Neural network based receiver

Abstract

In some examples, a node for a telecommunications network includes a neural network-based receiver for uplink communications. The node is configured to modify the neural network-based receiver to generate a set of modified receiver frames that define respective different versions for the receiver, use each modified receiver frame to generate a respective measurement representing bits encoded by a signal received at the node, calculate a value representing a variance of the measurement, and determine whether to select a signal received at the node for use as part of a training data set for the neural network-based receiver based on the value.

Description

Technical Field

[0001]

[0013] Aspects relate generally to telecommunication networks and, more particularly, although not limited thereto, to neural network based receivers. Background Art

[0002] In a simple representation, a wireless telecommunications network comprises a transmitting side and a receiving side, and a wireless channel between them. The transmitting side usually consists of a data source that generates data (bits) and a modulation system that includes a carrier wave that is modulated by the data to provide an output signal. The output (output signal) of the transmitting side is sent to the receiving side through the channel. In general, the channel corrupts the transmitted output signal with noise and any interference that may be present due to adverse channel conditions.

[0003] The receiver on the receiving side can be used to demodulate the signal received from the transmitting side through the channel for data recovery. The receiver is usually configured to demodulate the received signal based on algorithms for channel estimation, equalization, symbol detection, etc., which are pre-configured before the receiver is deployed in the network. As such, the receiver, while being able to demodulate a portion of the received signal from the transmitting side of the network, may not be able to demodulate all of the signal due to dominant channel conditions that may damage the output signal. Summary of the invention

[0004] According to an example, a node for a telecommunications network is provided, the node including a neural network-based receiver for uplink communication, wherein the node is configured to modify the neural network-based receiver to generate a set of modified receiver frames, the set of modified receiver frames defining respective different versions for the receiver, using each of the modified receiver frames to generate a respective measurement representing a bit encoded by a signal received at the node, calculating a value representing a variance of the measurement, and based on the value, determining whether to select the signal received at the node to be used as part of a training data set for the neural network-based receiver. Accordingly, the NN-based receiver can be trained and calibrated after deployment. This enables the NN-based receiver to adjust its operation according to the environment at the time. If the signal received at the receiver is similar to the samples used in the training data, the model can reliably detect the bits encoded by the received signal waveform and will have a lower uncertainty. However, if the sample is significantly different from the sample presented in the training data, the model must be extrapolated and there is typically a higher model uncertainty. Such model uncertainty can be captured by receiving a given waveform with several randomly manipulated versions of the neural network that supports the NN-based receivers in the network nodes, and using them to calculate the variance of the detected bits. High variance means high model uncertainty. Note that this variance / uncertainty is different from detection uncertainty, which is caused by the noisy information channel. In other words, model uncertainty is the uncertainty of detection uncertainty. Therefore, the collected data can be used in the cloud to retrain NN-based receivers that are configured in a similar way to the deployed receivers with the exact amount of data that needs to be used according to the NN architecture.

[0005] The neural network based receiver may be modified by applying a drop mask. The node may compare a value representing the variance to a threshold. The node may receive data representing the threshold and use the threshold to adjust the selection of signals received at the node for use as part of the training data set. In an example, the corresponding measurement is a log likelihood ratio value LLR.

[0006] The node may calculate a measure of variance of a plurality of LLR values ​​stored in a temporary storage device of the node, and determine a median of the variance. In an example, the neural network based receiver is a radio receiver.

[0007] According to an example, a method for selecting training samples for a neural network-based receiver configured for uplink communications in a selected deployment environment of a telecommunications network is provided, the method comprising generating a plurality of measurements using respective different neural network-based receiver frameworks, the plurality of measurements representing bits encoded by a signal received at the receiver, calculating a variance of the measurements, and determining whether to select the signal received at the receiver as part of a training data set based on a comparison of the variance with a threshold. The method may also include applying a random drop mask to the neural network-based receiver deployed in the selected environment so as to generate the different neural network-based receiver frameworks. The method may also include receiving a threshold from a core network entity of the telecommunications network. The method may also include sending the signal received at the receiver and the plurality of measurements to the core network entity of the telecommunications network. The plurality of metric values ​​may be LLR values.

