ABSTRACT
Abstract
Embodiments of this application disclose a neural network-based communication method and a related apparatus. Specifically, joint training optimization is performed on an encoding neural network used by a transmit end and a decoding neural network used by a receive end. A first neural network in the encoding neural network reuses the decoding neural network and a parameter of the decoding neural network.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
This application is a continuation of International Application No. PCT/CN20211/120839, filed on Sep. 27, 2021, which claims priority to Chinese Patent Application No. 202011062368.2, filed on Sep. 30, 2020. The disclosures of the aforementioned applications are hereby incorporated by reference in their entireties.
TECHNICAL FIELD
This application relates to the field of communication technologies, and in particular, to a neural network-based communication method and a related apparatus.
BACKGROUND
A wireless communication system may be divided into three parts: a transmitter, a channel, and a receiver, where the channel is used to transmit a waveform signal exchanged between the transmit end and the receive end.
In a conventional wireless communication system, a transmit end includes a source encoder and a channel encoder, and a receive end includes a source decoder and a channel decoder. Source encoding is source symbol transformation with a purpose of improving communication effectiveness and reducing or eliminating a source redundancy degree. By processing a symbol output through source encoding, channel encoding enables the communication system to have a specific error correction capability and anti-interference capability, to reduce a bit error rate during channel transmission.
In order to optimize performance of the wireless communication system, the foregoing encoders and decoders need to be optimized. Because different encoders and decoders have independent mathematical models, the encoders and decoders are usually optimized independently based on their own mathematical models. However, in an actual application scenario, the channel, the transmit end, and the receive end have non-idealization and nonlinear problems. Therefore, a mathematical model mismatch may exist, which affects performance of the wireless communication system.
SUMMARY
Embodiments of this application provide a neural network-based communication method, in which joint training optimization is performed on an encoding neural network used by a transmit end and a decoding neural network used by a receive end. A first neural network in the encoding neural network reuses the decoding neural network and a parameter of the decoding neural network. In this way, the decoding neural network provides a part of prior information for the encoding neural network, so that the encoding neural network can more accurately predict a benefit of an action, thereby reducing a variance of gradient estimation, accelerating a network convergence speed, and improving convergence performance.
According to a first aspect, an embodiment of this application proposes a neural network-based communication method, where the method is applied to the transmit end and includes: the transmit end obtains a first data stream; the transmit end processes the first data stream by using an encoding neural network, to obtain a first symbol stream, where the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network; and the transmit end outputs the first symbol stream.
Optionally, the encoding neural network may be deployed inside the transmit end, or may be deployed on a cloud server. When the encoding neural network is deployed on the cloud server, the transmit end requests a service of the encoding neural network from the cloud server.
Specifically, the transmit end obtains the first data stream, where the first data stream is a binary digit (BIT) stream. When communication interaction is performed between the transmit end and the receive end, the first data stream is normal data sent from the transmit end to the receive end. When neural network training is performed between the transmit end and the receive end, the first data stream is training sample data sent from the transmit end to the receive end.
When the first data stream is the training sample data, the first data stream may be preconfigured inside the transmit end, or the first data stream may be data input externally (relative to the transmit end), or the first data stream may be data negotiated between the transmit end and the receive end. This is not limited herein.
In an optional implementation, when the first data stream is preconfigured inside the transmit end, the first data stream may be further preconfigured inside the receive end.
In an optional implementation, the encoding neural network in this embodiment of this application is trained by using an âimproved Actor-Critic (Actor-Critic)â algorithm. Based on a function, the encoding neural network may be divided into two parts: a first neural network and a second neural network. The first neural network is also referred to as a âcritic neural networkâ, and the second neural network is also referred to as an âactor neural networkâ.
The first neural network may reuse all of the decoding neural network. For example, the first neural network uses all neuron structures of the decoding neural network (the neuron structure includes a quantity of neuron layers).
Alternatively, the first neural network may reuse a part of the decoding neural network (and/or a parameter of a part of the decoding neural network). For example, the first neural network uses a part of a neuron structure of the decoding neural network. A quantity of neuron layers in the first neural network may be consistent with a quantity of neuron layers in the decoding neural network, but only some neurons in the first neural network are consistent with neurons in the decoding neural network.
Alternatively, some neuron layers in the first neural network may be consistent with some neuron layers in the decoding neural network, but a quantity of neuron layers in the first neural network is inconsistent with a quantity of neuron layers in the decoding neural network.
Alternatively, some neurons in the first neural network may use a parameter of some neurons in the decoding neural network, or use a variation of a parameter of some neurons in the decoding neural network.
Alternatively, some neurons in the first neural network may use variations of some neurons in the decoding neural network. This is not limited herein.
It should be noted that âreusingâ may indicate that the first neural network uses exactly the same neuron or neuron parameter in the decoding neural network, or the first neural network includes the neuron or the neuron parameter in the decoding neural network, or may indicate that the first neural network uses the variation of the neuron or the variation of the neuron parameter in the decoding neural network, or the like. This is not limited herein.
In this embodiment of this application, the first neural network in the encoding neural network reuses the decoding neural network and the parameter of the decoding neural network. In this way, the decoding neural network provides a part of prior information for the encoding neural network, so that the encoding neural network can more accurately predict a benefit of an action, thereby reducing a variance of gradient estimation, accelerating a network convergence speed, and improving convergence performance.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a first weight, where the first weight is from the decoding neural network, and the first weight is used to train the first neural network.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the first weight from the decoding neural network. The first weight is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a first gradient, where the first gradient is from the decoding neural network, and the first gradient is used to train the first neural network.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the first gradient from the decoding neural network. The first gradient is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a first function, where the first function is from the decoding neural network, and the first function is used to train the first neural network.
Specifically, the first function may be the first function itself, or may be a function value (or a calculation result) of the first function, or may be the first function itself and the function value (or the calculation result) of the first function. This is not limited herein.
When the transmit end receives the first function, the transmit end performs calculation processing on the first function by using the encoding neural network (for example, the first neural network), to generate a new gradient. The new gradient is used to train (for example, update) the encoding neural network (for example, the first neural network).
In this embodiment of this application, the first function is originally a function used by the receive end side to train the decoding neural network. Therefore, after the first function is generated on the receive end side, the first function may be sent to the transmit end. The transmit end trains the first neural network by using the first function, to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a second function, where the second function is from the decoding neural network, and the second function is a loss function and/or a reward function; and the transmit end processes the second function by using the encoding neural network, to obtain a second gradient, where the second gradient is used to train the encoding neural network.
Specifically, the second function may be the second function itself, or may be a function value (or a calculation result) of the second function, or may be the second function itself and the function value (or the calculation result) of the second function. This is not limited herein.
When the transmit end receives the second function, the transmit end performs calculation processing on the second function by using the encoding neural network, to generate a second gradient. The second gradient is used to train (for example, update) the encoding neural network. Specifically, the encoding neural network performs calculation processing on the second function by using a gradient estimation (Gradient approximation) algorithm, to generate the second gradient. For example, the gradient approximation algorithm includes but is not limited to a stochastic gradient descent (SGD) algorithm, an âAdagrad algorithmâ, a root mean square prop (RMSProp) algorithm, an adaptive moment estimation (Adaptive moment estimation) algorithm, and a momentum (momentum) algorithm.
