Robust model architecture for ai CSI compression

Abstract

Methods and systems are described for designing a model which is robust (scalable) to different network configuration by leveraging transformer robustness to sequence length. Certain embodiments include methods to transform both input and output dimensions dependent on sequence lengths. Certain embodiments can include a token design which has a fixed size and when network configuration changes, the sequence length changes, instead of the token. In certain embodiments, to handle the different requirements of latent space, the self–attention in the encoder design can be replaced by a cross–attention one, which allows for the encoder output sequence to have a different length, matching the desired latent space dimension.

Description

P113000W001ROBUST MODEL ARCHITECTURE FOR Al CSI COMPRESSIONCROSS REFERENCE TO RELATED INFORMATION

[0001] This application claims the benefit of United States of America priority application No. 63 / 755,881 filed on February 7, 2025, titled “Robust Model Architecture for Al CSI Compression.”TECHNICAL FIELD

[0002] The present disclosure generally relates to systems and methods for Al enhanced CSI reporting.BACKGROUND

[0003] Artificial Intelligence (Al) and Machine Learning (ML) have been investigated, both in academia and industry, as promising tools to optimize the air interface design in wireless communication networks. One example of AI / ML implementation for the physical layer (Al PHY) is using AI / ML autoencoder to improve the channel compression accuracy and / or to reduce the channel state information (CSI) feedback overhead. This topic has been discussed throughout 3GPP Rel. 18 and continues to be discussed in Rel. 19.General Aspects for AI / ML-based CSI Compression

[0004] In the legacy mechanism (i.e., non- AI / ML-based CSI compression), the UE (user equipment) can be configured to report a suggested precoder to the NW (network), a so-called precoder matrix indicator (PMI). The PMI is based on measured CSI-RS (CSI Reference Signal) and sent to the NW as a CSI report, based on a certain mechanism, the so-called codebook. The codebook generally defines how the UE arranges the reported bits based on (the number of) beams and taps selected by the UE to be reported to the NW and how the UE quantizes the precoding matrix.

[0005] In AI / ML-based CSI compression, an AI / ML-based autoencoder (AE) replaces at least part of, the legacy mechanism. AEs can have different architectures. For example, AEs can be based on dense NNs (neural networks), multi-dimensional convolution NNs, variational, recurrent NNs, transformer networks, or any combination thereof. However,P113000W001all AE architectures possess an encoder-bottleneck-decoder structure illustrated in Figure 1. Figure 1 shows an illustration of a fully connected autoencoder.

[0006] The codeword’s size (denoted by Y in Figure 1) of an AE is smaller than the input data’s size (X in Figure 1). The AE encoder thus reduces the dimensionality of the input features X down to T’s. The decoder part of the AE tries to invert the encoder and reconstruct X with minimal error, according to some predefined loss function, also known as the target function in the general optimization literature.

[0007] Figure 2 illustrates using autoencoder for CSI compression. Figure 2 illustrates how an AE might be used for AI / ML-enhanced CSI reporting in NR. The UE measures the channel in the downlink using CSI- RS. The UE estimates the channel for each subcarrier (SC) from each base station TX (transmission) antenna to each UE RX (reception) antenna. The estimate can be viewed as a three-dimensional channel matrix. The 3D channel matrix represents the MIMO (multiple input multiple output) channel estimated over several SCs and is input to the encoder. However, there are different architectures where a processed version of the 3D MIMO channel, or a processed subset of the information, e.g., singular vectors of the MIMO channel, is input to the encoder.

[0008] The AE encoder is implemented in the UE, and the AE decoder is implemented in the NW, denoted BS for base station in the figure. The output of the AE encoder is signaled from the UE to the NW over the uplink. The codeword can be viewed as a learned latent representation of the channel. Properties of the data (e.g., CSI-RS channel estimates), the channel size, uplink feedback rate, and hardware limitations of the encoder and decoder need to be considered when optimizing the AE’s architecture.

[0009] The weights and biases of an AE (with a fixed architecture) are trained to minimize the reconstruction error (the error between the input X and output X) on some training datasets. For example, the weights and biases can be trained to minimize the mean squared error, MSE (mean(X — X) ). To achieve good performance during live operation, the training data set should represent the actual data the AE will encounter during live operation.

[0010] Figure 3 shows an illustration of quantization operation at the output of the encoder to fit the CSI payload over the air interface. In the two-sided CSI compression, the output of the UE-side encoder needs to be communicated over the air interface to the gNB (gNodeB) decoder with the assigned CSI reporting payload and, therefore, needs to be quantized to a finite number of bits (e.g., 1-4 bits per encoder output’s neuron) to obtain an efficient transmission as shown in Figure 3. Accordingly, a quantization layer is connected toP113000W001the output of the encoder or directly included in the encoder. In an example, the quantization layer may implement scalar quantization which quantizes the output of each neuron of the encoder output layer (the bottleneck layer of AE) to generate bits to fit the CSI reporting payload in the uplink control information (UCI). Other quantization methods, e.g., vector quantization, may also be used.Pre-Processing for Input Data to the AE

[0011] A proper pre-processing on the input to the encoder can greatly reduce the size and complexity of designing and / or training an AI / ML model, and, in the meantime, improve the scalability and transferability of the model. In addition, pre-processing may reduce the need for multiple models depending on bandwidth variation and variation in the number of antenna ports at the gNB. By using pre-processing, instead of directly compressing the channel (i.e., with dimensions of RX x TX x SC as in Figure 2, the channels are first processed into another representation.

[0012] Figure 4 illustrates a pre-process the channel into eigenvector before use is as inputs to the encode inputs. One example of the pre-processing may include transforming the channel into eigenvectors (Figure 4). Here, the UE may conduct the following steps:• Compute the covariance matrix of the channel and extract the relevant eigenvectors;• The covariance matrix is summed over 4 “f-units”, to get 13 “subchunks” (subbands) in frequency;• For each of the 13 averaged covariance matrices, compute an eigen- decomposition and extract the 4 eigenvectors corresponding to the 4 largest eigenvalues;• Normalize the phase and magnitude of the eigenvectors;• The number of eigenvectors to feedback should be the same for all subchunks and depends on the rank hypothesis testing with a value between 1 and 4.

[0013] In another example, pre-processing may include transforming the channel into the beam-delay domain. The feature extraction for beam-delay reduced eigenvector-based feedback is illustrated in Figure 5. This is also called a “W2 compressor” as the feature extraction is similar to the standardized 3GPP procedures and the jargon for the matrix left to compress for the AI / ME is “W2”. The steps are as follows:P113000W001• The UE selects an orthogonal basis, one of a set of oversampled / rotated, spatial domain, DFT (Discrete Fourier Transform) bases. On that basis, the UE selects E vectors. These vectors represent the spatial domain (SD) basis DFT. This basis is wideband applicable, i.e., valid for all subbands, and applicable for both polarizations. This selection may happen jointly, e.g., the UE does a spatial domain DFT on the 32 x 4 (TX x RX) matrix, per RB, and selects the L strongest beams out of 16 (for one polarization). The beam-space channel is computed by multiplying the channel with the selected SD basis. The covariance of the beamspace channel is summed over, e.g., 4 RBs to produce a covariance matrix for each subband.• For each covariance matrix (per subband) the UE extracts a number of eigenvectors and may select the rank, i.e. number of layers.• The UE does a frequency domain DFT per layer, transforming to a delay domain, whereafter it selects the M strongest taps. The resulting tensor of dimensions 2L x number of layers x M is called the linear combination coefficients, in 3GPP jargon “W2”. The W2 matrix can be used to reconstruct the, by the UE suggested, precoding matrices.• The data is used as input in the AI / ML model. This could be the raw linear combination coefficients, or it could be enhanced with information about the selected beams and taps, noise levels, etc.

[0014] Figure 5 illustrates a pre-process the channel into beam delay domain (W2-like format) before use is as inputs to the encode inputs.Scalability Aspects of Transformer Models as Autoencoders for CSI Compression

[0015] It is important to evaluate how a model or model structure generalizes for different network configuration. Specifically, we consider how the model can be applicable when the number of receiver or transmitter ports, physical resource blocks as well as slots changes.

[0016] Considering that the 'Agreements from RANI #119' stablishes transformer as the baseline, possible solutions include stablishing a model structure and training different models for each configuration, define adaptation layers for different configuration and use common internal model or define a model structure that is applicable to all configurations.P113000W001

[0017] Figure 6 illustrates a model structure for encoder model. Figure 7 illustrates a model structure for the decoder model.

[0018] The model structure presented in Figure 6 is the most common presented and evaluated by different companies in 3GPP. The input to the model has some dimensions which are reshaped so that the result has two dimensions, the sequence length and the tokens. In this case the input has as dimensions the number of receiver and transmitter antenna ports, RX and TX, respectively and physical resource blocks PRB. In this case, the tokens are chosen to be the TX dimensions, which, in turns implies the sequence length (number of tokens) to be RX * PRB. The tokens are then projected into another dimension (usually higher space, to produce overparameterization), denoted by E , using the fully connected network (FCN) embedding layer. This sequence goes through positional encoding and the transformer block maintaining its shape. Finally, it is flattened to be used in an FCN to be projected in the latent space L. Figure 7 shows the decoder which mirrors such encoder, the latent space is projected back into the sequence length with token dimension as the embedding, which goes through positional encoding and transformer, and finally the tokens are projected into the TX dimension to then be reshaped into the original encoder input.

[0019] If processing steps are implemented at the UE to derive precoding matrix in the angular and delay domain as the input (i.e., W2-based CSI compression), then, similarly to the eigenvector-based CSI compression example, assume that the size of the W2 precoding matrix is [ 2 / , M], where J is the number of selected beams with J < TX and M is the number of selected FD basis with M < SB. Then, each column of this precoding matrix can be treated as a token. In this way, the feature dimension of each token is 2 / and the sequence length (the number of tokens) is M.

[0020] In such example, the model input could as well be preprocessed, for example, into spatial beam, frequency basis and layer. Similarly to the eigenvector example, the special beam can be used as the token dimension and the sequence length given by frequency basis and layer.

[0021] Such model structure design is naturally applicable for different sequence lengths, thus, for the given examples of the raw-channel-based CSI compression model design and the W2-based CSI compression model design, changing the sequence size (value of RX , PRB or M ) does not need another model, presenting examples of such configuration to the model should be enough for it to learn. However, if the TX changes, the model would not be applicable, given that the input would not be compatible with theP113000W001dimension expected by the FCN embedding layer. Similarly, if a variable payload (L) were required the model might not be applicable given the FCN latent output. The decoder is analogous and would not be applicable for different values of L (in the FCN expansion) and TX (in the FCN 'de-embedding').