[0008] According to an example, a network entity for a telecommunications network is provided, the network entity being configured to receive a signal and a corresponding set of data bits from a node of the telecommunications network, train a neural network-based receiver using the signal and the corresponding set of data bits, the neural network-based receiver being configured for uplink communications to the node, determine a measure representing a degree of overfitting for a model, and based on the measure, provide an updated neural network-based receiver to the node, or generate a request for additional data from the node. The network entity may modify a threshold based on the measure, and send the modified threshold to the node. The network entity may determine a backhaul capacity associated with the node, and based on the determined backhaul capacity, determine a measure representing a degree of overfitting for the model. In an example, the network entity may be in the form of a cloud-based training entity configured to receive training data uploaded by a network node, such as a node described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] For a more illustrative understanding of the present disclosure, reference is now made, by way of example only, to the following description taken in conjunction with the accompanying drawings, in which:

[0010] Figure 1 is a schematic diagram of a node for a telecommunications network according to an example;

[0011] Figure 2 is the choice of example threshold Schematic diagram of the method;

[0012] Figure 3 is a schematic diagram of a process for data selection according to an example; and

[0013] Figure 4 is a schematic diagram of a node according to an example. DETAILED DESCRIPTION

[0014] The example embodiments are described in full detail below so that those of ordinary skill in the art can embody and implement the systems and processes described herein. It is important to understand that the embodiments can be provided in many alternative forms and should not be construed as being limited to the examples set forth herein. Therefore, although the embodiments can be modified in various ways and take various alternative forms, their specific embodiments are shown in the drawings and described in detail below as examples. It is not intended to be limited to the specific form disclosed. On the contrary, all modifications, equivalents, and alternatives falling within the scope of the appended claims should be included. The elements of the example embodiments are consistently represented by the same reference numerals throughout the drawings and the appropriate places of the specific embodiments.

[0015] The terms used to describe the embodiments herein are not intended to limit the scope. The articles "one", "an" and "the" are singular because they have a single referent, but the use of the singular form in this document should not exclude the occurrence of more than one referent. In other words, the elements referred to in the singular form can be one or more, unless the context clearly states otherwise. It will be further understood that the terms "include" and / or "comprise" specify the presence of the features, items, steps, operations, elements and / or components when used herein, but do not exclude the presence or addition of one or more other features, items, steps, operations, elements, components and / or their groups.

[0016] Unless otherwise defined, all terms (including technical and scientific terms) used in this article should be interpreted as common terms in the art. It should be further understood that, unless explicitly defined herein, common terms should also be interpreted as common meanings in the relevant art, rather than idealized or overly formal meanings.

[0017] Recent advances in deep learning in areas such as natural language processing, image processing, autonomous driving, and more have spurred interest in its application in the field of communication signal processing. For example, deep neural networks (DNNs) can be used to perform detailed waveform classification, and it is possible to implement digital receiver chains for receivers in telecommunication networks using neural networks (NNs). Thus, there is no need to implement explicit algorithms, such as for channel estimation, equalization, symbol detection, and demodulation, and NN-based receivers can be trained with simulated (or real) data prior to deployment to enable them to perform operations implicitly.

[0018] Nevertheless, if the NN-based receiver can specialize its operation according to the channel conditions in the deployment area, its performance will be improved, as the conditions in different deployment locations can vary significantly. However, in order to do this, the NN-based receiver needs to be retrained in the field to adapt to the channel conditions at the time. This may introduce the need for additional hardware in the network nodes implementing the NN-based receiver. This is not desirable as the cost efficiency of the radio components in such nodes is strictly controlled. In addition, the bandwidth in the network backhaul link is very limited as the actual data traffic must be prioritized. As a result, this means that nodes are often unable to send training data to the extent required to provide a fully trained model.

