Machine learning for wireless channel estimation

Abstract

Certain aspects of the present disclosure provide techniques and apparatus for wireless channel estimation using machine learning. In a first iteration of a machine learning model, a current sparsification dictionary is generated by processing a sensing matrix for a digital communication channel and a current channel observation using a posterior neural network, and generating a current sparse channel representation in the first iteration by processing the current sparsification dictionary, the sensing matrix, and the current channel observation using a likelihood neural network.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Greek Patent Application No. 20220100416 filed on May 19, 2022, the entire contents of which are incorporated herein by reference.

[0003] introduction

[0004] Aspects of the present disclosure relate to machine learning for wireless channel estimation.

[0005] Wireless communication channels may have a variety of dynamic characteristics that affect how signals propagate from a transmitter to a receiver. Such characteristics are often collectively referred to as channel state information (CSI), and determining or estimating such characteristics may be referred to as channel estimation. Accurate channel estimation allows wireless transmission to be adapted based on channel conditions, which can significantly improve communication reliability and throughput, particularly when multiple antennas are used, such as in multiple-input multiple-output (MIMO) systems. However, channel estimation is an underdetermined problem (e.g., a problem in which there are enough unknown quantities that the values ​​of all problem parameters cannot be determined conclusively) in many deployments (such as millimeter wave (mmWave) massive MIMO systems using analog beamforming). Therefore, channel estimation is often inaccurate or computationally expensive to perform or simply unavailable. Summary of the invention

[0006] Certain aspects provide a processor-implemented method for performing channel estimation of a digital communication channel using a machine learning model, comprising: generating a current sparsification dictionary in a first iteration of the machine learning model by processing a sensing matrix and current channel observations for the digital communication channel using a posteriori neural network; and generating a current sparse channel representation in the first iteration by processing the current sparsification dictionary, the sensing matrix, and the current channel observations using a likelihood neural network.

[0007] Certain aspects provide a processor-implemented method for training a machine learning model for channel estimation of a digital communication channel, comprising: receiving a sensing matrix and a current channel observation for the digital communication channel; generating a current sparsification dictionary by processing the sensing matrix and the current channel observation using a posteriori neural network; generating a current sparse channel representation by processing the current sparsification dictionary, the sensing matrix, and the current channel observation using a likelihood neural network; generating a first loss by processing the current sparsification dictionary using an a priori neural network; generating a second loss based on the current channel estimate; and refining the a posteriori neural network, the likelihood neural network, and the a priori neural network based on the first loss and the second loss.

[0008] Other aspects provide: a processing system configured to perform the aforementioned methods and those described herein; a non-transitory computer-readable medium comprising instructions that, when executed by one or more processors of the processing system, cause the processing system to perform the aforementioned methods and those described herein; a computer program product embodied on a computer-readable storage medium comprising code for performing the aforementioned methods and those further described herein; and a processing system comprising components for performing the aforementioned methods and those further described herein.

[0009] The following description and associated drawings set forth in detail certain illustrative features of the one or more aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings depict certain aspects of the disclosure and therefore are not to be considered limiting of the scope of the disclosure.

[0011] Figure 1 An example environment for using machine learning for improved channel estimation is depicted.

[0012] Figure 2 Depicted is an example architecture for a machine learning model to perform improved channel estimation.

[0013] Figure 3 Depicted is an example workflow for training a machine learning model to perform channel estimation.

[0014] Figure 4 Depicted is an example workflow for performing channel estimation using a trained machine learning model.

[0015] Figure 5 is a flow chart depicting an example method for training a machine learning model for channel estimation.

[0016] Figure 6 is a flow chart depicting an example method for performing channel estimation using a machine learning model.

[0017] Figure 7 is a flow chart depicting an example method for generating channel estimates.

[0018] Figure 8 is a flow chart depicting an example method for training a machine learning model for channel estimation.

[0019] Fig. 9 An example processing system configured to perform various aspects of the present disclosure is depicted.

[0020] To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are common to the figures. It is contemplated that elements and features of one aspect may be beneficially incorporated in other aspects without further recitation. DETAILED DESCRIPTION

[0021] Aspects of the present disclosure provide apparatus, methods, processing systems, and non-transitory computer-readable media for machine learning based channel estimation.

[0022] In various wireless communication systems, and in particular in MIMO wireless systems, channel estimation can be a significant technical problem. Although some examples described in this disclosure relate to using machine learning to improve channel estimation in mmWave (e.g., communications in frequency bands above 24 GHz, where wavelengths are on the order of millimeters) systems, beamforming systems, and / or MIMO systems, aspects of this disclosure can be readily applied to any estimation problem with underdetermined measurements. For example, aspects of this disclosure can be used to improve magnetic resonance imaging (MRI) and other imaging problems.

[0023] Various compressed sensing techniques have been used in the context of mmWave and massive MIMO channel estimation, but the channel estimation problem remains an underdetermined problem. For example, in systems using analog beamforming, channel observations are typically lower-dimensional representations of relatively high-dimensional channels. In order to provide improved estimates, some compressed sensing techniques attempt to exploit the sparsity of the channel in the angular domain (rather than the antenna concentration domain). That is, the channel can be relatively dense in the spatial domain (e.g., where most elements have similar values ​​and have little or no structure). However, in the angular domain, the channel is typically sparse (where most elements have low or null values) and structured (where a few elements have significant values). For example, the outer product of the antenna responses on the transmitting and receiving sides can be used as atoms or elements of a sparse dictionary that can be used to provide compressed sensing and channel estimation. In some aspects, the channel can be represented as the product between a sparse dictionary Ψ and a sparse vector representation of the signal, where the sparse dictionary corresponds to the angular domain of the transmitter and receiver. In some aspects of the present disclosure, a sparsified dictionary may be generated using a machine learning model based on observed channel information and a corresponding sensing matrix Φ (which may correspond to a beamforming matrix or codebook used to send / receive signals, as discussed in more detail below).

[0024] Typically, conventional methods for solving this compressive sensing problem involve many complexities or trade-offs, such as iterative methods like orthogonal matching pursuit (OMP) techniques, iterative hard threshold algorithm (IHTA)-based techniques, iterative soft threshold algorithm (ISTA) techniques, and approximate message passing (AMP) techniques. The hyperparameters of such algorithms (e.g., number of iterations and thresholds) can be manually tuned to try to optimize the performance / complexity trade-off, but at least some conventional methods are still significantly limited in their flexibility and accuracy.

[0025] Some machine learning methods have been used to try to learn sparse recovery algorithms from sample data. For example, methods like Learning ISTA (LISTA) and variants have focused on using fixed sensing matrices (also referred to as measurement matrices) and sparsification dictionaries to solve this problem. That is, such conventional methods use static sparsification dictionaries and static sensing matrices, while real deployments are typically significantly more dynamic. Using such fixed methods results in inherently limited estimates, thereby preventing accurate and reliable channel estimation in a variety of real deployments.

[0026] Conventional approaches to providing MIMO channel estimation typically suffer from various problems, including lack of a ground truth dictionary for channel recovery, inability to adapt the sparsified dictionary to different sensing matrices (also referred to in some aspects as measurement matrices), having a fixed number of iterations / layers (typically incurring increased computational overhead compared to more flexible approaches), inability to understand whether the dictionary remains valid or is out of distribution, inability to determine when retraining should be performed, etc.

[0027] In various aspects of the present disclosure, a technology for considering a channel (e.g., utilizing a planar dual-polarized antenna array in three-dimensional space) with its complete universality is provided. However, the dimensionality of such problems expands rapidly in these deployments, which makes the sparse recovery problem significantly more complicated. In some aspects, in order to reduce or handle such complexity, a variational-LISTA (V-LISTA) method is introduced. In some aspects of the present disclosure, an enhanced dictionary is used to learn ISTA (A-DLISTA) architecture. In some aspects, the A-DLISTA architecture can be implemented as a single block iterated multiple times, and / or an expanded version of ISTA (e.g., using one or more sequence blocks, each sequence block having one or more neural networks to perform processing operations). Various aspects described herein use Bayesian formulas for systems to define the distribution on dictionaries that can be updated at each iteration. To this end, in some aspects, the system generates and uses a priori distribution on a dictionary, a likelihood function for observed data, and a variational posteriori of a true posteriori on an approximate dictionary. In some aspects, a likelihood function is used to generate a distribution based on learned parameters and a sparsifying dictionary generated. This distribution may then be used as a channel estimate and / or a sparse channel representation (which may be used to generate a channel estimate, as discussed in more detail below).

[0028] In contrast to at least some conventional techniques, aspects of the present disclosure may enable sparsified dictionaries to be adaptive to specific input sensing matrices, have variable numbers of iterations (allowing for dynamic stopping during inference when certain criteria are met), recognize out-of-distribution dictionaries, understand when retraining is useful or advantageous, etc. Each of these may significantly improve the accuracy and performance of the model and better represent the real system, as well as significantly reduce the computational expense of the model.

[0029] Example environment for using machine learning for channel estimation

[0030] Figure 1 An example environment 100 for using machine learning for improved channel estimation is depicted. In the illustrated example, environment 100 is a MIMO system that uses analog beamforming to provide wireless communications. In at least one aspect, environment 100 corresponds to a mmWave MIMO system.

[0031] As illustrated, environment 100 includes a transmitter 105 and a receiver 115 communicating via a channel 110. In the depicted example, transmitter 105 uses a set of transmit antennas 125A-125C, while receiver 115 uses a set of receive antennas 130A-130C. Although three antennas are depicted on each system, transmitter 105 and receiver 115 may generally use any number of antennas. Additionally, although the illustrated example depicts a single transmitter 105 and receiver 115, in some aspects, each device may operate as both a transmitter and a receiver (e.g., a transceiver). In addition, although not included in the illustrated example, in some aspects, some or all of the antennas of transmitter 105 and / or receiver 115 may be arranged into one or more discrete sub-arrays. In at least one aspect, transmitter 105 corresponds to a wireless base station, such as a 5G radio base station (e.g., gNodeB), and receiver 115 corresponds to user equipment.

[0032] In some aspects, transmitter 105 (acting as a base station or BS) can wirelessly communicate (e.g., send signals to or receive signals from user equipment (UE)) (such as receiver 115) via a communication link (e.g., via channel 110). The communication link between the BS and the UE may include uplink (UL) (also known as reverse link) transmissions from the UE to the BS and / or downlink (DL) (also known as forward link) transmissions from the BS to the UE. In various aspects, the communication link may use multiple-input multiple-output (MIMO) antenna technology, including spatial multiplexing, beamforming, and / or transmit diversity.

[0033] A BS (such as transmitter 105) may generally include: a NodeB, an enhanced NodeB (eNB), a next generation enhanced NodeB (ng-eNB), a next generation NodeB (gNB or gNodeB), an access point, a transceiver base station, a radio base station, a radio transceiver, a transceiver function, a transmission reception point, and / or the like. Each BS may provide communication coverage for a corresponding geographic coverage area, which may sometimes be referred to as a cell, and may overlap in some cases (e.g., a small cell may have a coverage area that overlaps with the coverage area of ​​a macro cell). For example, a BS may provide communication coverage for a macro cell (covering a relatively large geographic area), a pico cell (covering a relatively small geographic area, such as a stadium), a femto cell (a relatively small geographic area (e.g., a home)), and / or other types of cells.