In this embodiment of this application, the first function is originally a function used by the receive end side to train the decoding neural network. Therefore, after the first function is generated on the receive end side, the first function may be sent to the transmit end. The transmit end trains the first neural network by using the first function, to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, that the transmit end outputs the first symbol stream includes: the transmit end performs filtering processing on the first symbol stream, to obtain a first waveform signal, where an out-of-band signal is filtered out from the first waveform signal; and the transmit end outputs the first waveform signal.
In this embodiment, to introduce intersymbol interference in the first symbol stream, the transmit end uses the first filter to perform filtering processing on the first symbol stream, to obtain the first waveform signal.
Optionally, a low-pass filter may be selected as the first filter.
Optionally, the first filter may further filter out an out-of-band signal in the first symbol stream, to limit a sending bandwidth.
In this embodiment of this application, to improve signal transmission performance, the transmit end may use the first filter to perform filtering processing on the first symbol stream, to obtain the first waveform signal.
With reference to the first aspect, in a possible implementation of the first aspect, that the transmit end processes the first data stream by using an encoding neural network, to obtain a first symbol stream includes: the transmit end performs encoding processing on the first data stream, to obtain a first channel encoding code word; and the transmit end processes the first channel encoding code word by using the encoding neural network, to obtain the first symbol stream.
Specifically, to avoid a dimension curse problem (also referred to as a number of dimensions curse problem), related operations such as source encoding, channel encoding, modulation, and/or filtering may be further performed by another module in the transmit end, and the another module is referred to as a first module. The first module may be deployed on an input side of the encoding neural network, or may be deployed on an output side of the encoding neural network, or may be deployed on both the input side and the output side of the encoding neural network.
In this embodiment, the transmit end uses the first module to process the first data stream, to obtain the first channel encoding code word.
Optionally, the first module may be cascaded outside the communication system, to further improve error-free transmission performance of the system.
With reference to the first aspect, in a possible implementation of the first aspect, that the transmit end processes the first cha
CROSS-REFERENCE TO RELATED APPLICATIONS
This application is a continuation of International Application No. PCT/CN20211/120839, filed on Sep. 27, 2021, which claims priority to Chinese Patent Application No. 202011062368.2, filed on Sep. 30, 2020. The disclosures of the aforementioned applications are hereby incorporated by reference in their entireties.
TECHNICAL FIELD
This application relates to the field of communication technologies, and in particular, to a neural network-based communication method and a related apparatus.
BACKGROUND
A wireless communication system may be divided into three parts: a transmitter, a channel, and a receiver, where the channel is used to transmit a waveform signal exchanged between the transmit end and the receive end.
In a conventional wireless communication system, a transmit end includes a source encoder and a channel encoder, and a receive end includes a source decoder and a channel decoder. Source encoding is source symbol transformation with a purpose of improving communication effectiveness and reducing or eliminating a source redundancy degree. By processing a symbol output through source encoding, channel encoding enables the communication system to have a specific error correction capability and anti-interference capability, to reduce a bit error rate during channel transmission.
In order to optimize performance of the wireless communication system, the foregoing encoders and decoders need to be optimized. Because different encoders and decoders have independent mathematical models, the encoders and decoders are usually optimized independently based on their own mathematical models. However, in an actual application scenario, the channel, the transmit end, and the receive end have non-idealization and nonlinear problems. Therefore, a mathematical model mismatch may exist, which affects performance of the wireless communication system.
SUMMARY
Embodiments of this application provide a neural network-based communication method, in which joint training optimization is performed on an encoding neural network used by a transmit end and a decoding neural network used by a receive end. A first neural network in the encoding neural network reuses the decoding neural network and a parameter of the decoding neural network. In this way, the decoding neural network provides a part of prior information for the encoding neural network, so that the encoding neural network can more accurately predict a benefit of an action, thereby reducing a variance of gradient estimation, accelerating a network convergence speed, and improving convergence performance.
According to a first aspect, an embodiment of this application proposes a neural network-based communication method, where the method is applied to the transmit end and includes: the transmit end obtains a first data stream; the transmit end processes the first data stream by using an encoding neural network, to obtain a first symbol stream, where the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network; and the transmit end outputs the first symbol stream.
Optionally, the encoding neural network may be deployed inside the transmit end, or may be deployed on a cloud server. When the encoding neural network is deployed on the cloud server, the transmit end requests a service of the encoding neural network from the cloud server.
Specifically, the transmit end obtains the first data stream, where the first data stream is a binary digit (BIT) stream. When communication interaction is performed between the transmit end and the receive end, the first data stream is normal data sent from the transmit end to the receive end. When neural network training is performed between the transmit end and the receive end, the first data stream is training sample data sent from the transmit end to the receive end.
When the first data stream is the training sample data, the first data stream may be preconfigured inside the transmit end, or the first data stream may be data input externally (relative to the transmit end), or the first data stream may be data negotiated between the transmit end and the receive end. This is not limited herein.
In an optional implementation, when the first data stream is preconfigured inside the transmit end, the first data stream may be further preconfigured inside the receive end.
In an optional implementation, the encoding neural network in this embodiment of this application is trained by using an âimproved Actor-Critic (Actor-Critic)â algorithm. Based on a function, the encoding neural network may be divided into two parts: a first neural network and a second neural network. The first neural network is also referred to as a âcritic neural networkâ, and the second neural network is also referred to as an âactor neural networkâ.
The first neural network may reuse all of the decoding neural network. For example, the first neural network uses all neuron structures of the decoding neural network (the neuron structure includes a quantity of neuron layers).
Alternatively, the first neural network may reuse a part of the decoding neural network (and/or a parameter of a part of the decoding neural network). For example, the first neural network uses a part of a neuron structure of the decoding neural network. A quantity of neuron layers in the first neural network may be consistent with a quantity of neuron layers in the decoding neural network, but only some neurons in the first neural network are consistent with neurons in the decoding neural network.
Alternatively, some neuron layers in the first neural network may be consistent with some neuron layers in the decoding neural network, but a quantity of neuron layers in the first neural network is inconsistent with a quantity of neuron layers in the decoding neural network.
Alternatively, some neurons in the first neural network may use a parameter of some neurons in the decoding neural network, or use a variation of a parameter of some neurons in the decoding neural network.
Alternatively, some neurons in the first neural network may use variations of some neurons in the decoding neural network. This is not limited herein.
It should be noted that âreusingâ may indicate that the first neural network uses exactly the same neuron or neuron parameter in the decoding neural network, or the first neural network includes the neuron or the neuron parameter in the decoding neural network, or may indicate that the first neural network uses the variation of the neuron or the variation of the neuron parameter in the decoding neural network, or the like. This is not limited herein.
In this embodiment of this application, the first neural network in the encoding neural network reuses the decoding neural network and the parameter of the decoding neural network. In this way, the decoding neural network provides a part of prior information for the encoding neural network, so that the encoding neural network can more accurately predict a benefit of an action, thereby reducing a variance of gradient estimation, accelerating a network convergence speed, and improving convergence performance.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a first weight, where the first weight is from the decoding neural network, and the first weight is used to train the first neural network.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the first weight from the decoding neural network. The first weight is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a first gradient, where the first gradient is from the decoding neural network, and the first gradient is used to train the first neural network.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the first gradient from the decoding neural network. The first gradient is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a first function, where the first function is from the decoding neural network, and the first function is used to train the first neural network.