[0022] Note that it is a design choice to choose which dimension will be the token in this case, and, accordingly, not be adaptable to variation. We chose to exemplify with TX, but we could have chosen RX or PRB.

[0023] Thus, a possible way to apply such model structure to the CSI compression scheme, is to have a model for each configuration. In other words, a different model is trained for each applicable TX and L. Such approach might be beneficial and allows for optimization of model parameters for each specific case. However, it comes with the cost of increased overhead of life-cycle management (LCM) for different models.Adaptation Layers

[0024] Different adaptation techniques can be applied to match the input dimension of a FCN layer, for example for FCN (embedding) or FCN (expansion). The solution would be dependent on how the FCN layer is designed, i.e., it might be designed to accommodate the largest expected dimension (biggest TX in the example), another option could be to select a dimension of the most common size, or a least common option could be to select a minimum size. FCN and CNN are examples of ML layers.

[0025] Design FCN input to be the largest expected dimension: In this case we can select the tokens to have a maximum dimension, in the example case, where tokens are the TX dimension we could say that at maximum they could be 128. Then, different strategies can be used to match the dimension of the input which is expected to be smaller than such number.• Interpolation: Interpolating the samples can make it match the input resolution. Different techniques can be applied, such as polynomial, pricewise or spline interpolation.• Zero padding: Zero padding assigns a constant value (usually zero) to the dimensions needed to reach the input dimension.• Deconvolution networks: Deconvolution networks is a type of learnable interpolation, constant values are added in between samples to reach the desired dimension, which then are passed through a CNN to process and “learn” to fill up the constant values based on the original input.P113000W001

[0026] Design FCN input to be the smallest expected dimension: Similarly to designing the expected largest input, one can design the smallest expected input, for the example, we could assume the TX to be 2 and then find strategies to reduce the unmatching dimensions.• Downsampling: Filtering techniques can be used to reduce the dimension of the input, such as a lowpass filter followed by decimation to reach the desired dimension.• Pooling: Pooling is a strategy where a window is defined in terms of number of samples and an aggregation function is used to reduce its dimension. For example, a window of 4 samples with a max aggregation function would reduce the dimension by for by choosing the maximum of every set of 4 elements of a sequence.

[0027] Design FCN input to be a representative dimension: Choosing a representative dimension means that the input might be smaller or larger than the expected input. Thus, we can design a single FCN but depending on the input apply one of the methods presented in Sections

[0025] and

[0026] to either increase or decrease the dimension of the input.

[0028] Design an FCN for each expected dimension: A further design option that has been considered is where some layers are dependent on the dimension of the input, and only one of them is activated for each input. Figure shows such architecture. In the case shown, the TX has dimension 4, thus, the appropriate FCN embedding is selected. Similarly, for the output layer, depending on the desired latent space dimension (configured CSI payload size), the appropriate layer can be selected, in the example case, L = 8. Figure 8 illustrates a model structure with FCN for each expected dimension.

[0029] Puncture the output of the FCN: Another option for the output layer is to design it to produce the largest expected output dimension and when a smaller dimension is desired, some elements are dropped, either from specific locations, at random or following some specific pattern as proposed by dimension reducing methods described above regarding “Design FCN input to be the smallest expected dimension.”

[0030] There currently exist certain challenges. The tradeoff between complexity and performance gain with the introduction of Al for CSI compression has been a point for discussion within the community, with the argument that for Al to be introduced complexity must be reduced to justify the gains.P113000W001SUMMARY

[0031] One possible embodiment under the present disclosure is a method performed by a UE for training a ML model for autoencoding CSI reporting. The method includes: performing reshape of a dataset used by the ML model into a sequence of one or more atomic units of CSI; and projecting the one or more atomic units into an embedding space using a ML layer, wherein the ML model is scalable to multiple input and multiple output dimensions.

[0032] Another possible embodiment under the present disclosure is a method performed by a UE for training a ML model for autoencoding CSI reporting. The method can comprise performing a transformer encoder block, with input from a positional encoding and an output sequence initialization, wherein cross attention is used instead of self-attention; and performing a projection of an output sequence of the transformer encoder block into a latent space using a ML layer.

[0033] Another possible embodiment under the present disclosure is a method performed by a network node for decoding one or more signals from a UE. The method includes: performing expansion of a dataset from a latent space; and performing positional encoding, wherein the one or more signals were autoencoded using a ML model for CSI reporting.

[0034] Another possible embodiment under the present disclosure is s a method performed by a network node for decoding one or more signals from a UE. The method includes: performing de-embedding of one or more atomic units from an embedding space using a ML layer; and performing reshape to compose an output of one or more atomic units of CSI, wherein the one or more signals were autoencoded using a ML model for CSI reporting.

[0035] Another possible embodiment under the present disclosure is a UE for training a ML model for autoencoding CSI reporting. The UE can comprise: processing circuitry; and a memory storing instructions whereby the processing circuitry is operable to perform the steps of: performing reshape of a dataset used by the ML model into a sequence of one or more atomic units of CSI; and projecting the one or more atomic units into an embedding space using a ML layer, wherein the ML model is scalable to multiple input and multiple output dimensions.

[0036] Another possible embodiment under the present disclosure is a UE for training a ML model for autoencoding CSI reporting. The UE can comprise: processingP113000W001circuitry; and a memory storing instructions whereby the processing circuitry is operable to perform the steps of: performing a transformer encoder block, with input from a positional encoding and an output sequence initialization, wherein cross attention is used instead of selfattention; and performing a projection of an output sequence of the transformer encoder block into a latent space using a ML layer.

[0037] Another possible embodiment under the present disclosure is a network node for decoding one or more signals from a UE. The network node can comprise: processing circuitry; and a memory storing instructions whereby the processing circuitry is operable to perform the steps of: performing expansion of a dataset from a latent space; and performing positional encoding, wherein the one or more signals were autoencoded using a ML model for CSI reporting.

[0038] Another possible embodiment under the present disclosure is a network node for decoding one or more signals from a UE. The network node can comprise: processing circuitry; and a memory storing instructions whereby the processing circuitry is operable to perform the steps of: performing de-embedding of one or more atomic units from an embedding space using a ML layer; and performing reshape to compose an output of one or more atomic units of CSI, wherein the one or more signals were autoencoded using a ML model for CSI reporting.

[0039] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an indication of the scope of the claimed subject matter.BRIEF DESCRIPTION OF THE DRAWINGS

[0040] For a more complete understanding of the present disclosure, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, in which:

[0041] Fig. 1 illustrates an example of an autoencoder;

[0042] Fig. 2 illustrates an example of CSI compression with an autoencoder;

[0043] Fig. 3 illustrates an example of a quantization operation;

[0044] Fig. 4 illustrates an example of pre-processing a channel into eigenvectors;P113000W001

[0045] Fig. 5 illustrates an example of pre-processing a channel into beamdelay domain;

[0046] Fig. 6 illustrates an example of a model structure for an encoder;

[0047] Fig. 7 illustrates an example of a model structure for an decoder;

[0048] Fig. 8 illustrates an example of a model structure with FCN;

[0049] Fig. 9 illustrates an example of a model structure for an encoder under the present disclosure;

[0050] Fig. 10 illustrates an example of a model structure for an decoder under the present disclosure;

[0051] Fig. 11 illustrates a flow-chart of a method embodiment under the present disclosure;

[0052] Fig. 12 illustrates a flow-chart of a method embodiment under the present disclosure;

[0053] Fig. 13 illustrates a flow-chart of a method embodiment under the present disclosure;

[0054] Fig. 14 illustrates a flow-chart of a method embodiment under the present disclosure;

[0055] Fig. 15 shows a schematic of a communication system embodiment under the present disclosure;

[0056] Fig. 16 shows a schematic of a communication system embodiment under the present disclosure;

[0057] Fig. 17 shows a schematic of a user equipment or wireless device embodiment under the present disclosure;

[0058] Fig. 18 shows a schematic of a network node embodiment under the present disclosure; and

[0059] Fig. 19 shows a schematic of a virtualization environment embodiment under the present disclosure.DETAILED DESCRIPTION

[0060] Before describing various embodiments of the present disclosure in detail, it is to be understood that this disclosure is not limited to the parameters of the particularly exemplified systems, methods, apparatus, products, processes, and / or kits, which may, of course, vary. Thus, while certain embodiments of the present disclosure will beP113000W001described in detail, with reference to specific configurations, parameters, components, elements, etc., the descriptions are illustrative and are not to be construed as limiting the scope of the claimed embodiments. In addition, the terminology used herein is for the purpose of describing the embodiments and is not necessarily intended to limit the scope of the claimed embodiments. Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings. Embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art.

[0061] As described above, there currently exist certain challenges. The tradeoff between complexity and performance gain with the introduction of Al for CSI compression has been a point for discussion within the community, with the argument that for Al to be introduced complexity must be reduced to justify the gains. Thus, the proposed architectures have several inefficiencies that increase complexity, namely:• They suffer from sample inefficiency for training because they are permutation variant.• The solutions that create a model for each case or specialize parts for each dimension:i. Suffer from sample inefficiency, since the samples of one dimension are not used in the other model; and / orii. Create complexity for LCM, i.e., loading and unloading multiple models.• The presented solutions rely on designing a model with the maximum expected dimension:i. Are not future proof design; and / orii. The models might be too large when the input dimension is small. • The ones designed with the smallest or representative dimension might not be able to capture the higher dimension information.

[0062] Certain aspects of the disclosure and their embodiments may provide solutions to these or other challenges. The presented solution offers a strategy to design a model which is robust (scalable) to different network configuration by leveraging transformer robustness to sequence length. Certain embodiments include methods to transform both input and output dimensions dependent on sequence lengths. Certain embodiments can include a token design which has a fixed size and when network configuration changes, the sequence length changes, instead of the token. Transformers are naturally applicable to differentP113000W001sequence lengths. In certain embodiments, to handle the different requirements of latent space, the self-attention in the encoder design can be replaced by a cross-attention one, which allows for the encoder output sequence to have a different length, matching the desired latent space dimension. Certain embodiments include an initializer for latent space sequence, which is needed by the cross-attention mechanism.

[0063] Certain embodiments may provide one or more of the following technical advantages. Certain embodiments can improve sample efficiency because samples are used to train a single model. In certain embodiments, the complexity of the model is proportional to the input dimension, thus, not expending resources for smaller samples. In certain embodiments, improved LCM given a single model can be used. Certain embodiments can provide future-proof design given that no limitation is applied to the maximum dimensions.

[0064] With some adaptations to the standard transformer encoder presented in Figure 6, a model can be designed that to a certain extension is applicable to changes in dimensions of RX, TX, PRB and L. Such architecture is presented in Figure 8. One aspect is that this architecture assumes the definition of a token as an atomic unit of CSI, which can be repeated in a sequence of arbitrary length, this makes it robust to any dimension.

