Apparatus, method and computer program

Abstract

Apparatus, Method and Computer Program There is provided an apparatus comprising means for providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector, providing the latent vector to an adaptive delta modulation encoder, where-in the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension and providing the bit stream to a network entity for decoding by a spatial and frequency decoder.

Description

TITLEApparatus, Method and Computer ProgramTECHNICAL FIELD

[0001] Various embodiments of this disclosure relate generally to methods, apparatus and computer programs, and in particular, but not exclusively, to ADM-assisted recurrent latent space quantization scheme for AIML-enabled two-sided CSI compression.BACKGROUND

[0002] A communication system can be seen as a facility that enables communication sessions between two or more communication devices, or provides communication devices access to a network. A mobile or wireless communication network is one example of a communication network. A communication device may be provided with a service by an application server.

[0003] A mobile or wireless communication network may operate in accordance with standard(s), such as those provided by 3GPP (Third Generation Partnership Project) or ETSI (European Telecommunications Standards Institute). Examples of mobile or wireless communication network that operate in accordance with 3GPP standards are generally referred to as 4G (4th Generation) networks, 5G (5th Generation) network, 5G-Advanced networks and 6G networks.SUMMARY

[0004] Some embodiments of this disclosure will be described with respect to certain aspects. These aspects are not intended to indicate key or essential features of the various example embodiments of this disclosure, nor are they intended to be used to limit the scope of thereof. Other features, aspects, and elements will be readily apparent to a person skilled in the art in view of this disclosure. For example, it should be appreciated that further aspects may be provided by the combination of any two or more of the various aspects described herein.

[0005] In a first aspect there is provided a method comprising providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector, providing the latent vector to an adaptive delta modulation encoder, wherein the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector,wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension and providing the bit stream to a network entity for decoding by a spatial and frequency decoder.

[0006] The spatial and frequency encoder may comprise a machine learning model which is trained in combination with the spatial and frequency decoder and a scalar quantiser and de-quantiser.

[0007] The method may comprise obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information, wherein the initial step size is based on a standard deviation per dimension of the latent vectors of the data set.

[0008] The method may comprise obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information, wherein the initial step size is based on a maximum and minimum value per dimension of the latent vectors of the data set.

[0009] In a second aspect thee is provided a method comprising receiving a bit stream from a user equipment, wherein the bit stream comprises one or more bits per dimension of a latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension, providing the bit stream as input to an adaptive delta modulation decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value, providing the output of the adaptive delta modulation decoder to a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quantiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser and providing the estimated latent vector as input to a spatial and frequency decoder, wherein the output of the spatial and frequency decoder comprises reconstructed channel state information.

[0010] The spatial and frequency decoder may comprise a machine learning model which is trained in combination with a spatial and frequency encoder and the scalar quantiser and de-quantiser.

[0011] The scalar quantizer-dequantizer may be configured to map an adaptive delta modulation decoded value to a pre-defined scalar quantization level.

[0012] In a third aspect there is provided an apapratus comprising means for performing the method according to the first or second aspect.

[0013] In a fourth aspect there is provided an apparatus comprising at least one processor, and at least one memory storing instructions which, when executed by the at least one processor, cause the apparatus at least to perform a method according to the first or second aspect.

[0014] In a fifth aspect there is provided a non-transitory computer readable medium comprising instructions wherein the instructions when executed by at least one processor of an apparatus cause the apparatus to perform the method according to the first or second aspect.

[0015] In a sixth aspect there is provided a computer program comprising instructions which, when executed by an apparatus, cause the apparatus to perform at least the method according to the first or second aspect.

[0016] Some embodiments of the invention are defined in the dependent claims.

[0017] In the above, many different aspects have been described. As previously noted, it should be appreciated that further aspects may be provided by the combination of any two or more of the aspects described above (or otherwise in this disclosure).

[0018] Various other aspects are also described in the following detailed description and in the claims.BRIEF DESCRIPTION OF THE FIGURES

[0019] Some embodiments will be described, by way of non-limiting and illustrative example only, with reference to the figures, in which:

[0020] Fig. 1 shows an example of a communication network to which examples disclosed herein may be applied;

[0021] Fig. 2 shows a schematic diagram of architecture for two-sided CSI compression with recurrent SFT encoder / decoder;

[0022] Fig. 3 shows a schematic diagram of architecture for two-sided CSI compression with recurrent quantization;

[0023] Fig. 4 shows an example of a latent vector sequence;

[0024] Fig. 5 shows an example of a latent vector element distribution;

[0025] Fig. 6 shows an example of latent vector element distribution;

[0026] Fig. 7 shows a schematic diagram of an ADM encoder;

[0027] Fig. 8 shows a flowchart of a method according to an example embodiment;

[0028] Fig. 9 shows a flowchart of a method according to an exampleembodiment;

[0029] Fig. 10 shows a flowchart of a method according to an example embodiment;

[0030] Fig. 11 shows a schematic diagram of architecture for two-sided CSI compression with recurrent quantization employing adaptive delta modulation / demodulation followed by Scalar Quantizer-deQuantizer

[0031] Fig. 12 shows a flowchart of an example procedure;

[0032] Fig. 13 shows a schematic diagram of an example procedure;

[0033] Fig. 14 a schematic diagram of an example ADM encoder;

[0034] Fig. 15 shows an empirical distribution of pre-quantization (left) I postquantization (right) latent vector elements;

[0035] Fig. 16 shows SGCS performances versus overhead bits of ADM-assisted SQ compared to conventional SQ;

[0036] Fig. 17 shows an example of an apparatus.DETAILED DESCRIPTION

[0037] The following embodiments are provided by way of non-limiting and illustrative example. Although the specification may refer to “an”, “one”, or “some” embodiments) in several locations of the text, this does not necessarily mean that each reference is made to the same embodiment(s), or that a particular feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments. Further, when a particular feature, structure, or characteristic is described in connection of an embodiment, it intended such feature, structure, or characteristic may be applied in connection with other embodiments (whether or not explicitly described).

[0038] It shall be understood that although the terms “first,” “second” and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another.

[0039] For the purposes of this disclosure, the phrases “at least one of A or B”, “at least one of A and B”, and “A and / or B” means (A), (B), or (A and B). For the purposes of this disclosure, the phrase “A, B, and / or C” means (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C).

[0040] As used herein, the term “or” refers to a non-exclusive “or” unless otherwise indicated (e.g., use of “or else” or “or in the alternative”).

[0041] As used herein, unless stated explicitly, performing a respective feature, step, or functionality “in response to A” does not indicate that the respective feature, step, or functionality is performed immediately after “A” occurs as one or more intervening features, steps, or functionalities may be performed (at least in part) between an occurrence of the respective feature, step, or function and “A”. Analogously, performing a respective feature, step, or functionality “based on A” does not indicate that the respective feature, step, or functionality is performed solely based on “A” as the respective feature, step, or functionality may be further based on one or more other features, steps, or functionalities in addition to “A”.

[0042] Embodiments described herein may be implemented in a communication network, such as any of the following radio access technologies (RATs): Worldwide Interoperability for Micro-wave Access (WiMAX), Global System for Mobile communications (GSM, 2G), GSM EDGE radio access Network (GERAN), General Packet Radio Service (GRPS), Universal Mobile Telecommunication System (UMTS, 3G) based on basic wideband-code division multiple access (W-CDMA), high-speed packet access (HSPA), Long Term Evolution (LTE), LTE-Advanced, and enhanced LTE (eLTE), 5G (also called NR), or any future RAT such as 6G. Moreover, communication within the communication network may utilize any proper wireless communication technology, comprising but not limited to: Code Division Multiple Access (CDMA), Frequency Division Multiple Access (FDMA), Time Division Multiple Access (TDMA), Frequency Division Duplex (FDD), Time Division Duplex (TDD), Multiple-Input Multiple-Output (MIMO), Orthogonal Frequency Division Multiple (OFDM), and / or Discrete Fourier Transform spread OFDM (DFT-s-OFDM).

[0043] As used herein, the term “network device” or “network node” refers to a node in a communication network via which user equipment may access the network and / or which is configured to control radio communication and managing radio resources within a cell. The network node or network device may be referred to as a base station (BS), an access point (AP) or an access node. The network device may be, depending on the applied technology, for example, a node B (NodeB or NB), an evolved NodeB (eNodeB or eNB), an NR NB (also referred to as a gNB), a Remote Radio Unit (RRU), a radio head (RH), a remote radio head (RRH), a relay, an Integrated Access and Backhaul (IAB) node, a low power node, a non-terrestrial network (NTN) or nonground network device, such as a satellite network device, a low earth orbit (LEO) satellite and a geosynchronous earth orbit (GEO) satellite, or an aircraft network device.

[0044] Moreover, in connection of split radio access network (RAN), the networkdevice may refer to a centralised unit (CU) of a base station and / or a distributed unit (DU) of a base station. An interface between CU and DU may be referred to as an F1 interface in NR. In the split RAN architecture, node operations may be carried out, at least partly, in the central / centralized unit, CU, (e.g. server, host or node) operationally coupled to the DU, (e.g. a radio head / node). One CU may control one or more DUs, acting at least as transmit / receive (Tx / Rx) nodes. In some embodiments, the DUs may comprise e.g. a radio link control (RLC), medium access control (MAC) layer and a physical (PHY) layer, whereas the CU may comprise the layers above RLC layer, such as a packet data convergence protocol (PDCP) layer, a radio resource control (RRC) and an internet protocol (IP) layers. Other functional splits are possible too. In practice, any processing task may be performed in either the CU or the DU and the boundary where the responsibility is shifted between the CU and the DU may depend on the applied implementation.

[0045] The term “terminal device” refers to any end device that may be configured to perform wireless communication. By way of example, a terminal device may be referred to as a communication device, user equipment (UE), a Subscriber Station (SS), or a Mobile Station (MS). The terminal device may include a mobile phone, a cellular phone, a smart phone, voice over IP (VoIP) phones, wireless local loop phones a tablet, a wearable terminal device, a personal digital assistant (PDA), portable computers, desktop computer, image capture terminal devices such as digital cameras, gaming terminal devices, music storage and playback appliances, vehicle-mounted wireless terminal devices, USB dongles, an Internet of Things (loT) device, a watch or other wearable, a head-mounted display (HMD), a vehicle, a drone, a medical device and applications (e.g., remote surgery), an industrial device and applications (e.g., a robot and / or other wireless devices operating in an industrial and / or an automated processing chain contexts), a consumer electronics device, a device operating on commercial and / or industrial wireless networks, and the like.

[0046] A term “resource”, as used herein, may refer to radio resources in time domain, in frequency domain, in space domain, and / or in code domain. Some examples of resources may include, e.g., a physical resource block (PRB), a radio frame, a subframe, a time slot, a subband, a frequency region, a sub-carrier, a beam, etc. The term “transmission” and / or “reception” may refer to wirelessly transmitting and / or receiving via a wireless propagation channel on radio resources.

[0047] Fig. 1 illustrates an example of a communication network to which examples disclosed herein may be applied. The communication network or a cellularcommunication network may comprise a network node 110 configured to provide one or more cells, such as cell 100, and a network node 112 configured to provide one or more other cells, such as cell 102. Each cell may, for example, be a macro cell, a micro cell, femto, or a pico cell. The cell may define a coverage area or a service area of the corresponding access node.

[0048] The network node (110, 112) may be configured to provide a user equipment (UE) 120 (one or more UEs) with wireless access to the communication network. The wireless access may comprise downlink (DL) communication from the network node (110, 112) to the UE 120 and uplink (UL) communication from the UE 120 to the network node (110, 112). Examples of uplink channels may comprise physical uplink control channel (PUCCH) for transmitting control information and physical uplink shared channel (PUSCH) for transmitting data towards the network. Examples of downlink channels may comprise physical downlink control channel (PDCCH) for transmitting control information and physical downlink shared channel (PDSCH) for transmitting data towards the user equipment.

[0049] There may be a plurality of UEs (120, 122) in the system. Each of the plurality of UEs may be served by the same or by different network nodes (110, 112). UE may be configured with dual connectivity (DC), wherein the UE, for example UE 120, may be connected to multiple network nodes (110, 112). The UEs (120, 122) may communicate with each other, in case device-to-device (D2D) communication interface is established between them via a so-called sidelink (SL). Such D2D communications may be referred to as machine-to-machine, peer-to-peer (P2P) communications, or ve- hicle-to-vehicle (V2V), for example.

[0050] In the case of multiple network nodes in the communication network, the network nodes may be connected to each other via an interface. LTE specifications, for example, refer to such an interface as an X2 interface. An interface between an LTE node and a 5G node, or between two 5G nodes may be called an Xn interface.

[0051] The network nodes 110 and 112 may be further connected via another interface to a core network 116 of the communication network. The LTE specifications specify the core network as an evolved packet core (EPC), and the core network may comprise a plurality of entities (e.g. a mobility management entity (MME) and a gateway node). The MME may handle mobility of terminal devices in a tracking area encompassing a plurality of cells and handle signalling connections between the terminal devices and the core network. The gateway node may handle data routing in the core network and to / from the terminal devices. The 5G specifications specify the corenetwork as a 5G core (5GC). The 5GC may, for example, comprise an access and mobility management function (AMF) and a user plane function / gateway (UPF) and other functions. The AMF may handle termination of non-access stratum (NAS) signalling, NAS ciphering & integrity protection, registration management, connection management, mobility management, access authentication and authorization, security context management. The UPF node may, for example, support packet routing and forwarding, packet inspection and quality of service (QoS) handling.

[0052] AI / ML based channel state information (CSI) compression using two- sided model (two sides of an over the air (OTA) interface, e.g., UE and network entity) is being considered. One topic under discussion is temporal domain aspects on top of spatial and frequency domain compression.

[0053] Consideration of temporal domain aspect may lead to overall performance enhancement.

[0054] Two potential architectures for spatial / temporal / frequency (SFT) compression have been proposed.

[0055] Fig. 2 shows a potential architecture with recurrent SFT encoder / decoder. Here, history dependence is added to the encoder and decoder by feeding back some internal state variables with delay. This augmented encoder can be referred to as a recurrent SFT (spatial-frequency-time) encoder, and similarly for the decoder side.

[0056] An alternative architecture for exploiting the time correlation in CSI compression is to add recurrence to the quantizer / inverse quantizer (a.k.a. dequantizer) stage instead of the dimensionality reduction stage, as depicted in Fig. 3. In this approach, mapping to and from the latent space is done using an SF encoder decoder pair. The quantizer and inverse quantizer are augmented by feeding back state variables s and st' respectively, with delay. This structure may allow the quantizer to achieve lower mean squared error with a given number of bits in btwhen the CSI matrices Vtare correlated in time. This is due to the reason that the latent vectors ztinherit timedependence properties like those of the CSI matrices.

[0057] Recurrent quantizer-inverse quantizer operation may be enabled either by AI / ML based scheme or by an explicitly designed scheme. According to numerical evaluation results, a SF encoder / decoder with an explicitly designed recurrent quantizer may outperform both SF encoder / decoder with AI / ML based recurrent quantizer and recurrent SFT encoder / decoder with reduced complexity. The latent vector elementswhich, in an example, have been collected per 5 ms interval generally vary smoothly in time, which renders element-wise tracking appealing.

[0058] Current proposals for using a recurrent quantizer for AI / ML-enabled CSI compression (two-sided model) in the context of SFT involve an explicitly designed quantizer to track each individual latent vector element (or latent vector dimension) separately, with the predetermined number of 1- or 2-bit(s) resources.

[0059] A signal typically changes by less than a bounded amount (a typical delta) in each time step. The proposed algorithm keeps track of estimated upper and lower bounds on the signal at time instance t. This interval quantifies the uncertainty in the value of the signal. Quantization region boundaries are adapted to divide up the uncertainty interval. For example, in the case of 1 bit quantization the quantization boundary is put in the middle of the uncertainty interval.

[0060] Besides the coded bits per each feedback instance, a few additional bits (common to all latent dimensions; typical value in use is 4) are used to indicate a typical delta for each time step (common to all latent dimensions) out of the agreed set to use for uncertainty tracking.

[0061] It requires a few additional bits (e.g., 4 bits) per feedback instance for indication of “typical delta”, which is common to all latent dimensions. In the case of 64 latent vector dimension with 1 or 2 bit(s) per dimension recurrent quantization bit allocation to be assisted by 16 pre-defined common step size candidates, every feedback instance requires 68 bits (64*1+log2(16)) or 132 bits (64*2+ Iog2(16)), respectively.

[0062] All the individual latent dimensions are treated indifferently, whereas there exist “outlier dimensions” when looking into per-dimension latent vector data distribution. for indication of “typical delta” (common step size), pre-defined (offline prepared) set of candidate common step sizes is to be used per feedback of the entire latent vector, irrespective of latent vector element index. This implies that per each feedback, the chosen common step size should be applied to every latent vector element at this time instance. This may be acceptable so long as all the latent vector elements experience similar transient behavior with respect to their preceding values (as illustrated in Fig. 4 which shows the trajectories of latent vector sequences obtained by applying an SF encoder to a sequence of 50 CSI matrices in the top graph and the difference between successive latent vectors at consecutive time steps). However, it has also been observed in an independent study that overall distribution of the latent vector elements may show different behavior for some elements as can be observed in Fig. 5 (which shows an example of latent vector distribution for SF encoder-decoder without aquantizer-dequantizer inbetween). Even though most of the collected latent vector elements follows Gaussian distribution, there exist a few outliers as well. A similar trend can be observed in Fig. 6 for the case of quantization-aware training case (uniform scalar quantization (SQ)). Fig. 6 shows the distribution of the latent vector elements in case of model training with 1 / 2 / 3 bit(s) uniform scalar quantization per dimension (left: taken at the output of the encoder, right: taken at the output of scalar quantizer; top: Encoder-Decoder (ENC-DEC) trained with 1 bit uniform SQ in-between, middle: with 2 bits uniform SQ in-between, bottom: with 3 bits uniform SQ in-between). This implies that tracking performance (in terms of MSE) can be further improved by allowing dedicated step size per latent vector element / dimension. CSI encoder may be considered as CSI compression module, CSI decoder as CSI reconstruction module.

[0063] Adaptive delta modulation has been widely used in tracking of signals with correlation, e.g., coding TV, speech signals. It has also been proposed for feedback of the Givens parameters for Ml MO systems with slowly time-varying channels as a means of practical low-rate scalar coding scheme. There are variants of ADM implementation. One of the main features of the ADM is its adaptively changing step size in proportion to the input signal strength.

[0064] Assume that we want to construct a modulated (dequantized) signal <p[k] that tracks a signal <p[k]. This can be achieved according to the following construction. At each instant of time, we start with the value <p[k - 1] and update it to <p[k] so that this new value is closer to <p[k] than its old value. The update is based on the difference between <p[k] and <p[k - 1], defined by

[0065] ea[k] = p[k] ~ <p[k - l]. (Eq.1)

[0066] The signal <p[k - 1] is increased or decreased by a positive amount A[ ] depending on the encoder output history and the sign of the error (Eq.1). The step-size A[ ] of the one-bit quantization is adaptively changing to better track the dynamics of the signal. The step-size is increased if the consequent two encoded bits are same, and decreased otherwise, that is,

[0067] where A[ ] and c [Zc] = sign(ea[k]) e {-1, +1} are the step-size and the encoded bit for the -th sample, respectively, a is a system parameter, which satisfies a >1. The sign of the error (sign(ea[k]y) decides whether <p[k - 1] increases or decreases at each time instant. Thus the signal <p [ / c] is varied according to the adaptation rule:

[0068] <p[k] = <p[k — 1] + sign ea[k]~) ■ A[ / c]. (Eq.3)

[0069] An observation of the step-size A [Zc] reveals the following equivalent form.(Eq.4)

[0071] Here w[ / c] (w[0] = 0) can be considered as a weight factor which reflects the history of the delta modulation trend (Eq.5) and is used to decide the step-size (Eq.4). This alternative representation allows us to describe the scheme for updating <p[k] in a block diagram form, as shown in Fig. 7.

[0072] The upper and lower parts of the figure implement equations (Eq.4) and (Eq.3), respectively. System parameters a and A[0] (eventually A[Zc]) are assumed to be known to the ADM encoder and decoder, and only the encoded bit c [Zc] is required for the receiver to decode the quantized value <p[k]. The ADM is a low-rate scalar quantization scheme (as low as 1 bit per parameter), and it has inherently a channel tracking feature for the slowly varying channels over time.

[0073] Fig. 8 shows a flowchart of a method according to an example embodiment. The method may be performed at an apparatus. The apparatus may, e.g., be a UE or a network entity.

[0074] At 801 the method comprises obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information.

[0075] At 802, the method comprises determining a first parameter of an adaptive delta modulation scheme suitable for temporal encoding based on the data set, wherein the first parameter is an initial step size per dimension of a latent vector.

[0076] At 803, the method comprises providing the first parameter to a user equipment comprising the spatial and frequency encoder and an adaptive delta modulation encoder, wherein the output of the spatial and frequency encoder comprises a latent vector and wherein the output of the adaptive delta modulation encoder is a bit stream comprising adaptive delta modulation encoded values for each dimension of the latent vector based on the initial step size for the respective dimension.

[0077] Alternatively, or in addition, the method may comprise providing the first parameter to a network entity comprising an adaptive delta modulation decoder, and aspatial and frequency decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value and the output of the spatial and frequency decoder comprises reconstructed channel state information. The network entity may comprise a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quantiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser.

[0078] The initial step size may be based on a standard deviation per dimension of the latent vectors of the data set or on maximum and minimum value per dimension of the latent vectors of the data set.

[0079] The method as described with reference to Fig. 8 may comprise determining a second parameter of the adaptive delta modulation scheme based on the data set, wherein the second parameter is a step size scaling factor common to all dimensions of the latent vector and providing the second parameter to at least one of the user equipment and the network entity, wherein the output of the adaptive delta modulation encoder is a bit stream comprising adaptive delta modulation encoded values for each dimension of the latent vector based on the step size scaling factor for all dimensions of the latent vector.

[0080] The step size scaling factor may be a value which minimises a mean squared error over the data set.

[0081] Fig. 9 shows a flowchart of a method according to an example embodiment. The method may be performed at an apparatus. The apparatus may be, comprise or be comprised in a UE.

[0082] At 901 , the method comprises providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector.

[0083] At 902, the method comprises providing the latent vector to an adaptive delta modulation encoder, wherein the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension.

[0084] At 903, the method comprises providing the bit stream to a network entity for decoding by a spatial and frequency decoder.

[0085] Fig. 10 shows a flowchart of a method according to an example embodiment. The method may be performed at an apparatus. The apparatus may be, comprise or be comprised in a network entity.

[0086] At 1001 , the method comprises receiving a bit stream from a user equipment, wherein the bit stream comprises one or more bits per dimension of a latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension.

[0087] At 1002, the method comprises providing the bit stream as input to an adaptive delta modulation decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value.

[0088] At 1003, the method comprises providing the output of the adaptive delta modulation decoder to a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quantiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser.

[0089] At 1004, the method comprises providing the estimated latent vector as input to a spatial and frequency decoder, wherein the output of the spatial and frequency decoder comprises reconstructed channel state information.

[0090] The scalar quantizer-dequantizer may be configured to map an adaptive delta modulation decoded value to a pre-defined scalar quantization level.

[0091] The spatial and frequency encoder and the spatial frequency decoder may comprise a machine learning model which is trained in combination with the spatial and frequency decoder and the scalar quantiser and de-quantiser (e.g., trained in an SQ aware manner).

[0092] The method as described with reference to Figs. 8 to 10 incorporates an adaptive delta modulation (ADM) for recurrent quantization for a AI / ML-enabled CSI compression (two-sided model) use case, to exploit temporal correlation of the latent vector in a feedback resource efficient manner. ADM pertains to an innate feature of automatic scaling of step size, which has shown its effectiveness of tracking temporally correlated signal at the presence of sporadic sudden rapid level changes. Numerical validation validates its superior performance in terms of MSE and SGCS.

[0093] An example overall system architecture is depicted in Fig. 11. A scalar quantizer-dequantizer module at the output of Adaptive Delta Demodulator at NW-side, to map element-wise floating-point value to one of the pre-defined scalar quantization levels. ADM is used to keep track of the floating-point value of each latent vector dimension with a low cost of 1 bit / dim to reduce feedback overhead, and at the NW side it is followed by scalar quantizer-dequantizer to map a floating-point value to a quantized level at the input to CSI decoder (CSI reconstruction model) with which its underlining CSI encoder / decoder have been trained in a SQ-aware manner.

[0094] In an example embodiment, two system configuration parameters of ADM, e.g., the first parameter and the second parameter described with reference to Fig. 8, may be determined as follows.

[0095] The first parameter may be initial step size A;[0] (where i is latent vector element index). The initial step size may be determined by taking a standard deviation of the collected latent vector elements. Dedicated initial step size value may be designated per each latent vector element / dimension, to reflect possibly different statistical behaviour of individual latent vector elements (dimensions). This is an example of the initial step size is based on a standard deviation per dimension of the latent vectors of the data set.

[0096] In an example embodiment, may be determined by a rule-of-thumb, for example, by taking ft ■ (z™ax- z™in) where is a constant (typical value: 0.145, 0.20, etc.), and z’"az, z’"mis a maximum, minimum value of the collected latent vector elements per index i, respectively. This is an example where the initial step size is based on a maximum and minimum value per dimension of the latent vectors of the data set.

[0097] Table 1 indicates benefit of configuring dimension-wise initial step size, rather than single common value for all dimensions. One can take a standard deviation of the collected latent vector elements per dimension or take 20% of peak-to-peak value, etc.Table 1

[0098] Pre-defined clipping thresholds can be set to limit operational range when it comes to ADM-assisted tracking of latent vector elements.

[0099] In one embodiment, the clipping thresholds (±ZTh) may be set such that at least the certain portion (for example, 95%) of the collected latent vector element samples should be lying in the in-between region, i.e., (~ZTh, +ZTh). This can be done by examining the empirical probability density function on the collected latent vector element samples.

[0100] The second parameter may be step size scaling factor a (a > 1). Step size scaling factor may be determined by monitoring MSE when trying multiple candidate values over the collected latent vector data set (in a numerical analysis manner).The value which leads to the minimum MSE shall be chosen. This is an example where the step size scaling factor is a value which minimises a mean squared error over the data set. The step size scaling factor may be chosen as common value for all the latent vector elements. Recommended value which has been acquired by numerical experiments is 1.92.

[0101] The parameters may be configured offline prior to deployment of the models, e.g., encoder, decoder, and ADM (encoder / decoder) to be followed by quantizer- dequantizer as depicted in Fig. 11. For example, the first and second parameter may be pre-determined and aligned offline between UE and NW vendors, so it does not require additional overhead in the air interface signaling. The ADM-configuration parameters below may also be exchanged between NW and UE as additional information being associated with training dataset or encoder parameters.

[0102] The whole model training and deployment procedure follows two step approach, e.g., Step A: E2E CSI encoder / decoder quantization-aware model training stage (including determination of the first parameter and the second parameter) and Step B: Inference stage with ADM being incorporated.

[0103] Fig. 12 shows a flowchart of an example model training and deployment procedure. Fig. 13 shows a schematic diagram of E2E model training and inference (deployment) procedures for ADM-assisted recurrent quantizer scheme for AI / ML-ena- bled CSI compression use case.

[0104] At step 1 , E2E model training is performed with SF Encoder and SF decoder being directly connected with scalar quantizer - dequantizer in-between (quantization-aware training). Any training collaboration typemay be supported, e.g., joint training, joint sequential training, separate sequential training, etc., and any inter-vendor collaboration direction I option which allows UE-side offline engineering can be supported.

[0105] At step 2, once SF encoder I decoder model training is completed, the trained model parameters need to be saved. With trained model parameters (weights) being loaded, input target CSI data can be fed into the SF encoder model to generate corresponding latent vectors at the output of CSI encoder prior to quantizer, which are to be collected to form the dataset of the latent vectors.

[0106] At step 3, based on the collected latent vectors in step 2, the ADM system configuration parameters, e.g., step size scaling factor and initial step size, are determined.

[0107] ADM configuration parameters may be determined at the very last stage,i.e. , actual inference using the (SQ-aware) trained UE and NW models. With the trained UE model, output at the UE model can be collected by injecting UE side data (target CSI). These collected unquantized latent values may be used for determination of (common) step size scaling factor and (per dimension) initial step sizes. It is assumed that UE shall share these parameter values with NW prior to deployment.

[0108] Step 4 is the first step of Step B, the model inference / deployment stage. Load the trained model parameters / weights to SF encoder and SF decoder. The ADM encoder and ADM decoder being followed by scalar quantizer-dequantizer are plugged in at the output of SF encoder and at the input of SF decoder, respectively. ADM encoder and decoder work per dimension independently. A Bit stream of size of the latent vector dimension (e.g., 1 bit / dim) is transmitted over the air.

[0109] The chain with SF encoder - ADM encoder - ADM decoder - Quantizer- deQuantizer - SF decoder is run in an inference mode for E2E performance validation. Here, SF encoder and ADM encoder are located UE-side, while ADM decoder, Quan- tizer-deQuantizer, and SF decoder are located at the NW-side.

[0110] As ADM is designed to track (unquantized) floating point value with small (e.g., 1 or 2 bits) per dimension overhead, OTA overhead can be reduced.

[0111] At step 5, upon successful E2E performance validation, deploy the E2E model chain with ADM-assisted recurrent quantizer - dequantizer being plugged-in.

[0112] Fig. 14 depicts an example schematic of ADM encoder / decoder in more detail. ADM is a scalar encoding scheme which works on each individual dimension of the latent vector. The ADM encoder side requires the whole components of the schematic in Fig. 14 (see the modules enclosed by dashed line), whereas ADM decoder side needs only subset of it, which is indicated by solid line, tagged as “Common modules for encoder and decoder / Modules for decoder”.

[0113] As ADM encoder and decoder share the system configuration parameters, i.e., (common) step size scaling factor a and (element-wise) initial step size Aj[O], ADM encoder and decoder can keep track of the modulated (recurrently dequantized) value of the latent vector element.

[0114] ADM encoder modulates (quantizes) its current input value <p[k] by comparing it with the previously modulated value <p[k - 1] to generate an encoded binary value c [Zc] , which is to be transmitted over the air to ADM decoder side.

[0115] ADM encoder also updates the latest modulated value to be <p[k] via (Eq.3) for next instance operation, by keeping track of element-wise step size A k] via (Eq.2).

[0116] ADM decoder reconstructs (dequantizes) the currently modulated value <p[k] by keeping track of element-wise step size A k] via (Eq.2) independently.

[0117] For ADM-assisted recurrent quantizer for AI / ML-enabled CSI compression, the underlying CSI encoder-decoder should to be trained with a certain resolution of the scalar quantization in mind (Step A). As ADM’s tracking has been known to work well with data following a Gaussian-like data distribution, we recommend selecting a high-resolution scalar quantization, e.g., nq> 2 [bits / dim]. According to observation, as the resolution of the quantizer increases, the distribution of the unquantized elements appears to morph into the normal distribution, i.e. , the distribution of the quantization- free training case as shown in Figs. 5 and 6. Motivated by this, nq= 4 [bits / dim] for the numerical validation.

[0118] Fig. 15 shows the empirical distribution of pre-quantization (left) I postquantization (right) latent vector elements (aggregated / flattened over dimensions; uniform scalar quantization with 4 bits / dimension for 128 dimensions / feedback.

[0119] Table 2 describes that ADM could reduce MSE greatly with respect to the conventional SQ scheme, by exploiting temporal correlations at the latent domain.Table 2

[0120] Table 3 and its graphical version in Fig. 16 show that ADM-based recurrent quantization scheme could achieve -50% of overhead reduction (or up to -11 .4% SGCS gain per same overhead) compared to spatial-frequency (SF) domain compression with conventional SQ.Table 3

[0121] The proposed scheme shows better MSE improvements, and its E2E SGCS performance enhancements are on par with (in terms of overhead reduction in terms of %) / better than (in terms of % increase of SGCS at 128 dimensions) the previously proposed schemes with less feedback overhead (the method does not require extra feedback (e.g., 4 bits) for common delta size indication).

[0122] Methods as described above may provide advantages of the recurrent quantizer scheme with respect to SFT encoder / decoder architecture. These include modularity, e.g., design and standardization of the SF encoder I decoder pair can be essentially independent of the design and standardization of the recurrent quantizer. All of the previous works done on SF encoder / decoder can be maintained. One advantage is reduced complexity. Since the recurrent quantization is done in latent space, the timebased processing is done in lower dimension than time-based processing for SFT encoder / decoder. Recurrent quantizer can be done with many fewer parameters and with much simpler architectures than are needed for dimensionality reduction. This has operational benefits, but also standardization benefits. The recurrent quantizer I inverse quantizer can be tested separately from the SF encoder / decoder, using an intermediate KPI of mean squared error in latent space.

[0123] Methods as described above may be agnostic to training collaboration type or inter-vendor collaboration direction I options which allow UE-side offline engineering under 3GPP discussion.

[0124] There is no need of additional bits for indication of “typical delta” (common step size): ADM regulates the step size automatically. Only 1 bit per latent vector element is required for feedback, and no additional bits are needed.

[0125] Methods may provide reduced MSE. According to our independent numerical validation, MSE for latent vector dimension of 64 is 0.0069 in case of using ADM for recurrent quantization (1 bit per element; 64 bits overhead per feedback). This is lower than the case of “conventional 2-bit / dim (128 bits overhead per feedback)” (MSE: 0.0211 ; more in Table 2).

[0126] Methods may provide a means of treating each latent element differently via configuring a dedicated initial step size per each latent vector dimension: when taking a 20% of peak-to-peak value per dimension as the initial step size, trackingperformance improvement is -0.028 reduction in terms of MSE for latent vector dimension of 64 (Table 1).

[0127] Fig. 17 shows, by way of example, a block diagram of an apparatus 10. The apparatus 10 comprises, for example, at least one processor 12 and at least one memory 14 storing instructions 15 that, when executed by the at least one processor, cause the apparatus 10 at least to perform the method or methods (or portion(s) thereof) as disclosed herein, and any of the embodiments (or respective portion(s) thereof). In an example, the at least one memory and the instructions (e.g. a computer program code, software), are configured, with the at least one processor, to cause the apparatus 10 to perform the method or methods (or portion(s) thereof) as disclosed herein, and any of the embodiments (or respective portion(s) thereof).

[0128] A processor 12 may comprise circuitry, or be constituted as circuitry or circuitries, the circuitry or circuitries being configured to perform phases of methods in accordance with embodiments described herein.

[0129] As used herein, the term “circuitry” may refer to one or more or all of the following: (a) hardware-only circuit implementations, such as implementations in only analog and / or digital circuitry, and (b) combinations of hardware circuits and software, such as, as applicable: (i) a combination of analog and / or digital hardware circuit(s) with software / fi rmware and (ii) any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a user equipment, to perform various functions) and (c) hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of a microprocessors), that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation. This definition of circuitry applies to all uses of this term herein, including in any claims. As a further example, as used herein, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.

[0130] The memory 14 may be implemented using any suitable data storage technology. The memory may comprise a database for storing data. The memory 14 may, for example, be at least in part external to apparatus 10 but accessible to apparatus 10.

[0131] The instructions 15 may be comprised in a computer readable medium or a non-transitory computer readable medium. A term non-transitory, as used herein, is a limitation of the medium itself (i.e. tangible, not a signal) as opposed to a limitation on data storage persistency (e.g. random access memory, RAM, vs. read only memory, ROM).

[0132] For example, the apparatus 10 is a terminal device, such as a UE. As another example, the apparatus is comprised in such a terminal device, e.g. as a chipset configured to control the terminal device. The apparatus 10 may be caused or configured or comprise means to perform at least the method of Figs. 8 to 10 and / or any one or more of the embodiments described herein.

[0133] As another example, the apparatus 10 is a network entity. In another embodiment, the apparatus is comprised in such a network entity, e.g. as a chipset configured to control the network entity. The apparatus 10 may be caused or configured or comprise means to perform at least the method of Figs. 8 to 10 and / or any one or more of the embodiments described herein.

[0134] The apparatus may comprise one or more entities of any of protocol layers, such as a MAC entity, an RRC entity, an RLC entity, a PDCP entity or a PHY entity. In some embodiments, the entity is configured to perform at least the method of Figs. 5, and / or any one or more of the embodiments described.

[0135] The apparatus 10 comprises a radio interface 16. The radio interface16 may provide the apparatus 10 with communication capabilities. The radio interface 16 may comprise a receiver configured to receive information in accordance with at least one cellular or non-cellular standard. The radio interface 16 may comprise a transmitter configured to transmit information in accordance with at least one cellular or non- cellular standard. The receiver may comprise more than one receiver. The transmitter may comprise more than one transmitter. The radio interface 16 may comprise a transceiver configured to receive and transmit information in accordance with at least one cellular or non-cellular standard. The transceiver may comprise more than one transceiver.

[0136] The apparatus 10 may comprise a user interface 18 comprising, for example, at least one of a keypad, a microphone, a touch display, a display, a speaker, etc. The user interface 18 may be used to control the apparatus by the user. The user interface 18 may be external to the apparatus 10. For example, the apparatus 10 may be connected to another device, such as a computer, either via wireless or wired connection, and the apparatus 10 is controlled by the user via the computer.

[0137] In an embodiment, at least some of the processes described herein may be carried out by an apparatus comprising means for carrying out at least some of the described processes. Means for performing method steps as disclosed herein may include software and / or hardware components of the apparatus 10. For example, the at least one processor 12, the memory 14, and the computer program code form means for carrying out the method or methods (or portion(s) thereof) as disclosed herein, and any of the embodiments (or respective portion(s) thereof). As used herein the term “means” is to be construed in singular form, i.e. referring to a single element, or in plural form, i.e. referring to a combination of single elements. Therefore, terminology “means for [performing A, B, C]”, is to be interpreted to cover an apparatus in which there is only one means for performing A, B and C, or where there are separate means for performing A, B and C, or partially or fully overlapping means for performing A, B, C. Further, terminology “means for performing A, means for performing B, means for performing C” is to be interpreted to cover an apparatus in which there is only one means for performing A, B and C, or where there are separate means for performing A, B and C, or partially or fully overlapping means for performing A, B, C.

[0138] Even though this disclosure has been described above with reference to non-limiting and illustrative examples according to the accompanying figures, it is clear that the scope of this disclosure is not restricted thereto - but can be modified in many different ways. As technology advances, it will become apparent to a person skilled in art as to how the disclosure can be further implemented and / or modified in various ways. Further, it is clear to a person skilled in the art that the embodiments described herein may, but are not required to, be combined in various ways with other embodiments described herein.

Claims

22WE CLAIM:1 . An apparatus comprising means for: providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector; providing the latent vector to an adaptive delta modulation encoder, wherein the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; and providing the bit stream to a network entity for decoding by a spatial and frequency decoder.

2. The apparatus according to claim 1 , wherein the spatial and frequency encoder comprises a machine learning model which is trained in combination with the spatial and frequency decoder and a scalar quantiser and dequantiser.

3. The apparatus according to claim 1 or claim 2 comprising means for: obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information, wherein the initial step size is based on a standard deviation per dimension of the latent vectors of the data set.

4. The apparatus according to claim 1 or claim 2, comprising means for: obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information, wherein the initial step size is based on a maximum and minimum value per dimension of the latent vectors of the data set.

5. An apparatus comprising means for: receiving a bit stream from a user equipment, wherein the bit stream comprises one or more bits per dimension of a latent vector, wherein the one ormore bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; providing the bit stream as input to an adaptive delta modulation decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value; providing the output of the adaptive delta modulation decoder to a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quan- tiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser; and providing the estimated latent vector as input to a spatial and frequency decoder, wherein the output of the spatial and frequency decoder comprises reconstructed channel state information.

6. The apparatus according to claim 5, wherein the spatial and frequency decoder comprises a machine learning model which is trained in combination with a spatial and frequency encoder and the scalar quantiser and dequantiser.

7. The apparatus according to claim 6, wherein the scalar quantizer-dequan- tizer is configured to map an adaptive delta modulation decoded value to a pre-defined scalar quantization level.

8. An apparatus comprising at least one processor, and at least one memory storing instructions which, when executed by the at least one processor, cause the apparatus at least to perform: providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector; providing the latent vector to an adaptive delta modulation encoder, wherein the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; andproviding the bit stream to a network entity for decoding by a spatial and frequency decoder.

9. An apparatus comprising at least one processor, and at least one memory storing instructions which, when executed by the at least one processor, cause the apparatus at least to perform: receiving a bit stream from a user equipment, wherein the bit stream comprises one or more bits per dimension of a latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; providing the bit stream as input to an adaptive delta modulation decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value; providing the output of the adaptive delta modulation decoder to a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quantiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser; and providing the estimated latent vector as input to a spatial and frequency decoder, wherein the output of the spatial and frequency decoder comprises reconstructed channel state information.

10. A method comprising: providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector; providing the latent vector to an adaptive delta modulation encoder, wherein the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; and providing the bit stream to a network entity for decoding by a spatial and frequency decoder.2511 . The method according to claim 10, wherein the spatial and frequency encoder comprises a machine learning model which is trained in combination with the spatial and frequency decoder and a scalar quantiser and dequantiser.

12. The method according to claim 10 or claim 11 comprising means for: obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information, wherein the initial step size is based on a standard deviation per dimension of the latent vectors of the data set.

13. The method according to claim 10 or claim 11 , comprising means for: obtaining a data set from a spatial and frequency encoder, the data set comprising latent vectors determined by the spatial and frequency encoder based on channel state information, wherein the initial step size is based on a maximum and minimum value per dimension of the latent vectors of the data set.

14. A method comprising: receiving a bit stream from a user equipment, wherein the bit stream comprises one or more bits per dimension of a latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; providing the bit stream as input to an adaptive delta modulation decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value; providing the output of the adaptive delta modulation decoder to a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quantiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser; and providing the estimated latent vector as input to a spatial and frequency decoder, wherein the output of the spatial and frequency decoder comprises reconstructed channel state information.2615. A computer program comprising instructions which, when executed by an apparatus, cause the apparatus to perform: providing channel state information as input to a spatial and frequency encoder, wherein the output of the spatial and frequency encoder is a latent vector; providing the latent vector to an adaptive delta modulation encoder, wherein the output of the adaptive delta modulation encoder is a bit stream comprising one or more bits per dimension of the latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; and providing the bit stream to a network entity for decoding by a spatial and frequency decoder.

16. A computer program comprising instructions which, when executed by an apparatus, cause the apparatus to perform: receiving a bit stream from a user equipment, wherein the bit stream comprises one or more bits per dimension of a latent vector, wherein the one or more bits per dimension indicate an adaptive delta modulation encoded value determined based on an initial step size for the respective dimension; providing the bit stream as input to an adaptive delta modulation decoder, wherein the output of the adaptive modulation decoder is an adaptive delta modulation decoded value; providing the output of the adaptive delta modulation decoder to a scalar quantiser and de-quantiser, wherein the output of the scalar quantiser and de-quantiser is an estimated latent vector of which each dimension has a quantized discrete scalar value as defined by the scalar quantiser; and providing the estimated latent vector as input to a spatial and frequency decoder, wherein the output of the spatial and frequency decoder comprises reconstructed channel state information.

Images