Signal demapping

Abstract

Aspects and embodiments provide apparatus and methods that can be used to help select one of multiple available receiver configurations to demap received signals at a receiver in a wireless communication network. The described apparatus may include trainable logic that may take the form of a neural network. The trainable logic is configured to quantize multiple receiver configurations. The trainable logic includes: determination logic configured to determine multiple available receiver configurations and associate each receiver configuration with a context matrix; codebook logic configured to select a quantization codebook to be used by the trainable logic for the context matrix; and learning logic configured to learn based on a training data set, the training data set including multiple received signal samples related to reconstruction of a transmitted message; the learning logic is configured to generate, based on the training data set: a set of overlay parameters and context matrices corresponding to the multiple available receiver configurations, and a set of context parameters for each context, the context parameters being generated by combining the overlay parameters with the context matrix associated with each receiver configuration; quantizing the context matrix according to the quantization codebook; and repeating the generation of the overlay parameters, the context matrix, and the quantization of the context matrix until a stopping criterion is met. The arrangement may recognize that by providing a context matrix for each receiver configuration and aligning those context matrices with a simple codebook, performing context (receiver configuration) switching becomes computationally more efficient.

Description

Technical Field

[0001] Various example embodiments are directed to methods and apparatus that may be used to facilitate selection of one of a plurality of available receiver configurations for demapping received signals at a receiver in a wireless communication network. Background Art

[0002] Neural networks (NNs) can be used in communication systems to improve system performance. For example, for a receiver or transceiver within a communication network, a neural network can be trained on input parameters so that for a given system configuration, the neural network selects a desired operating mode (output parameters).

[0003] Neural networks, which are configured to make decisions or identify outcomes based on a set of input examples, generally perform well. One of the reasons for the success of NNs is that they generalize well to examples that they did not see during training but were drawn from the distribution from which the training dataset was sampled.

[0004] One drawback of NNs is their lack of configurability once trained. In particular, a NN trained for a specific task often performs poorly on a different task, even if the two tasks are similar (e.g., two different computer vision tasks).

[0005] In the case of implementing a neural network for use in a communications receiver, it is desirable that the receiver work well under a variety of settings (e.g., different modulation schemes and / or different channel conditions). Ensuring that the receiver works well across such settings typically requires training and implementing multiple NNs. Training a dedicated NN for each possible system configuration (e.g., one for each modulation scheme, such as BSPK, QPSK, 16QAM, 64QAM) is straightforward, but results in a significant increase in the hardware resources required to implement these dedicated NNs. The increased hardware resources may include, for example: memory, surface area on a field programmable gate array (FPGA), etc. Such increased hardware requirements may make approaches that require multiple NNs undesirable in practice.

[0006] Bin Le ETAL, "MODULATION IDENTIFICATION USING NEURAL NETWORKS FORCOGNITIVE RADIOS," January 1, 2005 (2005-01-01) XP055689907, describes the design of a signal modulation classifier using an artificial neural network. A modulation classifier is simulated using seven modulations: AM, FM, BPSK, BFSK, OPSK, OAM8, and OAM16. Real-world signals are created and used to create the modulation classifier.

[0007] Aspects and embodiments relate to receiver algorithms with trainable parameters implemented by logic in a communication system, such as logic comprising, for example, a neural network (NN).Aspects and embodiments relate to efficient implementation of such receiver algorithms in practical systems. Summary of the Invention

[0008] The scope of protection sought by various embodiments of the present invention is defined by the independent claims. Embodiments, examples, arrangements and features described in this specification that do not fall within the scope of the independent claims (if any) are to be construed as examples useful for understanding various embodiments of the present invention.

[0009] According to various, but not necessarily all, embodiments of the present invention, there is provided an apparatus according to claim 1 .

[0010] The apparatus may be such that the stopping criteria comprises one or more of: a predefined number of iterations.

[0011] The apparatus may relate the stopping criterion to a performance metric associated with the reconstruction of the transmitted message. Such a performance metric may comprise, for example, a selected symbol error rate.

[0012] The apparatus may relate the stopping criterion to a performance metric associated with quantizing the context matrix according to the quantization codebook.

[0013] The apparatus may be such that selecting the quantization codebook comprises setting coefficients of the context matrix to match values of the finite matrix codebook.

[0014] The device may enable the finite matrix codebook to include one of the following items: {-1, 0, 1}; a value of a power of 2.