[0034] In some aspects, the receiver 115 is part of the UE. In various aspects, the UE may more generally include: a cellular phone, a smart phone, a session initiation protocol (SIP) phone, a laptop, a personal digital assistant (PDA), a satellite radio, a global positioning system, a multimedia device, a video device, a digital audio player, a camera, a game console, a tablet computer, a smart device, a wearable device, a vehicle, an electric meter, a gas pump, a large or small kitchen appliance, a health care device, an implant, a sensor / actuator, a display, an Internet of Things (IoT) device, an always-on (AON) device, an edge processing device, or other similar devices. The UE may also be more generally referred to as a mobile device, a wireless device, a wireless communication device, a station, a mobile station, a subscriber station, a mobile subscriber station, a mobile unit, a subscriber unit, a wireless unit, a remote unit, a remote device, an access terminal, a mobile terminal, a wireless terminal, a remote terminal, a mobile phone, and others.

[0035] In the depicted environment 100, the transmitter 105 and the receiver 115 each use a set of phase shifters 120 (e.g., 120A-120C) and 135 (e.g., 135A-135C) to provide analog beamforming according to beamforming codebooks 140A and 140B, respectively. In particular, each transmit antenna 125 of the transmitter 105 has a corresponding phase shifter 120, and each receive antenna 130 of the receiver 115 has a corresponding phase shifter 135. As discussed above, the use of multiple antennas on each side of the transmit (e.g., the fact that the environment 100 depicts a MIMO system) and the use of analog beamforming (using phase shifters 120 and 135) significantly increases the complexity of performing channel estimation.

[0036] The channel 110 generally represents the wireless medium through which the signal propagates from the transmitter 105 to the receiver 115. The characteristics of the channel 110 can significantly affect the manner in which the signal propagates. Therefore, channel estimation using machine learning can be used to determine, predict, or infer the state of the channel 110. In some aspects, the signal observed by the receiver 115 can be formulated as Y=AHB, where Y is the observed signal (also referred to as a channel observation in some aspects), A and B are the analog beamforming codebooks 140A and 140B (also referred to as beamforming matrices) used by the transmitter 105 and the receiver 115, respectively (which can be proprietary and / or change over time), and H represents the channel 110. In one such formulation, the goal of the channel estimation is to estimate or predict H based on the channel observation Y.

[0037] In the illustrated example, the estimation component 145 is communicatively coupled with the receiver 115. Although depicted as a discrete component for conceptual clarity, in some aspects, the estimation component 145 is integrated into the receiver 115. In some aspects, for example, the estimation component can use machine learning and Bayesian statistics to provide a channel estimate for the channel 110, as discussed in more detail below. Additionally, although depicted as being coupled with the receiver 115, in some aspects, the estimation component 145 can be integrated with or coupled to other components, such as the transmitter 105.

[0038] In some conventional systems, channel estimates are determined based on various data including element response matrices of receiver 115 and transmitter 105 calculated at two-dimensional angular grid points, where the response matrices can be used to define a sparsification dictionary for reconstructing the channel. That is, conventional methods typically rely on knowledge of the response matrix (e.g., the antenna element responses in any given direction) to define a fixed and static sparsification dictionary. In contrast, aspects of the present disclosure enable the use of machine learning to generate a sparsification dictionary, as discussed in more detail below. Additionally, as discussed above, conventional systems typically rely on the use of a fixed measurement or sensing matrix, while aspects of the present disclosure enable the sensing matrix to be changed.

[0039] In some aspects, the use of a varying sensing matrix may be particularly useful in the context of MIMO systems and beamforming systems. Since aspects of the present disclosure do not require the recovery of prior knowledge of a set of sparse basis vectors about a signal, and aspects of the present disclosure do not require a fixed sparse dictionary, the techniques described herein can significantly improve model performance and accuracy. In some aspects, the channel 110 generates a prior model, a posterior model, and / or a likelihood model by using a trained neural network, rather than estimating using a calculated dictionary. For example, the estimation component 145 may use a neural network to define the parameters of a prior distribution on a sparse dictionary generated at each iteration of the model, and similarly use a neural network to define the parameters of a variational posterior distribution on the dictionary. Similarly, a neural network (e.g., an A-DLISTA model) may be used as a likelihood model to generate a channel estimate (e.g., a distribution with parameters generated using A-DLISTA) based on the generated sparse dictionary at each iteration.

[0040] In some aspects, rather than operating directly on channel estimates, the disclosed models may be used to receive and / or generate sparse channel representations, as discussed in more detail below. In some aspects, the sparse channel representations and the channel estimates themselves are related or associated such that matrix multiplication may be used to generate the channel estimates based on the sparse representations. For example, in some such aspects, the channel estimates It can be represented by the sampled sparsification dictionary Ψ and the sparse channel Matrix multiplication is performed between them to generate , as discussed in more detail below.

[0041] Once the model is trained (as discussed below), the model can be used to generate sparse channel representations, which can then be used to generate channel estimates, which can drive various additional operations to improve the functionality of the network / environment 100. For example, by estimating the channel 110, the estimation component 145 can enable improved analog beamforming design, faster beam selection, improved spectral efficiency prediction, sensing or positioning (e.g., determining the location of an object in space), link adaptation, communication parameter selection, etc. For example, the estimated channel can be used to design a beamforming codebook used by a transmitter and / or receiver to achieve better end-to-end performance.

[0042] Although channel estimation is used as one example use case for the disclosed architecture and techniques, aspects of the disclosure are readily applicable to a wide variety of compressed sensing problems.

[0043] Example architecture for a machine learning model for channel estimation

[0044] Figure 2An example architecture 200 for performing improved channel estimation for a machine learning model is depicted. In some aspects, the architecture 200 is used by an inference system to perform channel estimation at runtime. For example, the architecture 200 can be used by an estimation component, such as Figure 1 The estimation component 145 is used.

[0045] As discussed in more detail below, architecture 200 may generally use a supervised training approach to refine the parameters of the model based on training examples, where each example includes a channel observation y (represented as observation 210 in the illustrated example (e.g., a particular observed data sample Y of the received signal)), a corresponding sensing matrix (Φ) 205, and a ground truth or known channel state h. Typically, channel observation 210 corresponds to a signal received at a receiver (e.g., Figure 1 The sensing matrix 205 is a function of the beamforming matrix (which is typically a proprietary design and is generally known to the receiver device and the transmitter device) used for measurement or signal observation. In some aspects, the sensing matrix 205 is the Kronecker product of the two beamforming codebooks. For example, in at least one aspect, the sensing matrix 205 is defined as Among them B T is the transpose of matrix B (which may correspond to the beamforming codebook of the receiver used to observe observation 210, as discussed above), and A is the beamforming codebook of the transmitter used to generate the transmission.

[0046] As discussed above, in contrast to at least some conventional approaches, architecture 200 enables sensing matrix 205 to be different or varied. That is, each observation 210 can have a corresponding sensing matrix 205, rather than using a fixed sensing matrix for all samples. This can improve the flexibility and accuracy of the model compared to at least some conventional approaches. Similarly, architecture 200 enables the sparsification dictionary to adapt to different sensing matrices, thereby improving model accuracy.

[0047] The architecture 200 generally depicts a first portion 202 corresponding to the posterior model and the likelihood model (as discussed in more detail below) and a second portion 248 corresponding to the prior model (as discussed in more detail below).

[0048] In the illustrated example, the first portion 202 includes a series of blocks 215A-215B (collectively referred to as blocks 215) (wherein the ellipsis before and after the block 215 indicates that any number of blocks can be used, including a single block / iteration). Each block 215 generally corresponds to the application of a posterior model and a likelihood model, as discussed in more detail below. Although depicted as a discrete sequence of blocks 215 for conceptual clarity, in some aspects, the first portion 202 can be implemented as a single block, where data is iteratively passed through the block. That is, the output of one iteration (e.g., block 215A) can be used as an input to the same block for the next iteration. In various aspects of the present disclosure, the block 215 may alternatively be referred to as a layer (e.g., of a machine learning model or neural network), an iteration, etc.

[0049] Similarly, in the illustrated example, the second portion 248 includes a series of blocks 255A-255B (255) (wherein the ellipsis before and after the block 255 indicates that any number of blocks can be used, including a single block / iteration). In some aspects, there may be one or more logic blocks or components (not depicted) to control the existence of the loop / sequence. Each block 255 generally corresponds to the application of the prior model, as discussed in more detail below. Although depicted as a discrete sequence of blocks 255 for conceptual clarity, in some aspects, the second portion 248 may be implemented as a single block, where data is iteratively passed through the block. That is, the output of one iteration (e.g., block 255A) can be used as an input to the same block for the next iteration. In various aspects of the present disclosure, the block 255 may alternatively be referred to as a layer (e.g., of a machine learning model or neural network), an iteration, etc. Additionally, although depicted as discrete portions 202 and 248 for conceptual clarity, in some aspects, second portion 248 (eg, block 255 ) may be incorporated into first portion 202 (eg, within block 215 ).

[0050] In some aspects of the present disclosure, a conditional prior is defined on the sparsified dictionary, where the prior for the dictionary at a given iteration is conditioned on the dictionary sampled from the previous iteration. That is, as illustrated, from (depicted as Ψ in block 255A) t-1 The first iteration of the sparsification dictionary may be used as a dictionary for (depicted as Ψ in block 255B) t In some aspects, as discussed in more detail below, a machine learning model (e.g., a neural network) f θ It is used to define the conditional prior on the dictionary. That is, the neural network f θ Parameters that can be used to generate a prior on the dictionary, where the model is characterized by a learnable parameter θ. In the illustrated example, this is depicted as a parameter 250 (illustrated as θ) being provided as an input to define the prior distribution.

[0051] In at least one aspect, in the case of the first iteration (e.g., when the dictionary from the previous iteration is not available), the inference system sets the prior to be an unconditional standard Gaussian distribution. For subsequent iterations, the prior may be given by a Gaussian distribution with a mean and variance generated using a neural network, as discussed in more detail below.

[0052] In some aspects, the joint prior (through T iterations) is defined using the following Equation 1, where p θ is the prior distribution, Ψ t is the sparse dictionary at iteration t, Ψ 1:T is the sparsified dictionary from the first to the Tth, and p(Ψ1) is the first iteration (e.g., ) and p θ (Ψ t |Ψ t-1 ) represents the conditional prior at iteration t based on the dictionary from the prior iteration at t-1.

[0053]

[0054] At any given iteration t, in some aspects, the conditional prior can be defined as p θ (Ψ t |Ψ t-1 ). In some aspects, as discussed above, the conditional prior is implemented as a Gaussian distribution with parameters (e.g., mean and standard deviation or variance) defined by a machine learning model (such as a neural network). For example, in some aspects, the conditional prior at iteration t is defined using the following Equation 2, where represents a Gaussian distribution, μ is the mean of the distribution, σ 2 is the variance, and f θ (Ψ t-1 ) indicates the use of neural network f θ To process the previously sparse dictionary Ψ t-1 That is, the prior p θ The mean μ and variance σ of 2 Generated by processing a previously sparsified dictionary using a trained neural network, as discussed in more detail below.