[0065] In the example, it is proposed to define the atomic unit as 2 TX and 4 PRB. Given that the tokens now have a fixed dimension (8), the embedding layer can be applied. The robustness to latent space dimension is achieved by replacing the transformer encoder self-attention mechanism to a cross-attention one. The difference is that the crossattention can handle sequence of different lengths, and we can take advantage of such fact to generate an output sequence of arbitrary length, L, in this case. The caveat is that the output sequence must be initialized, thus, an additional output sequence initialization is introduced. As an example, it can be initialized to all zero values. The transformer encoder block receives as input both the input and output sequence. As multiple layers of the transformer operate the output sequence is updated. Finally, the output of the transformer, which have a fixed size embedding dimension can be projected into a one-dimension value giving us the latent space representation.

[0066] The initialization of the output sequence can also use other strategies, for example: using the input sequence; applying some downsampling technique; initializing with random values; or use learnable parameters for the output token.

[0067] The definition of an atomic unit means that CSR-RS which contains less than that amount is not supported. The choice of this atomic unit is a design choice, andP113000W001although it causes a limitation, there is an argument that the atomic unit is so small that for reports smaller than it, Al might not be advantageous, and other strategies could be applied. However, we still propose an alternative adaptation to support it as described below regarding adaptation layers for token embedding.

[0068] In one embodiment, the token may be defined in an atomic unit, where the atomic unit may comprise one or more parameters related to CSI such as TX, RX, PRB, Layer, Subband, Spatial domain, etc. In an example of embodiment, the token may be built by splitting the input according to a 2-d window composed by a X number of samples of a first parameter and a Y number of samples of a second parameter. In the example in Figure 8, the atomic unit is defined using the first parameter TX and the second PRB while assigning X=2 and Y=4.

[0069] Figure 9 illustrates one possible embodiment of a model structure for an encoder 1200 robust to multiple input and output dimensions. Step 1210 can comprise reshaping. Step 1220 comprises FCN (embedding). Step 1230 can comprise positional encoding. The output of 1230 can be used at 1235 for output sequence initialization. The output of 1230 and 1235 can be input at 1240 for a transformer encoder block, which may utilize cross attention 1245. The output of 1240 can be used at 1250 for FCN (latent).

[0070] Different enhancements can be made to such architecture. For example, before the reshape functionality a convolution neural network (CNN) could be applied, so that information among tokens is further leveraged, even a 2d or 3d CNN can be considered. The reshape component 1210 could also be improved by adding a special token <new_rx> to be used when concatenating RX dimensions into the flattening, this could ease the learning for the transformer. A further option is to use a 2 or 3-d positional encoding to better capture the sequences, or yet, a positional encoding (1230) dependent on the antenna array design. Additionally, a self-attention mechanism could be applied after the positional encoding to allow for the input tokens to share information.

[0071] Analogous to such encoder architecture of Figure 9, a decoder can be designed. Figure 10 shows one such example of a possible embodiment of a decoder 1400. Figure 10 illustrates a model structure for decoder 1400 robust to multiple input and output dimensions. At 1410, the latent space is expanded back into an embedded dimension, such as by FCN. At 1420, positional encoding can be applied. At 1425, the output sequence is initialized to match the encoder’s sequence length. As with the sequence initialization of the encoder, multiple strategies can be used, among them, the ones presented elsewhere herein.P113000W001The output of positional encoding 1420 and output sequence initialization 1425 can be input to the transformer decoder 1430. The standard transformer decoder 1430 can be applied and then at 1440, FCN layer 1440 can be used to project the embedding space into the feature dimension of the tokens. The output is then reshaped at 1450 into the original encoder model input size.

[0072] Using such design allows for usage of a single model that is robust to changes in the different radio network configurations. It improves sample efficiency as a single model is trained and embeddings are reused. Additionally, having a single model simplified the LCM, while allowing for hardware optimization to leverage the scale of a utilizing a single model.

[0073] The architecture is also applicable to other use cases, such as the one where an additional slot dimension is added to report CSI for different time instances (prediction). Such dimension can also be flattened in the sequence, concatenating each report, possibly with a special token <new_slot> to mark each slot, analogous to the example for the <new_rx>.

[0074] Such architecture is also applicable to preprocessing cases, the dimensions would just be, for example, beam, frequency domain basis and layer. Where an atomic unit can be used to define a token and the same architecture applied.

[0075] Another advantage is that the encoder input scalability defining the tokens by the atomic unit can be adopted independently of the association of the latent space with the sequence length and vice versa. In other words, the parts of this architecture can be applied independently for encoder input scalability of latent space scalability.Adaptation layer for token embedding

[0076] If it is desirable to still handle tokens smaller than the atomic unit, strategies presented above could be applied to the token before the embedding, allowing the model to still be used. Such strategies can avoid dependency on the latent space or assumption about the maximum supported dimension of the input.Additional Embodiments

[0077] Another possible embodiment under the present disclosure is shown in Figure 11. Method 1500 comprises a method performed by a UE for training a ML model for autoencoding CSI reporting. Step 1510 can comprise performing reshape of a dataset used by the ML model into a sequence of one or more atomic units of CSI. Step 1520 can compriseP113000W001projecting the one or more atomic units into an embedding space using a ML layer, wherein the ML model is scalable to multiple input and multiple output dimensions. In one example, an ML layer is a fully connected network (FCN). In another example an ML layer is a convolutional neural network (CNN).

[0078] Method 1500 can comprise a variety of additional, optional, asynchronous, parallel, and / or alternative steps. For example, some variations can further comprise performing positional encoding of the one or more atomic units of CSI. Some variations can further comprise performing a transformer encoder block, with inputs from the positional encoding and from an output sequence initialization, wherein cross attention is used instead of self-attention; and performing a projection of an output sequence of the transformer encoder block into a latent space using a respective ML layer. In some embodiments, the projection of the output sequence of the transformer encoder block into the latent space is performed per token. In some variations, the projection of the output sequence of the transformer encoder block into the latent space is performed using a FCN. In some embodiments, the positional encoding leverages information of an antenna configuration of the UE. In some embodiment, the one or more atomic units of CSI represents one or more of: reception antenna; transmission antenna; PRB; beams; layer; sub-band; or spatial domain. In some variations, input and / or output of the ML model comprises one or more tokens, each token defined as an atomic unit of CSI that can be repeated in a sequence of scalable length. In some variations, the ML model is configured to initialize an output sequence based on at least one of: the sequence of one or more atomic units of CSI; one or more downsampling techniques; one or more random values; or one or more learnable parameters for an output token. Some embodiments can further comprise applying a CNN before performing the reshape. Some variations can further comprise applying a self-attention mechanism after the positional encoding to allow for one or more input tokens to share information. In some embodiments, projecting the one or more atomic units into the embedding space is performed using FCN embedding.

[0079] Another possible embodiment under the present disclosure is shown in Figure 12. Method 1700 comprises a method performed by a UE for training a ML model for autoencoding CSI reporting. Step 1710 can comprise performing a transformer encoder block, with input from a positional encoding and an output sequence initialization, wherein cross attention is used instead of self-attention. Step 1720 can comprise performing a projection of an output sequence of the transformer encoder block into a latent space using a ML layer.P113000W001

[0080] Method 1700 can comprise a variety of additional, optional, asynchronous, parallel, and / or alternative steps. For example, in some variations, performing a projection is performed using a FCN. In some embodiments, the positional encoding leverages information of an antenna configuration of the UE. In some variations, the one or more atomic units of CSI represents one or more of: reception antenna; transmission antenna; physical resource block, PRB; beams; layer; sub-band; spatial domain. In some embodiments, the ML model defines a token as an atomic unit of CSI which can be repeated in a sequence of arbitrary length. In some embodiments, the ML model is configured to initialize an output sequence based on at least one of: the input sequence; one or more downsampling techniques; one or more random values; one or more learnable parameters for an output token. Some variations can further comprise applying a CNN before performing reshape. In some variations the method can further comprise applying a self-attention mechanism after the positional encoding to allow for one or more input tokens to share information. In some embodiments, projecting the one or more atomic units into the embedding space is performed using FCN embedding.

[0081] Another possible embodiment under the present disclosure is shown in Figure 13. Method 1900 comprises a method performed by a network node for decoding one or more signals from a UE. Step 1910 can comprise performing expansion of a dataset from a latent space. Step 1920 can comprise performing positional encoding, wherein the one or more signals were autoencoded using a ML model for CSI reporting.

[0082] Method 1900 can comprise a variety of additional, optional, asynchronous, parallel, and / or alternative steps. For example, some variations can comprise performing output sequence initialization; and performing transformer decoding, using both positional encoding and output sequence initialization as inputs. Some embodiments can further comprise performing de-embedding of the one or more atomic units from an embedding space using a ML layer; and performing reshape to compose an output of one or more atomic units of CSI. In some variations, an additional slot dimension is added to an output dataset to report CSI for different time instances. In some embodiments, performing expansion comprises a FCN expansion. In some variations, input and / or output of the ML model comprises one or more tokens, each token defined as an atomic unit of CSI.

[0083] Another possible embodiment under the present disclosure is shown in Figure 14. Method 2100 comprises a method performed by a network node for decoding one or more signals from a UE. Step 2110 can comprise performing de-embedding of one or more atomic units from an embedding space using a ML layer. Step 2120 can comprise performingP113000W001reshape to compose an output of one or more atomic units of CSI, wherein the one or more signals were autoencoded using a ML model for CSI reporting.

[0084] Method 2100 can comprise a variety of additional, optional, asynchronous, parallel, and / or alternative steps. For example, in some embodiments, performing de-embedding is performed using a FCN. In some variations, an additional slot dimension is added to an output dataset to report CSI for different time instances. In some embodiments, performing expansion comprises a FCN expansion. In some variations, the ML model defines a token as an atomic unit of CSI.

[0085] Figure 15 shows an example of a communication system 4100 in accordance with some embodiments. In the example, the communication system 4100 includes a telecommunications network 4102 that includes an access network 4104, such as a radio access network (RAN), and a core network 4106, which includes one or more core network nodes 4108. The access network 4104 includes one or more access network nodes or base stations of various types, access network nodes 4110A and 4110B are depicted (which may be collectively referred to as network nodes 4110), or any other similar 3rdGeneration Partnership Project (3GPP) access nodes or non-3GPP access points (APs). Some embodiments of the access network 4104 may include more than one access network technology. The network nodes 4110 of access network 4104 facilitate direct or indirect connection of wireless devices, also referred to as user equipments (UEs), such as by connecting UEs 4112A, 4112B, 4112C, and 4112D (one or more of which may be generally referred to as UEs 4112) to the core network 4106 over one or more wireless connections.