[0019] Therefore, while it is desirable to be able to tailor NN-based receivers to the conditions experienced at the deployment site, being able to select, collect and transmit only essential data from a specific environment, use that data to retrain the NN receiver, and then upload the new model to the node is problematic. Moreover, this must be done in a scalable manner.

[0020] According to an example, a node for a telecommunications network is provided. The node includes a neural network-based receiver for uplink communications. The operation of the NN-based receiver can be adjusted according to the dominant channels, interference, and traffic statistics in the node's deployment environment. Thus, in an example, the NN-based receiver can be pre-configured with a non-location-specific reception model. This initial model can be configured to handle common non-location-specific reception scenarios, while location-specific phenomena can be autonomously learned using field data.

[0021] In implementation, a deployed NN-based receiver may collect a certain proportion of received waveforms and their corresponding bits for model training purposes. For example, a predefined proportion, such as a percentage (e.g., 1%) of a signal waveform received by the NN-based receiver can be stored as a sample.

[0022] Some of these samples can be randomly selected (forming a dataset D random ) to be selected, while the other component can include samples with high modeling uncertainty (forming dataset D uncertain ). In the example, samples with high modeling uncertainty may generally refer to received waveforms that differ in some way from waveforms represented in earlier training data. That is, it is possible that an NN-based receiver may be able to detect such signals accurately enough to facilitate successful decoding, but such detection may be, for example, "lucky" and based on extrapolation if the signal was not previously provided in the training data for the NN-based receiver. Alternatively, the neural network may perform poorly for such signals but still achieve sufficient accuracy for the decoder to correct all bit errors.

[0023] According to an example, there is uncertainty associated with the degree to which a bit can be detected from a noise signal. In an NN-based receiver, this uncertainty can be represented by a log-likelihood ratio (LLR) value and can be referred to as "detection uncertainty". Factors such as noise level have an impact on this uncertainty. On the other hand, there is another type of uncertainty (which can be referred to as "model uncertainty") that defines how well the model is prepared so that it can detect the signal. For example, if the signal samples are similar to those in the training data, the model can reliably detect the bits encoded by the received signal waveform and the uncertainty will be low. However, if the samples are very different from those presented in the training data, the model must be extrapolated and there is usually a higher model uncertainty. This uncertainty can be inferred, for example, using integration methods or variational methods.

[0024] In an example, this model uncertainty can be captured by receiving a given waveform using several randomly manipulated versions of a neural network that supports NN-based receivers in network nodes and calculating the variance of the detected bits. High variance means high model uncertainty. Note that this variance / uncertainty is different from the detection uncertainty caused by the noisy information channel. In other words, model uncertainty is the uncertainty of detection uncertainty.

[0025] The collected data can be used in the cloud to retrain a NN-based receiver that is configured in a similar manner to the deployed receiver using the exact amount of data that needs to be used according to the NN architecture. In the event that the cloud-based version overfits the data, more data can be requested from the receiver deployed in the node, which can be determined using any of the normal processes for overfitting detection. The cloud-based version that has been trained with a sufficient amount of data from the deployment environment can be uploaded to the node to replace the initial NN-based receiver in order to improve detection performance. Accordingly, the NN-based receiver can be trained and calibrated after deployment. This enables the NN-based receiver to adjust its operation according to the environment at the time.

[0026] Figure 1 1 is a schematic diagram of a node for a telecommunications network according to an example. The node 100 includes a NN-based receiver 101. A backhaul link 103 is provided between the node 100 and a network entity, such as a cloud component 105 of the core of the telecommunications network. In an example, the backhaul link 103 can be used to send data from a temporary buffer 107 of the NN-based receiver 101 to a training data controller 109 of the component 105. Figure 1In the example of , the NN-based receiver 101 is configured to receive a signal 111. The signal 111 may be received from, for example, a user equipment operating in a telecommunications network. The received signal 111 is decoded using a pre-trained NN model 113. That is, data in the form of bits that have been used to modulate a carrier signal is received by the NN-based receiver 101 and decoded using the NN 113 in order to recover the data. The received signal 111 may be stored in a temporary storage device 117 of the NN-based receiver 101.