[0055]

[0056] As discussed in more detail below, the second portion 248 (corresponding to the prior distribution) may be used during training of the model. During inference (e.g., when generating channel estimates at runtime), the second portion 248 may be unused, and the inference system may generate estimates using only the posterior model and the likelihood model, as discussed in more detail below.

[0057] In the illustrated architecture 200, the first portion 202 corresponds to the use of a posterior distribution (defined using another neural network) and a likelihood model. In the illustrated example, at each block 215, a set of parameters 220 (denoted as φ) corresponding to the neural network used to define the posterior distribution is received along with the input data (e.g., observations 210 and sensing matrix 205). Each block 215 also receives a corresponding set of parameters 230 and 235 that define the likelihood model, as discussed in more detail below. As further illustrated, each block 215 receives a sparse channel representation 240 (240A, 240B) generated by a previous iteration / block.

[0058] In particular, as illustrated, the sparse channel representation 240A from the first iteration or block 215A is processed (e.g., by the posterior neural network) to generate a sparsified dictionary 225B at a second iteration or block 215B based at least in part on corresponding posterior network parameters 220. In particular, in some aspects, the neural network generates parameters of the posterior distribution, which are sampled to generate the sparsified dictionary 225B at block 215B. In some aspects, the input observations 210 and the sensing matrix 205 are also used as inputs to the posterior neural network to generate corresponding posterior distribution parameters, as discussed in more detail below. In the first iteration / layer, there is no previous iteration, and therefore no previous sparse channel representation. In at least one aspect, in the first iteration, the inference system Define a value / vector as zero and use this value / vector as input to the posterior network.

[0059] As illustrated, the sampled sparsification dictionary 225B is then provided as an input to the likelihood model to generate a current sparse channel representation 240B at iteration or block 215B based at least in part on corresponding likelihood model parameters 230B and 235B. In at least one aspect, the likelihood model is an A-DLISTA model defined by learnable parameters 230 (e.g., soft threshold) and 235 (e.g., step size). In the illustrated example, each iteration or block 215 has a corresponding set of likelihood model parameters 230 and 235 (e.g., where the parameters 230A and 235A for iteration or block 215A may be different from the parameters 230B and 235B for iteration or block 215B). In some aspects, the parameters 230 and 235 are shared across iterations or blocks 215.

[0060] Additionally, in the illustrated example, a-posteriori parameters 220 are shared across iterations or blocks 215. In some aspects, each iteration or block 215 may use a corresponding unique set of learning parameters 220, as discussed in more detail below. Similarly, in the illustrated example, a-priori parameters 250 are shared across iterations or blocks 255. In some aspects, each iteration or block may use a corresponding unique set of learning parameters 250, as discussed in more detail below.

[0061] In some aspects, a likelihood model is used to generate the sparse channel representation 240 and defines the probability on the sparse representation at iteration t and conditioned on the previously sampled dictionary and input data (observations 210 and sensing matrix 205). In some aspects, the likelihood model is defined using the following equation 3, where p Θ represents a likelihood model parameterized by learnable parameters Θ (which may include parameters 230 and 235), is the sparse channel representation generated at iteration t, x gt is a ground truth sparse channel representation corresponding to the input observation 210 and the sensing matrix 205 (eg, generated based on a ground truth channel estimate reflecting the actual state of the channel at the time the observation 210 was recorded), Ψ 1:t is the sparsified dictionary generated in iterations one to t, y is the input observation 210 , and Φ is the corresponding sensing matrix 205 .

[0062]

[0063] In some aspects, as discussed above, the likelihood model is implemented using a machine learning model (e.g., a neural network) such as an A-DLISTA model. In some such aspects, the likelihood model is implemented using a Gaussian distribution with parameters (e.g., mean and variance) defined by the trained model. For example, the likelihood model can be defined using the following Equation 4, where represents a Gaussian distribution, μ is the mean of the distribution, σ 2 is the variance, and LISTA(Ψ 1:t ,y,Φ;Θ) denotes the use of the A-DLISTA machine learning model parameterized by the learnable parameter Θ to process the previously sparse dictionary Ψ 1:t , observation 210 and sensing matrix Φ. That is, the likelihood model p Θ The mean μ of can be generated by processing the previously sparse dictionary and input data using a trained neural network, as discussed in more detail below. In some aspects, the variance σ 2is a hyperparameter of the architecture 200. In other aspects, the variance may also be a learned parameter. Using Equation 4, the inference system may generate a distribution with a mean defined using a neural network (e.g., an A-DLISTA model) and a fixed (e.g., hyperparameter) or learned variance. This distribution may then be returned as the sparse channel representation 240 for the current iteration (and, in some aspects, may be used to generate a channel estimate for the current iteration, as discussed above).

[0064]

[0065] In some aspects, the joint log probability distribution of the channel and the dictionary at each iteration may be defined as and is given by the following equation 5, where Ψ 1:T is the set of sparse dictionaries generated in iterations 1 to T, is the sparse channel representation generated in iterations 1 to T, x gt is the ground truth sparse channel representation, and y and Φ are the channel observation and sensing matrices, respectively. In some aspects, the joint probability for the dictionary and the sparse channel representation represents the product between the likelihood and the prior distribution on the dictionary. In some aspects, this joint probability can be used in formulating the objective function for training the model.

[0066]

[0067] In some aspects, a variational posterior is used to sample the sparsified dictionary 225 and is defined as a Gaussian distribution with parameters (e.g., mean and variance) generated using a machine learning model such as a neural network (e.g., a posterior network). In some such aspects, the posterior distribution is defined as where q φ is the posterior distribution, Ψ t is the sparsified dictionary sampled at iteration t, is the sparse channel representation generated at iteration t-1, y is the observation, and Φ is the sensing matrix. In some aspects, the posterior is implemented using a Gaussian distribution with parameters (e.g., mean and variance or standard deviation) defined by the trained model. For example, the posterior can be defined using Equation 6 below, where represents a Gaussian distribution, μ is the mean of the distribution, β -1 is the precision of the distribution (e.g., the inverse of the variance), and Representation uses a machine learning model (e.g., a neural network) parameterized by a learnable parameter φ to process the prior sparse channel representation generated in the immediately preceding a priori iteration observation 210 and the sensing matrix Φ. That is, the parameters μ and β of the posterior distribution -1may be generated by processing a previous sparse channel representation and input data using a trained neural network, as discussed in more detail below.

[0068]

[0069] In some aspects, the sampled sparsified dictionary Ψ t implicitly depends on Ψ 1:t-1 , because the sparse dictionary Ψ at iteration t t is used (which in turn depends on Ψ 1:t-1 ) to generate / sample.

[0070] Example workflow for training a machine learning model for channel estimation

[0071] Figure 3 An example workflow 300 for training a machine learning model to perform channel estimation is depicted. In some aspects, the workflow 300 consists of an estimation component, such as Figure 1 The estimation component 145 is used. In some aspects, the workflow 300 is used to provide offline training of the machine learning model. The training can generally be performed by any suitable device or system, such as a dedicated training system.

[0072] As illustrated, the workflow 300 begins with a set of input data 305 including a channel observation y and a corresponding sensing matrix Φ. As illustrated, the exemplary input data 305 is provided as input to a posterior network 310 and a likelihood network 325. The posterior network 310 is typically a machine learning model (e.g., a neural network) that learns to generate distribution parameters for a variational posterior distribution, as discussed above. In the illustrated example, the posterior network 310 also receives as input a sparse channel representation 330 generated using the likelihood network 325 discussed in more detail below. In some aspects, the posterior network receives the input data 305 and the sparse channel representation 330 generated in a previous iteration, layer, or block of the model, as discussed above. In particular, as illustrated, the posterior network 310 processes the previous sparse channel representation and input data 305 (y and Φ) to generate distribution parameters (e.g., mean and variance), and the training system uses these values ​​to generate the posterior distribution As illustrated, the training system can then apply the distribution (denoted as ) to generate the current sparsification dictionary 315 (denoted as Ψ t ).

[0073] In some aspects, the posterior network 310 includes three input heads, one for each input (e.g., one for the prior sparse channel representation ), one input head for observations y of input data 305, and one input head for sensing matrix Φ of input data 305. In some aspects, posterior network 310 also includes one or more subsequent layers (e.g., linear layers) shared by each input head. In some aspects, posterior network 310 includes a set of branches after the shared layer, wherein each branch includes one or more layers (e.g., convolutional layers) and outputs a corresponding value. For example, a first branch may generate a mean value, while a second branch may generate a variance value of a distribution represented by posterior network 310.

[0074] In the illustrated example, this sampled sparsified dictionary 315 is provided to both the prior network 320 and the likelihood network 325. As depicted and discussed above, the prior network 320 may process the current sparsified dictionary Ψ t and / or a priori sparsified dictionary Ψ t-1 To generate the distribution parameters, and thus define the conditional prior distribution p θ (Ψ t |Ψ t-1 ). In some aspects, the prior distribution is used to help define an entropy loss 335 (as indicated by dashed lines 331 and 333) for refining the architecture / model, as discussed in more detail below.

[0075] In some aspects, the prior network 320 includes one input head for a single input (e.g., one input head for the sparse dictionary 315). In some aspects, in a manner similar to the posterior network 310, the prior network 320 also includes one or more subsequent layers (e.g., one or more convolutional layers), followed by a set of branches after the shared layer, wherein each branch includes one or more layers (e.g., convolutional layers) and outputs a corresponding value. For example, a first branch may generate a mean value, while a second branch may generate a variance value of a distribution represented by the prior network 320.

[0076] In the illustrated workflow 300, the likelihood network 325 also receives the sampled sparsification dictionary 315 and uses it together with the input 305 to generate a new sparse channel representation 330 (denoted as ). As illustrated, the likelihood network 325 is parameterized by learnable parameters Θ, which may include a soft threshold λ and a step size τ, as discussed above. In some aspects, the likelihood network 325 corresponds to an A-DLISTA network. In the illustrated example, the likelihood network 325 may use corresponding parameters Θ learned specifically for iteration t t That is, the model can learn a different set of parameters for each iteration / block. In other aspects, these parameters can be shared across iterations.

[0077] In some aspects, as discussed above, the likelihood network is used to generate one or more parameters (eg, mean and / or variance) of a likelihood distribution. This distribution can then be returned as the sparse channel representation 330 for the current iteration.

[0078] As illustrated, the sparse channel representation 330 generated in the current iteration may be provided to the posterior network 310 in a subsequent or subsequent iteration. In some aspects, the sparse channel representation 330 is also used to generate a reconstruction loss 340 (as indicated by dashed line 337), which is discussed in more detail below. In some aspects, the training system generates an entropy loss 335 and a reconstruction loss 340 at each iteration. That is, for each block or layer of the model, the training system may generate a corresponding entropy loss 335 and a reconstruction loss 340, and use these losses to refine the model. In other aspects, given the input data 305, the training system may generate a reconstruction loss 340 and an entropy loss 335 only after the final iteration or layer has been processed to generate the final sparse channel representation from the model.