Specifically, the first function may be the first function itself, or may be a function value (or a calculation result) of the first function, or may be the first function itself and the function value (or the calculation result) of the first function. This is not limited herein.
When the transmit end receives the first function, the transmit end performs calculation processing on the first function by using the encoding neural network (for example, the first neural network), to generate a new gradient. The new gradient is used to train (for example, update) the encoding neural network (for example, the first neural network).
In this embodiment of this application, the first function is originally a function used by the receive end side to train the decoding neural network. Therefore, after the first function is generated on the receive end side, the first function may be sent to the transmit end. The transmit end trains the first neural network by using the first function, to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, after the transmit end outputs the first symbol stream, the method further includes: the transmit end receives a second function, where the second function is from the decoding neural network, and the second function is a loss function and/or a reward function; and the transmit end processes the second function by using the encoding neural network, to obtain a second gradient, where the second gradient is used to train the encoding neural network.
Specifically, the second function may be the second function itself, or may be a function value (or a calculation result) of the second function, or may be the second function itself and the function value (or the calculation result) of the second function. This is not limited herein.
When the transmit end receives the second function, the transmit end performs calculation processing on the second function by using the encoding neural network, to generate a second gradient. The second gradient is used to train (for example, update) the encoding neural network. Specifically, the encoding neural network performs calculation processing on the second function by using a gradient estimation (Gradient approximation) algorithm, to generate the second gradient. For example, the gradient approximation algorithm includes but is not limited to a stochastic gradient descent (SGD) algorithm, an âAdagrad algorithmâ, a root mean square prop (RMSProp) algorithm, an adaptive moment estimation (Adaptive moment estimation) algorithm, and a momentum (momentum) algorithm.
In this embodiment of this application, the first function is originally a function used by the receive end side to train the decoding neural network. Therefore, after the first function is generated on the receive end side, the first function may be sent to the transmit end. The transmit end trains the first neural network by using the first function, to implement end-to-end joint optimization training.
With reference to the first aspect, in a possible implementation of the first aspect, that the transmit end outputs the first symbol stream includes: the transmit end performs filtering processing on the first symbol stream, to obtain a first waveform signal, where an out-of-band signal is filtered out from the first waveform signal; and the transmit end outputs the first waveform signal.
In this embodiment, to introduce intersymbol interference in the first symbol stream, the transmit end uses the first filter to perform filtering processing on the first symbol stream, to obtain the first waveform signal.
Optionally, a low-pass filter may be selected as the first filter.
Optionally, the first filter may further filter out an out-of-band signal in the first symbol stream, to limit a sending bandwidth.
In this embodiment of this application, to improve signal transmission performance, the transmit end may use the first filter to perform filtering processing on the first symbol stream, to obtain the first waveform signal.
With reference to the first aspect, in a possible implementation of the first aspect, that the transmit end processes the first data stream by using an encoding neural network, to obtain a first symbol stream includes: the transmit end performs encoding processing on the first data stream, to obtain a first channel encoding code word; and the transmit end processes the first channel encoding code word by using the encoding neural network, to obtain the first symbol stream.
Specifically, to avoid a dimension curse problem (also referred to as a number of dimensions curse problem), related operations such as source encoding, channel encoding, modulation, and/or filtering may be further performed by another module in the transmit end, and the another module is referred to as a first module. The first module may be deployed on an input side of the encoding neural network, or may be deployed on an output side of the encoding neural network, or may be deployed on both the input side and the output side of the encoding neural network.
In this embodiment, the transmit end uses the first module to process the first data stream, to obtain the first channel encoding code word.
Optionally, the first module may be cascaded outside the communication system, to further improve error-free transmission performance of the system.
With reference to the first aspect, in a possible implementation of the first aspect, that the transmit end processes the first channel encoding code word by using the encoding neural network, to obtain the first symbol stream includes: the transmit end processes the first channel encoding code word by using the encoding neural network, to obtain a first intermediate symbol; and the transmit end performs filtering processing on the first intermediate symbol by using a filtering neural network, to obtain the first symbol stream, where intersymbol interference is introduced into the first symbol stream.
Specifically, the filtering neural network may be introduced on a receive end side. Optionally, the filtering neural network is deployed at a transmit end, and the filtering neural network is also deployed on the receive end side. Optionally, the filtering neural network is deployed only on the receive end side. The filtering neural network may be deployed between a second filter and an analog-to-digital converter, or the filtering neural network may be deployed between the analog-to-digital converter and a decoding neural network. This is not limited herein.
A 1 bit communication system based on oversampling is used as an example. Due to an information loss caused by 1 bit quantization, effective channel estimation and equalization cannot be finished in a conventional manner such as a pilot. Therefore, in a fading channel scenario, performance of the communication system is relatively poor. Fading refers to a phenomenon that an amplitude of a received signal changes randomly due to the change of the channel, that is, signal fading. A channel that causes signal fading is referred to as a fading channel.
Based on this, a filtering neural network is disposed, and the filtering neural network performs filtering processing on a first symbol stream output from the encoding neural network, to obtain a first intermediate symbol. When the communication system further includes an upsampling module, the upsampling module performs upsampling processing on the first intermediate symbol, to obtain a second intermediate symbol, where the second intermediate symbol is obtained by performing zero insertion processing on the first intermediate symbol after a sampling point. Then, a filtering neural network at the transmit end performs filtering processing on the second intermediate symbol to obtain the first symbol stream, where the filtering processing performed by the filtering neural network includes introducing intersymbol interference (ISI).
Optionally, the encoding neural network, the decoding neural network, and the filtering neural network may be jointly trained; or a pre-training model (for example, the encoding neural network and the decoding neural network that are trained on an additive white Gaussian noise (AWGN) channel) may be used for initialization, and then the filtering neural network is trained individually.
In this embodiment of this application, the filtering neural network is added after the encoding neural network. In a fading channel scenario, the filtering neural network is combined with the encoding neural network and the decoding neural network for end-to-end training optimization. Specifically, the filtering neural network filters a symbol (the first symbol stream) mapped by the oversampled encoding neural network, to mitigate impact of a fading channel on a system. When quantization precision at the receive end is relatively high, the filtering neural network may play a pre-equalization effect after training. However, due to the information loss caused by low precision quantization, the filter neural network is more difficult to learn pre-equalization in the low precision quantization, and is more inclined to another local optimal solution. Through the joint training optimization of the encoding neural network, decoding neural network, and the filtering neural network, the transmit end and the receive end can use multipath interference of a channel to distinguish sent signals. This improves communication performance in the fading channel scenario.
According to a second aspect, an embodiment of this application proposes a neural network-based communication method, where the method is applied to a receive end and includes: the receive end obtains a first data stream; the receive end receives a first symbol stream, where the first symbol stream is from a transmit end, the first symbol stream is obtained by the transmit end by processing the first data stream by using an encoding neural network, the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network; the receive end processes the first symbol stream by using the decoding neural network, to obtain a second data stream; and the receive end trains the decoding neural network based on the first data stream and the second data stream.
Specifically, on a transmit end side, the filtering processing is performed on the first symbol stream to obtain a first waveform signal. The transmit end sends the first waveform signal to the receive end through a channel. For ease of description, the channel is referred to as a first channel. The first waveform signal is affected by channel interference in a transmission process of the first channel. In this case, the receive end processes, by using the decoding neural network, the first waveform signal (the first symbol stream) interfered by the first channel, to obtain the second data stream.