[0027] As described above, channel conditions may corrupt signal 111 such that the decoded data is different than the transmitted data.A cyclic redundancy check (CRC) 115 may be performed to determine if errors are present.

[0028] If the result of the CRC indicates a decoding error, the result of the CRC can be stored in a temporary storage device 117 of the NN-based receiver 101 and associated with the corresponding signal 111. If the CRC indicates a correct decoding, the NN-based receiver 101 can determine (119) whether the signal 111 that resulted in the correct CRC is a retransmission. If it is a retransmission, the signal 111 can be retrieved from the temporary storage device 117 and assigned (121) to the decoded data bits encoded by the signal 111. In other words, a signal that cannot be decoded from the first transmission can be stored in the temporary storage device 117 while waiting for a retransmission of the corresponding packet.

[0029] Once it is retransmitted and successfully decoded, the final bit can be associated with the original signal.Thus, the NN-based receiver 101 expects that the poorly detected signal can be more accurately detected for improving the performance of the NN-based receiver.

[0030] As will be described in more detail below, the NN-based receiver 101 can determine (123) whether to upload the signal 111 and the corresponding decoded data bits to the component 105. If not, they can be discarded (125). Otherwise, they can be provided to the temporary buffer 107 to be sent to the training data controller 109 of the component 105 via the backhaul 103. Figure 1 In the example of , the training data controller 109 can communicate with the training data storage device 127, which can store training data that can be used to train the NN model for decoding the signal 111. Figure 1 In the example of , data from temporary buffer 107 can be stored in storage device 127, and controller 109 can retrieve such data and provide it to temporary training data storage device 129, where, as long as it is not older than a pre-selected threshold, it can be used to train (131) the above-mentioned model.

[0031] Component 105 may determine (133) whether the model is overfitting the training data. If not, the model may be uploaded 137 to node 100 and thereby form an updated version of the pre-trained NN 113 for NN-based receiver 101. If overfitting is present, more data may be requested (135) from node 100. Figure 1 In the example of , the request for more data can inform a decision at the NN-based receiver 101 regarding whether to upload (123) data to the component 105. Thus, the result of the determination (133) regarding overfitting indicates that overfitting can result in sending (139) a request for data 135 to the node 100.

[0032] In order to generate training data from its own deployment environment, the node 100 can use different types of data, so that it can have sufficient variability in the training data. According to the example, and as mentioned above Figure 1 As described, the node 100 can randomly save a certain proportion of the RX signal 111 and combine these signals with the decoded bits to form a label for the signal. In an alternative implementation, the coded bits can be used as labels, and the NN-based receiver 101 can output, for example, LLRs, which can be decoded with a traditional LDPC decoder.

[0033] The proportion of data saved (compared to the total amount of data received at the node 100 ) may be of the order of 1-5% to avoid blocking the backhaul link 103 when transmitting data to the cloud component 105 .

[0034] According to an example, the node 100 may collect data that is not decodable when it is transmitted for the first time, for example. Figure 1 As described, this can be accomplished by holding the undecodable RX signal 111 in a buffer (107) while awaiting a retransmission of the corresponding packet. Once it is retransmitted and successfully decoded, for example with the aid of HARQ, the final bits can be associated with the initial RX signal and potentially uploaded to the cloud component 105. This allows the performance of the NN-based receiver 101 to be improved, as these poorly detected signals are signals that the NN should ideally be able to detect more accurately, as they are likely the result of channel conditions experienced due to the deployment environment of the node 100.

