML) modules in the transmitting and receiving modules. The corresponding units / modules can be implemented using hardware, one or more components of an executing software, or a combination thereof. For example, one or more units / modules can be an integrated circuit, such as an application-specific integrated circuit (ASIC) or a field programmable gate array (FPGA). It will be understood that if the above modules are implemented by software executing on a processor, the modules can be retrieved from memory by the processor as needed, individually or collectively, retrieved for processing by the processor as needed in one or more instances, and the modules themselves can include instructions for further deployment and instantiation. Other details of the EDs 160 and 180, etc., are known to those skilled in the art. Thus, for the sake of brevity, these details are omitted here.

[0059] Generally, the transmission of information from the EDs 110-114 will adhere to an information transmission scheme that defines characteristics of the transmitted signals (e.g., channel coding, modulation, and waveforms). The EDs 110-114 can have some degree of autonomy to adjust or select the transmission scheme. In other embodiments, the EDs 110-114 can rely on the BS 140 and / or other components of the core network 130 to issue commands or messages over a wireless downlink channel to adjust or select the information transmission scheme.

[0060] In one embodiment, one or more of the EDs 110-114 can be implemented as sensors in a machine-to-machine communication system for collecting information about the object of interest 154. Examples of sensors can include Internet of Things (IoT) devices, cameras, smart meters, wireless sensors, machine-type communication devices, or other information collecting devices. For example, the EDs 110-114 can be implemented as camera sensors that capture image information of the object of interest 154 from different angles. Thus, each of the sensors collects image information and transmits the information over a wireless channel to the base station 140 via the wireless connections 150, 152.

[0061] Reference is made to Figure 2A block diagram of a transmitter that can be implemented in one of the EDs 110-114 is shown generally at 200. The transmitter 200 includes an encoder 202 and a quantization block 204. The encoder 202 includes a source encoder 206 having an input 208 for receiving source information S M . The source information can be information captured by a sensor associated with one of the EDs 110-114, as described above. In other embodiments, the source information can be other information related to human-based communications, etc. The source encoder 206 has a plurality of outputs 210 (T1-T k ) for generating a plurality (k) of feature probability distributions 212. The feature probability distributions 212 represent aspects of the source information. In the illustrated embodiment, each output 210 generates an output pair T i (where i is an index of the i-th of the k features). Each output T i includes an expected value 214 and a variance 216 associated with the respective feature distribution 212 (i.e., T i = <m i , σ i >). Thus, the source encoder 206 outputs k output pairs as samples from the feature distributions 212. In the present disclosure, the value of the variance 216 is denoted by the value σ, which is the square root of the variance.

[0062] The encoder 202 also includes a distribution channel encoder 218 that includes n outputs 222 (C1-C n ) and k inputs 220 for receiving the plurality (k) of feature probability distributions 212. The distribution channel encoder 218 converts (or encodes) the plurality of feature probability distributions 212 into a plurality (n) of distributions 224 of the outputs at the outputs (C1-C n ) of the distribution channel encoder. Each distribution can be sampled to produce outputs, each output represented by an output pair that includes an expected value m and a variance σ for the single-dimensional distributions shown at 212. In other embodiments, the feature probability distributions 212 can be multi-dimensional probability distributions, and the outputs can be represented by vector outputs (or matrices) that include a plurality of expected values m and variances σ.

[0063] The transformation performed by the distributed channel encoder 218 is an increase dimensionality transformation (i.e., n > k). In the illustrated embodiment, the distributed channel encoder 218 receives k Gaussian feature distributions 212 at the input 220 and generates n Gaussian output distributions at the output 222. In other embodiments, the output plurality of distributions 224 can be non-Gaussian probability distributions. The distributed channel encoder 218 is essentially a signal manifold transformer that transforms an input signal manifold (i.e., the plurality of feature distributions 212) to a target manifold (i.e., the output plurality of distributions 224).

[0064] The output pairs <m, σ> associated with the plurality of distributions 224 are sampled and then combined into a real-valued block (as shown at 226). The quantization block 204 performs quantization of the real-valued block 226 and generates a transmission waveform 228 that represents the encoded source information S M The transmission waveform 228 is transmitted over a transmission channel 230. The transmission channel 230 can be a wireless transmission channel, a wired transmission channel, or other physical transmission channel. The transmission channel 230 is typically subject to noise that degrades the transmission waveform 228 during its propagation over the channel.