[0079] In the illustrated example, the training component 345 can generate a loss and / or use the loss to refine the learning parameters of the prior network 320, the posterior network 310, and the likelihood network 325. In at least one aspect, the model is trained by maximizing the evidence lower bound (ELBO) (also referred to as the variational lower bound or negative variational free energy in some aspects). The ELBO can generally be used as a lower bound on the log-likelihood of the observed data. In at least one aspect, the ELBO is defined using the following equation 7, where the ELBO is the total loss (which is maximized during training), Loss R is the mean square error for the reconstructed channel and / or sparse channel representation (assuming a Gaussian prior) defined using Equation 8 below, and Loss P1 and Loss P2 is the Kullback-Leibler (KL) divergence between the variational posterior and the conditional prior at each iteration t, defined in Equations 9 and 10 below, respectively. In some aspects, the training system may use a defined number of iterations for each input during training (e.g., using a defined number of blocks or layers to process the input data 305). The same or a different number of iterations may then be used during inference, as discussed in more detail below.

[0080] ELBO=Loss R -Loss P1 -Loss P2 (7)

[0081] In some aspects, the following Equation 8 is used to define the reconstruction loss 340. In some aspects, as discussed above, the reconstruction loss 340 is generated for the current iteration (as indicated by dashed line 337) based at least in part on the sparse channel representation 330 (or, in some aspects, a corresponding channel estimate generated based on the sparse channel representation). In Equation 8, Ψ 1:t is the sparse dictionary sampled in layers 1 to t, y is the channel observation, Φ is the corresponding sensing matrix, is the sparse channel representation from layer / iteration zero (e.g., the first iteration) to t-1, is the sparse channel representation generated at iteration t, and x gt is a ground truth sparse channel representation (eg, generated based on the ground truth channel). Equation 8 generates the sum across model layers / iterations from one to T. For each iteration, given a dictionary sampled from the posterior distribution, the training system can determine the expected value of the log-likelihood distribution.

[0082]

[0083] In some aspects, the following equation 9 is used to define a portion of the entropy loss 335. In some aspects, as discussed above, the entropy loss 335 is generated based at least in part on the variational posterior distribution (generated using the posterior network 310 and as indicated by the dashed line 333) and the generated prior distribution (generated using the prior network 320 and as indicated by the dashed line 331). In a manner similar to equation 8, the sum in equation 9 spans the model from the second iteration to the final layer. For each layer, equation 9 evaluates the expected value of the KL divergence between the variational posterior distribution (e.g., generated using the posterior network 310) and the prior distribution (e.g., generated using the prior network 320). In the example below, the summation may start from the second layer because the following equation 10 includes the KL divergence of the first layer.

[0084]

[0085] In some aspects, the following Equation 10 is used to define a portion of the entropy loss 335. For example, the entropy loss 335 can be defined as Loss P1 -Loss P2 Using Equation 10, the training system can determine the KL divergence between the posterior distribution and the prior distribution for the first layer / iteration of the model.

[0086]

[0087] In some aspects, using workflow 300, a training system can use a variational A-DLISTA method to learn an approximate sparsified dictionary for an inverse problem such as channel estimation, wherein the variational A-DLISTA method defines uncertainty in LISTA parameters such as dictionary, threshold, and step size. In some aspects, as discussed above, variational A-DLISTA enables a training system to adapt the dictionary to different measurements.

[0088] In some aspects, as discussed above, the training system may learn the sparsified dictionary Ψ in its entirety. In at least one aspect, the training system may determine components or subsets of the sparsified dictionary from other sources and learn the remainder. For example, in some such aspects, the training system may receive a sparsified dictionary used by a transmitting device (e.g., a base station) for communications. Since the overall sparsified dictionary Ψ may be the Kronecker product of a receiver dictionary and a transmitter dictionary, in some aspects, the training system may use the (known) transmitter dictionary to optimize the learning of the overall sparsified dictionary Ψ (e.g., learning only the receiver-side dictionary). Similarly, the training system may receive a sparsified dictionary for a receiving device and learn the transmitter's dictionary.

[0089] In some aspects, once a given input sample (including sensing matrices, observations, and ground truth channel states and / or sparse channel representations) has been used to refine a model, subsequent samples may be similarly used to generate corresponding sparse channel representations 330 and losses 335 and 340. In various aspects, this training workflow 300 may be performed using stochastic gradient descent (e.g., backpropagating each loss for each input sample individually) and / or batch gradient descent (e.g., refining a model based on losses generated for a batch of input samples).

[0090] Once training is complete, the model architecture can be used to generate a sparse channel representation based on the input channel observation and sensing matrices during inference (which, in some aspects, can then be used to generate channel estimates), as discussed in more detail below.

[0091] Example workflow for using a machine learning model for channel estimation

[0092] Figure 4 An example workflow for using a trained machine learning model to generate a sparse channel representation (e.g., to perform channel estimation) is depicted. In some aspects, workflow 400 consists of an estimation component, such as Figure 1 In some aspects, workflow 400 is used to provide runtime generation of a sparse channel representation (usable to provide channel estimation) using a trained machine learning model. In at least one aspect, workflow 400 is performed by a user equipment (e.g., a user equipment) participating in wireless communication. Figure 1In general, workflow 400 may be performed by any inference system.

[0093] Workflow 400 typically uses a trained machine learning model (e.g., using Figure 3 The network trained by workflow 300 of FIG. 4 generates a sparse channel representation 430 based on input data 405 (including channel observations and corresponding sensing matrices). As discussed above, channel observations typically correspond to signals received by a receiver (e.g., Figure 1 The sensing matrix is ​​typically defined based on a beamforming codebook used by the transmitter and receiver when observing the signal.

[0094] Workflow 400 largely mirrors workflow 300, except that no Figure 3 The prior network 320 is generated, and the management component 450 uses the generated sparse channel representation for various purposes (rather than the training component 345 using the predictions to refine the model).

[0095] As illustrated, workflow 400 begins with a set of input data 405 including channel observations y and corresponding sensing matrices Φ. As illustrated, this input data is provided as input to a posteriori network 410 (which, once trained, may correspond to Figure 3 The posterior network 310 of ) and the likelihood network 425 (once trained, it may correspond to Figure 3 As discussed above, the posterior network 410 is typically a machine learning model (e.g., a neural network) that learns to generate distribution parameters for a variational posterior distribution, as discussed above. In the illustrated example, the posterior network 410 also receives as input a sparse channel representation 430 generated by the likelihood network 425 discussed in more detail below. In some aspects, the posterior network receives the current input data 405 and the sparse channel representation 430 generated in a previous iteration, layer, or block of the model, as discussed above. In particular, as illustrated, the posterior network 410 processes the previous sparse channel representation and input data 405 (y and Φ) to generate distribution parameters (e.g., mean and variance), and the inference system uses these values ​​to generate the posterior distribution As illustrated, the inference system can then apply the distribution (represented as ) are sampled to generate the current sparsification dictionary 415 for the current iteration.

[0096] In the illustrated example, this sampled sparsification dictionary 415 is provided to a likelihood network 425. The likelihood network 425 uses the sampled sparsification dictionary 415 together with the input data 405 to generate a new sparse channel representation 430 for the current iteration. As illustrated, the likelihood network 425 is parameterized by learning parameters Θ, which may include a soft threshold λ and a step size τ, as discussed above. In some aspects, the likelihood network 425 corresponds to an A-DLISTA network. In the illustrated example, the likelihood network 425 may use the corresponding parameters Θ learned specifically for iteration t t That is, the model can learn a different set of parameters for each iteration / block. In other aspects, these parameters can be shared across iterations.

[0097] As illustrated, the sparse channel representation 430 generated in the current iteration can be provided to the a posteriori network 410 in a subsequent or subsequent iteration. As illustrated, the sparse channel representation 430 can also be selectively provided to a management component 450 discussed in more detail below. The sparse channel representation 430 can then be used for various purposes, such as generating corresponding channel estimates, thereby enabling operations such as improving analog beamforming, beam selection, spectral efficiency prediction, etc.

[0098] In the illustrated example, at inference time, for each iteration, the posterior is used to sample the dictionary, and the sparse channel representation is reconstructed by the likelihood model. In some aspects, the inference system uses the workflow 400 to iteratively generate a new sparse channel representation 430 at each iteration / block / layer of the model until one or more defined termination criteria are met. In some aspects, the termination criteria relate to the number of completed iterations or layers. In some such aspects, the inference system may generate a new sparse channel representation 430 until a defined number of iterations (or blocks) have been processed. For example, if the model is trained using ten iterations, the workflow 400 may iterate through an equal number (e.g., ten) of iterations. After the tenth iteration, the sparse channel representation 430 may be provided to the management component 450 (rather than the posterior network 410). As discussed above, the management component 450 (or another component) may then use the sparse channel representation to generate a channel estimate (e.g., where the channel estimate Equal to the sampled sparsification dictionary Ψ and the sparse channel representation multiply).

[0099] In at least one aspect, the inference system may use one or more dynamic termination criteria to immediately output the current sparse channel representation 430, rather than continuing processing, in order to reduce computational costs and latency. In some such aspects, the inference system may determine the entropy of the likelihood (e.g., of the sparse channel representation / distribution generated using the likelihood network), and use this entropy to define one or more stopping criteria for the iterative procedure in order to reduce the number of iterations. For example, if the determined entropy meets one or more defined criteria, the inference system may output the current sparse channel representation / distribution, or may output (e.g., to the management component 450) the current channel estimate generated based on the current sparse channel representation and suppress any additional processing. That is, the inference system may suppress providing the sparse channel representation 430 to the posterior network 410 for processing in subsequent iterations (or in subsequent blocks). In some aspects, one or more defined criteria may include determining whether the value of the entropy is equal to or less than a defined value, whether the difference between the entropies of the likelihoods generated in consecutive layers is less than a threshold, etc.

[0100] In this way, the inference system can use fewer iterations or blocks at runtime than it used to train at training time. In contrast, for at least some conventional methods, inference uses the same number of iterations / blocks as training. This dynamic online stopping can reduce the total number of iterations, thereby reducing latency and power consumption of the model.

[0101] Additionally, as discussed above, at each iteration, the distribution of the sparsified dictionary is updated using a machine learning model (e.g., a neural network). In some aspects, this network can be used to provide uncertainty estimates at the output. For example, the inference system can evaluate the generated variance of the posterior distribution (generated by the posterior network). This uncertainty estimate can then be used for out-of-distribution detection (e.g., where a sample is considered out-of-distribution if the uncertainty meets or exceeds a certain threshold), as well as making decisions about communication parameters and / or beam management procedures. Similarly, in some aspects, uncertainty can be used to indicate that retraining and / or fine-tuning the model can be advantageous. That is, when the uncertainty meets one or more defined criteria, the inference system can initiate retraining or fine-tuning of the model to ensure that high accuracy is maintained.