[0035] In an example, the data uploaded to component 105 for training may include the above two cases, thereby ensuring a balanced representation of low SINR and high SINR data samples. The randomly generated training data set is referred to herein as D random In addition, node 100 can save samples decoded with large model uncertainty. This can be performed with a small number of data samples. A dataset with such uncertain samples is called Duncertain . Model uncertainty prediction can be performed in node 100, but it is not necessary to perform training at node 100. Even though NN 113 outputs (log) likelihood ratios (LLRs), these essentially describe the effects of noise / interference, rather than the uncertainty of the model itself. According to the example, estimates of model uncertainty can be calculated using variational inference or over a collection of models. For example, dropout can be applied during training, which is a regularization method to reduce overfitting. In dropout, a portion of the model activations are randomly dropped (i.e., set to zero) at each training step. However, dropout is a form of variational inference, so model uncertainty can be calculated by applying dropout during inference. Therefore, even if the training itself is performed centrally in the core cloud 105, the method can be applied at node 100.

[0036] Therefore, according to an example, the modeling uncertainty for the i-th subcarrier on the j-th OFDM symbol can be expressed as:

[0037]

[0038] Where LLR ij，n is the nth sample of the LLR computed using either the randomly dropped activation (dropout) or the nth model (in the model ensemble), and N s is the number of dropouts / ensembles of model samples, and μ ij is the average

[0039] Thus, the samples to be transmitted back for inclusion in the dataset D can be determined by those samples that meet the following criteria: uncertain A sample of:

[0040] median(δ ij )>δ threshold

[0041] where the median is calculated over all indices i and j in a TTI, and δ threshold is the chosen threshold.

[0042] Figure 2 is the choice of example threshold In an example, the threshold calculation may be performed in the central core component 105 , which then forwards the threshold to the node 100 .

[0043] In block 201, a new training data sample is obtained from node 100 via backhaul 103 at training data controller 109. In block 203, component 105 determines whether backhaul 103 capacity exists between component 105 and node 100. If it is determined that there is no available backhaul capacity, then delta is increased in block 205. threshold , and transmits the increased value to node 100 in box 207. If in box 203, component 105 determines that there is backhaul capacity between component 105 and node 100, it is determined in box 209 whether the existing model is overfitted using the current data. If not, increase δ in box 205 threshold The value of , and transmits the increased value to the node 100 in box 207. If overfitting occurs, then reduce δ in box 211 threshold and transmits the reduced value to the node 100 in block 207. Figure 2 In the example of , a model uncertainty threshold for data collection can be determined in the core component 105 and transmitted to the node 100 in order to adjust the transmission of training data from the node 100 to the component 105 via the backhaul 103. Therefore, the threshold is increased to reduce the amount of training data transmitted. In the example, this is done when the backhaul resources are low and / or the model is no longer overfitted.

[0044] Figure 3 is a schematic diagram for the data selection process according to an example. In particular, Figure 3 It is a process for uncertainty calculation and data selection at a node (such as node 100) of a telecommunications network. Figure 3 In the example above, you can refer to Figure 1 The process is implemented in block 123 of the node 100 described in the embodiment of the present invention. Figure 3 The procedure for determining whether a single TTI should be added to a dataset D uncertain The initial criteria for using the signal for uncertainty calculations can be based on random selection (e.g., 1% of the data), or uncertainty calculations can be performed at fixed intervals. Figure 3 In the example of , the process represents the case where the signal can be decoded without retransmission. The process is essentially similar for the case where retransmission is required, but it is not included in the figure for simplicity and readability. In the example, retransmissions can be used to determine the correct information bits (which are used as labels during training), but the LLRs used for uncertainty calculation can be obtained using the initial transmission.