[0065] In a conventional variational autoencoder implementation, the feature distributions 212 would be sampled and the expected values can be directly combined into a real-valued block and quantized for transmission over the transmission channel 230. In embodiments that use the transmitter 200 to transmit IoT source information, the source information content is unlikely to change significantly over time. Thus, the source encoder 206 can be used to extract the k features for subsequent transmissions without reconfiguration. However, even if the source information content remains regularized, the extracted k features <m, σ> can change over time. For example, one output pair of extracted features can be <m = 3.5, σ = 0.4> and another output pair can be <m = 3.0, σ = 0.9>. Typically, the range of the output pairs is found by observing a large number of samples and quantizing the features according to the worst case. For example, the range of the expected values m can be 0.4 to 3.4 and the range of the variances σ can be 0.2 to 0.9. In this example, it is feasible to perform quantization according to the expected value m = 3.4 and the variance σ = 0.9. However, this worst case quantization has the effect of reducing the efficiency of the transmission. The worst case quantization also ignores the different variances associated with different extracted features.

[0066] Figure 2The illustrated embodiment enables the distributed channel encoder 218 to directly encode the feature distribution 212 extracted by the source encoder, thus avoiding the aforementioned problems. Therefore, the encoder 202 expands the k-dimensional distribution 212 into an n-dimensional output distribution 224.

[0067] refer to Figure 3 A block diagram of the receiver is generally shown at position 300. For example, the receiver 300 can be... Figure 1A The receiver 300 is implemented within one of the base stations 140 or 142 shown. The receiver 300 includes a waveform receiver 302 and a decoder 304. Figure 2 The waveform 228 shown is transmitted on the transmission channel 230 and received as waveform 306 at the waveform receiver 302. During transmission on the channel 230, waveform 306 may have degraded and therefore may differ from the transmitted waveform 228. The waveform receiver 302 generates a plurality of distributions 310 representing the source information S encoded in the transmitted waveform 228. M The distribution 310 typically corresponds to the output plurality of distributions 224. Due to noise associated with the transmission channel 230, the expected value m of the plurality of distributions 310 may differ from the received expected value of the output plurality of distributions 224.

[0068] The receiver 300 includes a distributed channel decoder 312, which has multiple input terminals 314 for receiving the plurality of distributions 310 and multiple output terminals 316 for reproducing the feature distribution 318. The distributed channel decoder 312 converts (or decodes) the plurality of distributions 310 into forms substantially corresponding to... Figure 2 The corresponding feature distribution shown is 212(C1to C). n The dimensionality reduction output of multiple feature distributions 318(C′1 to C′) n ).

[0069] The decoder 304 further includes a source decoder 320, which includes an input terminal 322 for receiving the plurality of feature distributions 318 and a source information S for generating the source information S. M The regenerated result is output at terminal 324. In response to receiving the plurality of feature distributions 318 at input terminal 322, the source decoder 320 generates regenerated source information S′ at output terminal 324. L The received expected value remains at the original variance σ1 to σ2 associated with the corresponding feature distribution 212. k Within the specified range, the regenerated source information S′ M It should roughly correspond to the source information S. M Therefore, the aforementioned pair <mi ,σ i The sample values m within the described distribution should be sufficient for the decoder to recover the source information. The feature distributions 212 and 318 are often referred to as being in the “latent space” because these distributions represent intermediate outputs associated with features that are extracted from the source information. Similarly, the distributions 224 and 318 represent latent signal manifolds.

[0070] Thus, the transmitter 200 extends the k-dimensional Gaussian feature manifold to an n-dimensional Gaussian manifold by adding (n-k) redundant dimensions, thereby implementing the dimension- increasing converter as a distributed channel encoder. The coding rate of the encoding scheme will be k / n, where n>k. The integration of the source encoder 206 and the distributed channel encoder 218 breaks the assumption of independent and identically distributed (i.i.d.) variables of the encoder 202. Removing the i.i.d. assumption enables improved encoder / decoder configurations, as detailed below.

[0071] The source encoder 206 is implemented as a probabilistic encoder that encodes the source information S M into the respective feature distribution 212. Similarly, the source decoder 320 is implemented as a probabilistic decoder that aims to recover the source information S M (as S' M ) from the plurality of feature distributions 318. The source encoder 206 and the source decoder 320 can be implemented using an encoder neural network such as a deep neural network (DNN). The neural network of the source encoder 206 can be trained to extract k elementary features from the source information S M , where the extracted features are typically independent of each other. The source decoder 320 can be similarly trained to recover the source information from the k elementary features. In one embodiment, the source encoder 206 and the source decoder 320 can be jointly optimized to maximize the likelihood between S M and S' M (for example by minimizing the L2 distance). The joint optimization can be performed using a machine learning based approach.