For example, the decoding neural network includes but is not limited to a fully connected neural network, a radial basis function neural network, a convolutional neural network, a cyclic neural network, a Hopfield neural network, a restricted Boltzmann machine or deep confidence network. The encoding neural network may be any one of the foregoing neural networks, or the encoding neural network may be a combination of the foregoing plurality of neural networks. This is not limited herein.
It should be noted that the decoding neural network and the encoding neural network may use a same neural network, or may use different neural networks, or may use some same neural networks. This is not limited herein.
Optionally, a loss function that may be selected for the decoding neural network (and/or the encoding neural network) includes but is not limited to a binary cross entropy, a classification cross entropy, or a mean square error. For example, if the binary cross entropy is selected, an objective of end-to-end joint optimization in this embodiment of this application is to minimize a cross entropy between input and output.
In this embodiment of this application, the first neural network in the encoding neural network reuses the decoding neural network and the parameter of the decoding neural network. In this way, the decoding neural network provides a part of prior information for the encoding neural network, so that the encoding neural network can more accurately predict a benefit of an action, thereby reducing a variance of gradient estimation, accelerating a network convergence speed, and improving convergence performance. End-to-end joint optimization training further improves communication efficiency of the system.
With reference to the second aspect, in a possible implementation of the second aspect, that the receive end trains the decoding neural network based on the first data stream and the second data stream includes: the receive end processes the first data stream and the second data stream by using the decoding neural network, to obtain a first function, where the first function is a loss function and/or a reward function; the receive end processes the first function by using the decoding neural network, to obtain a first gradient; and the receive end trains the decoding neural network by using the first gradient.
Specifically, the receive end trains the decoding neural network by using the first data stream and the second data stream. Specifically, the decoding neural network obtains a first function through calculation according to a first data stream and a second data stream, where the first function is a loss function and/or a reward function. The decoding neural network calculates a first gradient based on the first function, where the first gradient is used for further training (for example, updating) of the decoding neural network. Further, the decoding neural network may share a weight of the trained decoding neural network with the encoding neural network. For ease of description, the weight is referred to as a first weight.
With reference to the second aspect, in a possible implementation of the second aspect, after the receive end processes the first data stream and the second data stream by using the decoding neural network, to obtain a first function, the method further includes: the receive end sends a first function to the transmit end, where the first function is used to train the first neural network in the encoding neural network.
In this embodiment of this application, after the first function is generated on the receive end side, the first function may be sent to the transmit end. The transmit end trains the first neural network by using the first function, to implement end-to-end joint optimization training.
With reference to the second aspect, in a possible implementation of the second aspect, after the receive end trains the decoding neural network by using the first gradient, the method further includes: the receive end sends a first gradient to the transmit end, where the first gradient is used to train the first neural network in the encoding neural network.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the first gradient from the decoding neural network. The first gradient is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the second aspect, in a possible implementation of the second aspect, after the receive end trains the decoding neural network by using the first gradient, the method further includes: the receive end generates a first weight by using the decoding neural network, where the first weight is a weight of the decoding neural network; and the receive end sends the first weight to the transmit end, where the first weight is used to train the first neural network in the encoding neural network.
In an optional implementation, the receive end transmits the second function, the first gradient, or the first weight to the transmit end by using a reward feedback frame or a gradient/weight sharing frame. The transmit end checks the reward feedback frame or the gradient/weight sharing frame from the receive end. When it is found that a cyclic redundancy check (CRC) error exists, the transmit end discards corresponding data (which refers to training data, for example, a second function, a first gradient, or a first weight) in the reward feedback frame or the gradient/weight sharing frame. The reason is that if a data frame is transmitted over an air interface, an error may occur, and incorrect data (which refers to training data, for example, the second function, the first gradient, or the first weight) may cause negative impact on reinforcement learning training. The negative impact can be effectively eliminated by using the foregoing method.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the first weight from the decoding neural network. The first weight is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the second aspect, in a possible implementation of the second aspect, after the receive end trains the decoding neural network based on the first data stream and the second data stream, the method further includes: the receive end processes the first data stream and the second data stream by using the decoding neural network, to obtain a second function, where the second function is a loss function and/or a reward function; and the receive end sends the second function to the transmit end, where the second function is used to train the encoding neural network.
Specifically, the receive end processes the first data stream and the second data stream by using the decoding neural network, to obtain a function value of a second function, where the function value of the second function is used to train the encoding neural network (including the first neural network and a second neural network).
When the transmit end receives the second function, the transmit end performs calculation processing on the second function by using the encoding neural network, to generate a second gradient. The second gradient is used to train (for example, update) the encoding neural network. Specifically, the encoding neural network performs calculation processing on the second function by using a gradient estimation (Gradient approximation) algorithm, to generate the second gradient. For example, the gradient approximation algorithm includes but is not limited to a stochastic gradient descent (SGD) algorithm, an âAdagrad algorithmâ, a root mean square prop (RMSProp) algorithm, an adaptive moment estimation algorithm, and a momentum algorithm.
In this embodiment of this application, the transmit end may further train the encoding neural network by using the second function from the decoding neural network. The second function is from the decoding neural network at the receive end, and is used to implement end-to-end joint optimization training.
With reference to the second aspect, in a possible implementation of the second aspect, that the receive end processes the first symbol stream by using the decoding neural network, to obtain a second data stream includes: the receive end receives a first waveform signal, where the first waveform signal is obtained by the transmit end by performing filtering processing on the first symbol stream; the receive end performs quantization processing on the first waveform signal, to obtain a quantized symbol sequence, where one group of the first data streams is mapped to one or more of the quantized symbol sequences, bit flipping exists at most once in each symbol of the quantized symbol sequence, and the bit flipping is 0 to 1 or 1 to 0; and the receive end processes a first sequence in the quantized symbol sequence by using the decoding neural network, to obtain the second data stream.
Specifically, when the first symbol stream includes x symbol groups, and each of the symbol groups includes n symbols, the first waveform signal includes x symbol groups, where x is a positive integer greater than or equal to 1, and n is a positive integer greater than or equal to 1; and when the receive end performs quantization processing on n symbols in the first waveform signal, to obtain the quantized symbol sequence, a length of the quantized symbol sequence is nM, where M is an oversampling rate of the receive end.
The quantized symbol sequence has the following features: one group of first data streams is mapped to one or more quantized symbol sequences, bit flipping exists at most once in each symbol of the quantized symbol sequence, and the bit flipping is 0 to 1 or 1 to 0.
Optionally, the analog-to-digital converter may be an oversampling analog-to-digital converter. When an oversampling rate of the analog-to-digital converter is M, a symbol rate of the quantized symbol sequence is Mxf s , where M is an integer greater than 0. Oversampling (oversampling) is a process of sampling a signal at a frequency twice a signal bandwidth or at a highest frequency thereof.
Optionally, the analog-to-digital converter may alternatively be an analog-to-digital converter with limited sampling precision (quantization precision), for example, an oversampling analog-to-digital converter with the sampling precision of 1 bit (referred to as 1 bit oversampling ADC for short).