[0045] refer to Figure 3In block 301, a signal 111 is received at a node 100, for example, over a transmission time interval (TTI). In block 303, LLRs are detected from the signal received at 301 using the NN decoder 113. In block 305, a low density parity check (LDPC) is performed and data bits 307 encoded by the signal 111 are determined. In block 309, it is determined whether to use the signal 111 for uncertainty calculation. As described above, a predefined proportion of the signals may be selected, for example, randomly. If a signal is not selected in block 309, it is discarded for the purpose of uncertainty calculation (311).

[0046] If signal 111 is selected in box 309, the counter is set to zero in box 313. In box 315, the random drop mask is applied to the NN models 113 of the NN-based receiver 101 to generate a set of modified receiver frames that define corresponding different versions for the receiver 101. More specifically, the drop mask enables a set of NN models 113 to be generated, each of which can be used in box 317 to generate a corresponding measurement representing a bit encoded by the signal 111 received at the node 100. In an example, the measurement may include an LLR. The measurement is stored in a temporary storage device 319 and the counter is incremented in box 321. In box 323, a check is performed to determine whether the counter so incremented is less than N. s (Number of model samples dropped / set). If yes, then the process described from block 315 is repeated. Otherwise, in block 325, a value representing the variance of the measurement is calculated. For example, the variance of the LLRs in storage device 319 (in N s Thus, in block 327, the median (δ ij ), and then compare this value with δ threshold Compared (in block 329). If the median (δ ij )>δ threshold , the signal 111 and the corresponding data bits 307 can be uploaded to the training data controller 109 via the backhaul 103. Otherwise, in block 331, the sample can be discarded. Thus, the node 100 is able to determine whether to select a signal received at the node for use as part of a training data set for the neural network-based receiver 101 based on a value representing the variance of the measurement.

[0047] Whenever the node 100 selects a particular transport block (sample) as part of the training data set (either by random sampling or due to uncertainty criteria), the corresponding RX signal 111 and the information bits 307 are accordingly moved to the upload buffer 107. From there they are transmitted to the core cloud component 105. In an example, the transmission may be of low priority, since the transmission of the training data is not delay critical.

[0048] According to an example, the available backhaul capacity can be reflected in the proportion of data selected to be uploaded for training. Figure 2 The decision-making process collects the data required for NN retraining faster, which is beneficial to node performance. Figure 1 As shown and discussed in more detail below, component 105 can play an active role in deciding when to upload data and how much data to upload. In addition, the proportion of data used for training can be adjusted based on the time since the node was deployed. For example, a newly deployed NN-based receiver will have a more urgent need for an updated model to improve performance, while a more mature NN-based receiver may be closer to optimal performance, making slower data collection sufficient.

[0049] refer to Figures 1 to 3 , it can be seen that the training process is performed using a network entity such as cloud component 105. Therefore, the known hardware limitations of node 100 are not an issue, because the cloud training entity 105 receives the training data uploaded by the node 100, where it is first fed to the training data controller 109, while maintaining two data sets D random and D uncertain The training data controller 109 obtains the separation between random and D uncertain Random sampling generates a total training data set that also includes data samples previously stored in the database 127. However, in an example, newly collected field data can be over-represented in the total data set to ensure that the NN-based receiver 101 is specific to the environment at the time. This can be done by a specific fixed time window within which data must be collected to be included in the training set. Therefore, any data older than this is considered out of date and will not be used for training.

[0050] Such a combined training data set is fed to temporary storage device 129 where it is saved until training resources become available at component 105. In examples, model training may be performed using a number of processes depending on the NN architecture (e.g., backpropagation with Adam optimization). After training is complete, it is determined (block 133) whether the NN has overfit to the training data. Overfitting may be detected by validating the trained NN using a separate validation set. The validation set may be from a dataset D that has not been used for training. randomThe validation data is randomly formed from those samples of . This ensures that the validation data represents typical scenarios encountered by the NN when deployed. If overfitting occurs, more data is requested (135) from the node 100 to expand the training data set. In the special case that the link 139 between the node 100 and the network core component 105 is congested or temporarily disconnected, more data can also be requested from the central database if applicable.