[0015] The apparatus may be such that combining the overlay parameter with the context matrix comprises one of: a matrix multiplication or an element-wise product.

[0016] The apparatus may be such that the receiver configuration comprises one or more of: a receiver or a demodulation configuration.

[0017] According to various, but not necessarily all, embodiments of the invention, there is provided an apparatus according to claim 8.

[0018] The apparatus may be such that it further comprises: a receiver configuration selection logic configured to select a receiver configuration from the plurality of receiver configurations to apply to the received wireless communication system based on one or more factors, the one or more factors comprising: the generated overlay parameter and the quantized context matrix.

[0019] According to various, but not necessarily all, embodiments of the invention, there is provided a method according to claim 10.

[0020] According to various, but not necessarily all, embodiments of the present invention, there is provided a computer program product configured to perform the above method when executed on a computer.

[0021] Further particular and preferred aspects are set out in the accompanying independent and dependent claims. Features of the dependent claims may be combined with features of the independent claims as appropriate and may be combined with features other than those explicitly set out in the claims.

[0022] Where an apparatus feature is described as being operable to provide a function, it will be understood that this includes apparatus features that provide that function or that are adapted or configured to provide that function. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Some example implementation examples will now be described with reference to the accompanying drawings, in which:

[0024] Figure 1a illustrates a general receiver architecture according to the arrangement;

[0025] Figure 1b A typical outline of the neural network used is schematically illustrated;

[0026] Figure 2 schematically illustrates receiver operations using parameters specific to the selected context;

[0027] Figure 3 including visual representations of inference flow charts; and

[0028] Figure 4 Includes a flow chart of the training process according to the arrangement. DETAILED DESCRIPTION

[0029] Before discussing example embodiments in greater detail, an overview will first be provided.

[0030] Modulation and demodulation of wireless signals are fundamental building blocks of any wireless communication system. Typically, a receiver / demodulator / demapper within a communication system is configured to operate to receive and demap signals by taking into account factors such as Gaussian noise, channel state information, channel noise distribution, coding methods, and similar signal input parameters. It will be understood that conditions associated with ideal transmission and reception do not typically occur in the physical deployment of a wireless communication system. Wireless communication network deployments are often subject to unexpected noise, multipath and fast fading, and other complex interference. As a result, for example, it may be difficult to accurately know or estimate channel state information, etc. In addition, the receiver may not know the coding method used to transmit the signal, and therefore providing reliable and efficient methods and apparatus for providing signal demapping / demodulation is challenging.

[0031] NN and machine learning make it possible to provide methods and apparatus with improved demapping and demodulation performance. In particular, within a given encoding method and associated demodulation scheme, example received signals associated with known transmitted signals can be provided as a training data set to train the NN to perform efficient and effective demodulation within that scheme.

[0032] It will be appreciated that a receiver may be capable of receiving and demodulating signals according to a plurality of different available coding and channel schemes. Furthermore, the receiver may include the capability to provide different receiver configurations and may, for example, include a plurality of different receiver hardware elements or configurations. In each case, a separate NN may be provided to improve demodulation / demapping performance. In such an arrangement, it may become challenging to provide a mechanism for the receiving device to select or choose the most appropriate NN to apply to the received signal.

[0033] The described arrangement aims to provide the ability to embed different communication receivers or receiver configurations (demapping mechanisms) (implemented as a neural network) within a set of parameters. B. Cheung, A. Terekhov, Y. Chen, P. Agrawal, and B. Olshausen. "Superposition of many models into one". arXiv preprint arXiv:1902.05522, 2019, has proposed superimposing models within a set of weights. The described arrangement can provide quantization of the parameters used to change the NN configuration. Such an approach can reduce any memory requirements associated with providing a system that can provide a superposition of NN models and make switching between NNs trained for multiple results practical. Without quantization, the memory required to support the superposition of models can be as large as the memory consumed by each of the multiple independent models, calling into question the entire concept of superimposing multiple NNs. Based on the quantization method, a single NN can be effectively configured to implement multiple receivers or different receiver configurations, which are essentially NNs themselves, and those receivers or receiver configurations have been trained to share the same set of weights. In other words, the receivers (or receiver configurations) "learn" to use the same set of weights. Such an approach allows for superimposed NN results to be achieved within achievable overall system hardware requirements.

[0034] In general, an arrangement can allow a NN to represent multiple receivers (or receiver configurations). This can be achieved by providing the NN with receivers having additional inputs. In the described arrangement, the additional inputs can be referred to as "contexts." The contexts define a context matrix that is used to implement the superposition of multiple models in the NN. By selecting an appropriate context matrix, the corresponding model can be selected during training and inference.