[0102] Example method for training a machine learning model for channel estimation

[0103] Figure 5 is a flowchart depicting an example method 500 for training a machine learning model for generating a sparse channel representation (eg, for channel estimation). In some aspects, the method 500 comprises an estimation component, such as Figure 1 The estimation component 145 is executed. In some aspects, the method 500 is performed to provide offline training of one or more machine learning models, such as by a training device or system.

[0104] At block 505, the training system determines, identifies, receives, or otherwise accesses a channel including channel observations (e.g., Figure 2 Observation 210), sensing matrix (e.g., Figure 2 In some aspects, as discussed above, the channel observations correspond to wireless signals as observed or measured by a receiving device or system, the sensing matrix is ​​defined based on parameters used to transmit and / or receive signals (e.g., beamforming codebooks used by transmitters and receivers), and the channel state corresponds to characteristics of the wireless medium when the signal is transmitted.

[0105] In some aspects, channel observations, sensing matrices, and channel states (or sparse channel representations) may typically be determined or received as part of a training sample or example, as discussed above. Using such samples, a training system may iteratively refine a machine learning model (e.g., a posteriori network, a priori network, and likelihood model) to generate an improved and more accurate sparse channel representation and channel estimate.

[0106] At block 510, the training system generates a current sparsified dictionary for the current iteration. In some aspects, as discussed above, the training system generates the current sparsified dictionary using a variational posterior distribution derived from a machine learning model (e.g., Figure 3 The posterior network 310 of the training system may be defined by parameters generated by the posterior network 310 of the training system. In some aspects, as discussed above, the posterior network processes the input sensing matrix and observations together with the sparse channel representation generated in the previous iteration (if available) to generate distribution parameters (e.g., mean and / or variance values). This distribution may then be sampled to generate the current sparsified dictionary. In at least one aspect, as discussed above, in a first iteration (e.g., where no previous channel estimates are available), the training system may use a standard unconditional Gaussian distribution (e.g., ) as the posterior distribution, and sample this distribution to generate the sparsified dictionary for the first iteration.

[0107] At block 515, the training system may then generate a sparse channel representation based at least in part on the generated sparsification dictionary. In some aspects, the sparse channel representation is generated by a likelihood model (e.g., an A-DLISTA model). In some aspects, as discussed above, the training system generates the current sparse channel representation using a likelihood model and / or distribution that uses a distribution generated by a machine learning model (e.g., Figure 3In some aspects, as discussed above, the likelihood network processes the input sensing matrix and observations together with the current sparsification dictionary generated in block 510 to generate distribution parameters (e.g., mean and / or variance values). This distribution (defined by the generated parameters) can then be used as a sparse channel representation and / or can be sampled to generate a sparse channel representation.

[0108] At block 520, the training system determines whether there is at least one additional iteration, layer, or block remaining in the model. That is, the training system may determine whether the iteration processed at blocks 510 and 515 is the last or final iteration specified in the model architecture. If there is at least one additional iteration remaining, the method 500 returns to block 510, where the training system may generate a new sparsification dictionary based at least in part on processing the current sparse channel representation (generated using the a priori / initial sparsification dictionary at block 515) using the next layer, block, or iteration of the model. In this way, the method 500 may continue to process data through each iteration of the model, thereby iteratively refining or updating the sparse channel representation as it passes through each layer.

[0109] Although the illustrated examples depict one or more termination criteria as a number of iterations, in some aspects, as discussed above, the training system may use one or more other criteria (such as entropy of the likelihood model) to determine whether to process additional iterations.

[0110] If no additional layers remain (or one or more other termination criteria are met), method 500 continues to block 525. At block 525, the training system generates an entropy loss based on the input training samples, as discussed above. For example, as discussed above, the training system can use Equations 9 and 10 to generate a KL divergence loss between the variational posterior and the conditional prior at each iteration of the model.

[0111] At block 530, the training system may then calculate a reconstruction loss based on the sparse channel representation generated at block 515 (or based on a corresponding channel estimate generated from the sparse channel representation) and the known sparse channel representation (or known channel state, as discussed above) determined at block 505. For example, as discussed above, the training system may calculate a normalized mean square error between the sparse channel representation and / or the channel estimate and the channel state using Equation 8. Although the illustrated example depicts generating the loss based on a single training sample (e.g., a single observation with a corresponding sensing matrix and channel state), in some aspects, the training system may generate the loss based on multiple samples using batch training.

[0112] At block 535, the training system can then refine the various networks (e.g., posterior networks, prior networks, and / or likelihood networks) of the model based on the loss, as discussed above (e.g., using Equation 7). For example, starting with the final layer and moving toward the first layer, the training system can use backpropagation to refine the parameters of each layer of each model. In general, method 500 can be used to refine a neural network model using any number and variety of training samples.

[0113] In some aspects, as discussed above, the training system can train or refine a separate set of parameters for each iteration, layer, or block of the model. That is, each respective iteration can have a respective set of parameters for the likelihood model (e.g., learned step sizes and / or learned thresholds), a respective set of parameters for the posterior model, and / or a respective set of parameters for the a priori model. In other aspects, the training system can use a shared set of parameters for one or more of the networks across iterations, as discussed above.

[0114] By using method 500, the training system can train models to generate significantly improved sparse channel representations, resulting in more accurate channel estimates, particularly for MIMO systems and / or systems using analog beamforming. Additionally, as discussed above, the trained models can generally use sparsified dictionaries with lower dimensionality than at least some conventional methods, thereby enabling these models to be used by devices or systems with more limited computational or power resources, such as low-power user equipment.

[0115] Example method for channel estimation using a machine learning model

[0116] Figure 6 is a flowchart depicting an example method for performing channel estimation using a machine learning model. In some aspects, method 600 comprises an estimation component, such as Figure 1 The estimation component 145 is performed. In some aspects, the method 600 is performed by an inference device or system. In at least one aspect, the method 600 is performed by a user equipment participating in wireless communication. In some aspects, the method 600 may be additionally or alternatively performed by a base station participating in the communication.

[0117] At block 605, the inference system determines, identifies, receives, or otherwise accesses a channel including channel observations (e.g., Figure 2 of observation 210) and sensing matrix (e.g., Figure 2 In some aspects, as discussed above, the channel observations correspond to wireless signals as observed or measured by a receiving device or system, and the sensing matrix is ​​defined based on parameters used to transmit and / or receive signals (e.g., beamforming codebooks used by transmitters and receivers).

[0118] In some aspects, the channel observation and sensing matrices may generally be determined or received as part of data samples processed at runtime, as discussed above. By evaluating such samples, the inference system may generate an accurate channel estimate for the wireless channel.

[0119] At block 610, the inference system generates a current sparsified dictionary for the current iteration. In some aspects, as discussed above, the inference system generates the current sparsified dictionary using a variational posterior distribution derived from a machine learning model (e.g., Figure 3 The posterior network 310 of the present invention is defined by parameters generated by the posterior network 310 of the present invention. In some aspects, as discussed above, the posterior network processes the input sensing matrix and observations together with the sparse channel representation generated in the previous iteration (if available) to generate distribution parameters (e.g., mean and / or variance values). This distribution can then be sampled to generate the current sparsified dictionary. In at least one aspect, as discussed above, in a first iteration (e.g., where no previous sparse channel representation is available), the inference system can use a standard unconditional Gaussian distribution (e.g., ) as the posterior distribution, and sample this distribution to generate the sparsified dictionary for the first iteration.

[0120] At block 615, the inference system may then generate a sparse channel representation based at least in part on the generated sparsified dictionary. In some aspects, the sparse channel representation is generated by a likelihood model (e.g., an A-DLISTA model). In some aspects, as discussed above, the inference system generates the current sparse channel representation using a likelihood model and / or distribution that uses a probability distribution generated by a machine learning model (e.g., Figure 3 In some aspects, as discussed above, the likelihood network processes the input sensing matrix and observations together with the current sparsification dictionary generated in block 610 to generate distribution parameters (e.g., mean and / or variance values). This distribution (generated using the generated parameters) can then be used as a sparse channel representation (or can be sampled to generate a sparse channel representation).

[0121] At block 620, the inference system determines whether there is at least one additional iteration, layer, or block remaining in the model. That is, the inference system may determine whether the iteration processed at blocks 610 and 615 is the last or final iteration specified in the model architecture. If there is at least one additional iteration remaining, the method 600 returns to block 610, where the inference system may generate a new sparsification dictionary based at least in part on processing the current sparse channel representation (generated using the a priori / initial sparsification dictionary at block 615) using the next layer, block, or iteration of the model. In this way, the method 600 may continue to process data through each iteration of the model, thereby iteratively refining or updating the sparse channel representation as it passes through each layer.

[0122] Although the illustrated examples depict one or more termination criteria as a number of iterations, in some aspects, as discussed above, the inference system may use one or more other criteria (such as entropy of the likelihood model) to determine whether to process additional iterations.

[0123] If no additional iterations remain (or one or more other termination criteria are met), the method 600 continues to block 625. At block 625, the inference system may then use the sparse channel representation to generate or recover a channel estimate (as discussed above). The channel estimate may then be used to reconfigure, refine, or otherwise improve the wireless communication, such as by redefining the beamforming configuration. For example, as discussed above, the channel estimate may be used to inform or drive an analog beamforming design to enable faster and more accurate beam selection for the communication, and / or to provide improved spectral efficiency predictions.

[0124] Example Method for Generating Channel Estimates

[0125] Figure 7 is a flow chart depicting an example method 700 for generating channel estimates. In some aspects, the method 700 comprises an estimation component, such as Figure 1 In some aspects, the method 700 is performed by a user equipment participating in wireless communication. In some aspects, the method 700 may additionally or alternatively be performed by a base station participating in communication. In at least one aspect, the method 700 is performed by a processing system, such as Fig. 9 Executed by the processing system 900.

[0126] At block 705 , a current sparsified dictionary is generated in a first iteration of a machine learning model by processing a sensing matrix and current channel observations for a digital communication channel using an a posteriori neural network.

[0127] At block 710, a current sparse channel representation is generated in a first iteration by processing a current sparsified dictionary, a sensing matrix, and a current channel observation using a likelihood neural network.

[0128] In some aspects, generating a current sparsified dictionary using a posteriori neural network includes: generating mean and variance values ​​by processing a sensing matrix, a current channel observation, and a previous sparse channel representation generated in a previous iteration of a machine learning model using the posteriori neural network; and sampling a distribution having the mean and variance values.

[0129] In some respects, a distribution is defined as The posterior distribution of , where: t is the current sparse dictionary, is the previous sparse channel representation generated in the previous iteration of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0130] In some ways, the posterior distribution is equal to Where: μ is the mean, f φ is the posterior neural network, and β -1 is the precision value.

[0131] In some aspects, generating a current sparse channel representation using a likelihood neural network includes: generating a mean by processing a current sparsified dictionary, a sensing matrix, and a current channel observation using a likelihood neural network; and generating a distribution having a mean and a variance value.