[0072] Each of the source encoder 206 and the source decoder 320 can implement a variational autoencoder (VAE) algorithm. A VAE is a neural network configuration that operates on source information with a base probability distribution, such as a Gaussian distribution. Configurations suitable for VAE source encoders and source decoders are described in commonly owned U.S. Patent Application No. 16 / 428,396, entitled “METHODS AND APPARATUSES FOR FEATURE-DRIVEN MACHINE-TO-MACHINE COMMUNICATIONS,” filed May 31, 2019, the entire contents of which are incorporated by reference in the present application.

[0073] Referring to Figure 4 A block diagram of a training configuration is generally shown at 400. In the training configuration 400, each of the source encoder 206 and source decoder 402 can implement a VAE algorithm for encoding and decoding source information. The source decoder 402 is for the source decoder 300 corresponding to the receiver 320. However, for training the network, in the embodiment, the source decoder 402 is implemented at the transmitter 200.

[0074] The source encoder 206 is for receiving source information S M and has an output 406 for producing an output pair <m i ,σ i > for i = 1..3 in this training example. Thus, the source encoder 206 extracts three features to represent the source information training set. In other embodiments, a greater number of features can be extracted. The source decoder 402 has k inputs 408 corresponding to the k outputs 406 of the source encoder 206. Each of the outputs 406 of the source encoder 206 is connected directly to the inputs 408 of the source decoder 402 through summing blocks 410, thus ignoring the transmission channel 230 in terms of training. The source decoder 402 is for outputting a reproduced version of the source information S M as S' M .

[0075] The source encoder 206 is for extracting important features from the source information S M . As a result, the reproduced source information S' M may be degraded. In this training embodiment, the source information S MThe training set is based on the Modified National Institute of Standards and Technology (MNIST) database for handwritten digits, which are commonly used to train image processing systems. Each character provides 28x28 pixels as training input; therefore, in this embodiment, the source encoder 206 may include 784 input terminals for receiving the pixels as training input. Various other training sets may be used depending on the type of source information to be encoded.

[0076] The training of the encoder 206 and the decoder 402 can be performed to achieve the output S′ of the source decoder 402. M and the source information training set S M The primary objective is to maximize the likelihood (ML) between the source information and the source encoder 206. Training is performed based on likelihood maximization to ensure the reliability of the regenerated source information. The ML process includes determining a set of parameters for the source encoder 206 and the source decoder 402 to generate a training set S corresponding to the source information. M The regeneration of output S′ M In one embodiment, the ML process may include: causing the output S′ of the source decoder 402 to... M and the source information training set S M Minimize the L2 distance between them.

[0077] The training of the encoder 206 and the decoder 402 can also be performed to achieve a second objective of minimizing the relative entropy between the output of the feature encoder neural network and a target distribution. In one embodiment, this can comprise minimizing the Kullback-Leibler divergence (KL divergence) between the k latent features and a target Gaussian distribution. Minimizing the KL divergence between k features in a k-dimensional Gaussian manifold causes each feature to be as close as possible to the target Gaussian distribution. In this embodiment, the training configuration 400 comprises a multiplier block for the summation block 410. For example, the <m1, σ1> output is connected to a summation block 412, and the summation block 412 is connected to a multiplication block 414. For each of the outputs 406, the remaining summation and multiplication blocks are similarly connected. In this embodiment, for each output 406, the objective of minimizing the KL divergence is achieved by multiplying the variance σ by a standard normal distribution 416 (N(0, 1)). In this embodiment, the standard normal distribution provides the shape of the target distribution, and has a zero mean m and unit variance σ = 1. For the example of the output <m1, σ1>, the standard normal distribution 416 is first multiplied by σ1to expand the variance of the standard normal distribution to σ1. The resulting expanded normal distribution is then shifted by adding m1in the summation block 412. Adding m1shifts the mean of the standard normal distribution from zero to m1. The same operation is performed for the <m2, σ2> and <m3, σ3> outputs of the source decoder 402.

[0078] In this embodiment, the transmitter 200 performs the training of the source encoder 206 and the source decoder 402. The resulting neural network weights and biases determined for the source decoder are output as configuration information 418. The source decoder configuration information 418 can be transmitted in the transmission channel 230 or by another control channel to the receiver 300. The receiver 300 is then able to configure the source decoder 320 for decoding the received feature distribution 318. Thus, in the example where the transmitter is an IoT sensor, one of the EDs 110 to 114 can perform the training. The ED can then provide the necessary configuration information to the base station 140 to configure a source decoder for decoding the transmitted source information. In this embodiment, the configuration information can be transmitted securely to the base station 140, which has the effect of keeping the transmission of the source information private. Thus, the source information is only decodable by a receiver that has received the configuration information.