[0086] Moreover, a network node is not necessarily limited to an implementation in which a radio portion and a baseband portion are supplied and integrated by a single vendor. Thus, it will be understood that network nodes include disaggregated implementations or portions thereof. For example, in some embodiments, the telecommunications network 4102 includes one or more Open-RAN (ORAN) network nodes. An ORAN network node is a network node in the telecommunications network 4102 that supports an ORAN specification (e.g., a specification published by the O-RAN Alliance, or any similar organization) and may operate alone or together with other network nodes to implement one or more functionalities of any network node in the telecommunications network 4102, including one or more access network nodes 4110 and / or core network nodes 4108.

[0087] Examples of an ORAN network node include an open radio unit (O-RU), an open distributed unit (O-DU), an open central unit (O-CU), including an O-CU controlP113000W001plane (O-CU-CP) or an O-CU user plane (O-CU-UP), a RAN intelligent controller (near-real time or non-real time) hosting software or software plug-ins, such as a near-real time control application (e.g., xApp) or a non-real time control application (e.g., rApp), or any combination thereof (the adjective “open” designating support of an ORAN specification). An ORAN network node may support a specification by, for example, supporting an interface defined by the ORAN specification, such as an Al, Fl, Wl, El, E2, X2, Xn interface, an open fronthaul user plane interface, or an open fronthaul management plane interface. Moreover, an ORAN network node may be a logical node in a physical node. Furthermore, an ORAN network node may be implemented in a virtualization environment (described further below) in which one or more network functions are virtualized. For example, the virtualization environment may include an O-Cloud computing platform orchestrated by a Service Management and Orchestration Framework via an 0-2 interface defined by the 0-RAN Alliance or comparable technologies.

[0088] The network nodes 4110 facilitate direct or indirect connection of one or more UEs 4112 to the core network 4106 over one or more wireless connections. Example wireless communications over a wireless connection include transmitting and / or receiving wireless signals using electromagnetic waves, radio waves, infrared waves, and / or other types of signals suitable for conveying information without the use of wires, cables, or other material conductors. Moreover, in different embodiments, the communication system 4100 may include any number of wired or wireless networks, network nodes, UEs, and / or any other components or systems that may facilitate or participate in the communication of data and / or signals whether via wired or wireless connections. The communication system 4100 may include and / or interface with any type of communication, telecommunication, data, cellular, radio network, and / or other similar type of system.

[0089] The UEs 4112 may be any of a wide variety of communication devices, including wireless devices arranged, configured, and / or operable to communicate wirelessly with the network nodes 4110 and other communication devices. Similarly, the network nodes 4108, 4110 are arranged, capable, configured, and / or operable to communicate directly or indirectly (e.g., via other devices of telecommunications network 4102) with the UEs 4112 and / or with other network nodes or equipment in the telecommunications network 4102 to enable and / or provide network access, such as wireless network access, and / or to perform other functions, such as administration in the telecommunications network 4102. More specifically, UEs 4112 may send messages, data, and / or other signals to network nodes 4108, 4110 or otherP113000W001elements of the telecommunications network 4102 by transmitting such signals to the relevant device directly without the signals passing through any intervening devices or by transmitting such signals to the relevant device indirectly through an intervening device (or multiple intervening devices) that then transmit the signal to the relevant device. Similarly, network nodes 4108, 4110 may send messages, data, and other signals to UEs 41122, other network nodes 4108, 4110, and other devices in telecommunications network 4102 directly or indirectly. As one specific example, a core network node 108 may transmit a particular message to a UE 4112 by transmitting the message to an access network node 4110 that will then transmit the message to the intended UE 4112. Similarly, a core network node 108 may receive a particular message from a UE 4112 by receiving the message from an access network node 4110 that itself received the message from the UE 4112.

[0090] In the depicted example, the core network 4106 connects elements of the access network 4104 (e.g., one or more of the network nodes 4110) to one or more host computing systems, such as host 4116. These connections may be direct or indirect via one or more intermediary networks or devices. In other examples, network nodes may be directly coupled to hosts. The core network 4106 includes one or more core network nodes (e.g., core network node 4108) of various types, one or more of which may be generally referred to as network nodes 4108. Network nodes 4108 are structured with hardware and software components. Features of these components may be substantially similar to those described with respect to the UEs, access network nodes, and / or hosts, such that the descriptions thereof are generally applicable to the corresponding components of the core network node 4108. Example core network nodes provide functions of one or more of a Mobile Switching Center (MSC), Mobility Management Entity (MME), Home Subscriber Server (HSS), Access and Mobility Management Function (AMF), Session Management Function (SMF), Authentication Server Function (AUSF), Subscription Identifier De-concealing function (SIDE), Unified Data Management (UDM), Security Edge Protection Proxy (SEPP), Network Exposure Function (NEF), and / or a User Plane Function (UPF).

[0091] The host 4116 may be under the ownership or control of a service provider other than an operator or provider of the access network 4104 and / or the telecommunications network 4102. The host 4116 may be operated by the service provider or on behalf of the service provider. The host 4116 may host a variety of applications to provide one or more service. Examples of such applications include live and pre-recorded audio / video content, data collection services such as retrieving and compiling data on various ambientP113000W001conditions detected by a plurality of UEs, analytics functionality, social media, functions for controlling or otherwise interacting with remote devices, functions for an alarm and surveillance center, or any other such function performed by a server.

[0092] As a whole, the communication system 4100 of Figure 15 enables connectivity between the UEs, network nodes, and hosts. In that sense, the communication system 4100 may be configured to operate according to predefined rules or procedures, such as specific standards that include, but are not limited to: Global System for Mobile Communications (GSM); Universal Mobile Telecommunications System (UMTS); Long Term Evolution (LTE), and / or other suitable 2G, 3G, 4G, 5G standards, or any applicable future generation standard (e.g., 6G); wireless local area network (WLAN) standards, such as the Institute of Electrical and Electronics Engineers (IEEE) 802.11 standards (Wi-Fi); and / or any other appropriate wireless communication standard, such as the Worldwide Interoperability for Microwave Access (Wi-Max), Bluetooth, Z-Wave, Near Field Communication (NFC) ZigBee, Li-Fi, and / or any low-power wide-area network (LPWAN) standards such as LoRa and Sigfox. Moreover, the communication system 4100 may be configured to support multiple different standards, protocols, or other rule sets, with individual components supporting all of the relevant rule sets or with different components or sub-systems within the communication system 4100 supporting different standards, protocols, or rule sets.

[0093] As one example, in certain embodiments, access network 4104 may contain some access network nodes 4110 that support 3GPP radio access technologies (RAT), such as LTE or NR, while other access network nodes 4110 support (or the same access network nodes 4110 additionally support) non-3GPP RATs, such as Wi-Fi or a proprietary RAT. As another example, telecommunications network 4102 may support multiple generations of related communication standards (e.g., 4G and 5G 3GPP communication standards) and, as a result, may include an access network 104 and / or a core network 106 that supports multiple different standard generations or may include multiple access networks 104 and / or multiple core networks 106 with individual networks 104, 106 supporting different standard generations.

[0094] Telecommunications network 4102 may support network slicing to provide different logical networks to different devices that are connected to the telecommunications network 4102. For example, the telecommunications network 4102 may provide Ultra Reliable Low Latency Communication (URLLC) services to some UEs, whileP113000W001providing Enhanced Mobile Broadband (eMBB) services to other UEs, and / or Massive Machine Type Communication (mMTC) / Massive loT services to yet further UEs.

[0095] In some examples, one or more of the UEs 4112 are configured to transmit and / or receive information without direct human interaction. For instance, a UE may be designed to transmit information to the access network 4104 on a predetermined schedule, when triggered by an internal or external event, or in response to requests from the access network 4104. Additionally, a UE may be configured for operating in single- or multi-RAT or multi-standard mode. For example, a UE may operate with any one or combination of Wi-Fi, NR (New Radio) and LTE, i.e. being configured for multi-radio dual connectivity (MR-DC), such as E-UTRAN (Evolved-UMTS Terrestrial Radio Access Network) New Radio - Dual Connectivity (EN-DC).

[0096] In the example, the hub 4114 communicates with the access network 4104 to facilitate indirect communication between one or more UEs (e.g., UE 4112C and / or 4112D) and network nodes (e.g., network node 4110B). In some examples, the hub 4114 may be a controller, router, content source and analytics, or any of the other communication devices described herein regarding UEs. For example, the hub 4114 may be a broadband router enabling access to the core network 4106 for the UEs. As another example, the hub 4114 may be a controller that sends commands or instructions to one or more actuators in the UEs. Commands or instructions may be received from the UEs, network nodes 4110, or by executable code, script, process, or other instructions in the hub 4114.

[0097] As another example, the hub 4114 may be a data collector that acts as temporary storage for UE data and, in some embodiments, may perform analysis or other processing of the data. As another example, the hub 4114 may be a content source. For example, for a UE that is a VR headset, display, loudspeaker or other media delivery device, the hub 4114 may retrieve VR assets, video, audio, or other media or data related to sensory information via a network node, which the hub 4114 then provides to the UE either directly, after performing local processing, and / or after adding additional local content. In still another example, the hub 4114 acts as a proxy server or orchestrator for the UEs, in particular if one or more of the UEs are low energy loT devices.

[0098] The hub 4114 may have a constant / persistent or intermittent connection to the network node 4110B. The hub 4114 may also allow for a different communication scheme and / or schedule between the hub 4114 and UEs (e.g., UE 4112C and / or 4112D), and between the hub 4114 and the core network 4106. In other examples, the hub 4114 is connectedP113000W001to the core network 4106 and / or one or more UEs via a wired connection. Moreover, the hub 4114 may be configured to connect to an M2M service provider over the access network 4104 and / or to another UE over a direct connection. In some scenarios, UEs may establish a wireless connection with the network nodes 4110 while still connected via the hub 4114 via a wired or wireless connection. In some embodiments, the hub 4114 may be a dedicated hub - that is, a hub whose primary function is to route communications to / from the UEs from / to the network node 411 OB. In other embodiments, the hub 4114 may be a non-dedicated hub - that is, a device which is capable of operating to route communications between the UEs and network node 4110B, but which is additionally capable of operating as a communication start and / or end point for certain data channels.