[0132] In some respects, a distribution is defined as The likelihood model of , where: is the current sparse channel representation, x gt Corresponding to the ground truth channel state, ψ 1:t is the sparsified dictionary generated in one or more previous iterations in the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0133] In some respects, the likelihood model is equivalent to where: μ is the mean, ADLISTA(·) denotes the application of the likelihood neural network, Θ is the training parameter of the likelihood neural network, and σ 2 is the variance value, where the variance value is a hyperparameter.

[0134] In some aspects, method 700 further includes: generating an uncertainty measure using a likelihood neural network; determining that the uncertainty measure satisfies one or more defined criteria; and in response to determining that the uncertainty measure satisfies the one or more defined criteria, initiating retraining of the posterior neural network and the likelihood neural network.

[0135] In some aspects, method 700 also includes: determining entropy based on a likelihood neural network; determining that the entropy satisfies one or more defined criteria; and in response to determining that the entropy satisfies one or more defined criteria: suppressing subsequent iterations of using the machine learning model to process the sensing matrix and the current channel observation; generating a current channel estimate based on the sparse channel representation; and outputting the current channel estimate.

[0136] In some aspects, the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

[0137] In some aspects, each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

[0138] In some aspects, the digital communication channel comprises a millimeter wave digital communication channel.

[0139] In some aspects, the method 700 further comprises performing one of analog beamforming, beam selection, or spectral efficiency prediction based on the current sparse channel representation.

[0140] In some aspects, method 700 further includes: generating a first loss by processing the current sparsified dictionary using a priori neural network; generating a second loss based on the current sparse channel representation; and refining the posterior neural network, the likelihood neural network, and the prior neural network based on the first loss and the second loss.

[0141] In some aspects, generating the first loss using the prior neural network includes: generating mean and variance values ​​by processing a previously sparsified dictionary generated in a previous iteration of the machine learning model using the prior neural network; and generating a distribution having the mean and variance values.

[0142] In some aspects, the distribution is defined as p θ (Ψ t |Ψ t-1 ), where: t is the current sparse dictionary, and Ψ t-1 is the previously sparsified dictionary generated in a previous iteration of the machine learning model.

[0143] In some respects, the prior distribution is equal to σ 2 =f θ (Ψ t-1 ), where: μ is the mean, f θ is a priori neural network, and σ 2 is the variance value.

[0144] In some aspects, the previous iteration corresponds to an input iteration of the machine learning model, and the previous sparsified dictionary is drawn from a Gaussian distribution Sampled in.

[0145] In some aspects, the first loss and the second loss are collectively defined as in: is the expected value of the sparsified dictionary evaluated for up to iteration t sampling, is the sparse channel representation generated for up to iteration t-1, x gt is the ground truth sparse channel representation, y is the channel observation, Φ is the sensing matrix, and is the expected value of the sparsified dictionary evaluated for up to iteration t-1 sampling.

[0146] Example method for training a machine learning model for channel estimation

[0147] Figure 88 is a flowchart depicting an example method 800 for training a machine learning model for channel estimation. In some aspects, the method 800 comprises an estimation component, such as Figure 1 In some aspects, the method 800 is performed by a user equipment participating in wireless communication. In some aspects, the method 800 may be additionally or alternatively performed by a base station participating in communication. In some aspects, the method 800 is performed by a processing system, such as Fig. 9 Executed by the processing system 900.

[0148] At block 805 , a sensing matrix and current channel observations for a digital communication channel are received.

[0149] At block 810 , a current sparsified dictionary is generated by processing the sensing matrix and the current channel observation using an a posteriori neural network.

[0150] At block 815, a current sparse channel representation is generated by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using a likelihood neural network.

[0151] At block 820 , a first loss is generated by processing the current sparsified dictionary using a priori neural network.

[0152] At block 825, a second loss is generated based on the current sparse channel representation.

[0153] At block 830 , the posterior neural network, the likelihood neural network, and the prior neural network are refined based on the first loss and the second loss.

[0154] In some aspects, generating the first loss using the prior neural network includes: generating mean and variance values ​​by processing a previously sparsified dictionary generated in a previous iteration of the machine learning model using the prior neural network; and generating a distribution having the mean and variance values.

[0155] In some aspects, the distribution is defined as p θ (Ψ t |Ψ t-1 ), where: t is the current sparse dictionary, and Ψ t-1 is the previously sparsified dictionary generated in a previous iteration of the machine learning model.

[0156] In some respects, the prior distribution is equal to σ 2 =f θ (Ψ t-1 ), where: μ is the mean, f θ is a priori neural network, and σ 2 is the variance value.

[0157] In some aspects, the previous iteration corresponds to an input iteration of the machine learning model, and the previous sparsified dictionary is sampled from a Gaussian distribution N(0,1).

[0158] In some aspects, the first loss and the second loss are collectively defined as

[0159] in: is the expected value of the sparsified dictionary evaluated for up to iteration t sampling, is the sparse channel representation generated for up to iteration t-1, x gt is the ground truth sparse channel representation, y is the channel observation, Φ is the sensing matrix, and is the expected value of the sparsified dictionary evaluated for up to iteration t-1 sampling.

[0160] In some aspects, generating a current sparsified dictionary using a machine learning model including a posteriori neural network includes: generating mean and precision values ​​by processing a sensing matrix, a current channel observation, and a previous sparse channel representation generated in a previous iteration of the machine learning model using the posteriori neural network; and sampling a distribution having mean and variance values.

[0161] In some respects, a distribution is defined as The posterior distribution of , where: t is the current sparse dictionary, is the previous sparse channel representation generated in the previous iteration of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0162] In some ways, the posterior distribution is equal to Where: μ is the mean, f φ is the posterior neural network, and β -1 is the precision value.

[0163] In some aspects, generating a current sparse channel representation using a likelihood neural network includes: generating a mean by processing a current sparsified dictionary, a sensing matrix, and a current channel observation using a likelihood neural network; and generating a distribution having a mean and a variance value.

[0164] In some respects, a distribution is defined as The likelihood model of , where: is the current sparse channel representation, x gt is the ground truth sparse channel representation, Ψ t:t is the sparsified dictionary generated in one or more previous iterations in the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0165] In some respects, the likelihood model is equivalent to where: μ is the mean, ADLISTA(·) denotes the application of the likelihood neural network, Θ is the training parameter of the likelihood neural network, and σ 2 is the variance value, where the variance value is a hyperparameter.

[0166] In some aspects, the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

[0167] In some aspects, each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

[0168] In some aspects, the digital communication channel comprises a millimeter wave digital communication channel.

[0169] Example processing system for improving channel estimation

[0170] In some respects, reference Figures 1 to 8 The described workflows, techniques, and methods may be implemented on one or more devices or systems. Fig. 9 An example processing system 900 is depicted that is configured to perform various aspects of the present disclosure, including, for example, Figures 1 to 8 Techniques and methods described herein. In some aspects, processing system 900 may correspond to a computing system that trains a machine learning model (e.g., a training system) and / or a computing system that uses the trained model to perform inferences (e.g., an inference system). In some aspects, processing system 900 corresponds to a base station and / or user equipment participating in wireless communications. Although depicted as a single system for conceptual clarity, in at least some aspects, as discussed above, the operations described below with respect to processing system 900 may be distributed across any number of devices. For example, a first system may train a model while a second system uses the trained model to generate a sparse channel representation / channel estimate.

[0171] The processing system 900 includes a central processing unit (CPU) 902, which may be a multi-core CPU in some examples. Instructions executed at the CPU 902 may be loaded, for example, from a program memory associated with the CPU 902, or may be loaded from a partition of the memory 924.

[0172] The processing system 900 also includes additional processing components tailored for specific functions, such as a graphics processing unit (GPU) 904 , a digital signal processor (DSP) 906 , a neural processing unit (NPU) 908 , a multimedia processing unit 910 , and wireless connectivity components 912 .

[0173] An NPU, such as NPU 908, is typically a dedicated circuit configured to implement control and arithmetic logic for executing machine learning algorithms, such as algorithms for processing artificial neural networks (ANNs), deep neural networks (DNNs), random forests (RFs), etc. An NPU is sometimes alternatively referred to as a neural signal processor (NSP), a tensor processing unit (TPU), a neural network processor (NNP), an intelligence processing unit (IPU), a vision processing unit (VPU), or a graphics processing unit.

[0174] NPUs, such as NPU 908, are configured to accelerate the execution of common machine learning tasks, such as image classification, machine translation, object detection, and various other predictive models. In some examples, multiple NPUs may be instantiated on a single chip, such as a system on a chip (SoC), while in other examples, the NPU may be part of a dedicated neural network accelerator.

[0175] The NPU can be optimized for training or inference, or in some cases configured to balance performance between the two. For NPUs that can perform both training and inference, the two tasks can generally still be performed independently.

[0176] NPUs designed to accelerate training are generally configured to accelerate the optimization of new models, which is a highly computationally intensive operation that involves inputting an existing data set (usually labeled or tagged), iterating on the data set, and then adjusting model parameters (such as weights and biases) to improve model performance. Typically, optimization based on error predictions involves passing back through the layers of the model and determining the gradient to reduce the prediction error.

[0177] NPUs designed to accelerate inference are generally configured to operate on complete models. Such NPUs can thus be configured to input new data slices and quickly process the data through an already trained model to generate model output (e.g., inference).

[0178] In some implementations, the NPU 908 is part of one or more of the CPU 902 , the GPU 904 , and / or the DSP 906 .

[0179] In some examples, wireless connectivity component 912 may include, for example, subcomponents for third generation (3G) connectivity, fourth generation (4G) connectivity (e.g., 4G LTE), fifth generation connectivity (e.g., 5G or NR), Wi-Fi connectivity, Bluetooth connectivity, and other wireless data transmission standards. Wireless connectivity component 912 is further connected to one or more antennas 914.

[0180] The processing system 900 may also include one or more sensor processing units 916 associated with any manner of sensors, one or more image signal processors (ISPs) 918 associated with any manner of image sensors, and / or a navigation processor 920, which may include satellite-based positioning system components (e.g., GPS or GLONASS) and inertial positioning system components.

[0181] The processing system 900 may also include one or more input and / or output devices 922, such as a screen, a touch-sensitive surface (including a touch-sensitive display), physical buttons, speakers, microphones, and the like.

[0182] In some examples, one or more of the processors of processing system 900 may be based on the ARM or RISC-V instruction set.

[0183] The processing system 900 also includes a memory 924, which represents one or more static and / or dynamic memories, such as dynamic random access memory, flash-based static memory, etc. In this example, the memory 924 includes computer-executable components that can be executed by one or more of the aforementioned processors of the processing system 900.

[0184] Specifically, in this example, the memory 924 includes a posterior component 924A, a priori component 924B, a likelihood component 924C, a training component 924D, and a management component 924E. The memory 924 also includes a set of model parameters 924F. The model parameters 924F may generally correspond to parameters of all or a portion of a machine learning model trained to reconstruct a sparse channel representation (e.g., for channel estimation), such as one or more parameters (e.g., step size and / or threshold) of a likelihood model for one or more layers of a neural network, one or more parameters of a posterior network for one or more layers of a neural network, one or more parameters of a priori network for one or more layers of a neural network, etc.