[0079] In conventional implementations of the source encoder 206, the feature distributions 212 would be quantized into bits and combined into blocks of bits for transmission to the receiver 300. In Figure 2 In the illustrated encoder embodiment, rather than quantizing the feature distributions for transmission over the transmission channel 230, the feature distributions are directly decoded by the distribution channel encoder 218. Thus, the distribution channel encoder 218 takes as input a k-dimensional feature distribution and outputs an n-dimensional output distribution. As described above, the distribution channel encoder 218 performs an up- dimensional transformation (i.e., k > n) that transforms the input signal manifold (i.e., the plurality of feature distributions 212) to a target manifold (i.e., the output plurality of distributions 224).

[0080] The distribution channel encoder 218 can be implemented using a polar flow network. Details of configuration and training of a forward polar flow network and a reverse polar flow network are described in commonly owned U.S. Patent Application No. 16 / 562,045, filed September 5, 2019, entitled “A METHOD AND APPARATUS FOR WIRELESS COMMUNICATION USING POLARIZATION-BASED SIGNAL SPACE MAPPING,” the entirety of which is incorporated by reference herein. The polar flow network provides a framework for transformation between any two multidimensional signal manifolds of the same dimensionality. The polar flow network does not require a complete mathematical description of the source or target signal manifolds.

[0081] Referring to Figure 5 wherein configuration of a forward polar flow network 500 and a reverse polar flow network 520 are schematically illustrated in generalized form. Each of the forward polar flow network 500 and the reverse polar flow network 520 includes a cascade of n stages that transform between an input signal X and an output signal Y of the same dimensionality. The input signal X can include a plurality of information bits representing a complex-valued signal. Each stage in the stage cascade is similarly configured and includes a rearrangement block 502, a split block 504, a scaling function 506, and an offset function 508. The rearrangement block 502 implements a rearrangement or permutation function that receives the set of input bits X and outputs a set of rearranged or permuted bits. For example, rearrangement or permutation can be performed according to a permutation table. The split block 504 implements a split function that splits the rearranged bits into a first set of rearranged bits X1 (1) and a second set of rearranged bits X2 (1) By including the scaling function 506 and the offset function 508, each stage in the cascade can be written as:

[0082]

[0083] where the x operator is the dot product (or more colloquially, element-wise multiplication) and the + operator represents an element-wise addition operation between two operand vectors. Thus, each stage of the forward polarization stream network 500 rearranges the input signal Y and splits the rearranged information into a first set of information and a second set of information The scaling function v(·) 506 is applied to the second set of information to generate a scaling vector The offset function u(·) 508 is applied to the second set of information to generate an offset vector Both the scaling vector and the offset vector have a dimension size equal to the first set of information The first set of information is multiplied element-wise by the scaling vector and the offset vector is added element-wise to obtain the output of each stage The output is a copy of the second set of information

[0084] The backward polarization stream network 520 operates on an input signal Y and produces an output signal X. The x operator is replaced by a dot division (or more colloquially, element-wise division) and the + operator is replaced by an element-wise subtraction operation between two operand vectors. The scaling function v(·) 512 and the offset function u(·) 514 are unchanged from the forward polarization stream network 500. Thus, each stage in the backward polarization stream network 520 rearranges the input signal and splits the rearranged information into a first set of information and a second set of information The scaling function v(·) 512 is applied to the second set of information to generate a scaling vector and the offset function u(·) 514 is applied to the second set of information to generate an offset vector Both the scaling vector and the offset vector have a dimension size equal to the first set of information. Thus, the output of each stage is as follows:

[0085]

[0086] The second output is a copy of the set of information. The first set of output information is obtained by subtracting the offset vector element-wise from the first set of information and then dividing by the scaling vector element-wise.

[0087] ​Since common elements appear in both the forward polarization flow network 500 and the reverse polarization flow network 520, it is only necessary to train either the forward network or the reverse network. For example, if the forward network 500 is to be trained, the reverse network 520 can be readily obtained by swapping the multiplication and division operators, and the addition and subtraction operators, and interchanging the inputs and the outputs. Adding the scaling function v(·) and the offset function u(·) to the polarization flow networks 500 and 520 reduces the reliability of certain points or dimensions, while increasing the reliability of other points or dimensions. In summary, the scaling function v(·) and the offset function u(·) provide flexibility in polarizing the reliability in the signal space to shape the input X to the output Y.

[0088] In one embodiment, the scaling function v(·) and the offset function u(·) can be implemented as neural networks within the forward polarization flow network 500 and the reverse polarization flow network 520. Figure 6 An example of a neural network portion for implementing the scaling function v(·) and the offset function u(·) is shown at 600. The neural network 600 includes multiple sets of fully connected nodes 602 in multiple layers, including an input layer 604, an output layer 606, and multiple hidden layers 608. The input layer 604 includes nodes that receive the second set of vectors The output layer 606 includes output nodes that provide the scaling function v(·) and the offset function u(·) for each stage in the stage cascade of the forward polarization flow network 500 and the reverse polarization flow network 520. Each arrow connecting the nodes can have an associated weighting factor w i .