[0051] Once the NN is deemed to have generalized to the data, and it has demonstrated higher performance than the deployed NN receiver, it can be uploaded (137) to the node 100. Performance can be measured by calculating the uncoded bit error rate (BER) using a portion of the data that has not been used during training or during validation (for overfitting detection). This data can also be obtained from the dataset D random Once the performance of the newly trained NN is higher than that of the deployed receiver 101 on the same data, it is ready for deployment.

[0052] Following the example, one approach is to use a comparison metric to compare the average BER:

[0053]

[0054] Where N eval is the number of TTIs used for performance evaluation, and BER n is the uncoded BER obtained for the nth TTI.

[0055] However, in some cases, this metric may emphasize low SNR regions, favoring NNs that perform well at low SNRs, but not necessarily well at high SNRs. In the example, the geometric mean may therefore also be calculated and used as a comparison metric (by replacing zeros with values ​​less than the minimum observable BER).

[0056] In this case, the comparison metric becomes:

[0057]

[0058] where n b，max is the number of bits in a single TTI with the maximum number of subcarriers and the highest order modulation and coding scheme.

[0059] In an example, the transmission of the newly trained NN model can be performed during low activity times (e.g., at night, or during a service outage) to avoid loss of service to subscribers. Once the new model is executed at the node 100, it is also possible to start training a new model in the cloud component 105, whose performance can be compared to the deployed model. The new model can be trained in a similar manner by providing data to the new model from a central database, emphasizing the latest data from the node 100. Once it is observed that the new model candidate achieves higher performance than the deployed model, it can be uploaded to the node 100 to replace the old model.

[0060] Examples in the present disclosure may be provided as methods, systems, or machine-readable instructions, such as any combination of software, hardware, firmware, etc. Such machine-readable instructions may be included on a computer-readable storage medium (including but not limited to optical disk storage, CD-ROM, optical storage, etc.) with computer-readable program code therein or thereon.

[0061] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices and systems according to the examples of the present disclosure. Although the flowcharts described above show a specific execution order, the execution order may be different from that described. The boxes described in one flowchart may be combined with the boxes of another flowchart. In some examples, some boxes of the flowchart may not be necessary, and / or additional boxes may be added. It should be understood that each process and / or box in the flowchart and / or block diagram, and the combination of the processes and / or diagrams in the flowchart and / or block diagram can be implemented by machine-readable instructions.

[0062] Machine-readable instructions can be executed, for example, by a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to implement the functions described and described in the figure. Specifically, a processor or a processing device can execute machine-readable instructions. Therefore, a module of a device (e.g., SOC 123) can be implemented by a processor that executes machine-readable instructions stored in a memory or by a processor that operates according to instructions embedded in a logic circuit. The term "processor" should be broadly interpreted to include a CPU, a processing unit, an ASIC, a logic unit, or a set of programmable gates, etc. Both the method and the module can be executed by a single processor or distributed between multiple processors.

[0063] Such machine-readable instructions may also be stored in a computer-readable storage device that can guide a computer or other programmable data processing device to operate in a specific mode. For example, instructions may be provided on a non-transitory computer-readable storage medium encoded with processor-executable instructions.

[0064] Figure 4is a schematic diagram of a node according to an example. Node 400 includes a NN-based receiver 401. NN-based receiver 401 includes a NN model 402, e.g. Figure 1 113 and Figure 3 303. The node 400 also includes a processor 403 and a memory 405 storing instructions 407. A temporary storage device 409 may be provided for the node 400. The temporary storage device 409 may send data to, for example, the network core component 105 via a backhaul 411.

[0065] Instructions 407 may be executed by processor 403. Instructions 407 may include instructions for generating a plurality of measurements representing bits encoded by a signal received at receiver 401 using respective different neural network-based receiver frameworks, calculating variance 325 of measurements 317, and determining whether to select the signal received at receiver 401 as part of a training data set based on a comparison of the variance to a threshold. Accordingly, the node may implement a method for selecting training samples for a neural network-based receiver 401 configured for uplink communication in a selected deployment environment of a telecommunications network.