Optionally, a length of the first sequence in the quantized symbol sequence is [(n -1)M + 1, nM]. For example, n = 2. That is, when the first symbol stream includes x symbol groups, and each symbol group includes two symbols, a length of a first sequence used by the decoding neural network for processing (also referred to as detection) is [M + 1, 2 M]. A longer length of the first sequence indicates a higher oversampling rate, and better performance may be obtained. Because a plurality of first sequences may be corresponding to one piece of data, even if there is a difference among quantized symbol sequences generated after the oversampling process, data can be finally correctly restored. A plurality of sequences may be corresponding to one sequence, and one sequence cannot be corresponding to a plurality of pieces of data, so as to ensure a correctness percentage of decoding.
With reference to the second aspect, in a possible implementation of the second aspect, that the receive end performs quantization processing on the first waveform signal, to obtain a quantized symbol sequence includes: the receive end receives the first waveform signal interfered by a first channel, where the first channel is a channel between the receive end and the transmit end; and the receive end performs quantization processing on the first waveform signal interfered by the first channel, to obtain the quantized symbol sequence.
In this embodiment of this application, the receive end may perform training by using a waveform signal interfered by a channel, to improve applicability of a training result, and further improve performance of the communication system.
With reference to the second aspect, in a possible implementation of the second aspect, that the receive end processes the first symbol stream by using the decoding neural network, to obtain a second data stream includes: the receive end processes the first waveform signal by using the decoding neural network, to obtain a first channel decoding code word or a log-likelihood ratio; and the receive end processes the first channel decoding code word or the log-likelihood ratio, to obtain the second data stream.
Specifically, in an optional implementation, on the transmit end side, when related operations such as source encoding, channel encoding, modulation, and/or filtering of the first data stream are performed by another module (namely, the first module) in the transmit end, on the receive end side, processing such as source decoding, channel decoding, demodulation, and/or equalization may be performed by the decoding neural network, or may be performed by another processing module independent of the decoding neural network. The another processing module is referred to as a second module. The second module may be deployed on an input side of the decoding neural network, or may be deployed on an output side of the decoding neural network, or may be deployed on both the input side and the output side of the decoding neural network. Specifically, the receive end first processes the first waveform signal by using the decoding neural network, to obtain a first channel decoding code word or a log-likelihood ratio (Log-likelihood Ratio, LLR). The log-likelihood ratio is usually used for soft decoding in communication. Second, the receive end processes the first channel decoding code word or the log-likelihood ratio by using the second module, to obtain the second data stream.
With reference to the second aspect, in a possible implementation of the second aspect, that the receive end performs quantization processing on the first waveform signal, to obtain a quantized symbol sequence includes: the receive end performs filtering processing on the first waveform signal, to obtain a second waveform signal; and the receive end performs quantization processing on the second waveform signal, to obtain a quantized symbol sequence.
In this embodiment, the receive end uses a second filter to perform filtering processing on a first waveform signal transmitted through the first channel, to obtain the second waveform signal. Optionally, a matched filter may be selected as the second filter.
In an optional implementation, a same root-raised cosine filter (Root-raised cosine filter, RRC filter) may be selected as the first filter and the second filter. The RRC filter is configured to perform signal shaping.
According to a third aspect, an embodiment of this application provides a communication apparatus, including: a transceiver module, configured to obtain a first data stream; and a processing module, configured to process the first data stream by using an encoding neural network, to obtain a first symbol stream, where the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network, where the transceiver module is further configured to output the first symbol stream.
In some optional embodiments of this application, the transceiver module is further configured to receive a first weight, where the first weight is from the decoding neural network, and the first weight is used to train the first neural network.
In some optional embodiments of this application, the transceiver module is further configured to receive a first gradient, where the first gradient is from the decoding neural network, and the first gradient is used to train the first neural network.
In some optional embodiments of this application, the transceiver module is further configured to receive a second function, where the second function is from the decoding neural network, and the second function is a loss function and/or a reward function; and the transceiver module is further configured to process the second function by using the encoding neural network, to obtain a second gradient, where the second gradient is used to train the encoding neural network.
In some optional embodiments of this application, the processing module is specifically configured to perform filtering processing on the first symbol stream, to obtain a first waveform signal, where an out-of-band signal is filtered out from the first waveform signal; and the transceiver module is specifically configured to output the first waveform signal.
In some optional embodiments of this application, the processing module is specifically configured to perform encoding processing on the first data stream, to obtain a first channel encoding code word; and the processing module is specifically configured to process the first channel encoding code word by using the encoding neural network, to obtain the first symbol stream.
In some optional embodiments of this application, the processing module is specifically configured to process the first channel encoding code word by using the encoding neural network, to obtain a first intermediate symbol; and the processing module is specifically configured to perform filtering processing on the first intermediate symbol by using a filtering neural network, to obtain the first symbol stream, where intersymbol interference is introduced into the first symbol stream.
In some optional embodiments of this application, the processing module is specifically configured to perform upsampling processing on the first intermediate symbol, to obtain a second intermediate symbol; and the processing module is specifically configured to perform filtering processing on the second intermediate symbol by using the filtering neural network, to obtain the first symbol stream.
According to a fourth aspect, an embodiment of this application proposes a communication apparatus, including: a transceiver module, configured to obtain a first data stream, where the transceiver module is further configured to receive a first symbol stream, where the first symbol stream is from a transmit end, the first symbol stream is obtained by the transmit end by processing the first data stream by using an encoding neural network, the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network; and a processing module, configured to process the first symbol stream by using the decoding neural network, to obtain a second data stream, where the processing module is further configured to train the decoding neural network based on the first data stream and the second data stream.
In some optional embodiments of this application, the processing module is specifically configured to process the first data stream and the second data stream by using the decoding neural network, to obtain a first function, where the first function is a loss function and/or a reward function; the processing module is specifically configured to process the first function by using the decoding neural network, to obtain a first gradient; and the processing module is specifically configured to train the decoding neural network by using the first gradient.
In some optional embodiments of this application, the transceiver module is further configured to send the first gradient to the transmit end, where the first gradient is used to train the first neural network in the encoding neural network.
In some optional embodiments of this application, the processing module is further configured to generate a first weight by using the decoding neural network, where the first weight is a weight of the decoding neural network; and the processing module is further configured to send the first weight to the transmit end, where the first weight is used to train the first neural network in the encoding neural network.
In some optional embodiments of this application, the processing module is further configured to process the first data stream and the second data stream by using the decoding neural network, to obtain a second function, where the second function is a loss function and/or a reward function; and the transceiver module is further configured to send the second function to the transmit end, where the second function is used to train the encoding neural network.
In some optional embodiments of this application, the transceiver module is specifically configured to receive a first waveform signal, where the first waveform signal is obtained by the transmit end by performing filtering processing on the first symbol stream; the processing module is specifically configured to perform quantization processing on the first waveform signal, to obtain a quantized symbol sequence, where one group of the first data streams is mapped to one or more of the quantized symbol sequences, bit flipping exists at most once in each symbol of the quantized symbol sequence, and the bit flipping is 0 to 1 or 1 to 0; and the processing module is specifically configured to process a first sequence in the quantized symbol sequence by using the decoding neural network, to obtain the second data stream.
In some optional embodiments of this application, the transceiver module is specifically configured to receive the first waveform signal interfered by a first channel, where the first channel is a channel between the receive end and the transmit end; and the processing module is specifically configured to perform quantization processing on the first waveform signal interfered by the first channel, to obtain the quantized symbol sequence.