[0089] Referring back to Figure 7 , a block diagram of a training configuration for training the polarization flow network of the distributed channel encoder 218 is shown generally at 700. In this example, a trained source encoder 206 receives a source information training set S M The source encoder 206 extracts 3 feature distributions 704 for representing the source information S M (i.e., k = 3). The distributed channel encoder 218 includes a forward polarization flow network f(·) shown as block 710 in Figure 7 Generally, the forward polarization flow network f(·) can be trained by minimizing the mean squared error between the output of the forward polarization flow network f(·) and the source information S Figure 5As shown at position 500, the forward polarized flow network 710 is configured. In this example, the forward polarized flow network 710 is used for 6 input dimensions. To maintain reversibility, the forward polarized flow network 710 also has 6 output dimensions. The distributed channel encoder 218 further includes a box 706 for adding frozen dimensions to the feature distribution 704 to form the 6 input dimensions, as shown at position 708. In this example, box 706 adds three input dimensions folded to zero values ​​(or some other constant values). The folded dimensions and the extracted feature distribution 704 form the input 708 of the forward polarized flow network 710.

[0090] For the training of the forward polarized flow network 710, a target 6-dimensional signal manifold 712 is established. In some embodiments, the target manifold 712 may be thickened to train the distributed channel encoder 218 to tolerate some noise. The training process results in the determination of weights and biases for the neural network portion of the forward polarized flow network 710. In this embodiment, only the forward polarized flow network 710 of the distributed channel encoder 218 is trained.

[0091] After training, box 710 outputs distributed channel decoder configuration information 714. The configuration information 714 includes the weights and biases of the scaling function v(·) and the offset function u(·). Figure 5 The configuration information 714 includes the number and configuration of the polarization stages (as shown), and details of the implemented rearrangement function. The configuration information 714 may include the configuration of the forward polarization flow network 710, and allow the receiver 300 to reverse the forward polarization flow network. Alternatively, the configuration information 714 may include the derived configuration of the reverse polarization flow network. Block 706 also outputs an identifier of the frozen dimension to block 714. The configuration information can be transmitted to the source decoder 320 via the transmission channel 230 or via another existing communication channel between the transmitter 200 and the receiver 300.

[0092] refer to Figure 8 The distributed channel encoder 312 and the source decoder 320 are generally shown at 800 (e.g., Figure 3 The block diagram shows various aspects of the configuration. The distributed channel decoder 312 includes a reverse polarization flow neural network f -1 (·), shown as box 802. The reverse polarization flow network 802 can be roughly arranged as follows: Figure 5The output of the distribution channel encoder 218 is an n-dimensional Gaussian target manifold, including the output multiple distributions 224. The transmission waveform 228 is limited over the transmission channel 230 according to the Shannon-Hartley channel capacity theorem. The encoding step before the samples are combined into the information bit block at 226 is not subject to the Shannon-Hartley capacity limit, which assumes that each information bit is independent and identically distributed (i.i.d.). The encoder 202 is configured to accept the inherent unequal importance in the k feature probability distributions 212 by encoding the feature probability distributions directly as the output multiple distributions 224, rather than quantizing the feature probability distributions before encoding. Referring to FIG. 8, the distribution channel encoder configuration information 714 is shown configured at 520, and includes 6 input dimensions and 6 output dimensions in this example. The distribution channel encoder configuration information 714 received from the transmitter 200 is used to configure the reverse polar stream network 802. Because the weights and biases of the forward polar stream network 710 have been trained, there is no need for further training in the receiver. The distribution channel decoder 312 also includes a block 804 for removing the frozen dimensions identified in the configuration information 714. For example, the reverse polar stream network 802 converts a plurality of distributions 806 inputs to a plurality of outputs 808. The outputs include a “0” value and the three frozen dimensions represented by three feature distributions. The block 804 uses the configuration information 714 to identify and discard the frozen distributions, and outputs 3 remaining distributions as shown at 812.

[0093] The source decoder, also implemented using a neural network, is configured using the source decoder configuration information 418, which is sent by the transmitter at 416 above in connection with Figure 4 determined in the training exercise described above. Once configured, the distribution channel decoder 312 and the source decoder 320 are ready for operation in the receiver 300, as shown at 818. Figure 3 The source decoder 320 receives the 3 remaining dimensions 812 and decodes the distribution to recover the source information as S' = S M .