[0035] The arrangement recognizes that it is necessary to store as many context matrices as there are receiver configurations. By quantizing the context matrices, the memory requirements associated with storing those context matrices can be minimized and the computational complexity associated with switching receiver configurations can be reduced.

[0036] General Architecture

[0037] Figure 1a The diagram illustrates an arrangement of a receiver apparatus according to one arrangement. A data processor 10 is provided coupled to receiver hardware 20. Signal data 40 is provided to the data processor 10 in real time or from a memory 50. The signal data may be provided from a variety of different sources in any of a number of different formats. The signal data may include any of a variety of different types of signal data.

[0038] Machine Learning Algorithm Architecture

[0039] Figure 1b A typical outline of a neural network for optimizing the operation of one receiver configuration is schematically illustrated. According to this general arrangement, a neural network (NN) or model executed on a data processor 10 may select settings for a given receiver configuration based on received signal data. Figure 1b A typical outline of a neural network used to select settings for a given receiver configuration is illustrated. Figure 1b The network schematically illustrated in FIG has characteristics and parameters related to the received signal enter on the left, and the model created by the neural network has subsequent model layers generate increasingly abstract features as the parameters progress through the neural network.

[0040] exist Figure 1b In the example neural network 100 shown in FIG, a training data set 110 may be provided, in this example, a received signal associated with a known transmitted signal. An input layer 120 of the neural network may operate to identify features 130 associated with the training data set. As the NN moves toward the output layer 180, "hidden" layers 140, 160 operate to identify groups of "features" 150, 170, respectively, of the training data set, or features of the previous layer. The output layer 180 of the neural network may operate to identify specific optimal settings for a given receiver configuration, depending on the results of the recognition performed since the neural network has been trained.

[0041] It will be appreciated that it may be beneficial to be able to input a signal and use features or parameters that can be determined from the signal to select between various available receiver configurations, each of which may be associated with a separate NN to select specific optimal settings for a given receiver configuration. It will be further appreciated that some arrangements may provide a NN to select between the NNs associated with each receiver configuration. However, such an arrangement is computationally significant and may not represent the best use of available hardware.

[0042] Some arrangements may recognize that it may be beneficial to quantize the context matrices mapped to different receiver configuration NNs to achieve optimal system performance. The superposition of receiver configuration NNs may be performed as a balance or tradeoff related to the level of quantization performed to select between receiver configuration NNs. That is, in order to fully quantize the selection between receiver configuration NNs (context matrices), the performance of the receiver configuration NN may be compromised, or in order to achieve better receiver configuration NN performance, the quantization of different NNs associated with the receiver configuration may be compromised. The arrangement provides a mechanism to fully utilize available computing resources and provide more than one NN associated with the receiver configuration within a given hardware solution.

[0043] The arrangement may recognize that by providing a context matrix for each receiver configuration (as described in more detail below) and aligning these context matrices with a simple codebook, performing context (receiver configuration) switching becomes computationally more efficient.

[0044] The arrangement can be used to help decide which receiver configuration to use. The arrangement provides a mechanism that enables efficient construction, storage, and retrieval of multiple receiver configurations (quantized context matrices, overlay parameters, training algorithms). The logic that operates to decide which context (receiver configuration) to use at the receiver can be designed and run in conjunction with the various described arrangements and methods.

[0045] It will be appreciated that the apparatus and training methods according to some arrangements can be run on powerful hardware (GPUs, etc.) before being deployed in wireless communication networks. Once training is complete, the appropriately trained neural network architecture can be implemented on an actual receiver along with the determined overlay parameters and quantized context matrix, and an algorithm configured to help decide which receiver configuration (context) to use at any given time based on some input (e.g., channel state).

[0046] When used, the arrangement can be implemented such that a receiver configuration is selected in some manner based on the input signal. Once selected, the corresponding context matrix is retrieved, the superposition parameters are combined with the selected context matrix to generate context parameters, and finally the implemented NN parameters are updated with the context parameters thus generated.

[0047] Example training of a stacked receiver configuration NN

[0048] Throughout this description, the following notation will be used:

[0049] For vector x,x i represents its i-th element, that is, x=[x0,...,x n-1 ].

[0050] In a communication system where a receiver is provided with NN capabilities, arrangements may be implemented to improve receiver operation.