In some optional embodiments of this application, the processing module is specifically configured to process the first waveform signal by using the decoding neural network, to obtain a first channel decoding code word or a log-likelihood ratio; and the processing module is specifically configured to process the first channel decoding code word or the log-likelihood ratio, to obtain the second data stream.
In some optional embodiments of this application, the processing module is specifically configured to perform quantization processing on the first waveform signal, to obtain the quantized symbol sequence; and the processing module is specifically configured to process the first sequence in the quantized symbol sequence by using the decoding neural network, to obtain the first channel decoding code word or the log-likelihood ratio.
In some optional embodiments of this application, when the first symbol stream includes x symbol groups, and each of the symbol groups includes n symbols, the first waveform signal includes x symbol groups, where x is a positive integer greater than or equal to 1 , and n is a positive integer greater than or equal to 1; and the processing module performs quantization processing on n symbols in the first waveform signal, to obtain the quantized symbol sequence, a length of the quantized symbol sequence is nM, and a length of the first sequence in the quantized symbol sequence is [(n - 1)M + 1, nM], where M is an oversampling rate of the receive end.
In some optional embodiments of this application, the processing module is specifically configured to perform filtering processing on the first waveform signal, to obtain a second waveform signal; and the processing module is specifically configured to perform quantization processing on the second waveform signal, to obtain the quantized symbol sequence.
According to a fifth aspect, an embodiment of this application proposes a communication apparatus, including: a transceiver, configured to obtain a first data stream; and a processor, configured to process the first data stream by using an encoding neural network, to obtain a first symbol stream, where the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network, where the transceiver is further configured to output the first symbol stream.
In some optional embodiments of this application, the transceiver is further configured to receive a first weight, where the first weight is from the decoding neural network, and the first weight is used to train the first neural network.
In some optional embodiments of this application, the transceiver is further configured to receive a first gradient, where the first gradient is from the decoding neural network, and the first gradient is used to train the first neural network.
In some optional embodiments of this application, the transceiver is further configured to receive a second function, where the second function is from the decoding neural network, and the second function is a loss function and/or a reward function; and the transceiver is further configured to process the second function by using the encoding neural network, to obtain a second gradient, where the second gradient is used to train the encoding neural network.
In some optional embodiments of this application, the processor is specifically configured to perform filtering processing on the first symbol stream, to obtain a first waveform signal, where an out-of-band signal is filtered out from the first waveform signal; and the transceiver is specifically configured to output the first waveform signal.
In some optional embodiments of this application, the processor is specifically configured to perform encoding processing on the first data stream, to obtain a first channel encoding code word; and the processor is specifically configured to process the first channel encoding code word by using the encoding neural network, to obtain the first symbol stream.
In some optional embodiments of this application, the processor is specifically configured to process the first channel encoding code word by using the encoding neural network, to obtain a first intermediate symbol; and the processor is specifically configured to perform filtering processing on the first intermediate symbol by using a filtering neural network, to obtain the first symbol stream, where intersymbol interference is introduced into the first symbol stream.
In some optional embodiments of this application, the processor is specifically configured to perform upsampling processing on the first intermediate symbol, to obtain a second intermediate symbol; and the processor is specifically configured to perform filtering processing on the second intermediate symbol by using the filtering neural network, to obtain the first symbol stream.
According to a sixth aspect, an embodiment of this application proposes a communication apparatus, including: a transceiver, configured to obtain a first data stream, where the transceiver is further configured to receive a first symbol stream, where the first symbol stream is from a transmit end, the first symbol stream is obtained by the transmit end by processing the first data stream by using an encoding neural network, the encoding neural network includes a first neural network, and the first neural network reuses a part or all of a decoding neural network and/or a parameter of a part or all of the decoding neural network; and a processor, configured to process the first symbol stream by using the decoding neural network, to obtain a second data stream, where the processor is further configured to train the decoding neural network based on the first data stream and the second data stream.
In some optional embodiments of this application, the processor is specifical
CLAIMS
Claims ( 21 )
What is claimed is:
1 - 20 . (canceled)
21 . A method, applied to a transmit end, the method comprising:
obtaining, by the transmit end, a first data stream; processing, by the transmit end, the first data stream using an encoding neural network, to obtain a first symbol stream, wherein the encoding neural network reuses a part or all of a neural network that is the same as a decoding neural network that corresponds to the first data stream; and outputting, by the transmit end, the first symbol stream.
22 . The method according to claim 21 , further comprising:
after outputting, by the transmit end, the first symbol stream, receiving, by the transmit end, a first weight, wherein the first weight is from the decoding neural network, and the first weight is used to train the encoding neural network.
23 . The method according to claim 21 , further comprising:
after outputting, by the transmit end, the first symbol stream, receiving, by the transmit end, a first gradient, wherein the first gradient is from the decoding neural network, and the first gradient is used to train the encoding neural network.
24 . The method according to claim 21 , further comprising:
after outputting, by the transmit end, the first symbol stream, receiving, by the transmit end, a second function, wherein the second function is from the decoding neural network, and the second function is a loss function or a reward function; and processing, by the transmit end, the second function using the encoding neural network, to obtain a second gradient, wherein the second gradient is used to train the encoding neural network.
25 . The method according to claim 21 , wherein outputting, by the transmit end, the first symbol stream comprises:
performing, by the transmit end, filtering processing on the first symbol stream, to obtain a first waveform signal, wherein an out-of-band signal is filtered out from the first waveform signal; and outputting, by the transmit end, the first waveform signal.
26 . The method according to claim 21 , wherein processing, by the transmit end, the first data stream using the encoding neural network, to obtain the first symbol stream comprises:
performing, by the transmit end, encoding processing on the first data stream, to obtain a first channel encoding code word; and processing, by the transmit end, the first channel encoding code word using the encoding neural network, to obtain the first symbol stream.
27 . The method according to claim 21 , wherein the encoding neural network reusing a part or all of the neural network that is the same as the decoding neural network that corresponds to the first data stream comprises:
reusing, by the encoding neural network, a part or all of a model of the decoding neural network that corresponds to the first data stream, a loss function of the decoding neural network that corresponds to the first data stream, a reward function of the decoding neural network that corresponds to the first data stream, or a parameter of the decoding neural network that corresponds to the first data stream.
28 . A method, applied to a receive end, the method comprising:
obtaining, by the receive end, a first data stream; receiving, by the receive end, a first symbol stream, wherein the first symbol stream is received from a transmit end, the first symbol stream is obtained by the transmit end by processing the first data stream by using an encoding neural network, and the encoding neural network reuses a part or all of a neural network that is the same as a decoding neural network; processing, by the receive end, the first symbol stream using the decoding neural network, to obtain a second data stream; and training, by the receive end, the decoding neural network based on the first data stream and the second data stream.
29 . The method according to claim 28 , wherein training, by the receive end, the decoding neural network based on the first data stream and the second data stream comprises:
processing, by the receive end, the first data stream and the second data stream using the decoding neural network, to obtain a first function, wherein the first function is a loss function or a reward function; processing, by the receive end, the first function using the decoding neural network, to obtain a first gradient; and training, by the receive end, the decoding neural network by using the first gradient.
30 . The method according to claim 29 , further comprising:
after training, by the receive end, the decoding neural network using the first gradient, sending, by the receive end, the first gradient to the transmit end, wherein the first gradient is used to train the encoding neural network; or generating, by the receive end, a first weight using the decoding neural network, wherein the first weight is a weight of the decoding neural network, and sending, by the receive end, the first weight to the transmit end, wherein the first weight is used to train the encoding neural network.