[0094] In the above embodiment, the frozen dimensions are discarded. However, in an alternative embodiment, a successive cancellation algorithm can be used to decode the n-dimensional vector including the frozen dimensions. Thus, rather than discarding the (n-k) frozen dimensions, these dimensions can be used to help decode the k useful dimensions.

[0095] Referring again to Figure 2 , the output 222 of the distribution channel encoder 218 is an n-dimensional Gaussian target manifold, including the output multiple distributions 224. The transmission waveform 228 is limited over the transmission channel 230 according to the Shannon-Hartley channel capacity theorem. The encoding step before the samples are combined into the information bit block at 226 is not subject to the Shannon-Hartley capacity limit, which assumes that each information bit is independent and identically distributed (i.i.d.). The encoder 202 is configured to accept the inherent unequal importance in the k feature probability distributions 212 by encoding the feature probability distributions directly as the output multiple distributions 224, rather than quantizing the feature probability distributions before encoding. Referring to FIG. 8, the distribution channel encoder configuration information 714 is shown configured at 520, and includes 6 input dimensions and 6 output dimensions in this example. The distribution channel encoder configuration information 714 received from the transmitter 200 is used to configure the reverse polar stream network 802. Because the weights and biases of the forward polar stream network 710 have been trained, there is no need for further training in the receiver. The distribution channel decoder 312 also includes a block 804 for removing the frozen dimensions identified in the configuration information 714. For example, the reverse polar stream network 802 converts a plurality of distributions 806 inputs to a plurality of outputs 808. The outputs include a “0” value and the three frozen dimensions represented by three feature distributions. The block 804 uses the configuration information 714 to identify and discard the frozen distributions, and outputs 3 remaining distributions as shown at 812. Figure 9A simplified block diagram of an alternative encoder that extends the non-Shannon processing of the source information S M is shown generally at 900. The encoder 900 includes the source encoder 206 to extract k features from the source information S M . A distributed channel encoder 902 is used to implement a forward polar stream network configured as shown generally at 500 of Figure 5 . The distributed channel encoder 902 is used to convert the k extracted features 904 to an n-dimensional output 906. However, in this embodiment, the forward polar stream network of the distributed channel encoder 902 is trained to generate a non-Gaussian target manifold. An example of a standard orthogonal non-Gaussian target manifold is shown at 908. An example of a non-standard orthogonal non-Gaussian target manifold is shown at 910.

[0096] The target manifold 908 or 910 can be used to adapt to a transmission channel 230 having non-Gaussian channel conditions. For example, the attenuation of the transmission channel 230 can be distributed unevenly. The target manifold 908 or 910 can be used to avoid areas of the channel having greater attenuation and transmit the information over areas of the transmission channel having less attenuation.

[0097] In the training example shown at Figure 7 , the transmitter 200 trains the forward polar stream network 710 of the distributed channel encoder 218 by itself and transmits the configuration information 714 to the receiver 300. Referring to Figure 10 , in an alternative embodiment, a polar stream network 1000 can be trained to generate multiple distributed channel encoder configurations having different k and n. These configurations can be trained by the receiver 300 or other network devices. Configuration information for the trained configurations, including the identification of the frozen dimensions and the k and n values, can be saved to a storage location 1002. Thus, the receiver 300 can select a trained configuration from the storage location according to k and n and transmit the configuration information of the selected configuration to the transmitter 200 for configuring the distributed channel encoder 218.

[0098] Figure 10 The embodiment shown at Figure 1A may have advantages in situations where there are multiple IoT transmitters that extract similar features and the receiver 300 is implemented on a base station (e.g., the base station 140 shown at Figure 1AIf two or more IoT sensors capture information about the object of interest 154 from different perspectives, the associated transmitters can cooperate to provide optimal communication performance. Thus, the training configuration can be repeated for multiple IoT transmitters. For example, in a training configuration with k = 3 extracted features, one of the three extracted features is more important than the other two. However, for each particular IoT transmitter, the most important extracted feature can be different among the members of the multiple IoT transmitters. Thus, from the perspective of the receiver, when generating the multiple distributed channel encoder configurations with k = 3, each of the 3 features needs to be treated as equally important. In this embodiment, the training can be performed according to a spherical Gaussian manifold rather than an ellipsoidal Gaussian manifold. This makes the training more generally applicable to all of the multiple IoT transmitters.

[0099] As disclosed above, the integration of the source encoder and the distributed channel encoder breaks the i.i.d. assumption and enables an encoder / decoder configuration similar to a traditional Turbo coding scheme. The traditional Turbo coding channel coding algorithm implements an iterative or successive cancellation decoding scheme between two convolutional decoders to achieve optimal coding performance. A Turbo encoder uses two convolutional encoders, one of which encodes the original source bits and the other of which encodes the interleaved source bits.