[0051] One general arrangement involves a receiver with K possible receiver configurations. Let K = {1...K} denote the set of possible receiver configurations.

[0052] The receiver is configured such that it aims to reconstruct a message s selected from a finite set M_k=\{1...M_k\}(k∈K) from a vector of received samples y∈C^n (where n is the number of channels used).

[0053] Assume that the receiver is implemented with a trainable parameter vector θ k (which depends on the context k) The receiver is usually referred to first as NN.

[0054] Vector of samples that the receiver will receive Mapped to dimension M k The probability vector p of pi can be interpreted as the message The probability of being sent.

[0055] The context overlay operation enables configuration of Define the context matrix C k A combination operator ⊙ is also defined, which can be implemented in various ways. For example, the combination operator can be a matrix or element-wise product. If element-wise product is chosen, the context matrix C k is related to θ k Vectors of the same dimension.

[0056] According to some arrangements, a set of parameters is stored, which is represented as and can be interpreted as a superposition of parameters corresponding to K receiver configurations. When , the corresponding context matrix C is combined as follows k and To retrieve the corresponding parameter θ k :

[0057]

[0058] Figure 2 The receiver operation using parameters specific to the selected context is schematically illustrated.

[0059] Figure 3 Includes a visual representation of the inference flow graph. Figure 3 The following algorithm steps are illustrated:

[0060] S1: Retrieval context parameters θ k

[0061] S2: Using parameter θ k Infer a probability vector p from the received signal y.

[0062] Quantization of the context matrix

[0063] Since the arrangement requires the storage of K context matrices (one for each receiver configuration), it will be appreciated that the amount of memory required to store the context matrices will quickly prohibit the practical use of more than a few contexts. The arrangement will recognize that the context matrix C k Quantization of can provide a reduction in required memory. According to some arrangements, the elements of the context matrix are forced to approach or match the finite codebook The values in the finite codebook may include, for example, {-1, 0, 1} or a set of powers of 2. By enforcing such a restriction, the amount of memory required to store the context matrix can be greatly reduced. In addition, the use of a quantized codebook reduces the computational power required to retrieve the configuration. For example, if the codebook is selected Then computing the element-wise product reduces to a sign change.

[0064] Superposition of neural networks

[0065] Assuming the combinatorial operator ⊙ is differentiable (which is the case if one chooses matrix or element-wise product), then and the context matrix {C k} k＝1...K can be jointly trained. The training complexity may come from quantizing the context matrix into a predefined codebook The value in the requirement.

[0066] One arrangement is to have the training algorithm follow the method outlined below:

[0067] T1: Initialize (e.g., randomly) the context matrix C k , k=1...K.

[0068] T2: Calculate the quantized context matrix in It's a codebook The projection operator on .

[0069] T3: Initialize the Lagrange coefficient λ k (k=1...K) and penalty coefficient μ

[0070] T4: Initialization parameters

[0071] T5: Update by solving the following formula and C1...C K

[0072]

[0073] where L is a loss function that measures the performance of the receiver, such as categorical cross entropy.

[0074] T6: Calculation

[0075] T7: Update the Lagrange coefficient to:

[0076] T8: Increase μ according to a predefined schedule.

[0077] T9: If If it is small enough, stop or repeat from step 4.

[0078] Figure 4 A flowchart of a training process according to an arrangement such as that outlined above is included. Step 5 cannot usually be solved exactly, but can be approximately solved using stochastic gradient descent or a variant thereof. Steps 2 and 6 can be solved simply by selecting the closest value in a predefined codebook for each element of the context matrix.

[0079] Figure 4 Illustrated is a method by which the system jointly and simultaneously trains a receiver demapper and quantizes a context matrix. Figure 4 The method of the arrangement in [ 1 ] operates such that training starts by finding the best possible demapping performance associated with each receiver configuration (with a low value of μ). However, the best possible demapping performance may not provide a good quantization / match with the selected codebook, so a step-wise training is performed with increasing values of μ. As the value of μ increases, the training becomes more focused on the quantization of the context matrix rather than the reconstruction of the transmitted message. Given a selected codebook, according to Figure 4The method of arrangement described in seeks to find a compromise between demapping performance and quantization. It will be appreciated that, given a set of input parameters, the choice of codebook (e.g., the choice of codebook size) can be a factor in determining how to quantize or "stack" the various context matrices mapped to a particular receiver configuration.