[0099] Figure 16 is another example of a communication system 4200 according to some embodiments. As used herein, the communication system 4200 includes multiple access points (APs) 4210 (with four exemplary APs 4210A, 4210B, 4210C, and 4210D being depicted) and multiple wireless devices, referred to in the context of communication system 4200 as stations (STAs) 4212 (referred to individually as STA 4212A, STA 4212B, STA 4212C, STA 4212D, and STA 4212E). STA 4212A is served by AP 4210A in a first basic service set (BSS) 4220A. STA 4210B and STA 4210C are served by AP 4210B in a second BSS, BSS 4220B. STA 4212D is served by AP 4210C in a third BSS, BSS 4220C. STA 4212E is served by AP 4210D in a fourth BSS, BSS 4220D. Stations 4212 may be non-AP STAs and correspond to various kinds of wireless devices, for example, user terminals, such as mobile or stationary computing devices like smartphones, laptop computers, desktop computers, tablet computers, gaming devices, head-mounted displays (HMDs) for Augmented Reality (AR) or Virtual Reality (VR), or the like. Further, stations 4212 could, for example, correspond to other kinds of equipment like smart home devices, printers, multimedia devices, data storage devices, or the like.[000100] Each of STAs 4212 may connect through a radio link to one of APs 4210. For example, depending on location or channel conditions experienced by a given STA 4212, the STA may select an appropriate AP and BSS for establishing the radio link. The radio link may be based on one or more orthogonal frequency-division multiplexing (OFDM) carriers from a frequency spectrum that is shared on the basis of a contention-based mechanism, e.g., an unlicensed or license exempt band like 2.4 GHz Industrial, Scientific, and Medical (ISM) band, the 5 GHz band, the 6 GHz band, or the 60 GHz band.P113000W001[000101] Each AP 4210 may provide data connectivity to ST As 4212 connected to a particular AP 4210. As illustrated, APs 4210 may be connected to a data network 4230. In this way, APs 4210 may also provide data connectivity between STAs 4212 and other entities, e.g., to one or more servers, service providers, data sources, data sinks, user terminals, or the like. Accordingly, the radio link established between a given STA 4212 and its serving AP 4210 may be used for providing various kinds of services to STA 4212, e.g., a voice service, a multimedia service, or other data service. Such services may be based on applications that are executed on STA 4212 and / or on a device linked to STA 4212. By way of example, Figure 16 illustrates an application service platform 4232 provided in data network 4230. The application(s) executed on STA 4212 and / or on one or more other devices linked to STA 4212 may use the radio link for data communication with one or more other STA 4212 and / or the application service platform 4232, thereby enabling utilization of the corresponding service(s) at STA 4212.[000102] Figure 17 shows a wireless device 4300, which may be configured to operate in communication system 4100 of Figure 15 or in communication system 4200 of Figure 16. The wireless device 4300 may be alternatively referred to as a UE 4300, like a UE 4112 within the context of communication system 4100, or as a station (STA) 4300 or as a non-access-point station (non-AP STA) 4300, like a STA 4212 within the context of the communication system 4200, in accordance with respective embodiments. As used herein, a wireless device refers to a device capable, configured, arranged and / or operable to communicate wirelessly with network nodes and / or other wireless devices. Examples of a wireless device include, but are not limited to, a smart phone, mobile phone, cell phone, voice over IP (VoIP) phone, wireless local loop phone, desktop computer, personal digital assistant (PDA), wireless cameras, gaming console or device, music storage device, playback appliance, wearable terminal device, wireless endpoint, mobile station, tablet, laptop, laptop-embedded equipment (FEE), laptop-mounted equipment (LME), smart device, wireless customer-premise equipment (CPE), vehicle, vehicle-mounted or vehicle embedded / integrated wireless device, and wireless terminal. Other examples include any type of UE identified by the 3rd Generation Partnership Project (3GPP), including a narrow band internet of things (NB-IoT) UE, a machine type communication (MTC) UE, and / or an enhanced MTC (eMTC) UE.[000103] A wireless device 4300 may support device-to-device (D2D) communication, for example by implementing a 3GPP standard for sidelink communication, Dedicated Short-Range Communication (DSRC), vehicle-to-vehicle (V2V), vehicle-to-P113000W001infrastructure (V2I), or vehicle-to-everything (V2X). In other examples, wireless device 4300 may not necessarily have a user in the sense of a human user who owns and / or operates the relevant device. Instead, wireless device 4300 may represent a device that is intended for sale to, or operation by, a human user but which may not, or which may not initially, be associated with a specific human user (e.g., a smart sprinkler controller). Alternatively, wireless device 4300 may represent a device that is not intended for sale to, or operation by, an end user but which may be associated with or operated for the benefit of a user (e.g., a smart power meter).[000104] In particular embodiments, wireless device 4300 includes processing circuitry 4302 that is operatively coupled via a bus 4304 to an input / output interface 4306, a power source 4308, a memory 4310, a communication interface 4312, and / or any other component, or any combination thereof. Certain embodiments of wireless device 4300 may include all or a subset of the components shown in Figure 17. The level of integration between the components may vary from one embodiment of wireless device 4300 to another. In general, in a particular embodiment of wireless device 4300, processing circuitry 4302, input / output interface 4306, power source 4308, memory 4310, and communication interface 4312 may, in whole or in part, represent or include physical components common to or shared by one or more of the other elements of wireless device 4300. Further, certain embodiments of wireless devices 4300 may contain multiple instances of a component, such as multiple processors, memories, transceivers, transmitters, receivers, etc.[000105] The processing circuitry 4302 is configured to process instructions and data and may be configured to implement any sequential state machine operative to execute instructions stored as machine-readable computer programs in the memory 4310. The processing circuitry 4302 may be implemented as one or more hardware-implemented state machines (e.g., in discrete logic, field-programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), etc.); programmable logic together with appropriate firmware; one or more stored computer programs, general-purpose processors, such as a microprocessor or digital signal processor (DSP), together with appropriate software; or any combination of the above. For example, the processing circuitry 4302 may include multiple central processing units (CPUs).[000106] In the example, the input / output interface 4306 may be configured to provide an interface or interfaces to an input device, output device, or one or more input and / or output devices. Examples of an output device include a speaker, a sound card, a video card, a display, a monitor, a printer, an actuator, an emitter, a smartcard, another output device, or anyP113000W001combination thereof. An input device may allow a user to capture information into wireless device 4300. Examples of an input device include a touch-sensitive or presence-sensitive display, a camera (e.g., a digital camera, a digital video camera, a web camera, etc.), a microphone, a sensor, a mouse, a trackball, a directional pad, a trackpad, a scroll wheel, a smartcard, and the like. The presence-sensitive display may include a capacitive or resistive touch sensor to sense input from a user. A sensor may be, for instance, an accelerometer, a gyroscope, a tilt sensor, a force sensor, a magnetometer, an optical sensor, a proximity sensor, a biometric sensor, etc., or any combination thereof. An output device may use the same type of interface port as an input device. For example, a Universal Serial Bus (USB) port may be used to provide an input device and an output device.[000107] In some embodiments, the power source 4308 is structured as a battery or battery pack. Other types of power sources, such as an external power source (e.g., an electricity outlet), photovoltaic device, or power cell, may be used to supply power to circuitry or to charge an associated battery. The power source 4308 may further include power circuitry for delivering power from the power source 4308 itself, and / or an external power source, to the various parts of wireless device 4300 via input circuitry or an interface such as an electrical power cable. Power source 4308 may perform any formatting, converting, or other modification to make accessible power suitable for the respective components of the wireless device 4300 to which power is supplied.[000108] The memory 4310 may be or be configured to include memory such as random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic disks, optical disks, hard disks, removable cartridges, flash drives, and so forth. In one example, the memory 4310 includes one or more programs 4314, such as an operating system, web browser application, a widget, gadget engine, or other application, and corresponding data 4316. The memory 4310 may store, for use by wireless device 4300, any of a variety of various operating systems or combinations of operating systems.[000109] The memory 4310 may be configured to include a number of physical drive units, such as redundant array of independent disks (RAID), flash memory, USB flash drive, external hard disk drive, thumb drive, pen drive, key drive, high-density digital versatile disc (HD-DVD) optical disc drive, internal hard disk drive, Blu-Ray optical disc drive, holographic digital data storage (HDDS) optical disc drive, external mini-dual in-line memoryP113000W001module (DIMM), synchronous dynamic random access memory (SDRAM), external microDIMM SDRAM, smartcard memory such as tamper resistant module in the form of a universal integrated circuit card (UICC) including one or more subscriber identity modules (SIMs), such as a USIM and / or ISIM, other memory, or any combination thereof. The UICC may for example be an embedded UICC (eUICC), integrated UICC (iUICC) or a removable UICC commonly known as ‘SIM card.’ The memory 4310 may allow wireless device 4300 to access instructions, programs and the like, stored on transitory or non-transitory memory media, to off-load data, or to upload data. An article of manufacture, such as one utilizing a communication system may be tangibly embodied as or in the memory 4310, which may be or comprise a device -readable storage medium.[000110] The processing circuitry 4302 may be configured to communicate with an access network or other network via or using the communication interface 4312. The communication interface 4312 may comprise one or more communication subsystems and may include or be communicatively coupled to an antenna 4322. The communication interface 4312 may include one or more transceivers used to communicate, such as by communicating with one or more remote transceivers of another device capable of wireless communication (e.g., another wireless device or a network node in an access network). Each transceiver may include a transmitter 4318 and / or a receiver 4320 appropriate to provide network communications (e.g. , optical, electrical, frequency allocations, and so forth). Moreover, the transmitter 4318 and receiver 4320 may be coupled to one or more antennas (e.g., antenna 4322) and may share circuit components, software or firmware, or alternatively be implemented separately.[000111] In the illustrated embodiment, communication functions of the communication interface 4312 may include cellular communication, Wi-Fi communication (e.g., according to an IEEE 802.11 family standard), LPWAN communication, data communication, voice communication, multimedia communication, short-range communications such as Bluetooth, near-field communication, location-based communication such as the use of the global positioning system (GPS) to determine a location, another like communication function, or any combination thereof. Communications may be implemented according to one or more communication protocols and / or standards, such as IEEE 802.11, Code Division Multiplexing Access (CDMA), Wideband Code Division Multiple Access (WCDMA), GSM, LTE, New Radio (NR), UMTS, WiMax, Ethernet, transmission control protocol / internet protocol (TCP / IP), synchronous optical networking (SONET), Asynchronous Transfer Mode (ATM), QUIC, Hypertext Transfer Protocol (HTTP), and so forth.P113000W001[000112] In particular embodiments, wireless device 4300 may provide an output of data captured via a sensor, through its communication interface 4312, via a wireless connection to a network node, and / or in any appropriate manner. Data captured by sensors of a wireless device 4300 can be communicated through a wireless connection to a network node via another wireless device 4300. In particular embodiments, such output may be periodic (e.g., once every 15 minutes if it reports the sensed temperature), random (e.g., to even out the load from reporting from several sensors), in response to a triggering event (e.g., when moisture is detected an alert is sent), in response to a request (e.g., a user initiated request), or a continuous stream (e.g., a live video feed of a patient).[000113] As another example, wireless device 4300 comprises an actuator, a motor, or a switch, related to a communication interface configured to receive wireless input from a network node via a wireless connection. In response to the received wireless input the states of the actuator, the motor, or the switch may change. For example, wireless device 4300 may comprise a motor that adjusts the control surfaces or rotors of a drone in flight according to the received input or to a robotic arm performing a medical procedure according to the received input.[000114] Wireless device 4300, when in the form of an Internet of Things (loT) device, may be a device for use in one or more application domains, these domains comprising, but not limited to, wearable technology, extended industrial application and healthcare. Nonlimiting examples of such an loT device are a device which is or which is embedded in: a connected refrigerator or freezer, a TV, a connected lighting device, an electricity meter, a robot vacuum cleaner, a voice controlled smart speaker, a home security camera, a motion detector, a thermostat, a smoke detector, a door / window sensor, a flood / moisture sensor, an electrical door lock, a connected doorbell, an air conditioning system like a heat pump, an autonomous vehicle, a surveillance system, a weather monitoring device, a vehicle parking monitoring device, an electric vehicle charging station, a smart watch, a fitness tracker, a wearable for tactile augmentation or sensory enhancement, a water sprinkler, an animal- or item-tracking device, a sensor for monitoring a plant or animal, an industrial robot, an Unmanned Aerial Vehicle (UAV), and any kind of medical device, like a heart rate monitor or a remote controlled surgical robot. In particular embodiments, wireless device 4300 represents an loT device that comprises circuitry and / or software in dependence of the intended application of the loT device in addition to other components as described in relation to the example embodiment of wireless device 4300 shown in Figure 17.P113000W001[000115] As yet another specific example, in an loT scenario, wireless device 4300 may represent a machine or other device that performs monitoring and / or measurements, and transmits the results of such monitoring and / or measurements to another wireless device and / or a network node. Wireless device 4300 may in this case be an M2M device, which may in a 3GPP context be referred to as an MTC device. As one particular example, wireless device 4300 may implement the 3GPP NB-IoT standard. In other scenarios, wireless device 4300 may represent a vehicle, such as a car, a bus, a truck, a ship and an airplane, or other equipment that is capable of monitoring and / or reporting on its operational status or other functions associated with its operation.