[0185] Although not included in the illustrated example, in some aspects, memory 924 also includes training data, which generally corresponds to the training samples or examples discussed above, such as channel observation and sensing matrix pairs (and corresponding ground truth channel states). The depicted components, as well as other components not depicted, can be configured to perform various aspects of the techniques described herein. Although not shown for conceptual clarity, Fig. 9 924A, prior component 924B, likelihood component 924C, training component 924D, and inference component 924E may be implemented jointly or separately.

[0186] Processing system 900 also includes a posteriori circuit 926, a priori circuit 927, likelihood circuit 928, training circuit 929, and management circuit 930. The depicted circuits, as well as other circuits not depicted, may be configured to perform various aspects of the techniques described herein.

[0187] For example, the posterior component 924A and the posterior circuit 926 may be used to generate variational posterior distributions using the trained neural network and to sample the sparsified dictionary from these distributions, as discussed above. The prior component 924B and the prior circuit 927 may be used to generate the prior distribution using the trained neural network, as discussed above. The likelihood component 924C and the likelihood circuit 928 may use a likelihood model (e.g., an A-DLISTA model) to generate a sparse channel representation based on the generated sparsified dictionary, as discussed above. The training component 924D and the training circuit 929 may correspond to Figure 3 The training component 345 of the network can be used to calculate the loss and / or refine the machine learning model, as discussed above. The management component 924E and the management circuit 930 can be used to coordinate the entire workflow (e.g., training and / or inference) using the above components, and / or drive network reconfiguration based on the channel estimate (generated based on the sparse channel representation), as discussed above.

[0188] Although for the sake of clarity Fig. 9 Although depicted as separate components and circuits, a posteriori circuit 926, a posteriori circuit 927, likelihood circuit 928, training circuit 929, and management circuit 930 may be implemented collectively or individually in other processing devices of processing system 900, such as in CPU 902, GPU 904, DSP 906, NPU 908, etc.

[0189] Generally, the processing system 900 and / or its components may be configured to perform the methods described herein.

[0190] It is worth noting that in other aspects, aspects of the processing system 900 may be omitted, such as where the processing system 900 is a server computer, etc. For example, in other aspects, the multimedia processing unit 910, the wireless connectivity component 912, the sensor processing unit 916, the ISP 918, and / or the navigation processor 920 may be omitted. Furthermore, aspects of the processing system 900 may be distributed among multiple devices.

[0191] Sample Clauses

[0192] Specific implementation examples are described in the following numbered clauses:

[0193] Item 1: A processor-implemented method for performing channel estimation of a digital communication channel using a machine learning model, comprising: generating a current sparsification dictionary in a first iteration of the machine learning model by processing a sensing matrix and current channel observations for the digital communication channel using a posteriori neural network; and generating a current sparse channel representation in the first iteration by processing the current sparsification dictionary, the sensing matrix, and the current channel observations using a likelihood neural network.

[0194] Clause 2: A method according to clause 1, wherein using the posterior neural network to generate the current sparsification dictionary includes: generating mean and variance values ​​by processing the sensing matrix, the current channel observations, and previous sparse channel representations generated in previous iterations of the machine learning model using the posterior neural network; and sampling a distribution having the mean and the variance values.

[0195] Clause 3: A method according to any one of clauses 1 to 2, wherein the distribution is defined as The posterior distribution of , where: t is the current sparsified dictionary, is the previous sparse channel representation generated in a previous iteration of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0196] Clause 4: A method according to any one of clauses 1 to 3, wherein the posterior distribution is equal to Where: μ is the mean, f φ is the posterior neural network, and β -1 is the precision value.

[0197] Clause 5: A method according to any one of clauses 1 to 4, wherein using the likelihood neural network to generate the current sparse channel representation includes: generating a mean by processing the current sparsification dictionary, the sensing matrix and the current channel observation using the likelihood neural network; and generating a distribution having the mean and variance values.

[0198] Clause 6: A method according to any one of clauses 1 to 5, wherein the distribution is defined as The likelihood model of , where: is the current sparse channel representation, x gt is the ground truth sparse channel representation, Ψ 1:t is the sparsified dictionary generated in one or more previous iterations of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0199] Clause 7: A method according to any one of clauses 1 to 6, wherein the likelihood model is equal to Where: μ is the mean, ADLISTA(·) represents the application of the likelihood neural network, Θ is the training parameter of the likelihood neural network, and σ 2 is the variance value, where the variance value is a hyperparameter.

[0200] Clause 8: The method according to any one of clauses 1 to 7 further includes: using the likelihood neural network to generate an uncertainty measure; determining that the uncertainty measure satisfies one or more defined criteria; and in response to determining that the uncertainty measure satisfies the one or more defined criteria, initiating retraining of the posterior neural network and the likelihood neural network.

[0201] Clause 9: The method according to any one of clauses 1 to 8, further comprising: determining entropy based on the likelihood neural network; determining that the entropy satisfies one or more defined criteria; and in response to determining that the entropy satisfies the one or more defined criteria: suppressing the use of subsequent iterations of the machine learning model to process the sensing matrix and the current channel observation; generating a current channel estimate based on the current sparse channel representation; and outputting the current channel estimate.

[0202] Clause 10: A method according to any one of clauses 1 to 9, wherein the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

[0203] Clause 11: A method according to any one of clauses 1 to 10, wherein each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

[0204] Clause 12: The method of any one of clauses 1 to 11, wherein the digital communication channel comprises a millimeter wave digital communication channel.

[0205] Clause 13: The method of any one of clauses 1 to 12, further comprising: performing one of analog beamforming, beam selection, or spectral efficiency prediction based on the current sparse channel representation.

[0206] Clause 14: The method according to any one of clauses 1 to 13, further comprising: generating a first loss by processing the current sparsification dictionary using a priori neural network; generating a second loss based on the current sparse channel representation; and refining the posterior neural network, the likelihood neural network and the prior neural network based on the first loss and the second loss.

[0207] Clause 15: A method according to any one of clauses 1 to 14, wherein using the prior neural network to generate the first loss includes: generating a mean and a variance value by processing a previously sparsified dictionary generated in a previous iteration of the machine learning model using the prior neural network; and generating a distribution having the mean and the variance value.

[0208] Clause 16: A method according to any one of clauses 1 to 15, wherein the distribution is defined as p θ (Ψ t |Ψ t-1 ), where: t is the current sparsified dictionary, and Ψ t-1 is the previously sparsified dictionary generated in the previous iteration of the machine learning model.

[0209] Clause 17: A method according to any one of clauses 1 to 16, wherein the prior distribution is equal to σ 2 =f θ (Ψ t-1 ), where: μ is the mean, f θ is the prior neural network, and σ 2 is the variance value.

[0210] Clause 18: A method according to any one of clauses 1 to 17, wherein: the previous iteration corresponds to an input iteration of the machine learning model, and the previous sparsified dictionary is derived from a Gaussian distribution Sampled in.

[0211] Clause 19: A method according to any one of clauses 1 to 18, wherein the first loss and the second loss are collectively defined as

[0212] in: is the expected value of the sparsified dictionary evaluated for up to iteration t sampling, is the sparse channel representation generated for up to iteration t-1, x gt is the ground truth sparse channel representation, y is the channel observation, Φ is the sensing matrix, and is the expected value of the sparsified dictionary evaluated for up to iteration t-1 sampling.

[0213] Item 20: A processor-implemented method for training a machine learning model for channel estimation of a digital communication channel, comprising: receiving a sensing matrix and a current channel observation for the digital communication channel; generating a current sparsification dictionary by processing the sensing matrix and the current channel observation using a posteriori neural network; generating a current sparse channel representation by processing the current sparsification dictionary, the sensing matrix, and the current channel observation using a likelihood neural network; generating a first loss by processing the current sparsification dictionary using a priori neural network; generating a second loss based on the current sparse channel representation; and refining the posteriori neural network, the likelihood neural network, and the priori neural network based on the first loss and the second loss.

[0214] Clause 21: A method according to clause 20, wherein using the prior neural network to generate the first loss includes: generating a mean and a variance value by processing a previously sparsified dictionary generated in a previous iteration of the machine learning model using the prior neural network; and generating a distribution having the mean and the variance value.

[0215] Clause 22: A method according to any one of clauses 20 to 21, wherein the distribution is defined as p θ (Ψ t |Ψ t-1 ), where: t is the current sparsified dictionary, and Ψ t-1 is the previously sparsified dictionary generated in the previous iteration of the machine learning model.

[0216] Clause 23: A method according to any one of clauses 20 to 22, wherein the prior distribution is equal to σ 2 =f θ (Ψ t-1 ), where: μ is the mean, f θ is the prior neural network, and σ 2 is the variance value.

[0217] Clause 24: A method according to any one of clauses 20 to 23, wherein: the previous iteration corresponds to an input iteration of the machine learning model, and the previous sparsified dictionary is derived from a Gaussian distribution Sampled in.

[0218] Clause 25: A method according to any one of clauses 20 to 24, wherein the first loss and the second loss are collectively defined as

[0219] in: is the expected value of the sparsified dictionary evaluated for up to iteration t sampling, is the sparse channel representation generated for up to iteration t-1, x gt is the ground truth sparse channel representation, y is the channel observation, Φ is the sensing matrix, and is the expected value of the sparsified dictionary evaluated for up to iteration t-1 sampling.

[0220] Clause 26: A method according to any one of clauses 20 to 25, wherein generating the current sparsification dictionary using the machine learning model including the posteriori neural network comprises: generating a mean and a precision value by processing the sensing matrix, the current channel observations and a previous sparse channel representation generated in a previous iteration of the machine learning model using the posteriori neural network; and sampling a distribution having the mean and the variance values.

[0221] Clause 27: A method according to any one of clauses 20 to 26, wherein the distribution is defined as The posterior distribution of , where: t is the current sparsified dictionary, is the previous sparse channel representation generated in a previous iteration of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0222] Clause 28: A method according to any one of clauses 20 to 27, wherein the posterior distribution is equal to Where: μ is the mean, f φ is the posterior neural network, and β -1 is the precision value.

[0223] Clause 29: A method according to any one of clauses 20 to 28, wherein using the likelihood neural network to generate the current sparse channel representation includes: generating a mean by processing the current sparsification dictionary, the sensing matrix and the current channel observation using the likelihood neural network; and generating a distribution having the mean and variance values.

[0224] Clause 30: A method according to any one of clauses 20 to 29, wherein the distribution is defined as The likelihood model of , where: is the current sparse channel representation, x gt is the ground truth sparse channel representation, Ψ 1:t The sparsified dictionary generated in one or more previous iterations of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

[0225] Clause 31: A method according to any one of clauses 20 to 30, wherein the likelihood model is equal to Where: μ is the mean, ADLISTA(·) represents the application of the likelihood neural network, Θ is the training parameter of the likelihood neural network, and σ 2 is the variance value, where the variance value is a hyperparameter.