[0100] Reference is made to Figure 11 , 1100 illustrates a configuration that extends the Turbo coding concept to the current context in a simplified block diagram. Configuration 1120 includes transmitter 1102, transmission channel 1104, and receiver 1106. Source information input S M is received at a source encoder 1108 of the transmitter 1102. The source information input S M is converted by a pseudo-random conversion block 1110, where the pseudo-random conversion is denoted by "T". As a simplified example, in this embodiment, the conversion T rotates the source information input S M by 180° and then provides input to a source encoder 1112. In practice, the conversion T would perform a more random conversion on the source information than a simple rotation. Thus, the conversion block 1110 is analogous to a bit interleaver implemented in the encoder / decoder of a traditional Turbo coding, which breaks the correlation between the original information and the converted or interleaved information. In an IoT camera embodiment, the converted source information can be alternatively provided by a second camera with a different perspective.

[0101] The source encoder 1108 encodes the source information input S MThe features are extracted in block 1106 and provided as input to a distributed channel encoder 1114. The distributed channel encoder 1114 implements a forward polar stream network. The source encoder 1108 and the distributed channel encoder 1114 (labeled "A" in block Figure 11 The source encoder 1112 is trained in block 1108 as described above. The source encoder 1112 receives as input the transformed source information S M The features are extracted in block 1106 and provided as input to a distributed channel encoder 1114. The distributed channel encoder 1114 implements a forward polar stream network. The source encoder 1108 and the distributed channel encoder 1114 (labeled "A" in block Figure 11 The source encoder 1112 is trained in block 1108 as described above. The source encoder 1112 receives as input the transformed source information S

[0102] After propagation over the transmission channel 1104, the noisy combined codeword is received at block 1120. The received combined codeword is then split into codeword inputs 1122 and 1124 corresponding to the respective outputs of the distributed channel encoders 1114 and 1116. The split codeword inputs 1122 and 1124 are received at respective averaging blocks 1126 and 1128. Each of the averaging blocks 1126 and 1128 includes a respective second feedback input 1130 and 1132. The averaging block 1126 outputs an average between the split codeword input 1122 and the feedback input 1130, which is initially empty. The role of the feedback inputs 1130 and 1132 will be described below. The averaging block 1128 outputs an average between the split codeword 1124 and the feedback input 1132.

[0103] The average output of the block 1126 is received by a distributed channel decoder 1134 configured to implement a reverse polar stream network according to the training configuration performed on the distributed channel encoder A (1114). Similarly, the average output of the block 1128 is received by a distributed channel decoder 1136 configured to implement a reverse polar stream network according to the training configuration performed on the distributed channel encoder B (1116). The feature inputs recovered at the distributed channel decoders 1134 and 1136 are input to respective source decoders 1138 and 1140. In the above description of the training exercise, the source decoder B (1138) would be trained with the source encoder A (1108). Similarly, in the training exercise, the source decoder B (1140) would be trained with the source encoder A (1112). Thus, the source decoder 1138 recovers the source information S Figure 4 In the training exercise generally described, the source decoder B (1138) would be trained with the source encoder A (1108). Similarly, in the training exercise, the source decoder B (1140) would be trained with the source encoder A (1112). Thus, the source decoder 1138 recovers the source information SM As mentioned at 1142, the output S′ is regenerated. M In this embodiment, the feedback input 1132 to the average box 1128 is generated by a set of boxes 1144. The transformation T of the box 1110 is applied to the recovered source information S. M This process generates a 180° rotated copy of the recovered source information. The transformed source information is then output to a source encoder B, which in turn outputs the extracted features from the source information to a distributed channel encoder B. Both the source encoder B and the distributed channel encoder B have been trained and configured as boxes 1112 and 1116. The output of the distributed channel encoder B provides the feedback input 1132 to the average box 1128.

[0104] The feedback input 1130 to the average box 1128 is generated by a set of boxes 1146. The transformation T of the box 1110 is reversed to T. -1 The conversion T -1 The source information is applied and recovered at the output of the source decoder 1140 to generate the original source information S. M A copy is then generated. The transformed source information is then output to source encoder A, which in turn outputs the extracted features from the source information to distributed channel encoder A. Source encoder A and distributed channel encoder A have been trained and configured as boxes 1108 and 1114. The output of distributed channel encoder A provides the feedback input 1130 to the average box 1126.