[0066] Such machine-readable instructions may also be loaded onto a computer or other programmable data processing device to cause the computer or other programmable data processing device to perform a series of operations to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable data processing device provide operations for implementing the functions specified by the (multiple) processes in the flowchart and / or the blocks in the (multiple) block diagrams.

[0067] Furthermore, the teachings herein may be implemented in the form of a computer software product that is stored in a storage medium and includes a plurality of instructions for causing a computer device to implement the methods described in the examples of the present disclosure.

Claims

1. A node for a telecommunications network, the node comprising a neural network based receiver for uplink communication, wherein the node is configured to: modifying the neural network-based receiver to generate a set of modified receiver frames, the set of modified receiver frames defining respective different versions for the receiver; generating, using each of the modified receiver frames, a respective measurement representing bits encoded by a signal received at the receiver; calculating a value representing a variance of the measurement; determining whether to select the signal received at the receiver for use as part of a training data set for the neural network-based receiver based on a comparison of the value of the variance to a threshold; in response to the value being greater than the threshold, sending the signal received at the receiver and the measurement to a core network device of the telecommunications network; as well as A neural network based receiver updated based at least on the signals and the measurements, or a request for additional signals and measurements for the training data set is received from the core network device.

2. A node according to claim 1, wherein the neural network based receiver is modified by applying a dropping mask.

3. The node of claim 1, wherein the node is configured to receive data representing the threshold value and to use the threshold value to adjust a selection of the signal received at the node for use as part of the training data set.

4. The node according to claim 1 or 2, wherein the corresponding measurement is a log-likelihood ratio (LLR).

5. The node according to claim 4, wherein the node is configured to: calculating a measure of variance of a plurality of LLR values ​​stored in a temporary storage device of the node; and The median of the variances is determined.

6. A node according to claim 1 or 2, wherein the neural network based receiver is a radio receiver.

7. A method for selecting training samples for a neural network based receiver configured for uplink communications in a selected deployment environment of a telecommunications network, the method comprising: generating a plurality of measurements representing bits encoded by a signal received at the receiver, the receiver using respective different neural network-based receiver frameworks; calculating a variance of the measurement; determining whether to select the signal received at the receiver as part of a training data set based on a comparison of the variance to a threshold; as well as In response to the variance being greater than the threshold, the signal received at the receiver and the plurality of measurements are sent to a core network device of the telecommunications network.

8. The method according to claim 7, further comprising: A random dropout mask is applied to the neural network based receiver deployed in the selected environment to generate different neural network based receiver frameworks.

9. The method according to claim 7 or 8, further comprising: The threshold value is received from a core network device of the telecommunications network.

10. The method of claim 7 or 8, wherein the plurality of measurements are LLR values.

11. A core network device for a telecommunications network, the core network device being configured to: receiving a signal and a corresponding set of data bits from a node of a telecommunications network; training a neural network-based receiver using the signal and the corresponding set of data bits, the neural network-based receiver configured for uplink communications to the node; determining a measure representing a degree of overfitting for the neural network-based receiver; In response to determining based on the measurements that the neural network-based receiver is not overfitting, providing an updated neural network-based receiver to the node; as well as A request for an additional signal and a corresponding set of data bits from the node for training the neural network based receiver is generated in response to determining that the neural network based receiver is overfitting based on the measurements.

12. The core network device according to claim 11, wherein the core network device is configured as: modifying a threshold selected for regulating the signal received at the node based on the measurement; and The modified threshold is sent to the node.

13. The core network device according to claim 11 or 12, wherein the core network device is configured as: determining a backhaul capacity associated with the node; and Based on the determined backhaul capacity, the measure representing a degree of overfitting for the neural network based receiver is determined.

Images