[000116] In practice, any number of wireless devices 4300 may be used together with respect to a single use case. For example, a first wireless device 4300 might be or be integrated in a drone and provide the drone’s speed information (obtained through a speed sensor) to a second wireless device 4300 that is a remote controller operating the drone. When a user makes changes from the remote controller, the first wireless device 4300 may adjust the throttle on the drone (e.g. by controlling an actuator) to increase or decrease the drone’s speed. The first and / or the second wireless device 4300 can also include more than one of the functionalities described above. For example, wireless device 4300 might comprise the sensor and the actuator, and handle communication of data for both the speed sensor and the actuators.[000117] Figure 18 shows a network node 4400 in accordance with some embodiments. As used herein, network node refers to equipment capable, configured, arranged and / or operable to communicate directly or indirectly with a UE and / or with other network nodes or equipment, in a telecommunications network. In accordance with respective embodiments, network node 4400 may be configured to operate in communication system 4100 of Figure 15, like network nodes 4108 or 4110, or in communication system 4200 of Figure 16, like an AP 4210 or a station 4212. Examples of network nodes include, but are not limited to, access points (APs) (e.g., radio access points), base stations (BSs) (e.g., radio base stations, Node Bs, evolved Node Bs (eNBs) and NR NodeBs (gNBs)), O-RAN nodes or components of an O-RAN node (e.g., O-RU, O-DU, O-CU).[000118] Network nodes 4400 may be categorized based on the amount of coverage they provide (or, stated differently, their transmit power level) and so, depending on the provided amount of coverage, may be referred to as femto base stations, pico base stations, micro base stations, or macro base stations. Network node 4400 may be a relay node or a relay donor node controlling a relay. Network nodes 4400 may also include one or more (or all) partsP113000W001of a distributed radio base station such as centralized digital units, distributed units (e.g., in an O-RAN access node) and / or remote radio units (RRUs), sometimes referred to as Remote Radio Heads (RRHs). Such remote radio units may or may not be integrated with an antenna as an antenna integrated radio. Parts of a distributed radio base station may also be referred to as nodes in a distributed antenna system (DAS).[000119] Other examples of network nodes 4400 include multiple transmission point (multi-TRP) 5G access nodes, multi-standard radio (MSR) equipment such as MSR BSs, network controllers such as radio network controllers (RNCs) or base station controllers (BSCs), base transceiver stations (BTSs), transmission points, transmission nodes, multi-cell / multicast coordination entities (MCEs), Operation and Maintenance (O&M) nodes, Operations Support System (OSS) nodes, Self-Organizing Network (SON) nodes, positioning nodes (e.g., Evolved Serving Mobile Location Centers (E-SMLCs)), and / or Minimization of Drive Tests (MDTs).[000120] In particular embodiments, network node 4400 includes a processing circuitry 4402, a memory 4404, a communication interface 4406, and a power source 4408. In general, in a particular embodiment of network node 4400, processing circuitry 4402, memory 4404, communication interface 4406, and power source 4408 may, in whole or in part, represent or include physical components common to or shared by one or more of the other elements of network node 4400.[000121] The network node 4400 may be composed of multiple distinct network entities (e.g., a NodeB entity and a RNC entity, or a BTS entity and a BSC entity, etc.), which may each have or utilize their own respective physical components. In certain scenarios in which the network node 4400 comprises multiple such entities (e.g., BTS and BSC), one or more of the separate entities may be shared among several network nodes. For example, a single RNC may control multiple NodeBs. In such a scenario, each unique NodeB and RNC pair, may in some instances be considered a single separate network node. In some embodiments, the network node 4400 may be configured to support multiple radio access technologies (RATs). In such embodiments, some components may be duplicated (e.g., separate memories 4404 or portions of memory 4404 for different RATs) and some components may be reused (e.g., a same antenna 4410 may be shared by different RATs). The network node 4400 may also include multiple sets of the various illustrated components for different wireless technologies integrated into network node 4400, for example GSM, WCDMA, LTE, NR, Wi-Fi (e.g., according to an IEEE 802.11 family standard), Zigbee, Z-P113000W001wave, LoRaWAN, Radio Frequency Identification (RFID) or Bluetooth wireless technologies. These wireless technologies may be integrated into the same or different chip or set of chips and other components within network node 4400.[000122] The processing circuitry 4402 may comprise a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application-specific integrated circuit, field programmable gate array, or any other suitable computing device, resource, or combination of hardware, software and / or encoded logic operable to provide, either alone or in conjunction with other components, such as the memory 4404, to provide network node 4400 functionality.[000123] In some embodiments, the processing circuitry 4402 includes a system on a chip (SOC). In some embodiments, the processing circuitry 4402 includes one or more of radio frequency (RF) transceiver circuitry 4412 and baseband processing circuitry 4414. In some embodiments, the RF transceiver circuitry 4412 and the baseband processing circuitry 4414 may be on separate chips (or sets of chips), boards, or units, such as radio units and digital units. In alternative embodiments, part or all of RF transceiver circuitry 4412 and baseband processing circuitry 4414 may be on the same chip or set of chips, boards, or units.[000124] The memory 4404 may comprise any form of volatile or non-volatile computer-readable memory including, without limitation, persistent storage, solid-state memory, remotely mounted memory, magnetic media, optical media, random access memory (RAM), read-only memory (ROM), mass storage media (for example, a hard disk), removable storage media (for example, a flash drive, a Compact Disk (CD) or a Digital Video Disk (DVD)), and / or any other volatile or non-volatile, non-transitory device-readable and / or computer-executable memory devices that store information, data, and / or instructions that may be used by the processing circuitry 4402. The memory 4404 may store any suitable instructions, data, or information, including a computer program, software, an application including one or more of logic, rules, code, tables, and / or other instructions capable of being executed by the processing circuitry 4402 and utilized by the network node 4400. The memory 4404 may be used to store any calculations made by the processing circuitry 4402 and / or any data received via the communication interface 4406. In some embodiments, the processing circuitry 4402 and memory 4404 is integrated.[000125] The communication interface 4406 is used in wired or wireless communication of signaling and / or data with UEs, other network nodes, and / or any other network equipment. In the illustrated embodiment, communication interface 4406 comprisesP113000W001port(s) / terminal(s) 4416 to send and receive data, for example to and from a network over a wired connection. In particular embodiments, network node 4300 may be capable of wireless communication and communication interface 4406 may also include radio front-end circuitry 4418 that may be coupled to, or in certain embodiments a part of, an antenna 4410. Particular embodiments of radio front-end circuitry 4418 include filter(s) 4420 and amplifier(s) 4422. The radio front-end circuitry 4418 may be connected to an antenna 4410 and processing circuitry 4402. The radio front-end circuitry may be configured to condition signals communicated between antenna 4410 and processing circuitry 4402. The radio front-end circuitry 4418 may receive digital data that is to be sent out to other network nodes or UEs via a wireless connection. The radio front-end circuitry 4418 may convert the digital data into a radio signal(s) having the appropriate channel and bandwidth parameters using a combination of filters 4420 and / or amplifiers 4422. The radio signal(s) may then be transmitted via the antenna 4410. Similarly, when receiving data, the antenna 4410 may collect radio signals which are then converted into digital data by the radio front-end circuitry 4418. The digital data may be passed to the processing circuitry 4402. In other embodiments, the communication interface may comprise different components and / or different combinations of components.[000126] In certain alternative embodiments, network node 4400 may be capable of wireless communication but does not include separate radio front-end circuitry 4418, instead, the processing circuitry 4402 includes radio front-end circuitry and is connected to the antenna 4410. Similarly, in some embodiments, all or some of the RF transceiver circuitry 4412 is part of the communication interface 4406. In still other embodiments, the communication interface 4406 includes one or more ports or terminals 4416, the radio front-end circuitry 4418, and the RF transceiver circuitry 4412, as part of a radio unit (not shown), and the communication interface 4406 communicates with the baseband processing circuitry 4414, which is part of a digital unit (not shown).[000127] The antenna 4410 may include one or more antennas, or antenna arrays, configured to send and / or receive wireless signals. The antenna 4410 may be coupled to the radio front-end circuitry 4418 and may be any type of antenna capable of transmitting and receiving data and / or signals wirelessly. In certain embodiments, the antenna 4410 is separate from the network node 4400 and connectable to the network node 4400 through one or more interfaces or ports.[000128] The antenna 4410, communication interface 4406, and / or the processing circuitry 4402 may be configured to perform some or all of the receiving operations and / orP113000W001obtaining operations described herein as being performed by the network node 4400. Any information, data and / or signals may be received from a UE, another network node and / or any other network equipment. Similarly, the antenna 4410, the communication interface 4406, and / or the processing circuitry 4402 may be configured to perform some or all of the transmitting or sending operations described herein as being performed by the network node 4400. Any information, data and / or signals may be transmitted to a UE, another network node and / or any other network equipment.[000129] The power source 4408 provides power to the various components of network node 4400 in a form suitable for the respective components (e.g., at a voltage and current level needed for each respective component). The power source 4408 may further comprise, or be coupled to, power management circuitry to supply the components of the network node 4400 with power for performing the functionality described herein. For example, the network node 4400 may be connectable to an external power source (e.g., the power grid, an electricity outlet) via an input circuitry or interface such as an electrical cable, whereby the external power source supplies power to power circuitry of the power source 4408. As a further example, the power source 4408 may comprise a source of power in the form of a battery or battery pack which is connected to, or integrated in, power circuitry. The battery may provide backup power should the external power source fail.[000130] Embodiments of the network node 4400 may include additional components beyond those shown in Figure 18 for providing certain aspects of the network node’s functionality, including any of the functionality described herein and / or any functionality necessary to support the subject matter described herein. For example, the network node 4400 may include user interface equipment to allow input of information into the network node 4400 and to allow output of information from the network node 4400. This may allow a user to perform diagnostic, maintenance, repair, and other administrative functions for the network node 4400.[000131] Figure 19 is a block diagram illustrating a virtualization environment 4500 in which functions implemented by some embodiments may be virtualized. In the present context, virtualizing means creating virtual versions of apparatuses or devices which may include virtualizing hardware platforms, storage devices and networking resources. As used herein, virtualization can be applied to any device described herein, or components thereof, and relates to an implementation in which at least a portion of the functionality is implemented as one or more virtual components. Some or all of the functions described herein may beP113000W001implemented as virtual components executed by one or more virtual machines (VMs) implemented in one or more virtual environments 4500 hosted by one or more of hardware nodes, such as a hardware computing device that operates as an access network node, UE, core network node, or host. Further, in embodiments in which a virtual node does not require radio connectivity (e.g., a core network node or host), then the node may be entirely virtualized. In some embodiments, the virtualization environment 4500 includes components defined by the O-RAN Alliance, such as an O-Cloud environment orchestrated by a Service Management and Orchestration Framework via an 0-2 interface.[000132] Applications 4502 (which may alternatively be called software instances, virtual appliances, network functions, virtual nodes, virtual network functions, etc.) are run in the virtualization environment Q400 to implement some of the features, functions, and / or benefits of some of the embodiments disclosed herein.