[0105] As mentioned above Figure 4 and Figure 7 The configuration information of the source encoders 1108 and 1112 and the distributed channel encoders 1114 and 1116 can be generated by the transmitter 1102 and transmitted to the receiver 1106. Alternatively, the receiver can retrieve the configuration information from the storage location (e.g., for the distributed channel encoders 1114 and 1116) of the source encoders 1108 and 1112. Figure 10 (As shown at position 1002), the trained configuration is selected. In the first iteration, the receiver 1106 pairs represent S. M and S MThe split codewords of T are decoded and an initial regenerated output is provided at 1142. As described above, during the first iteration, the feedback inputs 1130 and 1132 are empty and the averages at blocks 1126 and 1128 are based only on the respective inputs 1122 and 1124. In the second and subsequent iterations, the feedback generated by blocks 1144 and 1146 is averaged with the respective received codewords at blocks 1126 and 1128 and a new regenerated output is produced. The regenerated output 1142 is compared to the transformed output 1148 to determine if sufficient iterations have been performed. In practice, the number of iterations can be set to a predetermined number (e.g., 6 iterations) determined from the allowable time delay. After the 6 iterations, the regenerated output 1142 is compared to the transformed output 1148 to determine if the outputs are sufficiently similar to terminate the iterations.

[0106] Thus, Figure 11 The illustrated embodiments exploit the natural geometric properties of the source information to reduce the impact of noise in the transmission channel 1104.

[0107] In general, although the present application describes examples with reference to DNNs, any suitable neural network architecture can be used to implement the source encoder and the source decoder. For example, a graph neural network (GNN) can be a useful architecture for extracting features (or textures) as such neural networks can benefit from some relevant prior knowledge of the structure of the source information. A recurrent neural network (RNN) can also be a useful architecture for extracting features (or textures) as it can benefit from temporal associations in the structure of the source information.

[0108] While the application has been described with reference to specific features and embodiments thereof, it is evident that various modifications and combinations can be made thereto without departing from the scope of the application. The specification and drawings are, accordingly, to be regarded as illustrative rather than restrictive.

Claims

1. A communication method characterized by comprising: The method comprises: The first communication device determines at least one artificial intelligence model configuration; The first communication device sends configuration information carrying the at least one artificial intelligence model configuration.

2. The method of claim 1, wherein, The first communication device is a network device or a terminal device.

3. The method according to claim 1 or 2, characterized in that, The first communication device selects or online trains from a predefined set to generate the artificial intelligence model configuration based on channel state information or network state.

4. The method according to any one of claims 1 to 3, characterized in that, The configuration information includes parameters for defining the reverse polarized flow network, and identification of excess input dimensions.

5. The method according to any one of claims 1 to 4, characterized in that, The first communication device selects the configuration information from a predefined set containing the at least one artificial intelligence model configuration according to communication conditions.

6. The method according to any one of claims 1 to 5, characterized in that, Receive auxiliary information indicating the decoding capability or performance preference of the first communication device to assist the first communication device in determining the artificial intelligence model configuration.

7. A communication method characterized by comprising: The method comprises: a second communication device receiving configuration information carrying at least one artificial intelligence model configuration; The second communication device configures at least one corresponding functional entity based on the configuration information.

8. The method of claim 7, wherein, The second communication device is a terminal device or a network device.

9. The method according to claim 7 or 8, characterized in that, The functional entity is a reverse polarized flow network for channel decoding, and the configuration information includes parameters for defining the reverse polarized flow network and identification of excess input dimensions.

10. The method according to any one of claims 7 to 9, characterized in that, The functional entity is a neural network decoder for source decoding, and the configuration information includes network weights and biases for defining the neural network decoder.

11. The method according to any one of claims 7 to 10, characterized in that, Send auxiliary information indicating the decoding capability or performance preference of the first communication device to assist the first communication device in determining the artificial intelligence model configuration.

12. An encoding method for communication, characterized by, The method comprises: Extracting a plurality of feature probability distributions from source information using a source encoder; Input the plurality of feature probability distributions into a flow-based generative model; Using the flow-based generative model, convert the plurality of feature probability distributions into a plurality of output distribution code words generated to adapt to the characteristics of the transmission channel.

13. The method of claim 12, wherein, The flow-based generative model implements a reversible nonlinear transformation.

14. The method according to claim 12 or 13, characterized in that, The flow-based generative model includes a polarized flow network.

15. The method according to any one of claims 12 to 14, characterized in that, The conversion includes mapping a k-dimensional feature probability distribution space to an n-dimensional output distribution code word space, where k is less than n.

16. A communications device, characterized by Comprise a unit or module for performing the method as claimed in any one of claims 1-15.

17. A communications device, characterized by Comprise a processor for executing computer programs or instructions to implement the method as claimed in any one of claims 1-15.

18. A computer-readable storage medium, characterized in that, The storage medium has instructions stored therein, and when the instructions are executed by a communication device, the method as claimed in any one of claims 1-15 is implemented.

Images