[0080] Those skilled in the art will readily recognize that the step of various said methods can be performed by a programmed computer. In this article, some embodiments are also intended to contain program storage devices (for example, digital data storage media), which are machine or computer readable and machine executable programs or computer executable programs of instructions are encoded, wherein instructions are performed some steps or all steps in the step of said method. Program storage devices can be for example digital memories, magnetic storage media such as disks and tapes, hard drives or optically readable digital data storage media. Embodiments are also intended to contain computers programmed to perform the step of said method.

[0081] Although embodiments of the present invention have been described in the preceding paragraphs with reference to various examples, it should be appreciated that modifications to the examples given can be made without departing from the scope of the invention as claimed.

[0082] Features described in the preceding description may be used in combinations other than the combinations explicitly described.

[0083] Although functions have been described with reference to certain features, those functions may be performable by other features whether described or not.

[0084] Although features have been described with reference to certain embodiments, those features may also be present in other embodiments whether described or not.

[0085] While in the foregoing description an effort has been made to draw attention to those features of the present invention which are regarded as particularly important, it will be understood that the applicants claim protection for any patentable feature or combination of features mentioned above and / or shown in the drawings whether or not specifically emphasized therein.

Claims

1. A device comprising: Trainable logic configured to quantize a plurality of receiver configurations, the trainable logic comprising: determination logic configured to determine a plurality of available receiver configurations and associate each receiver configuration with a context matrix; codebook logic configured to select a quantization codebook to be used by the trainable logic for the context matrix; learning logic configured to learn from a training data set comprising a plurality of received signal samples associated with a reconstruction of a transmitted message; The learning logic is configured to: generating, based on the training data set: a set of overlay parameters and context matrices corresponding to the plurality of available receiver configurations, and a set of context parameters for each context, the context parameters being generated by combining the overlay parameters with the context matrix associated with each receiver configuration; quantizing the context matrix according to the quantization codebook; repeating said generating of superimposition parameters, a context matrix, and quantizing of the context matrix until a stopping criterion related to a performance metric associated with said reconstruction of said transmitted message is satisfied; and Selection logic is configured to select a receiver configuration based on the generated superposition parameters, the context matrix, and the quantization of the context matrix satisfying the stopping criterion, and an input wireless communication signal.

2. The apparatus according to claim 1, wherein the stopping criterion further comprises: A predefined number of iterations. 3 . The apparatus according to claim 1 , wherein the stopping criterion is further related to a performance metric associated with quantizing the context matrix according to the quantization codebook.

4. The apparatus of claim 1 , wherein selecting the quantization codebook comprises: The coefficients of the context matrix are set to match the values of the finite matrix codebook. The apparatus according to claim 4 , wherein the finite matrix codebook comprises one of the following: {−1, 0, 1}; a value of a power of 2. The apparatus of claim 1 , wherein combining the overlay parameters with the context matrix comprises one of: a matrix multiplication or an element-wise product. The apparatus of claim 1 , wherein the receiver configuration comprises a demodulation configuration.

8. A device comprising: receiving logic configured to receive an incoming wireless communication signal; Processing logic is configured to process the received input wireless communication signal and demap the received signal, the processing logic comprising: a receiver control module, the receiver control module comprising the apparatus according to any one of claims 1 to 7.

9. The apparatus according to claim 8, further comprising: Receiver configuration selection logic is configured to select a receiver configuration from the plurality of receiver configurations to apply to the received wireless communication signal based on one or more factors, the one or more factors including the generated superposition parameter and the quantized context matrix.

10. A computer-implemented method comprising: Multiple receiver configurations are quantized as follows: determining a plurality of available receiver configurations, and associating each receiver configuration with a context matrix; Selecting a quantization codebook to be used for the context matrix; providing a training data set comprising a plurality of received signal samples associated with a reconstruction of a transmitted message; generating, based on the training data set: a set of overlay parameters and context matrices corresponding to the plurality of available receiver configurations, and a set of context parameters for each context, the context parameters being generated by combining the overlay parameters with the context matrix associated with each receiver; quantizing the context matrix according to the quantization codebook; repeating said generating of superimposing parameters, a context matrix, and quantizing of the context matrix until a stopping criterion related to a performance metric associated with said reconstruction of said transmitted message is satisfied; as well as A receiver configuration is selected based on the generated superposition parameters, the context matrix, and the quantization of the context matrix satisfying the stopping criterion and an input wireless communication signal.

11. A computer program product, configured to perform the method according to claim 10 when executed on a computing device.

Images