[000133] Hardware 4504 includes processing circuitry, memory that stores software and / or instructions executable by hardware processing circuitry, and / or other hardware devices as described herein, such as a network interface, input / output interface, and so forth. Software may be executed by the processing circuitry to instantiate one or more virtualization layers 4506 (also referred to as hypervisors or virtual machine monitors (VMMs)), provide VM 4508A and VM 4508B (which may be collectively referred to as VMs 4508), and / or perform any of the functions, features and / or benefits described in relation with some embodiments described herein. The virtualization layer 4506 may present a virtual operating platform that appears like networking hardware to one or more of the VMs 4508.[000134] The VMs 4508 comprise virtual processing, virtual memory, virtual networking or interface and virtual storage, and may be run by virtualization layer 4506. Different embodiments of the instance of a virtual appliance 4502 may be implemented on one or more of VMs 4508, and the implementations may be made in different ways. Virtualization of the hardware is in some contexts referred to as network function virtualization (NFV). NFV may be used to consolidate many network equipment types onto industry standard high volume server hardware, physical switches, and physical storage, which can be located in data centers, and customer premise equipment.[000135] In the context of NFV, each of the VMs 4508 may be a software implementation of a physical machine that runs programs as if they were executing on a physical, non-virtualized machine. Each of the VMs 4508, and that part of hardware 4504 that executes that VM, be it hardware dedicated to that VM and / or hardware shared by that VMP113000W001with others of the VMs, forms separate virtual network elements. Still in the context of NFV, a virtual network function is responsible for handling specific network functions that run in one or more of the VMs 4508 on top of the hardware 4504 and corresponds to an application 4502.[000136] Hardware 4504 may be implemented in a standalone network node with generic or specific components. Hardware 4504 may implement some functions via virtualization. Alternatively, hardware 4504 may be part of a larger cluster of hardware (e.g. such as in a data center or CPE) where many hardware nodes work together and are managed via management and orchestration 4510, which, among others, oversees lifecycle management of applications 4502. In some embodiments, hardware 4504 is coupled to one or more radio units that each include one or more transmitters and one or more receivers that may be coupled to one or more antennas. Radio units may communicate directly with other hardware nodes via one or more appropriate network interfaces and may be used in combination with the virtual components to provide a virtual node with radio capabilities, such as a radio access node or a base station. In some embodiments, some signaling can be provided with the use of a control system 4512 which may alternatively be used for communication between hardware nodes and radio units.[000137] Although the computing devices described herein (e.g., UEs, network nodes, hosts) may include the illustrated combination of hardware components, other embodiments may comprise computing devices with different combinations of components. It is to be understood that these computing devices may comprise any suitable combination of hardware and / or software needed to perform the tasks, features, functions and methods disclosed herein. Determining, calculating, obtaining or similar operations described herein may be performed by processing circuitry, which may process information by, for example, converting the obtained information into other information, comparing the obtained information or converted information to information stored in the network node, and / or performing one or more operations based on the obtained information or converted information, and as a result of said processing making a determination. Moreover, while components are depicted as single boxes located within a larger box, or nested within multiple boxes, in practice, computing devices may comprise multiple different physical components that make up a single illustrated component, and functionality may be partitioned between separate components. For example, a communication interface may be configured to include any of the components described herein, and / or the functionality of the components may beP113000W001partitioned between the processing circuitry and the communication interface. In another example, non-computationally intensive functions of any of such components may be implemented in software or firmware and computationally intensive functions may be implemented in hardware.[000138] In certain embodiments, some or all of the functionality described herein may be provided by processing circuitry executing instructions stored on in memory, which in certain embodiments may be a computer program product in the form of a non-transitory computer-readable storage medium. In alternative embodiments, some or all of the functionality may be provided by the processing circuitry without executing instructions stored on a separate or discrete device-readable storage medium, such as in a hard-wired manner. In any of those particular embodiments, whether executing instructions stored on a non-transitory computer-readable storage medium or not, the processing circuitry can be configured to perform the described functionality. The benefits provided by such functionality are not limited to the processing circuitry alone or to other components of the computing device, but are enjoyed by the computing device as a whole, and / or by end users and a wireless network generally.Example Embodiments[000139] Below are provided various possible example embodiments under the present disclosure. Examples are provided by way of example only, to show possible functionalities and elements of various possible implementations.[000140] Group A Embodiments[000141] Embodiment 1: A method performed by a wireless device for pre-processing / designing / training an AI / ML model for autoencoding for AI / ML-cnhanccd CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, the method comprising at least one of: leveraging transformer robustness to sequence length; transforming input and / or output dimensions dependent on sequence lengths; including a token design which has a fixed size and when network configuration changes, the sequence length changes, instead of the token; wherein transformers are naturally applicable to different sequence lengths; the self-attention in the encoder design can be replaced by a crossattention one, which allows for the encoder output sequence to have a different length, matching the desired latent space dimension; and / or including an initializer for latent space sequence, which may be needed by the cross-attention mechanism.P113000W001[000142] Embodiment 2: A method performed by a wireless device for pre-processing / designing / training an AI / ML model for autoencoding for AI / ML-cnhanccd CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, the method comprising at least one of: performing reshape, e.g., 2TX and 4 PRB are tokens; performing FCN (embedding); performing positional encoding; performing output sequence initialization; performing a transformer encoder block, with input from the positional encoding and the output sequence initialization, wherein cross attention is used instead of self attention; and performing FCN (latent).[000143] Embodiment 3: A method performed by a wireless device for pre-processing / designing / training an AI / ME model for autoencoding for AI / ML-enhanced CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, the method comprising: performing the flow chart of Figure 9.[000144] Embodiment 4: The method of any of embodiments 1 to 3, wherein the model is designed that is applicable to changes in dimensions of RX, TX, PRB and E.[000145] Embodiment 5: The method of any of embodiments 1 to 4, wherein this architecture assumes the definition of a token as an atomic unit of CSI, which can be repeated in a sequence of arbitrary length, this makes it robust to any dimension.[000146] Embodiment 6: The method of any of embodiments 1 to 5, wherein the initialization of the output sequence uses other strategies, for example, using the input sequence, applying some downsampling technique, initialize with random values, or use learnable parameters for the output token.[000147] Embodiment 7: The method of any of embodiments 1 to 6, wherein the definition of an atomic unit means that CSR-RS which contains less than that amount is not supported.[000148] Embodiment 8: The method of any of embodiments 1 to 7, wherein the token is defined in an atomic unit, where the atomic unit may consist of one or more parameters related to CSI such as TX, RX, PRB, Layer, Subband, Spatial domain, etc.[000149] Embodiment 9: The method of embodiment 8, wherein the token is built by splitting the input according to a 2-d window composed by a X number of samples of a first parameter and a Y number of samples of a second parameter.[000150] Embodiment 10: The method of any of embodiments 1 to 9, wherein the atomic unit is defined using the first parameter TX and the second PRB while assigning X=2 and Y=4.P113000W001[000151] Embodiment 11: The method of any of embodiments 1 to 10, wherein before the reshape functionality a convolution neural network (CNN) is applied, so that information among tokens is further leveraged, even a 2d or 3d CNN can be considered.[000152] Embodiment 12: The method of any of embodiments 1 to 11, wherein the reshape component is improved by adding a special token <new_rx> to be used when concatenating RX dimensions into the flattening, this could ease the learning for the transformer.[000153] Embodiment 13: The method of any of embodiments 1 to 12, wherein a 2 or 3-d positional encoding is used to better capture the sequences, or yet, a positional encoding dependent on the antenna array design.[000154] Embodiment 14: The method of any of embodiments 1 to 13, wherein a self-attention mechanism is applied after the positional encoding to allow for the input tokens to share information.[000155] Embodiment 15: The method of any of the previous embodiments, further comprising: providing user data; and forwarding the user data to a host via the transmission to the network node.[000156] Group B Embodiments[000157] Embodiment 16: A method performed by a network node for decoding signals from a wireless device, the signals autoencoded using an AI / ML model for AI / ML enhanced CSI reporting in NR, the method comprising at least one of: performing FCN (expansion); performing positional encoding; performing output sequence initialization; performing transformer decoder block, using both positional encoding and output sequence initialization as inputs; performing FCN (de-embedding); and performing reshape.[000158] Embodiment 17: A method performed by a network node for decoding signals from a wireless device, the signals autoencoded using an AI / ML model for AI / ML enhanced CSI reporting in NR, the method comprising: performing the flow chart of Figure 10.[000159] Embodiment 18: The method of embodiment 16 or 17, wherein an additional slot dimension is added to report CSI for different time instances (prediction).[000160] Embodiment 19: The method of embodiment 18, wherein such dimension can also be flattened in the sequence, concatenating each report, possibly with a special token <new_slot> to mark each slot, analogous to the example for the <new_rx>.P113000W001[000161] Embodiment 20: The method of any of embodiments 16 to 19, wherein such architecture is also applicable to preprocessing cases, the dimensions would just be, for example, beam, frequency domain basis and layer.[000162] Embodiment 21: The method of any of embodiments 16 to 20, wherein an atomic unit is used to define a token and the same architecture applied.[000163] Embodiment 22: The method of any of embodiments 16 to 21, wherein the encoder input scalability defining the tokens by the atomic unit is adopted independently of the association of the latent space with the sequence length and vice versa (e.g. the parts of this architecture can be applied independently for encoder input scalability of latent space scalability).[000164] Embodiment 23: The method of any of the previous embodiments, further comprising: obtaining user data; and forwarding the user data to a host or a user equipment.[000165] Group C Embodiments[000166] Embodiment 24: A wireless device for pre-processing / designing / training an AI / ML model for autoencoding for AI / ML-enhanced CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, comprising: processing circuitry configured to perform any of the operations of any of the Group A embodiments; and a power source configured to supply power to the processing circuitry.[000167] Embodiment 25: A network node for decoding signals from a wireless device, the signals autoencoded using an AI / ML model for AI / ML enhanced CSI reporting in NR, the network node comprising: processing circuitry configured to perform any of the operations of any of the Group B embodiments; a power source circuitry configured to supply power to the processing circuitry.[000168] Embodiment 26: A wireless device for pre-processing / designing / training an AI / ML model for autoencoding for AI / ML-enhanced CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, the wireless device comprising: one or more antennas; communication interface connected to the one or more antennas and to processing circuitry; the processing circuitry being configured to perform any of the operations of any of the Group A embodiments; an input interface connected to the processing circuitry and configured to allow input of information into the UE to be processed by the processing circuitry; an output interfaceP113000W001connected to the processing circuitry and configured to output information from the UE that has been processed by the processing circuitry; and a power source connected to the processing circuitry and configured to supply power to the UE.[000169] Embodiment 27: A computer-implemented method for pre-processing / designing / training an AI / ML model for autoencoding for AI / ML-cnhanccd CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, the method comprising at least one of: performing reshape, e.g., 2TX and 4 PRB are tokens; performing FCN (embedding); performing positional encoding; performing output sequence initialization; performing a transformer encoder block, with input from the positional encoding and the output sequence initialization, wherein cross attention is used instead of self attention; and performing FCN (latent).[000170] Embodiment 28: A computer-implemented method for pre-processing / designing / training an AI / ML model for autoencoding for AI / ML-enhanced CSI reporting in NR, the model structure for encoder robust to multiple input and output dimensions, the method comprising: performing the flow chart of Figure 9.[000171] Embodiment 29 : A computer-implemented method for decoding signals from a wireless device, the signals autoencoded using an AI / ML model for AI / ML enhanced CSI reporting in NR, the method comprising at least one of: performing FCN (expansion); performing positional encoding; performing output sequence initialization; performing transformer decoder block, using both positional encoding and output sequence initialization as inputs; performing FCN (de-embedding); and performing reshape.[000172] Embodiment 30: A computer-implemented method performed for decoding signals from a wireless device, the signals autoencoded using an AI / ML model for AI / ML enhanced CSI reporting in NR, the method comprising: performing the flow chart of Figure 10.