[0226] Clause 32: A method according to any one of clauses 20 to 31, wherein the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

[0227] Clause 33: A method according to any one of clauses 20 to 32, wherein each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

[0228] Clause 34: The method of any one of clauses 20 to 33, wherein the digital communications channel comprises a millimeter wave digital communications channel.

[0229] Clause 35: A processing system comprising: a memory including computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to perform a method according to any one of clauses 1 to 34.

[0230] Clause 36: A processing system comprising: means for performing the method according to any one of clauses 1 to 34.

[0231] Clause 37: A non-transitory computer readable medium comprising computer executable instructions which, when executed by one or more processors of a processing system, cause the processing system to perform the method of any of clauses 1 to 34.

[0232] Clause 38: A computer program product embodied on a computer readable storage medium, the computer program product comprising code for performing the method according to any one of clauses 1 to 34.

[0233] Additional considerations

[0234] The foregoing description is provided to enable any person skilled in the art to practice the various aspects described herein. The examples discussed herein do not limit the scope, applicability or aspects set forth in the claims. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects. For example, without departing from the scope of the present disclosure, the functions and arrangements of the elements discussed may be changed. Each example may ignore, replace or add various procedures or components as appropriate. For example, the described method may be performed in a different order than described, and various steps may be added, omitted or combined. In addition, the features described for some examples may be combined in some other examples. For example, any number of aspects set forth herein may be used to implement a device or practice method. In addition, the scope of the present disclosure is intended to cover such devices or methods practiced using other structures, functions, or structures and functions that are supplementary or alternative to the various aspects of the present disclosure set forth herein. It should be understood that any aspect of the disclosure disclosed herein may be embodied by one or more elements of the claims.

[0235] As used herein, the word “exemplary” means “serving as an example, instance, or illustration.” Any aspect described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other aspects.

[0236] As used herein, a phrase referring to "at least one of" a list of items refers to any combination of those items (including single members). As an example, "at least one of a, b, or c" is intended to cover a, b, c, ab, ac, bc, and abc, as well as any combination with multiple identical elements (e.g., aa, aaa, aab, aac, abb, acc, bb, bbb, bbc, cc, and ccc, or any other ordering of a, b, and c).

[0237] As used herein, the term "determining" encompasses a wide variety of actions. For example, "determining" may include calculating, computing, processing, deriving, investigating, searching (e.g., searching in a table, database, or other data structure), ascertaining, and the like. Furthermore, "determining" may include receiving (e.g., receiving information), accessing (e.g., accessing data in a memory), etc. Furthermore, "determining" may include resolving, selecting, choosing, establishing, etc.

[0238] The method disclosed herein includes one or more steps or actions for implementing the method. The steps and / or actions of the method can be interchangeable with each other without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of specific steps and / or actions can be modified without departing from the scope of the claims. In addition, the various operations of the method described above can be performed by any appropriate component capable of performing the corresponding function. The component may include various hardware and / or software components and / or modules, including but not limited to circuits, application specific integrated circuits (ASICs) or processors. Typically, in the case of operations illustrated in the accompanying drawings, those operations may have corresponding corresponding components with similar numbers plus functional components.

[0239] The following claims are not intended to be limited to the various aspects shown herein, but should be granted the full scope consistent with the language of the claims. Within the claims, unless otherwise specified, reference to an element in the singular form is not intended to mean "one and only one", but "one or more". Unless otherwise specified, the term "some" refers to one or more. No claim element should be interpreted according to the provisions of 35 U.S.C. § 112 (f), unless the phrase "parts for..." is used to expressly state the element, or in the case of a method claim, the phrase "step for..." is used to state the element. All structural and functional equivalents of the elements of the various aspects described throughout the present disclosure that are known or will be known later to a person of ordinary skill in the art are expressly incorporated herein by reference and are intended to be covered by the claims. In addition, regardless of whether such disclosure is explicitly stated in the claims, anything disclosed herein is not intended to be disclosed to the public.

Claims

1. A processor-implemented method, comprising: generating a current sparsified dictionary by processing a sensing matrix and a current channel observation for a digital communication channel using an a posteriori neural network in a first iteration of the machine learning model; as well as A current sparse channel representation is generated in the first iteration by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using a likelihood neural network.

2. The processor-implemented method of claim 1 , wherein using the a posteriori neural network to generate the current sparsified dictionary comprises: generating mean and variance values ​​by processing the sensing matrix, the current channel observation, and a previous sparse channel representation generated in a previous iteration of the machine learning model using the posterior neural network; as well as A distribution having the mean and the variance value is sampled.

3. The processor-implemented method of claim 2, wherein the distribution is defined as The posterior distribution of , where: ψ t is the current sparsified dictionary, is the previous sparse channel representation generated in a previous iteration of the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

4. The processor-implemented method of claim 1 , wherein using the likelihood neural network to generate the current sparse channel representation comprises: generating a mean by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using the likelihood neural network; as well as Generates a distribution with the stated mean and variance values.

5. The processor-implemented method of claim 4, wherein the distribution is defined as The likelihood model of , where: is the current sparse channel representation, x gt Corresponding to the ground truth channel state, Ψ 1:t is a sparsified dictionary generated in one or more previous iterations in the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

6. The processor-implemented method of claim 1 , further comprising: generating an uncertainty measure using the likelihood neural network; determining that the uncertainty measure satisfies one or more defined criteria; as well as Responsive to determining that the uncertainty measure satisfies the one or more defined criteria, retraining of the posterior neural network and the likelihood neural network is initiated.

7. The processor-implemented method of claim 1 , further comprising: determining entropy based on the likelihood neural network; determining that the entropy satisfies one or more defined criteria; as well as In response to determining that the entropy satisfies the one or more defined criteria: refraining from using a subsequent iteration of the machine learning model to process the sensing matrix and the current channel observation; generating a current channel estimate based on the current sparse channel representation; and The current channel estimate is output.

8. The processor-implemented method of claim 1 , wherein the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

9. The processor-implemented method of claim 1 , wherein each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

10. The processor-implemented method of claim 1 , further comprising: One of analog beamforming, beam selection, or spectral efficiency prediction is performed based on the current sparse channel representation.

11. The processor-implemented method of claim 1 , further comprising: generating a first loss by processing the current sparsified dictionary using a priori neural network; generating a second loss based on the current sparse channel representation; as well as The posterior neural network, the likelihood neural network, and the prior neural network are refined based on the first loss and the second loss.

12. The processor-implemented method of claim 11 , wherein using the prior neural network to generate the first loss comprises: generating mean and variance values ​​by processing a previously sparsified dictionary generated in a previous iteration of the machine learning model using the prior neural network; as well as Generates a distribution with the mean and the variance value.

13. A processor-implemented method comprising: receiving a sensing matrix and a current channel observation for a digital communication channel; generating a current sparsified dictionary by processing the sensing matrix and the current channel observation using a machine learning model including a posteriori neural network; generating a current sparse channel representation by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using a likelihood neural network; generating a first loss by processing the current sparsified dictionary using a priori neural network; generating a second loss based on the current sparse channel representation; as well as The posterior neural network, the likelihood neural network, and the prior neural network are refined based on the first loss and the second loss.

14. The processor-implemented method of claim 13, wherein using the prior neural network to generate the first loss comprises: generating mean and variance values ​​by processing a previously sparsified dictionary generated in a previous iteration of the machine learning model using the prior neural network; as well as Generates a distribution with the mean and the variance value.

15. The processor-implemented method of claim 14, wherein the distribution is defined as p θ (Ψ t |Ψ t-1 ), where: Ψ t is the current sparsified dictionary, and Ψ t-1 is the previously sparsified dictionary generated in the previous iteration of the machine learning model.

16. The processor-implemented method of claim 15, wherein: The previous iteration corresponds to an input iteration of the machine learning model, and The previously sparse dictionary is derived from a Gaussian distribution Sampled in.

17. The processor-implemented method of claim 13, wherein generating the current sparsification dictionary using the machine learning model including the posterior neural network comprises: generating mean and variance values ​​by processing the sensing matrix, the current channel observation, and a previous sparse channel representation generated in a previous iteration of the machine learning model using the posterior neural network; as well as A distribution having the mean and the variance value is sampled.

18. The processor-implemented method of claim 13, wherein using the likelihood neural network to generate the current sparse channel representation comprises: generating a mean by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using the likelihood neural network; as well as Generates a distribution with the stated mean and variance values.

19. The processor-implemented method of claim 13, wherein the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

20. The processor-implemented method of claim 13, wherein each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

21. A processing system comprising: a memory including computer executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to perform operations comprising: generating a current sparsified dictionary by processing a sensing matrix and a current channel observation for a digital communication channel using an a posteriori neural network in a first iteration of the machine learning model; as well as A current sparse channel representation is generated in the first iteration by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using a likelihood neural network.

22. The processing system of claim 21 , wherein using the posterior neural network to generate the current sparsification dictionary comprises: generating mean and variance values ​​by processing the sensing matrix, the current channel observation, and a previous sparse channel representation generated in a previous iteration of the machine learning model using the posterior neural network; as well as Samples a distribution with the stated mean and variance values.

23. The processing system of claim 21 , wherein using the likelihood neural network to generate the current sparse channel representation comprises: generating a mean by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using the likelihood neural network; as well as Generates a distribution with the stated mean and variance values.

24. The processing system of claim 23, wherein the distribution is defined as The likelihood model of , where: is the current sparse channel representation, x gt Corresponding to the ground truth channel state, Ψ 1:t is a sparsified dictionary generated in one or more previous iterations in the machine learning model, y is the current channel observation, and Φ is the sensing matrix.

25. The processing system of claim 21, the operations further comprising: generating an uncertainty measure using the likelihood neural network; determining that the uncertainty measure satisfies one or more defined criteria; as well as Responsive to determining that the uncertainty measure satisfies the one or more defined criteria, retraining of the posterior neural network and the likelihood neural network is initiated.

26. The processing system of claim 21, the operations further comprising: determining entropy based on the likelihood neural network; determining that the entropy satisfies one or more defined criteria; as well as In response to determining that the entropy satisfies the one or more defined criteria: refraining from using a subsequent iteration of the machine learning model to process the sensing matrix and the current channel observation; generating a current channel estimate based on the current sparse channel representation, and The current channel estimate is output.

27. The processing system of claim 21, wherein the posterior neural network and the likelihood neural network are shared across each iteration of the machine learning model.

28. The processing system of claim 21, wherein each respective iteration of the machine learning model has a respective posterior neural network and a respective likelihood neural network.

29. The processing system of claim 21, the operations further comprising: One of analog beamforming, beam selection, or spectral efficiency prediction is performed based on the current sparse channel representation.

30. A system comprising: means for generating a current sparsified dictionary by processing a sensing matrix and a current channel observation for a digital communication channel using an a posteriori neural network in a first iteration of the machine learning model; and means for generating, in the first iteration, a current sparse channel representation by processing the current sparsified dictionary, the sensing matrix, and the current channel observation using a likelihood neural network.

Images