31 . The method according to claim 29 , further comprising:
after training, by the receive end, the decoding neural network based on the first data stream and the second data stream, processing, by the receive end, the first data stream and the second data stream using the decoding neural network, to obtain a second function, wherein the second function is a loss function or a reward function; and sending, by the receive end, the second function to the transmit end, wherein the second function is used to train the encoding neural network.
32 . The method according to claim 28 , wherein processing, by the receive end, the first symbol stream using the decoding neural network, to obtain the second data stream comprises:
receiving, by the receive end, a first waveform signal, wherein the first waveform signal is obtained by the transmit end by performing filtering processing on the first symbol stream; performing, by the receive end, quantization processing on the first waveform signal, to obtain a quantized symbol sequence, wherein a group of the first data streams is mapped to one or more quantized symbol sequence, bit flipping exists at most once in each symbol of the quantized symbol sequence, and the bit flipping is 0 to 1 or 1 to 0; and processing, by the receive end, a first sequence in the quantized symbol sequence using the decoding neural network, to obtain the second data stream.
33 . The method according to claim 32 , wherein performing, by the receive end, quantization processing on the first waveform signal, to obtain the quantized symbol sequence comprises:
receiving, by the receive end, the first waveform signal interfered by a first channel, wherein the first channel is a channel between the receive end and the transmit end; and performing, by the receive end, quantization processing on the first waveform signal interfered by the first channel, to obtain the quantized symbol sequence.
34 . An communication apparatus, comprising:
a processor; and a transceiver connected to the processor; wherein the processor is configured to execute program code stored in a memory, and when the program code is executed, the apparatus is enabled to:
obtain a first data stream;
process the first data stream using an encoding neural network, to obtain a first symbol stream, wherein the encoding neural network reuses a part or all of as neural network that is the same as a decoding neural network that corresponds to the first data stream; and
output the first symbol stream.
35 . The communication apparatus according to claim 34 , wherein when the program code is executed, the apparatus is further enabled to:
receive a first weight, wherein the first weight is from the decoding neural network, and the first weight is used to train the encoding neural network.
36 . The communication apparatus according to claim 34 , wherein when the program code is executed, the apparatus is further enabled to:
receive a first gradient, wherein the first gradient is from the decoding neural network, and the first gradient is used to train the encoding neural network.
37 . The communication apparatus according to claim 34 , wherein when the program code is executed, the apparatus is further enabled to:
receive a second function, wherein the second function is from the decoding neural network, and the second function is a loss function or a reward function; and process the second function using the encoding neural network, to obtain a second gradient, wherein the second gradient is used to train the encoding neural network.
38 . The communication apparatus according to claim 34 , wherein when the program code is executed, the apparatus is enabled to:
perform filtering processing on the first symbol stream, to obtain a first waveform signal, wherein an out-of-band signal is filtered out from the first waveform signal; and output the first waveform signal.
39 . The communication apparatus according to claim 34 , wherein when the program code is executed, the apparatus is enabled to:
perform encoding processing on the first data stream, to obtain a first channel encoding code word; and process the first channel encoding code word using the encoding neural network, to obtain the first symbol stream.
40 . The communication apparatus according to claim 34 , wherein when the program code is executed, the apparatus is enabled to:
reuse a part or all of a model of the decoding neural network, a loss function of the decoding neural network, a reward function of the decoding neural network, or a parameter of the decoding neural network.
US18/192,276
2020-09-30
2023-03-29
Neural Network-Based Communication Method and Related Apparatus
Pending
US20230299872A1
( en )
Applications Claiming Priority (3)
Application Number
Priority Date
Filing Date
Title
CN202011062368.2A
CN114337911A
( en )
2020-09-30
2020-09-30
Communication method based on neural network and related device
CN202011062368.2
2020-09-30
PCT/CN2021/120839
WO2022068757A1
( en )
2020-09-30
2021-09-27
Neural network-based communication method and related apparatus
Related Parent Applications (1)
Application Number
Title
Priority Date
Filing Date
PCT/CN2021/120839
Continuation
WO2022068757A1
( en )
2020-09-30
2021-09-27
Neural network-based communication method and related apparatus
Publications (1)
Publication Number
Publication Date
US20230299872A1
true
US20230299872A1 ( en )
2023-09-21
Family
ID=80949652
Family Applications (1)
Application Number
Title
Priority Date
Filing Date
US18/192,276
Pending
US20230299872A1
( en )
2020-09-30
2023-03-29
Neural Network-Based Communication Method and Related Apparatus
Country Status (4)
Country
Link
US
( 1 )
US20230299872A1
( en )
EP
( 1 )
EP4210252A4
( en )
CN
( 1 )
CN114337911A
( en )
WO
( 1 )
WO2022068757A1
( en )
Cited By (7)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
CN115441993A
( en )
*
2022-09-01
2022-12-06
ä¸å½äººæ°è§£æ¾åå½é²ç§æ大å¦
Channel coding and decoding method, device, equipment and storage medium
US20230055263A1
( en )
*
2021-08-21
2023-02-23
Dynam.Ai, Inc.
Stratification in non-classified heterogeneous object labels
US20230308317A1
( en )
*
2020-08-20
2023-09-28
Nokia Technologies Oy
Neural-Network-Based Receivers
US20230389057A1
( en )
*
2020-10-21
2023-11-30
Interdigital Patent Holdings, Inc.
Methods, apparatus, and systems for artificial intelligence (ai)-enabled filters in wireless systems
WO2025077623A1
( en )
*
2023-10-09
2025-04-17
Guangdong Oppo Mobile Telecommunications Corp., Ltd.
Systems and methods for obtaining neural networks for data compression
WO2025107925A1
( en )
*
2023-11-22
2025-05-30
å为ææ¯æéå ¬å¸
Communication method and apparatus
EP4525335A4
( en )
*
2022-05-31
2025-08-27
Huawei Tech Co Ltd
SIGNAL PROCESSING METHOD AND DEVICE
Families Citing this family (12)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
CN113313233A
( en )
*
2021-05-17
2021-08-27
æé½æ¶è¯ç§ææéå ¬å¸
Neural network configuration parameter training and deploying method and device for dealing with device mismatch
CN115186797B
( en )
*
2022-06-06
2023-05-30
山西大å¦
A kind of pragmatic communication method and system
WO2024007189A1
( en )
*
2022-07-06
2024-01-11
Nokia Shanghai Bell Co., Ltd.