Claims

P113000W001ClaimsWhat is claimed is:

1. A method (1500) performed by a user equipment, UE (4300), for training a machine learning, ML, model for autoencoding channel state information, CSI, reporting, the method comprising:performing (1510) reshape of a dataset used by the ML model into a sequence of one or more atomic units of CSI; andprojecting (1520) the one or more atomic units into an embedding space using a ML layer;wherein the ML model is scalable to multiple input and multiple output dimensions.

2. The method of claim 1, further comprising:performing positional encoding of the one or more atomic units of CSI.

3. The method of claim 2, further comprising:performing a transformer encoder block, with inputs from the positional encoding and from an output sequence initialization, wherein cross attention is used instead of self-attention; andperforming a projection of an output sequence of the transformer encoder block into a latent space using a respective ML layer.

4. The method of claim 3, wherein the projection of the output sequence of the transformer encoder block into the latent space is performed per token.

5. The method of any of claims 3-4, wherein the projection of the output sequence of the transformer encoder block into the latent space is performed using a fully connected network, FCN.

6. The method of any of claims 2-5, wherein the positional encoding leverages information of an antenna configuration of the UE.

7. The method of any of claims 1 to 6, wherein the one or more atomic units of CSIP113000W001represents one or more of: reception antenna; transmission antenna; physical resource block, PRB; beams; layer; sub-band; or spatial domain.

8. The method of any of claims 1 to 7, wherein input and / or output of the ML model comprises one or more tokens, each token defined as an atomic unit of CSI that can be repeated in a sequence of scalable length.

9. The method of any of claims 1 to 8, wherein the ML model is configured to initialize an output sequence based on at least one of: the sequence of one or more atomic units of CSI; one or more downsampling techniques; one or more random values; or one or more learnable parameters for an output token.

10. The method of any of claims 1 to 9, further comprising:applying a Convolutional Neural Network, CNN, before performing the reshape.

11. The method of any of claim 2 or claims 3 to 10 incorporating claim 2, further comprising applying a self-attention mechanism after the positional encoding to allow for one or more input tokens to share information.

12. The method of any of claims 1 to 11, wherein projecting the one or more atomic units into the embedding space is performed using fully connected network, FCN, embedding.

13. A method (1700) performed by a user equipment, UE (4300), for training a machine learning, ML, model for autoencoding channel state information, CSI, reporting, the method comprising:performing (1710) a transformer encoder block, with input from a positional encoding and an output sequence initialization, wherein cross attention is used instead of self-attention; andperforming (1720) a projection of an output sequence of the transformer encoder block into a latent space using a ML layer.

14. The method of claim 13, wherein the performing a projection is performed using a fully connected network, FCN.P113000W00115. A method (1900) performed by a network node (4400) for decoding one or more signals from a user equipment, UE, the method comprising:performing (1910) expansion of a dataset from a latent space; andperforming (1920) positional encoding;wherein the one or more signals were autoencoded using a machine learning, ML, model for channel state information, CSI, reporting.

16. The method of claim 15, further comprising:performing output sequence initialization; andperforming transformer decoding, using both positional encoding and output sequence initialization as inputs.

17. The method of claim 15 or 16, further comprising:performing de-embedding of the one or more atomic units from an embedding space using a ML layer; andperforming reshape to compose an output of one or more atomic units of CSI.

18. The method of any of claims 15 to 17, wherein an additional slot dimension is added to an output dataset to report CSI for different time instances.

19. The method of any of claims 15 to 18, wherein performing expansion comprises a fully connected network, FCN, expansion.

20. The method of any of claim 15 to 19, wherein input and / or output of the ML model comprises one or more tokens, each token defined as an atomic unit of CSI.

21. A method (2100) performed by a network node (4400) for decoding one or more signals from a user equipment, UE (4300), the method comprising:performing (2110) de-embedding of one or more atomic units from an embedding space using a machine learning, ML, layer; andperforming (2120) reshape to compose an output of one or more atomic units of channel state information, CSI;P113000W001wherein the one or more signals were autoencoded using a ML model for CSI reporting.

22. The method of claim 21, wherein the performing de-embedding is performed using a fully connected network, FCN.

23. A user equipment, UE (4300), for training a machine learning, ML, model for autoencoding channel state information, CSI, reporting,processing circuitry (4302) configured to perform any of the operations of any of claims 1 to 14; anda power source (4308) configured to supply power to the processing circuitry.

24. A network node (4400) for decoding one or more signals from a user equipment, UE (4300), wherein the one or more signals were autoencoded using a machine learning, ML, model for channel state information, CSI, reporting, the network node comprising:processing circuitry (4402) configured to perform any of the operations of any of claims 15 to 22;a power source circuitry (4408) configured to supply power to the processing circuitry.

25. A user equipment, UE (4300), for training a machine learning, ML, model for autoencoding channel state information, CSI, reporting,processing circuitry (4302); anda memory storing (4310) instructions whereby the processing circuitry is operable to perform the steps of:performing reshape of a dataset used by the ML model into a sequence of one or more atomic units of CSI; andprojecting the one or more atomic units into an embedding space using a ML layer; wherein the ML model is scalable to multiple input and multiple output dimensions.

26. A user equipment, UE (4300), for training a machine learning, ML, model for autoencoding channel state information, CSI, reporting,processing circuitry (4302); anda memory storing (4310) instructions whereby the processing circuitry is operable toP113000W001perform the steps of:performing a transformer encoder block, with input from a positional encoding and an output sequence initialization, wherein cross attention is used instead of self-attention; andperforming a projection of an output sequence of the transformer encoder block into a latent space using a ML layer.

27. A network node (4400) for decoding one or more signals from a user equipment, UE (4300), wherein the one or more signals were autoencoded using a machine learning, ML, model for channel state information, CSI, reporting, the network node comprising:processing circuitry (4402); anda memory (4408) storing instructions whereby the processing circuitry is operable to perform the steps of:performing fully connected network, FCN, expansion; andperforming positional encoding;wherein the one or more signals were autoencoded using a machine learning, ML, model for channel state information, CSI, reporting.

28. A network node (4400) for decoding one or more signals from a user equipment, UE (4300), wherein the one or more signals were autoencoded using a machine learning, ML, model for channel state information, CSI, reporting, the network node comprising:processing circuitry (4402); anda memory (4408) storing instructions whereby the processing circuitry is operable to perform the steps of:performing de-embedding of one or more atomic units from an embedding space using a ML layer; andperforming reshape to compose an output of one or more atomic units of CSI; wherein the one or more signals were autoencoded using a ML model for CSI reporting.

Images