Scalable and quick waveform learning in multi-user communication system
CN115801509B
( en )
*
2022-11-04
2023-09-05
ä¸å½äººæ°è§£æ¾åå½é²ç§æ大å¦
Method for directly estimating single communication signal symbol sequence under interference condition
CN118075120A
( en )
*
2022-11-14
2024-05-24
ç»´æ²ç§»å¨éä¿¡æéå ¬å¸
Method, device and communication equipment for updating AI network model
CN116545804B
( en )
*
2023-05-16
2026-03-31
ä¸æµ·äº¤é大å¦
End-to-end optimization method and system for optical communication systems based on noise adaptation networks
EP4701107A4
( en )
*
2023-06-29
2026-05-13
Huawei Tech Co Ltd
DATA PROCESSING METHOD AND ASSOCIATED EQUIPMENT
CN121713525A
( en )
*
2023-09-05
2026-03-20
å为ææ¯æéå ¬å¸
Communication method and related equipment
CN119583017A
( en )
*
2023-09-07
2025-03-07
å为ææ¯æéå ¬å¸
Characteristic data transmission method and related equipment
CN119946611A
( en )
*
2023-11-03
2025-05-06
å为ææ¯æéå ¬å¸
A communication method and related equipment
CN120263341A
( en )
*
2024-01-03
2025-07-04
ä¸å ´é讯è¡ä»½æéå ¬å¸
Communication method, device and storage medium
CN119728373B
( en )
*
2024-12-17
2026-01-06
ä¸æµ·å®éªå®¤
A real-signal OFDM underwater acoustic communication method and system
Family Cites Families (7)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
KR102165645B1
( en )
*
2017-05-03
2020-10-14
ë²ì§ëì í í¬ ì¸í°ë ì¶ì¼ íë¼í¼í°ì¤, ì¸í¬.
Learning and deployment of adaptive wireless communication
CN111434049B
( en )
*
2017-06-19
2021-06-11
å¼åå°¼äºç§æç¥è¯äº§ææéå ¬å¸
Encoding and decoding of information transmitted wirelessly using a multi-antenna transceiver
EP3418948B1
( en )
*
2017-06-19
2020-06-03
Nokia Technologies Oy
Data transmission network configuration
WO2020092391A1
( en )
*
2018-10-29
2020-05-07
Board Of Regents, The University Of Texas System
Low resolution ofdm receivers via deep learning
US20220083870A1
( en )
*
2019-01-18
2022-03-17
Nokia Technologies Oy
Training in Communication Systems
CN110460402B
( en )
*
2019-07-15
2021-12-07
åå°æ»¨å·¥ç¨å¤§å¦
End-to-end communication system establishing method based on deep learning
CN111327381A
( en )
*
2020-02-04
2020-06-23
æ¸ å大å¦
Joint optimization method of wireless communication physical layer transceiver based on deep learning
2020
2020-09-30
CN
CN202011062368.2A
patent/CN114337911A/en
active
Pending
2021
2021-09-27
EP
EP21874417.5A
patent/EP4210252A4/en
active
Pending
2021-09-27
WO
PCT/CN2021/120839
patent/WO2022068757A1/en
not_active
Ceased
2023
2023-03-29
US
US18/192,276
patent/US20230299872A1/en
active
Pending
Cited By (8)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20230308317A1
( en )
*
2020-08-20
2023-09-28
Nokia Technologies Oy
Neural-Network-Based Receivers
US12267189B2
( en )
*
2020-08-20
2025-04-01
Nokia Technologies Oy
Neural-network-based receivers
US20230389057A1
( en )
*
2020-10-21
2023-11-30
Interdigital Patent Holdings, Inc.
Methods, apparatus, and systems for artificial intelligence (ai)-enabled filters in wireless systems
US20230055263A1
( en )
*
2021-08-21
2023-02-23
Dynam.Ai, Inc.
Stratification in non-classified heterogeneous object labels
EP4525335A4
( en )
*
2022-05-31
2025-08-27
Huawei Tech Co Ltd
SIGNAL PROCESSING METHOD AND DEVICE
CN115441993A
( en )
*
2022-09-01
2022-12-06
ä¸å½äººæ°è§£æ¾åå½é²ç§æ大å¦
Channel coding and decoding method, device, equipment and storage medium
WO2025077623A1
( en )
*
2023-10-09
2025-04-17
Guangdong Oppo Mobile Telecommunications Corp., Ltd.
Systems and methods for obtaining neural networks for data compression
WO2025107925A1
( en )
*
2023-11-22
2025-05-30
å为ææ¯æéå ¬å¸
Communication method and apparatus
Also Published As
Publication number
Publication date
WO2022068757A1
( en )
2022-04-07
CN114337911A
( en )
2022-04-12
EP4210252A4
( en )
2024-02-28
EP4210252A1
( en )
2023-07-12
Similar Documents
Publication
Publication Date
Title
EP4210252A1
( en )
2023-07-12
Neural network-based communication method and related apparatus
US20230342593A1
( en )
2023-10-26
Neural network training method and related apparatus
CN110741553B
( en )
2023-11-03
Neural network for forward error correction decoding
JP7228042B2
( en )
2023-02-22
Encoding method and device and decoding method and device
EP4220484A1
( en )
2023-08-02
Communication method, apparatus and system
CN117083839A
( en )
2023-11-17
Channel estimation method and device based on artificial intelligence
Gao et al.
2020
Learning to decode polar codes with one-bit quantizer
CN109831281B
( en )
2020-10-02
A low-complexity sparse code multiple access system multi-user detection method and device
Cao et al.
2020
Learning to denoise and decode: A novel residual neural network decoder for polar codes
WO2026007544A1
( en )
2026-01-08
Communication method and related apparatus
CN111769975A
( en )
2020-10-13
MIMO system signal detection method and system
WO2023283785A1
( en )
2023-01-19
Method for processing signal, and receiver
CN110474798A
( en )
2019-11-19
A method of wireless communication future signal is predicted using echo state network
Anderson et al.
2019
Deep modulation (Deepmod): A self-taught PHY layer for resilient digital communications
KR20210133687A
( en )
2021-11-08
Apparatus for transmitting data based on mimi-auto encoder
WO2023097645A1
( en )
2023-06-08
Data acquisition method and apparatus, and device, medium, chip, product and program
Boiko et al.
2025
Intelligent DRL-assisted decoding of error-correcting codes for 5G/6G telecommunication channels
US12101102B2
( en )
2024-09-24
Decoding method and apparatus
EP4550175A1
( en )
2025-05-07
Expansion or compression of transport block(s) based on a reverse autoencoder neural network
EP4550174A1
( en )
2025-05-07
Expansion or compression of transport block(s) based on a reverse autoencoder neural network
CN116961822A
( en )
2023-10-27
A communication system and device
EP4521662A1
( en )
2025-03-12
Channel encoding/decoding method and apparatus
US20250247271A1
( en )
2025-07-31
Communication Systems
Murugan et al.
2025
Neural Network-Based Adaptive Error Correction for High-Capacity Massive MIMO Systems
WO2024123224A1
( en )
2024-06-13
Hybrid quantum autoencoder for decoder in the physical layer
Legal Events
Date
Code
Title
Description
2023-07-27
STPP
Information on status: patent application and granting procedure in general
Free format text : DOCKETED NEW CASE - READY FOR EXAMINATION
2025-08-26
STPP
Information on status: patent application and granting procedure in general
Free format text : NON FINAL ACTION MAILED
2025-10-21
STPP
Information on status: patent application and granting procedure in general
Free format text : RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER
Free format text : NON FINAL ACTION COUNTED, NOT YET MAILED
2025-10-30
STPP
Information on status: patent application and granting procedure in general
Free format text : RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER
2026-02-05
STPP
Information on status: patent application and granting procedure in general
Free format text : NON FINAL ACTION COUNTED, NOT YET MAILED
2026-02-18
STPP
Information on status: patent application and granting procedure in general
Free format text : NON FINAL ACTION MAILED
2026-04-30
STPP
Information on status: patent application and granting procedure in general
Free format text : RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER