Methods and apparatuses relating to compression of downlink control information

Abstract

A terminal device (UE) comprising: means for transmitting, to a network, capability information which indicates that the terminal device is capable of handling compressed downlink control information,

Description

Field Various example embodiments relate to efficient radio resource utilization in control information transmissions. More specifically, various example embodiments relate to measures (including methods, apparatuses and computer program products) for realizing efficient radio resource utilization in control information transmissions. Background The present specification generally relates to transmission of control information, the compression thereof, as well as signalling in relation thereto. With a growing demand for wireless communication systems, radio resources become precious. A higher efficiency of using limited bandwidth provides a larger capacity. In particular, it is presumed that 3rd Generation Partnership Project (3GPP) 6th Generation (6G) networks will need to handle massive numbers of devices, and it is expected that even more device types will connect to the 6G network than to the current 5th Generation (5G) system. To precisely control these large number of devices, control channels need to efficiently utilize the limited resources. Summary The scope of protection sought for various embodiments of the invention is set out by the independent claims. The embodiments and features, if any, described in this specification that do not fall under the scope of the independent claims are to be interpreted as examples useful for understanding various embodiments of the invention. According to a first aspect, this specification describes a terminal device. The terminal device according to the first aspect comprises: means for transmitting, to a network, capability information which indicates that the terminal device is capable of handling compressed downlink control information, DCI; means for receiving, from a network node of the network and subsequent to transmitting the capability information, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI; means for storing the trained model at the terminal device; means for receiving a first instance of compressed DCI from the network node; and means for using the stored trained model to decompress the first compressed DCI. In some examples, the terminal device comprises means for, subsequent to storing the trained model at the terminal device and prior to receiving the first compressed DCI, transmitting, to the network node, an indication that the network node can begin transmitting compressed DCI to the terminal device. In some examples, the first information includes an identifier associated with the trained model. In some such examples, the terminal device comprises means for downloading the trained model from a remote storage system based on the identifier. In some examples, the first information includes the trained model. In some such examples the terminal device comprises means for, prior to receiving the first information including the trained model, receiving, from the network node, an indication that the trained model is available for delivery to the terminal device. In some examples, the terminal device comprises means for receiving second information identifying which previously-received instances of DCI are to be used by the terminal device when decompressing the first compressed DCI using the stored trained model. In some examples, the terminal device comprises means for receiving an indication that the stored trained model is to be updated. The terminal device may comprise means for, subsequent to receiving the indication that the stored trained model is to be updated, receiving, from the network node, multiple instances of uncompressed DCI. The terminal device may further comprise: means for storing the multiple instances of uncompressed DCI in a training buffer at the terminal device; and means for updating the trained model using the multiple instances of uncompressed DCI stored in the training buffer. In addition, the terminal device may comprise: means for transmitting, to the network, capability information which indicates that the terminal device is capable of updating a model that is to be used by the terminal device for decompressing compressed DCI; and means for receiving, from the network node, an indication that the terminal device should begin updating the trained model, wherein the trained model is updated using the multiple instances of uncompressed DCI stored in the training buffer in response to receiving the indication that the terminal device should begin updating the trained model. The terminal device may comprise: means for, subsequent to receiving the indication that the trained model is to be updated, receiving, from the network node, third information relating to an updated model that is to be used by the terminal device for decompressing compressed DCI; means for storing the updated model at the terminal device; means for receiving a further instance of compressed DCI from the network node; and means for using the stored updated model to decompress the further instance of compressed DCI. In some examples, the first instance of compressed DCI includes a first data portion of one or more consecutive bits, which indicate that the first instance of compressed DCI is compressed DCI. In some such examples, the first instance of compressed DCI includes a second data portion of one or more consecutive bits, which indicate a length of the payload of the first instance of compressed DCI. The first instance of compressed DCI may further include a third data portion which carries the payload of the first instance of compressed DCI, wherein the third data portion has the length indicated by the second data portion. In some examples, the terminal device comprises means for storing the decompressed first instance of DCI in a memory at the terminal device for use in decompressing a subsequently-received instance of compressed DCI. According to a second aspect, this specification describes a network node. The network node according to the second aspect comprises: means for receiving capability information which indicates that a terminal device served by the network node is capable of handling compressed downlink control information, DCI; means for transmitting, to the terminal device, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI; and means for transmitting a first instance of compressed DCI to the terminal device, the first instance of compressed DCI having been compressed using the trained model. In some examples, the network node further comprises means for, prior to transmitting the first instance of compressed DCI to the terminal device, receiving, from the terminal device, an indication that the node can begin transmitting compressed DCI to the terminal device. In some examples, the first information may include an identifier associated with the trained model, the identifier associated with the trained model being for use by the terminal device in retrieving a copy of the trained model from a remote storage system. In other examples, the first information may include a copy of the trained model. In some examples, the network node further comprises: means for providing, to the terminal device, second information which identifies which previously-received instances of DCI are to be used by the terminal device when decompressing the first instance of compressed DCI using the copy of the trained model. In some examples, the network node further comprises means for transmitting, to the terminal device, an indication that the trained model is to be updated. In some such examples, the network node further comprises: means for, subsequent to transmitting, to the terminal device, the indication that the trained model is to be updated, transmitting, to the terminal device, multiple instances of uncompressed DCI; and means for storing the multiple instances of uncompressed DCI in a training buffer at the network node, wherein the trained model is updated by the network node using the multiple instances of uncompressed DCI stored in the training buffer. In such examples, the network node may further comprise: means for receiving capability information which indicates that the terminal device is capable of updating a trained model that is to be used by the terminal device for decompressing compressed DCI; and means for transmitting, to the terminal device, an indication that the terminal device should begin updating the copy of the trained model that is stored at the terminal device. According to a third aspect, this specification describes a method that may be performed by a terminal device. The method comprises: transmitting, to a network, capability information which indicates that the terminal device is capable of handling compressed downlink control information, DCI; receiving, from a network node of the network and subsequent to transmitting the capability information, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI; storing the trained model at the terminal device; receiving a first instance of compressed DCI from the network node; and using the stored trained model to decompress the first compressed DCI. In some examples, the method further comprises, subsequent to storing the trained model at the terminal device and prior to receiving the first compressed DCI, transmitting, to the network node, an indication that the network node can begin transmitting compressed DCI to the terminal device. In some examples, the first information includes an identifier associated with the trained model. In some such examples, the terminal device comprises means for downloading the trained model from a remote storage system based on the identifier. In some examples, the first information includes the trained model. In some such examples the method comprises, prior to receiving the first information including the trained model, receiving, from the network node, an indication that the trained model is available for delivery to the terminal device. In some examples, the method further comprises receiving second information identifying which previously-received instances of DCI are to be used by the terminal device when decompressing the first compressed DCI using the stored trained model. In some examples, the method further comprises receiving an indication that the stored trained model is to be updated. The method may comprise, subsequent to receiving the indication that the stored trained model is to be updated, receiving, from the network node, multiple instances of uncompressed DCI. The method may further comprise: storing the multiple instances of uncompressed DCI in a training buffer at the terminal device; and updating the trained model using the multiple instances of uncompressed DCI stored in the training buffer. In addition, the method may comprise: transmitting, to the network, capability information which indicates that the terminal device is capable of updating a model that is to be used by the terminal device for decompressing compressed DCI; and receiving, from the network node, an indication that the terminal device should begin updating the trained model, wherein the trained model is updated using the multiple instances of uncompressed DCI stored in the training buffer in response to receiving the indication that the terminal device should begin updating the trained model. The method may further comprise: subsequent to receiving the indication that the trained model is to be updated, receiving, from the network node, third information relating to an updated model that is to be used by the terminal device for decompressing compressed DCI; storing the updated model at the terminal device; receiving a further instance of compressed DCI from the network node; and using the stored updated model to decompress the further instance of compressed DCI. In some examples, the first instance of compressed DCI includes a first data portion of one or more consecutive bits, which indicate that the first instance of compressed DCI is compressed DCI. In some such examples, the first instance of compressed DCI includes a second data portion of one or more consecutive bits, which indicate a length of the payload of the first instance of compressed DCI. The first instance of compressed DCI may further include a third data portion which carries the payload of the first instance of compressed DCI, wherein the third data portion has the length indicated by the second data portion. In some examples, the method comprises storing the decompressed first instance of DCI in a memory at the terminal device for use in decompressing a subsequently-received instance of compressed DCI. According to a fourth aspect, this specification describes method that may be performed by a network node. The method according to the fourth aspect comprises: receiving capability information which indicates that a terminal device served by the network node is capable of handling compressed downlink control information, DCI; transmitting, to the terminal device, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI; and transmitting a first instance of compressed DCI to the terminal device, the first instance of compressed DCI having been compressed using the trained model. In some examples, the method further comprises, prior to transmitting the first instance of compressed DCI to the terminal device, receiving, from the terminal device, an indication that the node can begin transmitting compressed DCI to the terminal device. In some examples, the first information may include an identifier associated with the trained model, the identifier associated with the trained model being for use by the terminal device in retrieving a copy of the trained model from a remote storage system In other examples, the first information may include a copy of the trained model. In some examples, the method further comprises: providing, to the terminal device, second information which identifies which previously-received instances of DCI are to be used by the terminal device when decompressing the first instance of compressed DCI using the copy of the trained model. In some examples, the method further comprises transmitting, to the terminal device, an indication that the trained model is to be updated. In some such examples, the method further comprises: subsequent to transmitting, to the terminal device, the indication that the trained model is to be updated, transmitting, to the terminal device, multiple instances of uncompressed DCI; and storing the multiple instances of uncompressed DCI in a training buffer at the network node, wherein the trained model is updated by the network node using the multiple instances of uncompressed DCI stored in the training buffer. In such examples, the method may further comprise: receiving capability information which indicates that the terminal device is capable of updating a trained model that is to be used by the terminal device for decompressing compressed DCI; and transmitting, to the terminal device, an indication that the terminal device should begin updating the copy of the trained model that is stored at the terminal device. According to a fifth aspect, this specification describes a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to perform a method according to either of the third and fourth aspects. According to a sixth aspect, this specification describes a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to perform a method according to either of the third and fourth aspects. According to a seventh aspect, this specification describes an apparatus, for example terminal device, comprising at least one processor and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to perform a method according to the third aspect. According to an eighth aspect, this specification describes an apparatus, for example a network node, comprising at least one processor and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to perform a method according to the fourth aspect. According to a ninth aspect, this specification describes a terminal device comprising: means for determining to initiate updating of a trained model that is stored at the terminal device and that is to be used by the terminal device for decompressing compressed DCI; means for receiving, from a network node, multiple instances of uncompressed DCI; means for storing the multiple instances of uncompressed DCI in a buffer at the terminal device; means for receiving, from the network node, an indication to begin updating the trained model stored at the terminal device; means for updating the trained model stored at the terminal device using the multiple instances of uncompressed DCI stored in the buffer; means for receiving, from the network node, a first instance of compressed DCI; means for decompressing the first instance of compressed DCI using the updated model. In some examples, the terminal device further comprises: means for, when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI messages includes uplink scheduling information, transmitting uplink data to the network node in accordance with the uplink scheduling information, wherein storing the given instance of uncompressed DCI in the buffer is performed responsive to receiving, at the terminal device, an acknowledgment from the network node that the uplink data has been successfully received. In addition or alternatively, the terminal device may further comprise: means for, when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI includes downlink scheduling information, receiving, from the network node and in accordance with the downlink scheduling information, a signal carrying downlink data, wherein storing the given instance of uncompressed DCI in the buffer is performed responsive to determining that the downlink data has been successfully received; and means for transmitting, to the network node, an acknowledgment that the downlink data has been successfully received. In some examples, the terminal device further comprises means for storing the decompressed first instance of compressed DCI in a second buffer at the terminal device, and means for using the stored decompressed first instance of compressed DCI to decompress a subsequently-received instance of compressed DCI. In some examples, determining to initiate the updating of the trained model that is stored at the terminal device is based on an indication, received from the network node, that the trained model is to be updated. In other examples, determining to initiate the updating of the trained model that is stored at the terminal device is based on a determination that a property associated with the trained model no longer satisfies a particular criterion. In such examples, the terminal may comprise means for, based on determining that the property associated with the trained model no longer satisfies the particular criterion, transmitting, to the network node, an indication that the trained model is to be updated. In some examples, the terminal device further comprises means for transmitting, to the network, capability information which indicates that the terminal device is capable of updating a trained model that is to be used by the terminal device for decompressing compressed DCI. According to a tenth aspect, this specification describes a network node comprising: means for determining to initiate updating of a trained model that is stored at the network node and that is to be used by the network node for compressing DCI; means for transmitting, to a terminal device, an indication that a copy of the trained model, which is stored at the terminal device, is to be updated; means for transmitting multiple instances of uncompressed DCI to the terminal device; means for storing the multiple instances of uncompressed DCI in a buffer at the network node; means for transmitting, to the terminal device, an indication to begin updating the copy of the trained model; means for, after the trained model that is stored at the network node has been updated using the multiple instances of uncompressed DCI, using the updated model to compress a first instance of DCI to form a first instance of compressed DCI; and means for transmitting, to the terminal device, the first instance of compressed DCI. In some examples, the network node may further comprise: means for, when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI includes uplink scheduling information, receiving a signal carrying uplink data from the terminal device in accordance with the uplink scheduling information, wherein storing the given instance of uncompressed DCI in the buffer is performed responsive to determining that the uplink data has been successfully received; and means for transmitting an acknowledgment to the terminal device responsive to determining that the uplink data has been successfully received. In some such examples, the network node may further comprise: means for, when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI includes downlink scheduling information, transmitting downlink data to the terminal device in accordance with the downlink scheduling information, wherein storing the given instance of uncompressed DCI in the buffer is performed based on receiving, from the terminal device, an acknowledgment that the downlink data has been successfully received at the terminal device. In some examples, transmitting, to the terminal device, the indication to begin updating the copy of the trained model is performed responsive to a pre-defined number of instances of uncompressed DCI being stored in the buffer at the network. In some examples, determining to initiate the updating of the trained model is based on a determination that a property associated with the trained model no longer satisfies a particular criterion. According to an eleventh aspect, this specification describes a method comprising: determining, by a terminal device, to initiate updating of a trained model that is stored at the terminal device and that is to be used by the terminal device for decompressing compressed DCI; receiving, by the terminal device and from a network node, multiple instances of uncompressed DCI; storing, by the terminal device, the multiple instances of uncompressed DCI in a buffer at the terminal device; receiving, by the terminal device and from the network node, an indication to begin updating the trained model stored at the terminal device; updating, by the terminal device, the trained model stored at the terminal device using the multiple instances of uncompressed DCI stored in the buffer; receiving, by the terminal device and from the network node, a first instance of compressed DCI; decompressing, by the terminal device, the first instance of compressed DCI using the updated model. The method may further comprise: when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI messages includes uplink scheduling information, transmitting, by the terminal device, uplink data to the network node in accordance with the uplink scheduling information, wherein storing the given instance of uncompressed DCI in the buffer is performed responsive to receiving, at the terminal device, an acknowledgment from the network node that the uplink data has been successfully received. In addition or alternatively, when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI includes downlink scheduling information, the method may comprise: receiving, by the terminal device and from the network node and in accordance with the downlink scheduling information, a signal carrying downlink data, wherein storing the given instance of uncompressed DCI in the buffer is performed responsive to determining that the downlink data has been successfully received; and transmitting, by the terminal device and to the network node, an acknowledgment that the downlink data has been successfully received. In some examples, determining to initiate the updating of the trained model that is stored at the terminal device is based on an indication, received from the network node, that the trained model is to be updated. In other examples, determining to initiate the updating of the trained model that is stored at the terminal device is based on a determination that a property associated with the trained model no longer satisfies a particular criterion. In such examples, the method further comprises: based on determining that the property associated with the trained model no longer satisfies the particular criterion, transmitting, by the terminal device and to the network node, an indication that the trained model is to be updated. In some examples, the method may further comprise: transmitting, by the terminal device and to the network, capability information which indicates that the terminal device is capable of updating a trained model that is to be used by the terminal device for decompressing compressed DCI. According to an twelfth aspect, this specification describes a method comprising: determining, by a network node, to initiate updating of a trained model that is stored at the network node and that is to be used by the network node for compressing DCI; transmitting, by the network node and to a terminal device, an indication that a copy of the trained model, which is stored at the terminal device, is to be updated; transmitting, by the network node, multiple instances of uncompressed DCI to the terminal device; storing, by the network node the multiple instances of uncompressed DCI in a buffer at the network node; transmitting, by the network node and to the terminal device, an indication to begin updating the copy of the trained model; after the trained model that is stored at the network node has been updated using the multiple instances of uncompressed DCI, using, by the network node, the updated model to compress a first instance of DCI to form a first instance of compressed DCI; and transmitting, by the network node and to the terminal device, the first instance of compressed DCI. The method may further comprise: i) when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI includes uplink scheduling information, receiving, by the network node, a signal carrying uplink data from the terminal device in accordance with the uplink scheduling information, and transmitting, by the network node, an acknowledgment to the terminal device responsive to determining that the uplink data has been successfully received, wherein storing the given instance of uncompressed DCI in the buffer is performed responsive to determining that the uplink data has been successfully received; and / or ii) when a given instance of uncompressed DCI of the multiple instances of uncompressed DCI includes downlink scheduling information, transmitting, by the network node, downlink data to the terminal device in accordance with the downlink scheduling information, wherein storing the given instance of uncompressed DCI in the buffer is performed based on receiving, from the terminal device, an acknowledgment that the downlink data has been successfully received at the terminal device. In some examples, the method may further comprise determining to initiate the updating of the trained model is based on a determination that a property associated with the trained model no longer satisfies a particular criterion. According to a thirteenth aspect, this specification describes a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to perform a method according to either of the eleventh and twelfth aspects. According to a fourteenth aspect, this specification describes a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to perform a method according to either of the eleventh and twelfth aspects. According to a fifteenth aspect, this specification describes an apparatus, for example terminal device, comprising at least one processor and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to perform a method according to the eleventh aspect. According to a sixteenth aspect, this specification describes an apparatus, for example a network node, comprising at least one processor and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to perform a method according to the twelfth aspect. According to a seventeenth aspect, this specification describes a network-side entity comprising: means for performing a feature importance assessment in respect of a first model that is trained to estimate a next instance of downlink control information, DCI, when provided a model input that is based on a sequence of DCI instances that precede the next instance, the feature importance assessment being performed using a data set, the data set comprising a plurality of sequences of DCI instances and, for each sequence of DCI instances, a target DCI instance which is the next DCI instance that follows the sequence of DCI instances; means for identifying, based on the feature importance assessment, K positions in the sequences of DCI instances of the plurality of sequences that make the greatest contribution to performance of the first model; and means for providing information indicative of the K positions for use by a terminal device when decompressing instances of compressed DCI received from a network node. In some examples, the information indicative of the K positions is for use by the terminal device when decompressing instances of compressed DCI using a second model, wherein the second model is trained to estimate, for a present sequence of DCI bits of a present DCI instance, a probability distribution, wherein the probability distribution is indicative of, for each bit of the present sequence of DCI bits of the present DCI instance, a respective probability of the bit having a first predetermined value, wherein, when the second model is in use to estimate the probability distribution, the second model is fed with an input that is derived from sequences of bits of DCI instances prior to the present DCI instance that are indicated by the information indicative of the K positions. In such examples, network-side entity may comprise means for providing the information indicative of the K positions for use in training the second model. The network-side entity may further comprise training the second model in accordance with the information indicative of the K positions. In some examples, providing, for use by the terminal device, the information indicative of the K positions that make the greatest contribution comprises transmitting the information to the terminal device. In other examples, providing, for use by the terminal device, the information indicative of the K positions that make the greatest contribution comprises uploading the information to a remote storage system for retrieval by the terminal device. In some examples, the network-side entity may further comprise means for transmitting the DCI instances of the plurality of sequences to the terminal device; and mean for storing the DCI instances in a buffer for use in performing the feature importance assessment. In some examples, performing the feature importance assessment in respect of the first model comprises: a) processing the data set with the first model to determine a first loss; and b) for each given position in the sequences of DCI instances of the data set, i) in each sequence, ii) corrupting the data representing the DCI instance at the given position to produce a corrupted data set, iii) processing the corrupted data set using the first model to determine a second loss, and iv) determining an importance metric for the given position based on the first loss and the second loss. In some such examples, identifying the K positions in the sequences of DCI instances of the plurality of sequences that make the greatest contribution to performance of the first model comprises identifying the K positions having the respective importance metrics which indicate the highest importance. In addition or alternatively, the plurality of sequences of DCI instances of the data set may be provided in a three-dimensional matrix, wherein a size of the first dimension of the matrix corresponds to a number of sequences in the data set, a size of a second dimension of the matrix corresponds to the number of DCI instances in each sequence, and a size of the third dimension of the matrix corresponds to the length of the DCI instances, and wherein corrupting the data representing the DCI instance at the given position to produce a corrupted data set comprises: swapping the first and second dimensions of the matrix; after swapping the first and second dimensions of the matrix, shuffling the data of the second dimension that is at a position corresponding to the given position; and reswapping the first and second dimensions of the matrix. In some examples, the network-side entity further comprises: means for training the first model, wherein the training is performed using a training data set comprising i) a second plurality of sequences of DCI instances, and ii) for each sequence of DCI instances of the second plurality of sequences, a target DCI instance which is a next DCI instance that follows the sequence of DCI instances of the second plurality of sequences. According to an eighteenth aspect, this specification describes a method comprising: performing, by a network-side entity, a feature importance assessment in respect of a first model that is trained to estimate a next instance of downlink control information, DCI, when provided a model input that is based on a sequence of DCI instances that precede the next instance, the feature importance assessment being performed using a data set, the data set comprising a plurality of sequences of DCI instances, and for each sequence of DCI instances, a target DCI instance which is the next DCI instance that follows the sequence of DCI instances; identifying, by the network-side entity and based on the feature importance assessment, K positions in the sequences of DCI instances of the plurality of sequences that make the greatest contribution to performance of the first model; and providing, by the network-side entity, information indicative of the K positions for use by a terminal device when decompressing instances of compressed DCI received from a network node. In some examples, the information indicative of the K positions is for use by the terminal device when decompressing instances of compressed DCI using a second model, wherein the second model is trained to estimate, for a present sequence of DCI bits of a present DCI instance, a probability distribution, wherein the probability distribution is indicative of, for each bit of the present sequence of DCI bits of the present DCI instance, a respective probability of the bit having a first predetermined value, wherein, when the second model is in use to estimate the probability distribution, the second model is fed with an input that is derived from sequences of bits of DCI instances prior to the present DCI instance that are indicated by the information indicative of the K positions. In such examples, the method may comprise providing, by the network-side entity, the information indicative of the K positions for use in training the second model. In addition, the method may comprise training, by the network-side entity, the second model in accordance with the information indicative of the K positions. In some examples, providing, for use by the terminal device, the information indicative of the K positions that make the greatest contribution comprises transmitting the information to the terminal device. In other examples, providing, for use by the terminal device, the information indicative of the K positions that make the greatest contribution comprises uploading the information to a remote storage system for retrieval by the terminal device. In some examples, the method further comprises: transmitting, by the network-side entity, the DCI instances of the plurality of sequences to the terminal device; and storing, by the network-side entity, the DCI instances in a buffer for use in performing the feature importance assessment. In some examples, performing the feature importance assessment in respect of the first model comprises: a) processing the data set with the first model to determine a first loss; and b) for each given position in the sequences of DCI instances of the data set, i) in each sequence, corrupting the data representing the DCI instance at the given position to produce a corrupted data set, ii) processing the corrupted data set using the first model to determine a second loss, and iii) determining an importance metric for the given position based on the first loss and the second loss. In such examples, identifying the K positions in the sequences of DCI instances of the plurality of sequences that make the greatest contribution to performance of the first model may comprise identifying the K positions having the respective importance metrics which indicate the highest importance. In addition or alternatively, the plurality of sequences of DCI instances of the data set may be provided in a three-dimensional matrix, wherein a size of the first dimension of the matrix corresponds to a number of sequences in the data set, a size of a second dimension of the matrix corresponds to the number of DCI instances in each sequence, and a size of the third dimension of the matrix corresponds to the length of the DCI instances, and corrupting the data representing the DCI instance at the given position to produce a corrupted data set may comprise: swapping the first and second dimensions of the matrix; after swapping the first and second dimensions of the matrix, shuffling the data of the second dimension that is at a position corresponding to the given position; and reswapping the first and second dimensions of the matrix. In some examples, the method may further comprise training, by the network-side entity, the first model, wherein the training is performed using a training data set comprising: a second plurality of sequences of DCI instances, and for each sequence of DCI instances of the second plurality of sequences, a target DCI instance which is a next DCI instance that follows the sequence of DCI instances of the second plurality of sequences. According to a nineteenth aspect, this specification describes a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to perform a method according to the eighteenth aspect. According to a twentieth aspect, this specification describes a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to perform a method according to the eighteenth aspect. According to a twenty-first aspect, this specification describes an apparatus, for example network-side entity such as a network node, comprising at least one processor and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to perform a method according to the eighteenth aspect. Brief Description of the Drawings Example embodiments will now be described by way of non-limiting example, with reference to the accompanying drawings, in which: FIG. 1 is a diagram schematically illustrating compression of downlink control information, DCI. FIG. 2 is a block diagram illustrating an example processing chain on a transmitter and receiver side. FIG. 3 is a block diagram which also illustrates example processing chains on a transmitter and receiver side. FIG. 4 illustrates communications that may take place between, and operations that may be performed by, a terminal device and a network node according to various examples described herein. FIG. 5 illustrates example communications that may take place between, and example operations that may be performed by, a terminal device and a network node in connection with training of a compression model for use in compressing DCI. FIG. 6 illustrates example communications that may take place between, and example operations that may be performed by, a terminal device and a network node in connection with sharing a compression model for use in compressing DCI. FIG. 7 illustrates example communications that may take place between, and example operations that may be performed by, a terminal device and a network node in connection with sharing a compression model for use in compressing DCI. FIG. 8 illustrates example communications that may take place between, and example operations that may be performed by, a terminal device and a network node in connection with updating a compression model for use in compressing DCI. FIG. 9 illustrates example communications that may take place between, and example operations that may be performed by, a terminal device and a network node in connection with updating a compression model for use in compressing DCI. FIG. 10 is a flow chart illustrating example operations that may be performed by a network entity in connection with DCI compression. FIG. 11 is a graph illustrating benefits that may be provided by techniques such as that described with reference to FIG. 10. FIG. 12 is a schematic illustration of aspects of telecommunications network in which the techniques described herein may performed. FIGS. 13 to 15 are schematic illustrations of a terminal device, a network node and a computer readable medium such as may be involved in the techniques described herein. Detailed Description The present disclosure is described herein with reference to particular non-limiting examples and to what are presently considered to be conceivable embodiments. A person skilled in the art will appreciate that the disclosure is by no means limited to these examples, and may be more broadly applied. It is to be noted that the following description of the present disclosure and its embodiments mainly refers to specifications being used as non-limiting examples for certain exemplary network configurations and deployments. Namely, the present disclosure and its embodiments are mainly described in relation to 3GPP specifications being used as non-limiting examples for certain exemplary network configurations and deployments. As such, the description of example embodiments given herein specifically refers to terminology which is directly related thereto. Such terminology is only used in the context of the presented non-limiting examples, and does naturally not limit the disclosure in any way. Rather, any other communication or communication related system deployment, etc. may also be utilized as long as compliant with the features described herein. Hereinafter, various embodiments and implementations of the present disclosure and its aspects or embodiments are described using several variants and / or alternatives. It is generally noted that, according to certain needs and constraints, all of the described variants and / or alternatives may be provided alone or in any conceivable combination (also including combinations of individual features of the various variants and / or alternatives). As used herein, "at least one of the following: " and "at least one of " and similar wording, where the list of two or more elements are joined by "and" or "or", mean at least any one of the elements, or at least any two or more of the elements, or at least all the elements. Described herein are techniques for enabling and / or realizing efficient radio resources utilization in control information transmissions. The Downlink Control Information (DCI), e.g. in 5GNR, conveys critical control information from the network (e.g. a node such as a Base Transceiver Station, BTS, gNB or the like) to User Equipments (UEs, also referred to herein as terminal devices). This includes PUSCH scheduling details, power control commands, link adaptation decisions, etc. The length of a 5GNR DCI instance (or message) can vary from 12 bits to around 150 bits. For robustness against errors, a notable amount of redundancy is added to the DCI payload via channel coding, thus increasing the signaling overhead for transmitting this message. This motivates the application of lossless compression techniques to be applied to the DCI prior to channel coding. Such compression reduces the DCI length and facilitates reliability enhancements and PDCCH overhead reduction. Lossless compression techniques reduce the length of binary sequences without losing information. This makes them an ideal choice for control channels, which demand robust decoding capabilities. Machine-learning (ML)-based and non-ML-based compression techniques exist for lossless compression. Many ML-based methods involve a training phase, where a model learns the statistical patterns of the data, and an inference phase, where the trained model is used to compress new data. Due to the sophistication of cellular networks, which enlarge the number of RAN features in each generation, the length of DCI messages also increases. As a result, the spectral occupation of DCI transmissions becomes significant, which limits the number of addressable UEs per time slot. FIG. 1 illustrates at a high-level the use of a machine-learned model-based compressor for DCI. As can be seen, an original DCI instance (or message) is compressed using a machine-learned model-based compressor (a compression model) from an original number of bits into a compressed DCI instance having fewer bits. Aspects described herein relate to compression and decompression of downlink control information, DCI. Information exchanges between the network and terminal devices are described via which the entities can coordinate DCI compression and aspects related thereto (including, for instance, compression model training, compression model deployment, compression model updating and compression model inference). The aspects applied herein may be utilized with 5G networks, 6G networks and future generations. FIG. 2 is a block diagram illustrating example processing chains on a transmitter and receiver side, which include DCI compression and decompression respectively. For DCI, the transmitter side is the network and the receiver side is the terminal device. Compared to traditional DCI generation and transmission steps over a PDCCH, examples described herein utilized machine learned-based lossless DCI compression, which reduces the DCI length. Symmetrically, at the receiver side, a lossless DCI decompression is utilized to recover the DCI message, without loss of information. FIG. 3 is a block diagram illustrating, in more detail, an example technique for DCI compression and decompression that may be utilized in conjunction with techniques described therein. Such techniques for compressing and decompressing downlink control information are described in co-pending application PCT / US2023 / 074994 ("Efficient radio resources utilization in control information transmissions") filed on September 25, 2023, which is hereby incorporated by reference in its entirety. As can be seen from FIG. 3, which is discussed in more detailed below under the "Example Compression Models" subheading, both the transmitter and receiver side utilize a machine-learned model (a compression model) for compression and decompression respectively. For lossless compression, as is utilized by various examples described herein, the two models utilized by the network (the transmitter side) and the terminal device (the receiver side) are the same. As can also be seen from FIG. 3, the compression model on both sides of the system takes as input features generated from previously-received / sent DCI instances. The compression model is trained to estimate, for a present DCI instance to be encoded or decoded and based on a number of foregoing DCIs (input features), a probability distribution (for each DCI bit, a probability of being 1 (or 0)). At the compression side, the probability distribution (for each DCI bit, a probability of being 1 (or 0)) is basis for arithmetic coding of a present DCI to be compressed and, at the decompression side, the probability distribution is basis for arithmetic decoding of a present DCI to be decompressed. Information Exchange relating to DCI Compression Techniques described in this specification facilitate use of such DCI compression techniques, in which the same machine-learned "compression model" is used by both the network and the terminal device. More specifically, the techniques relate to the information exchange between the terminal device and the network, e.g. at the Radio Resource Control (RRC) layer, which may enable such DCI compression to be employed. One such technique is illustrated in FIG. 4. In operation 4A of FIG. 4, the terminal device may transmit, to a network, capability information which indicates that the terminal device is capable of handling compressed downlink control information. The capability information may be included as an information element (IE), e.g. in a UE Capability Information message. In some examples, this capability information may be stored at the network as part of the UE context. It may then be shared with the base station(s) (e.g. gNBs) to which the terminal device is handed over as the terminal device moves around the network. In other examples, however, the terminal device may provide this capability information each time it connects to a new base station. In the example illustrated in FIG. 4, the terminal device transmits the capability information to the network node that performs the later operations of the illustrated technique. However, as will be appreciated from the above discussion, in some examples, the network node may receive the capability information which indicates that the terminal device is capable of handling compressed downlink control information, not from the terminal device directly but from another network entity, e.g. from an Access and Mobility Management Function (AMF) as part of the UE context, when the terminal device is handed over to the network node. As will be appreciated, in some examples, the network node may also inform the terminal device of its capabilities, for instance, whether the network node is capable of transmitting compressed DCI. This may be provided to the terminal device as part of the connection (or handover) process when the terminal device connects to the network node. In some such examples, it may be part of an exchange of capability information with the terminal device, which provides its capabilities in operation 4A. The transmission or exchange of such capability information may enable the network to serve terminal devices which are capable of handling compressed DCI, as well as those "legacy" devices that are not capable of such. In operation 4B, terminal device receives, from the network node, first information relating to a trained machine-learned model that is to be used by the terminal device for decompressing compressed DCI. Such a model may be referred to as a compression model. This first information may be transmitted to the terminal device subsequent to the terminal device having transmitted to the network the capability information which indicates that the terminal device is capable of handling compressed downlink control information. In addition, in some examples, the first information may be sent by the network node based on, or responsive to, the network node having received, from the terminal device or another network entity, the capability information which indicates that the terminal device is capable of handling compressed downlink control information. In operation 4C, the terminal device stores, in memory at the terminal device, the compression model to which the first information relates. This operation is discussed in more detail with reference to FIGS. 5 and 6. In operation 4D, having stored the compression model at the terminal device, the terminal device may, in some examples, transmit to the network node an indication that the node can begin transmitting compressed DCI to the terminal device. At a future point in time, in operation 4G, the network node transmits a first instance of compressed DCI to the terminal device, the first instance of compressed DCI having been compressed using the compression model. Specifically, it is compressed using a copy or replica of the trained compression model that was stored at the terminal device in operation 4C, the copy or replica being stored in the network, for instance at the network node. As illustrated by operation 4G, the compression of the DCI to generate the first instance of compressed DCI may be performed by the network node. In other examples, it may be performed by another network-side entity that is in communication with the network node. In addition, whichever entity performs compression may additionally store, in a first memory buffer, which may be referred to as a delay buffer, the (uncompressed) DCI which was compressed to form the first instance of compressed DCI. This is illustrated by operation 4H. As discussed above with reference to FIG. 3, this stored DCI is then utilised by the network, e.g. the network node, to form the model input for the compression model when compressing subsequent instances of DCI before transmission to the terminal device. The delay buffer (at the network node and the terminal device) may be a FIFO buffer, which stores a pre-determined number of DCI instances. As discussed below, this number of DCI instances may be determined during the model training or validation phase, and may be communicated to the terminal device by the network node. In other examples, the size of the delay buffer may be fixed. In operation 4J, the terminal device uses the compression model stored at the terminal device to decompress the first instance of compressed DCI. This is then used by the terminal device. In operation 4K, the terminal device stores the decompressed first instance of DCI, in a memory buffer (a delay buffer) at the terminal device. Similarly to as described with respect to the network node, this stored decompressed DCI can then used as input to the compression model when decompressing one or more subsequently received instances of compressed DCI. If the delay buffer is full, the decompressed first instance of DCI is stored in the delay buffer and the oldest DCI instance in the delay buffer is removed. Although not illustrated in FIG. 4, it will be appreciated that the network node may then continue to transmit compressed DCI instances, which are then decompressed by the terminal device using the stored compression model. In some examples, the network node and / or the terminal device may be configured to cause the network node to cease transmitting compressed DCI. For instance, the network node may, at some future time, transmit to the terminal device an indication that DCI compression will be disabled. In some examples, the network may then empty its delay buffer. Similarly, responsive to receiving the indication from the network node, the terminal device may also empty its buffer. In some examples, the decision to disable DCI compression may occur because it is determined that the compression model is outdated (see the below discussion of operations 8A, 8B, 9A and 9B with respect to FIGS 9 and 10). In some such examples, the indication that DCI compression will be disabled may be the indication described with reference to operations 8B or 9B. As illustrated by operation 4E, the network node, prior to transmitting the first instance of compressed DCI, transmits multiple instances of uncompressed DCI. These are received and used by the terminal device. As will be appreciated, receiving the uncompressed DCI involves decoding the downlink channel (the PDCCH) over which they are sent. The same is true for receiving the compressed instances of DCI, and in this case, the control channel should be decoded before the DCI can be decompressed. In operation 4F, the terminal device stores the received instances of uncompressed DCI in a memory buffer (the "delay buffer") for use in decompressing the instances of compressed DCI when the network node begins to transmit them. The terminal device may be configured to store received instances of DCI in the delay buffer, so that they are available for use when the network node begins transmitting compressed DCI. In some examples, the terminal device may begin storing the instances of uncompressed DCI after (or responsive to) having sent the indication in operation 4D or (responsive to) having received the first information in operation 4B. The number of instances of uncompressed DCI that are transmitted by the network node and stored by the terminal device after receipt of the first information in operation 4B (or the transmission of the indication of 4D) may be fixed. Alternatively, as described in more detail later in the specification, the number of instances of uncompressed DCI that are transmitted by the network node and stored by the terminal device after transmission of the indication in operation 4B may be determined during the training or validation phase of the machine learned model. In such examples, an indication of this number may be received in operation 4B or retrieved based on information included in the first information transmitted in operation 4B. In other examples, the indication that the node can begin transmitting compressed DCI to the terminal device (operation 4D) may be transmitted in response to the fixed or indicated number of instances of uncompressed DCI being stored in the delay buffer. In such examples, the terminal may begin storing instances of uncompressed DCI responsive to, for instance, operation 4B or 4C. In other such examples, terminal devices may begin storing instances of DCI responsive to receiving, from the network node, an indication that the network node is capable of transmitting compressed DCI messages. Training the Compression Model FIG. 5 illustrates examples of information exchange and operations that may be performed during a training phase of the compression model. The operations of FIG. 5 may occur, for instance, between operations 4A and 4B of FIG 4. That is to say, the operations of FIG. 5 may be performed after operation 4A has been performed but before operation 4B is performed. In operation 5A, the network node activates a model training phase. This may include initiating a training buffer stored at the network node, or another network-side entity that is in communication with the network node. During the model training phase, the network node transmits to the terminal device uncompressed DCI instances. This may be referred to as a legacy mode of operation. This is illustrated in operation 5B, in which the network node transmits an instance of uncompressed DCI to the terminal device. In operation 5C, which may take place before or after the transmission in operation 5B, the network node stores (or causes to be stored) the instance of uncompressed DCI in the training buffer. In operation 5D, the terminal device decodes the control channel (e.g. the PDCCH) to recover the instance of uncompressed DCI, which is then utilized by the terminal device. Operations 5B, 5C and 5D are then repeated until, in operation 5E, the training buffer has stored therein a sufficient number of instances of uncompressed DCI (or put another way, until the training buffer limit is reached). As will be appreciated, repetition of operations 5B, 5C and 5D means that multiple instances of uncompressed DCI are transmitted to the terminal device. Finally, in operation 5F, the machine learned compression model is trained using the DCI instances stored in the training buffer. As is discussed in more detail below, the training of the compression model may be in accordance with information identifying which DCI instances stored in the training buffer should be used to generate the training inputs for training the model. For instance, the information, which may be determined via a "Delay Length Selection" process as described below, and may indicate positions (e.g. via indices) within the buffer. Possible ways in which the compression model may be trained will be understood by the skilled person, not least from the discussion in the "Example Compression Models" section below. However, the precise way in which the model may be trained is not the primary focus of this specification. After the model has been trained, and as will be described with reference to FIGS. 6 and 7, the trained compression model can be shared with the terminal device. For instance, the trained model may be transmitted to the terminal device by the network node (FIG.7) or may be stored in remote storage for retrieval by the terminal device (FIG.6). In FIG. 5, the operations are shown as being performed by the network node. However, in other examples, some of the operations, such as 5A, 5C, 5E, and 5F may be performed by another network-side entity that is in communication with the network node. Sharing the Compression Model with the Terminal Device FIGS. 6 and 7 illustrate alternative examples of information exchange and operations that may be performed when sharing the compression model from the network to the terminal device. The operations illustrated in FIG. 6 may be performed at some point between operations 4A and 4D illustrated in FIG. 4 and include examples of operations 4B and 4C. The operations illustrated in FIG. 6 may be performed after the model is trained, for instance in operation 5F of FIG. 5. In operation 6A, the network node (or another network side entity that trained the machine learned model) may upload the compression model to remote storage. Such remote storage may be referred to as a model repository. The remote storage may store multiple models that have been trained based on DCI sent to multiple respective terminal devices. Put another way, the remote storage may store respective models for retrieval and use by multiple different terminal devices. The compression model uploaded in operation 6A is stored in the remote storage in association with an identifier based on which the model can be retrieved. In the example of FIG. 6, the first information transmitted in operation 4B of FIG. 4 includes an identifier that can be used by the terminal device for retrieving the compression model from the remote storage. In FIG. 6, transmission of this identifier is shown as operation 4B-1. In some other examples, the first information transmitted in operation 4B-1 may not include an identifier of the model but may simply inform the terminal device that a compression model is available for retrieval. In such examples, an identifier of the mobile terminal may be used for retrieving the compression model from remote storage. For instance, the compression model may be stored in association with the identifier of the mobile terminal. This identifier may be provided to the remote storage along with the compression model. In the examples described with reference to FIG. 6, storage of the compression model (operation 4C) comprises downloading the compression model from the remote storage. In examples in which the first information (of operation 4B-1) includes an identifier associated with the compression model, the terminal device may use this to download the compression model. For instance, it may be provided in a compression model retrieval request that is sent from the terminal device to the remote storage system. In other examples, as described above, the terminal device may provide an identifier, which identifies the terminal device, in the retrieval request. In some examples, as well as retrieving the compression model, the terminal device may also retrieve additional (or second) information. At least some of the additional information may be used when decompressing compressed DCI. The additional information may also be stored in association with the identifier that is usable for retrieving the compression model. The additional information may, for instance, indicate a number of instances of DCI that should be stored in the delay buffer at the terminal device for use when decompressing compressed DCI. Alternatively, as discussed above, this number may be fixed, e.g. pre-stored at the terminal device. In some examples, the additional information may identify specific indices associated with specific positions in the delay buffer. This information may be the same as that discussed in connection with operation 5F. It may be determined via a "Delay Length Selection" process as described below. The terminal device may then use the DCI instances stored at those positions in the delay buffer when decompressing a given compressed DCI instance. As explained with reference to the table below, in some examples, the additional information may include information such as the training finish time and / or in information for use by the terminal device when performing local updating of the model (see e.g. FIG. 9). This may include precision and gradient threshold. After having downloaded the compression model, the terminal device may perform operation 4D in which the terminal device transmits the indication that the network node can begin transmitting compressed DCI to the terminal device. As will be appreciated, the various operations described with reference to FIG. 4, but not depicted in FIG. 6, may be performed in conjunction with the operations depicted in FIG. 6. The remote storage may be a cloud database. To reliably share the compression models via such a remote storage, the following information may be beneficial. As such, the remote storage may store some or any combination of the following information, at least some of which may be provided to the terminal device. ID A unique identifier for the database instance. This may enable the terminal device to identify the remote storage to which to direct the request. This ID may be provided in the first information provided by the network node or may be prestored at the terminal device. Model metadata and description This could include the unique identifier (e.g. a universally unique ID), which allows the model to be identified. This could be an ID assigned to the model and provided in the first information, or it could be an identifier associated of the terminal device (which may already be known to the termina device). In addition or instead, this could include a use case for the trained model. For instance, the information could identify the type of device for which the model is intended, e.g. whether the model is for a smartphone or an loT device. Additional useful information such as one or any combination of RAN technology, training finish time, model size might also be added. Such information may be provided by the network node (other network-side entity) when it uploads the model to the remote storage. It may be retrieved by the terminal device along with the compressed model. Repository taxonomy This could include the type of content, format, and / or datatype for the model. It may also include additional information for use by the terminal device when performing local updating of the model (see e.g. FIG. 9). This may include precision and gradient threshold. This information may be uploaded by the network node (or other network-side entity) when uploading the compression model. Repository access Regarding the access level and permissions, since the terminal may not upload any new models to the cloud, the terminal device may only have read access to the remote storage. The network node, by contrast, may have full control over the repository. Repository governance and security This could include information related to the data quality, privacy and lifecycle management. Data quality may refer to the relevance of the training data, including the information of a latest update to the model. Privacy consideration could be similar to repository access, where certain groups of terminal devices are allowed access to the repository. Lifecycle management may involve the machine learning model's training, deployment, monitoring and retraining processes. Information such as the number of training data and target average compression ratio may be useful. Regarding security, the machine learning model's parameters may be encrypted and the terminal device may need an authentication to be granted by a gNB. As with FIG. 6, the operations illustrated in FIG. 7 may be performed at some point between operations 4A and 4D illustrated in FIG. 4 and include examples of operations 4B and 4C. The operations illustrated in FIG. 7 may be performed after the model is trained, e.g. in operation 5F of FIG. 5. In operation 7A, the terminal device receives, from the network node, an indication that the compression model is available for delivery to the terminal device. In some examples, in operation 7B, the terminal device may respond to the indication in operation 7A, by indicating, to the network node, that the compression model can be sent to the terminal device (or, put another way, that the terminal device is ready to receive the trained model). In operation 4B-2, the network node transmits the trained model to the terminal device. As such, in example implementations that are in accordance with FIG. 7, the first information described with reference to operation 4B of FIG. 4, includes the compression model. The compression model may be provided from the network node to the terminal device over a shared channel, such as the PDSCH. Transmission of the compression model to the terminal device may be performed responsive to the network node having received the indication in operation 7B. In other examples, however, the network node may proceed to transmit the compression model without receiving such an indication. For instance, the network node may transmit to the terminal device some scheduling information, e.g. as part of operation 7A, which indicates when the network node is going to transmit the compression model (e.g. over a shared channel). The terminal device then receives the compression model at the scheduled time. In some examples, in addition to receiving the compression model, the terminal device may also receive from the network node additional (or second) information. This additional information may be the same as the additional information described with reference to operation 4C-1 of FIG. 6. The additional information may be received in operation 4B-2, or as part of another operation. In operation 4C-2, the terminal device stores the compression model (and any additional information) received from the network node for use in the later operations described with reference to FIG. 4. As will be appreciated, the various operations described with reference to FIG. 4, but not depicted in FIG. 7, may be performed in conjunction with the operations depicted in FIG. 7. Updating the Trained Compression Model At some point the compression model may become outdated. This may be because the effectiveness of the compression model in compressing DCI has reduced. This may occur, for instance, because the DCI being transmitted to the terminal device has changed significantly since the model was originally trained (e.g. as described with reference to FIG. 5). It may therefore be beneficial to update, or re-train, the compression model. FIGS. 8 and 9 illustrate alternative approaches for updating the compression model, and associated information exchange between the network node and terminal device. In operation 8A of FIG. 8, the network node determines that the model is to be updated. This may occur, for instance, because the network node has determined that a property associated with the compression model no longer satisfies a particular criterion. The property may for instance be an average compression ratio resulting from use of the compression model. As such, in some examples, the network node may determine that the model is to be updated because an average compression ratio resulting from use of the compression model has fallen below a threshold. In addition or alternatively, the network node may make such a determination in response to determining that a predetermined duration has passed since the model was previously trained. This may, for example, be based on training finish time, discussed above. When the determination is based on the time since the since the model was previously trained, the property may be time since the since the model was previously trained, and the criterion may be that it should be less that a predetermined duration. Responsive to the determination of operation 8A, the network node transmits (in operation 8B) an indication that the compression model currently being utilised by the terminal device is to be updated. In some examples, in operation 8C, the terminal device transmits an acknowledgement back to the network node. This may be a simple acknowledgement of the indication sent in operation 8B. In other examples, it may be an indication as to whether or not the terminal device accepts that the model is to be updated. In such examples, commencement of the next operation (operation 8D) may occur only when the terminal device has accepted. If the terminal device does not accept, the network node may continue to use the outdated model or may revert to the legacy mode of operation in which compression is not performed.. In operation 8D, subsequent to transmitting the indication that the stored compression model is to be updated, network node transmits to the terminal device an instance of uncompressed DCI. As such, the network node stops compressing the DCI using the compression model. In operation 8E, which may take place before or after the transmission in operation 8D, the network node stores (or causes to be stored) the uncompressed instance of DCI in the training buffer. In operation 8F, the terminal device decodes the channel over which the instance of uncompressed DCI is transmitted (e.g. the PDCCH) in order to recover the DCI. This DCI is then used by the terminal device in the usual manner. Operations 8D, 8E and 8F are then repeated until, in operation 8G, the training buffer has stored therein a sufficient number of instances of uncompressed DCI, or, put another way, until the training buffer limit is reached. As will be appreciated, repetition of operations 8D, 8E and 8F means that multiple instances of uncompressed DCI are transmitted to the terminal device. Finally, in operation 8H, the compression model is updated (or re-trained) using the data stored in the training buffer. In other examples, a version of the model that has not previously been trained may be trained using the data stored in the training buffer. In some examples, a "Delay Length Selection" process may be performed as described below and the information derived therefrom may be used when training or re-training the model. In other examples, the buffer positions used when originally training the model may be used to when training or retraining in operation 8H. After the model has been updated, the model is then shared with the terminal device, for example as described with reference to FIGS. 6 and 7. After the model has been shared, the terminal device and network node may utilise the model in the manner described with reference to FIG. 4. In implementations in which the updated model is transmitted to the remote storage (as in FIG. 6), it may replace the outdated model in the remote storage. Although not illustrated in FIG. 8, prior to storing the instances of uncompressed DCI in the training buffer at the network, the training buffer may be emptied / cleaned and / or activated. This may be performed, for instance, in response to one of operations 8B or 8C. In some implementations in which the network updates the compression model, such as described with reference to FIG. 8, the network may update the compression model in the background. Put another way, the network may continue to transmit instances of DCI that have been compressed using the initial compression model, but may store those instances of DCI (in their uncompressed form) in a training buffer, and may then use them for updating the compression model. When the updated compression model is ready to be shared, the network node and terminal device may proceed in the manner described with reference to FIGS. 4 and 6 or 7. FIG. 8 illustrates that it is the network node that makes the initial determination to update the model. In other examples, however, the terminal device may determine independently from the network node that the model should be updated. For instance, the terminal device may determine that a property associated with the compression model no longer satisfies a particular criterion. In such examples, the terminal device may transmit a request for updating of the model to the network node. The network node may then accept this request. The network node may then begin transmitting instances of uncompressed DCI (as in operation 8D). FIG. 9 illustrates an alternative approach for updating the machine learned model. In the approach of FIG. 9, the respective copies of the compression model stored at the network and at the terminal device are updated in parallel. Although not depicted in FIG. 4, the terminal device may additionally, for instance, in operation 4A, or another such operation, inform the network as to whether the terminal device is capable of locally updating its compression model in parallel with the network updating its copy of the model. Likewise, in some examples, the network node may additionally inform the terminal device as to whether it is supports such functionality. In some examples, the operations depicted in FIG. 9 may be performed only if the terminal device and the network have indicated that they have such capability. In operation 9A, the network node may determine that the model is to be updated. This may occur, for instance, similarly to as described with reference to operation 8A. Responsive to the determination of operation 9A, the network node may transmit (in operation 9B) an indication that the compression model currently being utilised by the terminal device is to be updated. In some examples, in operation 9C, the terminal device transmits an acknowledgement back to the network node. This may be a simple acknowledgement of the indication sent in operation 9B. In other examples, it may be an indication as to whether or not the terminal device accepts that the model is to be updated. In such examples, commencement of the next operation (operation 9D) may occur responsive to the terminal device transmitting an acceptance. In operation 9D, network node transmits to the terminal device an instance of uncompressed DCI. As such, the network node stops compressing the DCI using the trained model. In operation 9E, which may take place before or after the transmission in operation 9D, the network node stores (or causes to be stored) the uncompressed instance of DCI in a training buffer stored at the network or elsewhere on the network-side. In operation 9F, the terminal device decodes the channel over which the instance of uncompressed DCI is transmitted (e.g. the PDCCH) in order to recover the DCI. This DCI is then used by the terminal device in the usual manner. In addition, in operation 9G, the terminal device stores the decompressed instance of DCI in a training buffer stored at the terminal device. Operations 9D, 9E, 9F and 9G are then repeated until, in operation 8G, the training buffer(s) has stored therein a sufficient number of instances of uncompressed DCI, or, put another way, until the training buffer limit is reached. In FIG. 9, the determination that the training buffer limit is reached is performed by the network (e.g. the network node). However, in some examples, it may be performed by the terminal device. In operation 91, an indication to begin updating the trained model based on the data in the training buffer is transmitted. This may occur responsive to the determination that the training buffer limit has been reached. In the example of FIG. 9, the indication of operation 91 is transmitted by the network node to terminal device but, in other examples, it may be transmitted from the terminal device node. In operation 9J, the network node (or other network side entity) updates (or re-trains) the compression model using the multiple instances of uncompressed DCI stored in the training buffer stored at the network. Likewise, in operation 9K, the terminal device retrains / updates, the copy of the compression model stored at the terminal device using the multiple instances of decompressed DCI in the training buffer stored at the terminal device. The training or updating in operations 91 and 9J may be similar to as described with reference to operation 8H. When updating the model, both the terminal device and network may perform the same number of training iterations. The number of training iterations used for updating the model may be indicated to the terminal device by the network, e.g. by a transmission from the network node to the terminal device or by the terminal device retrieving the information from the remote storage. As explained above, for lossless compression and decompression, the trained model should be the same at both the network and the terminal device. As such, when performing the update, trainable parameters of the model may be updated according to rV «- PF — awhere VK denotes the trainable parameters, a refers to the learning rate, and refers to the to the derivative from the loss function to the 3 dW trainable parameters. The updating may not include any random process that introduces a variation between the models at the network and the terminal device. In operation 9L, after both copies of the model have been updated (i.e. at the network and at the terminal device), the network node and the terminal device may agree that the network node will begin transmitting instances of DCI that have been compressed using the updated model. This may be through an exchange of information, for instance, in which the terminal device and the network node indicate to each other that their respective copies of the model have been updated. In other examples, just one of the terminal device and the network node may transmit an indication to the other of the terminal device and the network node indicating that the network node will (or can) begin transmitting instances of DCI that have been compressed using the updated model (e.g. similar to operation 4D). Once the model has been updated at both sides of the system and, for instance, the indication(s) that that the network node will (or can) re-start transmitting instances of compressed DCI, the network node may proceed in the manner described with reference to operations 4E to 4K of FIG. 4. As such, the network node uses the updated model to compress an instance of DCI to form an instance of compressed DCI, and transmits this to terminal device. Correspondingly the terminal device receives, from the network node, the instance of compressed DCI, and uses the updated model stored at the terminal device to decompress the received instance of compressed DCI. As is also explained with reference to FIG. 4, the network node may store the instance of DCI in its delay buffer for using in compressing future instances of DCI. Likewise, after performing decompression of the DCI, the terminal device may store the decompressed DCI in its delay buffer, for use in decompressing subsequently-received instances of compressed DCI. As noted above, for lossless compression the models at the terminal device and the network should be the same. Accordingly, the instances of DCI that are used to update the model on either side should be the same. As such, the instances of DCI may be stored in the training buffer at the network only when the network knows that the DCI has been correctly received by the terminal device. Likewise, the terminal device may only store instances of DCI in its training buffer, when it knows that the DCI has been correctly received. If it transpires that a DCI instance has not been correctly received, it is not stored in either of the training buffers. For instance, when a given uncompressed instance of DCI of the multiple uncompressed DCI messages includes downlink scheduling information, the terminal device may, subsequent to receiving the given uncompressed instance of DCI, receive, from the node and in accordance with the downlink scheduling information, a signal carrying downlink data. The terminal device may be configured to store the given uncompressed DCI message in its training buffer responsive to determining that the downlink data has been successfully received (which indicates that the DCI was also correctly received). The terminal device then transmits to the network node an acknowledgment that the downlink data has been successfully received. Responsive to this acknowledgment, the network node (or other network-side entity) may then store the given uncompressed instance of DCI in its training buffer. When a given instance of uncompressed DCI of the multiple uncompressed DCI messages includes uplink scheduling information, the terminal device may transmit uplink data to the network node in accordance with the uplink scheduling information. The terminal device may then store the given instance of uncompressed DCI in its training buffer responsive to receiving, at the terminal device, an acknowledgment from the network node that the uplink data has been successfully received at the network. Correspondingly, when the given instance of uncompressed DCI includes uplink scheduling information, the network node may store the given instance of uncompressed DCI in its training buffer at the network responsive to determining that a signal carrying uplink data transmitted, by the terminal device and in accordance with the uplink scheduling information, has been successfully received. The network node may then transmit an acknowledgment to the terminal device responsive to determining that the uplink data has been successfully received. In some examples, receipt of the indication transmitted in operation 9B may be considered to be the terminal device determining to initiate updating (or re-training) of the compression model that is stored at the terminal device. In other examples, for instance when the terminal device can accept or decline the updating of the model (e.g. as discussed with reference to operation 8C), the determining to initiate the updating may correspond to the terminal device making a decision to accept the updating of the model. In other examples, however, the terminal device may determine independently from the network node that the model should be updated. For instance, the terminal device may determine that a property associated with the model does not satisfy a particular criterion. In such examples, the terminal device may transmit an update request to the network node. The network node may then accept this request. The network node may then begin transmitting instances of uncompressed DCI (operation 9D). In these examples, receipt of the request by the network node or acceptance of the request by the network node may be considered to be the network node (or other network-side entity) determining to initiate updating of the compression model that is stored at the network. Although not illustrated in FIG. 9, in some examples, the terminal device and / or the network node (or other entity) may empty or clean their respective training buffers, if it is required, prior to storing new instances of DCI in their buffers. In some examples, this may not be required, for instance because the training buffers were emptied after a previous instance of updating the model. When it is required, the terminal device may empty its buffer, for instance, in response to receiving the indication (in operation 9B) that the trained model is to be updated. In some such examples, the terminal device may empty its buffer after having acknowledged or accepted that the model is to be updated (in operation 9C). The network node (or other network-side entity) may, for instance, empty its buffer responsive to one of operations 9A, 9B or 9C. In addition or alternatively, one or both of the terminal device and the network node may empty their training buffers after having updated their copies of the model in operations 9J and 9K. By utilizing the approach of FIG. 9, the same model is produced on both sides the system, without requiring any additional communication between them, such as transmission ordownloading of the model to the terminal device and any associated signaling. In addition, such an approach to updating the model might be useful in private network deployments in specific scenarios. For instance, training the models in parallel may offer advantages over the model sharing approaches when there are a large number of devices that may need a machine learning model for decompression. This is because the online training provides a lower latency without the need to transfer large amounts of data for the model parameters. In some other examples, rather than updating the model (e.g. as described in FIGS. 8 or 9) when a property associated with the model does not satisfy a particular criterion, the network node may identify another previously trained model for which the property would satisfy a particular criterion in the terminal device's current context. This previously-trained model may then be shared with the terminal device in the manners described above. For instance, the network node may, in the background, utilize one or more previously-trained models to compress current DCI for terminal and may determine the property (e.g. the compression ratio) for each model. A model for which the property does satisfy the criterion may then be selected. Where multiple previously-trained models are being tested, the model which yields the best performance (e.g. compression ratio) may be selected. New DCI Format for compressed DCI The current 3GPP NR technical specification defines various different DCI formats. At present these are: DCI Format 0 0 = 'fallback' DCI format for uplink resource allocations on the PUSCH. DCI Format 0_l = "standard' DCI format for uplink resource allocations on the PUSCH. DCI Format l_0 = 'fallback' DCI format for downlink resource allocations on the PDSCH. DCI Format 1_1 = "standard' DCI format for downlink resource allocations on the PDSCH. DCI Format 2_0 = provision of Slot Format Indicators (SFI). DCI Format 2_1 provision of Pre-emption Indications. DCI Format 2_2 provision of closed loop power control commands applicable to the PUCCH and PUSCH. DCI Format 2 3 provision of closed loop power control commands applicable to the SRS. In various examples described herein, new DCI formats for signaling lossless compression may be introduced. For instance, there may be a new format (e.g. DCI format 0 3), corresponding to lossless compressed 'standard' DCI for uplink resource allocations on the PUSCH). There may also be a new format (e.g. DCI format 1_3) corresponding to lossless compressed 'standard' DCI for downlink resource allocations on the PDSCH. According to the new formats, instances of compressed DCI may include a first data portion of one or more consecutive bits, which indicates that the instance of DCI is compressed DCI. They may then include a second data portion of one more consecutive bits which indicates a length of the payload of the instance of compressed DCI. They may also include a third data portion which carries the payload of the instance of compressed DCI (i.e. this is the bits representing the compressed DCI). The third data portion has the length indicated by the second data portion. The compressed DCI according to such formats may include CRC scrambled by C-RNTI or CS-RNTI or MCS-C-RNTI. In some examples, the first data portion, which indicates that the instance of DCI, may be one bit in length. For instance a 1 may be used to indicate that the DCI is compressed DCI. The second data portion is N bits in length. For instance, N may be equal to 2. The third data portion may be one of a predetermined number of different lengths. For instance, if N=2, the third data portion may be one of four different lengths plf to p4 (because 2 bits can have four different permutations). In downlink control information, the time domain (TD) scheduling and frequency domain (FD) scheduling tends to occupy most of the control bits in the DCI. As such, the predetermined lengths of the third data portion (in the above example, plf to p4) may be computed based on the number of control bits of TD and FD scheduling. For example, given sT and sF as the number of control bits for TD and FD scheduling, the possible lengths of the third data portion (that is the possible lengths of the compressed DCI) may be computed by pt = ^(sT + %)] for i e {1,2,3,4}. Example Compression Models As mentioned previously, examples of a compression model that may be used by techniques described herein are described in co-pending application PCT / US2023 / 074994. The reader is also referred to "DeepZip: Lossless Data Compression using Recurrent Neural Networks" by Mohit Goyal, Kedar Tatwawadi, Shubham Chandak and Idoia Ochoa, which is incorporated by reference herein. As depicted in FIG. 3, the compressor may involve the machine-learned compression model and an arithmetic encoder stage. Similarly, the decompressor may involve the machine-learned compression model and an arithmetic decoder stage. The arithmetic encoder and decoder utilize a probability distribution (generated by the compression model) over the DCI bits as their input. The probability distribution refers to the probability estimate of each bit in the DCI being a 0 or 1. The compression model is trained to estimate this probability distribution. More specifically, the compression model may be trained to estimate, for a present sequence of DCI bits of a present DCI instance, a probability distribution, wherein the probability distribution is indicative of, for each bit of the present sequence of DCI bits of the present DCI instance, a respective probability of the bit having a first predetermined value (e.g. "1"). In order to estimate the probability distribution, the model may be fed with the sequences of bits derived from a number of DCI instances prior to the present DCI instance. The prior DCI instances that are fed to the model may, in some examples, be a predetermined number, M, of instances preceding the present instance (e.g. the 5 or 10 DCI instances preceding the present instance). The model may be a neural network (NN), such as a DNN, and RNN or a transformer. For instance, as but one example, the NN may be deep feedforward NN that has two hidden layers with 256 and 128 neurons. A rectified linear unit may be used as the non-linear activation function and a binary cross-entropy loss function may be used to compare the label (i-th control bit in the DCI message) to the prediction (estimated probability for the i-th control bit of being 0 or 1). The prediction from NN serves as the input for the arithmetic encoding algorithm. Let b1 e be the DCI with a bit length of A at a scheduling time index t and / (%) = be the neural network function that maps input features x to the estimated probability for the / -th control bit bi of being a 1. Correspondingly, 1 - / (%) = Pbt,Q gives the probability for the / -th control bit bi of being a 0. Note that DCI bit lengths can vary as per standards specifications, but for a model according to this example, all instances of DCI (DCI messages) are assumed to have the length A, where A is the maximum DCI length and all shorter DCI instances are zero padded to have length A. Given a memory buffer with a size of M, the input features for the / -th control bit contain two portions of features to explore the correlations in the time and spatial domains, respectively. Denote xb as the input features for i e {1,2,..., A} at time index t, the two portions of features are denoted as %{ = {%{;1,Xt;2}- To exploit the temporal correlation, in some examples, the delay buffer may store the most recent M DCI messages which give = [b^^b^2, that has e NOte that jS not jnc|uc|ec| jn as £ / iS the pci message to be encoded and all the control bits at time index t share the common information from previous DCI messages. The machine-learned model does not have any prior knowledge on b^ Since the DCI is encoded sequentially by arithmetic coding, to exploit the spatial correlation a Toeplitz matrix may be set up. Each row of the Toeplitz matrix may be concatenated with %{lz which refers to the temporal information, to form the final input features for the compression model. -1 is used to distinguish the null information from the binary control bit. As a result, the first row of the Toeplitz matrix has all -Is with a length of A — 1 and the first column of the Toeplitz matrix contains [—1,dpIn the end, the Toeplitz matrix becomes: r -1 -1 -1 -r bl -1 -1 -i Xt,2 — bl -1 -i -i 7 Ki -cr -O ht °A-2 bll and each row of xt2, which serves as xlt2, is finally concatenated with x^to form the final input features for the machine learned model. Once the input features for each control bit is determined, the model takes the temporal and spatial features sequentially from xj to xj to provide the estimate on probability distribution. Eventually an encoded binary sequence is generated as the output of the DCI compressor. The DCI decompressor follows the sequential encoding structure from the compressor / transmitter side. The input features for the machine-learned model at the decompressor / receiver side, which are derived from previously-received DCI instances, which have been successfully decoded and are stored in the delay buffer at the terminal device are the same as the for the machine-learned model at the compressor / transmitter side. Since the arithmetic coding behaves as a lossless compression, the control bit starting from index i = 1 is decoded first without any error until the last control bit is correctly decoded. In the above discussion, it is described that the model is provided input that is generated based on the M most recent DCI instances. However, in other examples described herein, not least the in connection with the Delay Length Selection section below, the prior DCI instances that are used to generate the input to the model may be DCI instances at specific positions (or delay lengths) prior to the present DCI instance. In some cases, these DCI instances at specific positions may not be a consecutive sequence of DCI instances. For instance, and as an illustrative example only, it could be the most recent DCI instance, the 2nd most recent DCI instance, the 5th most recent DCI instance and the 10th most recent DCI instance. The skilled person would readily appreciate from the above discussion in connection with examples in which the M most recent DCI instances are used to generate the input features to the model, how the input features for the machine learned model can be instead be generated using the DCI instances at specific positions (or delay lengths) prior to the present DCI instance. As regards training, the skilled person would readily appreciate how the compression model may be trained based on previous instances of DCI to estimate the probability distribution. However, for completeness an example of the training process is discussed briefly below. A training input that is generated based on a number of previous DCI instances is provided to the model, which estimates a probability distribution. The probability distribution is then compared to a target output corresponding to the training input, the target output representing the bits of a current DCI that the model is trying to estimate. A loss is then determined based on the comparison. Stochastic gradient descent, SGD, (or some other optimizer) and the loss may then be used to modify the weights of the model. This is then repeated using other training input / output pairs for a number of iterations until the model is sufficiently trained. Delay Length Selection As mentioned above, in some examples, the M most recent instances of DCI may be used by the model for the compression and decompression. However, in other examples described herein, which may enhance the compression ratio, a delay length selection process can be utilized to determine "delay length information". This delay length information may identify specific positions in the delay buffer that result in improved compression performance. More specifically, the "delay length information" may identify specific positions in the delay buffer (e.g. via their indices) that should be used by the terminal device when decompressing compressed DCI (and which are also used when compressing the DCI). DCI instances at other positions in the buffer may not be used. As a non-limiting illustrative example, the delay length information might indicate that the instances of DCI at the 1st, 2nd, 10th, 15th and 20th positions in the buffer should be used to form the input features to the machine learned model. In such an example, the terminal may store all instances of DCI up to that with the greatest "delay" in this case the 20th. So, the delay buffer would store the 20 most-recently received instances of DCI, but only a select subset of those would be used to generate the input features to the machine learned model. Thus, the "delay length information" may be indicative of how many instances of DCI to store in the delay buffer, and also which specific instances to use for generating the input to the model. As mentioned above, for instance with reference to operation 4B-2 (in FIG. 7) and 4C-1 (FIG. 6) the delay length information may be shared as part of the additional (or second) information along with the machine learned model. FIG. 10 is a flow chart illustrating operations that may be performed by a network-side entity, such as a network node (e.g. a gNB), or combination of entities, such as a network node in combination with another network-side entity or function. In operation S10.1, a first model is trained to estimate a next instance of DCI when provided a model input that is based on a sequence of consecutive DCI instances that precede the next instance. The first model may be a different model to the compression model discussed elsewhere herein. Training the first model may be performed using a training data set comprising i) a plurality of training sequences of consecutive DCI instances and ii) for each training sequence of DCI instances, a target DCI instance which is a next DCI instance that follows the training sequence of DCI instances. The training sequences of DCI instances and corresponding target DCI instances may be representative of sequences of DCI instances that have been transmitted from a network node to a given terminal device (i.e. the terminal device to which the information is provided in operation S10.5). In an illustrative example, a deep feedforward neural network may be used as the first model. This neural network may be represented by fomt). The neural network, fDNN(-), may be trained using a training set XG and YE IR-6^. Here, X denotes a training set of DCI instances with temporal features which concatenate M latest DCI messages. Y denotes the "next" DCI instances to be estimated. B' represents the size of training set . N represents the length of one DCI instance. As will be appreciated, for a given DCI instance in Y, X includes a corresponding set of the M DCI instances, that immediately preceded (in time) the given DCI instance. In some examples, a cross-entropy loss, CE(-) can be used as the loss function to update fawiS). However, other loss functions may alternatively be used. In operation S10.2, a feature importance assessment may be performed in respect of the first model. The feature importance assessment is performed using a data set (which may be referred to as a validation data set). The data set comprises a plurality of sequences of consecutive DCI instances and, for each sequence of DCI instances, a target DCI instance which is the next DCI instance that follows the sequence of DCI instances. Similarly to the date in the training data set, the data in the data set may be derived from DCI sent from the network node to the terminal device. The data set may, in some examples, be entirely different to the training data set (i.e. there may be no overlap). In other examples, however, there may be some overlap between the two data sets. Returning to the above illustrative example, the data set may be represented as PyG BBxMxN and VyE1RBxN. Here, B represents the size of data set. Krand Vydenote the input features and target DCIs (or output labels). The M latest DCI instances form the temporal information for the estimate by the first model for the current DCI instance. In operation S10.3, based on the feature importance assessment, the K positions in the sequences of DCI instances of the plurality of sequences of the data set that make the greatest contribution to performance of the model may be identified. Performing the feature importance assessment in respect of the first model may comprise processing the data set with the first model to determine a first loss. This first loss may be represented as Xvaiid = CE(Tdnn( 1^), Vy). After determining the first loss, for each given position in the sequences of DCI instances of the data set, the following operations may be performed. In each sequence, the data representing the DCI instance at the given position may be corrupted to produce a corrupted data set. The corrupted data set may be processed using the first model to determine a second loss £s, and an importance metric for the given position may be determined based on the first loss and the second loss. For instance, the importance metric for the i-th position in the sequence may be represented as 6 = Zs, -Zvaiid. Identifying the K positions in the sequences of DCI instances of the plurality of sequences that make the greatest contribution to performance of the first model may comprise identifying the K positions having the respective importance metrics which indicate the highest importance metric. In some examples, the plurality of sequences of DCI instances of the data set may be provided in a three-dimensional matrix. For instance, a size of the first dimension of the matrix may correspond to a number of sequences in the validation set, a size of a second dimension of the matrix may correspond to the number of DCI instances in each sequence, and a size of the third dimension of the matrix may corresponds to the length of (i.e. the number of bits in) the DCI instances. In such examples, corrupting the data representing the DCI instance at the given position to produce a corrupted data set may comprises i) swapping the first and second dimensions of the matrix, ii) after swapping the first and second dimensions of the matrix, shuffling the data of the second dimension that is at a position corresponding the given position, and iii) reswapping the first and second dimensions of the matrix. In operation S10.5, information indicative of the K positions is provided for use by a terminal device when decompressing instances of compressed DCI received from a network node. The information indicative of the K positions may be referred to as the "delay length information" discussed above. The information may be provided to the terminal device as the additional (or second) information discussed above. More specifically, the information indicative of the K positions is for use by the terminal device when decompressing instances of compressed DCI using second model that is trained to estimate, for a present sequence of DCI bits of a present DCI instance, a probability distribution, wherein the probability distribution is indicative of, for each bit of the present sequence of DCI bits of the present DCI instance, a respective probability of the bit having a first predetermined value. The second model may be the compression model discussed above and, when the second model is in use to estimate the probability distribution, the second model may be fed with an input that is derived from sequences of bits of DCI instances that are indicated by the information indicative of the K positions. DCI instances that do not correspond to the K positions may not be used to generate the input to the model. In some examples, providing the information indicative of the K positions for use by the terminal device may comprise transmitting the information to the terminal device (e.g. as the second or additional information discussed with in connection with operation 4B-2 of FIG. 6). In other examples, providing the information indicative of the K positions for use by the terminal device may comprise uploading the information to a remote storage system for retrieval by the terminal device (e.g. in operation 4C-1 of FIG. 7). In some examples, the information indicative of the K positions may also be used for training the second model (e.g. as discussed above under the heading "Examples of a compression model"). That is, the second model may be trained using training sequences of DCI instances that comprise (or consist of) DCI instances corresponding to the K positions / delays (and not other DCI instances that do not correspond to the K positions). In addition, in some examples, the entity that identifies the K positions may also perform the training of the second model. An example process for determining the K positions is discussed below: 1. Process the data set using the trained network and save the loss Cvaiid = CE (fDNN(Kr), Ky). 2. Initialize a position index 1=1. 3. Swap the first and second dimension of 14, i.e., 14 becomes a size of 146 JJMxBxN 4. Shuffle the data of the second dimension (B) in the / -th postion: Vx / <-5(14,,2), where Vxi denotes the temporal feature from the i-th delay length and 5(14,,2) represents a shuffling function over the 2nd dimension of 14, / - All the other memory delay lengths features remain unchanged. The effect of this is that the DCI instance in the i-th position in every temporal feature set (of the M most-recent DCIs) is shuffled. 5. Swap the first and second dimension of I4and concatenate the second and third dimensions of Vx’, i.e., 14'becomes a size of IRBxNM. 6. Process the data set I4'and Prwith the network and determine the importance metric for the i-th position (or DCI instance): e, = £s. — £valid. 7. Reset validation set back to original and increment the position index i by 1. 8. Repeat operations 3 to 7 until i = M+l. 9. Choose the K position indices with the largest importance metrics to be used to construct the temporal features for the compression model (at inference and, in some examples, also in training). FIG. 11 is a graph which illustrates benefits that may be provided by using the DCIs at the "most important" K positions in the delay buffer to generate the input to the compression model. Specifically, FIG. 11 is shows two graphs of the number training batches against Binary Cross-Entropy (BCE) to illustrate convergence speed and performance. The top curve represents training loss using the most recent DCIs to generate the input to the compression model and the bottom curve represents training loss when the DCIs at the "most important" K positions in the delay buffer are used to generate the input to the compression model. The graphs were generated based on DCI traces from a Matlab system level simulator with 3 UEs (terminal devices), where the scheduling logs were collected to generate DCI messages. It can be seen that, when the DCIs at the "most important" K positions in the delay buffer are used to generate the input to the compression model, the model for DCI compression converges to a better performance with a faster convergence speed. Example Systems / Apparatuses FIG. 12 is a schematic illustration of an example system within which the abovedescribed techniques may be implemented. The system comprises a network node NN, which in this example is a base station (e.g. gNB) or some other TRP. It also comprised one or more terminal devices UE1 to UE4, which are being served by the network node NN. In addition, it shows another network-side entity or function NF, with which the network node is able to communicate. For instance, as described in some examples herein, the network function may be responsible fortraining and possibly also updating the compression models described herein, and where applicable transmitting them to the remote storage system RS. In such examples, in order to train (and update) the compression model for a given terminal device, the remote storage has access to the DCI instances transmitted by the network node to the UE. In other examples, however, training and updating may be performed by the network node. As described above, in some implementations, the terminal devices UE 1 to UE4 may communicate with the remote storage system to obtain a compression model. The remote storage system may store a respective compression model that has been trained for each terminal device. Other aspects, functions and capabilities associated with the entities depicted in FIG. 12 will be apparent from the discussed elsewhere herein. Example Configurations of Apparatuses FIG. 13 is a schematic illustration of an example configuration of a terminal device UE which may be configured to perform various operations described with reference to FIGS. 1 to 12. The terminal device UE may communicate, e.g. with a network node or base station, via an appropriate radio interface arrangement 805. The interface arrangement 805 may be provided for example by means of a radio part 805-2 (e.g. a transceiver) and an associated antenna arrangement 805-1. The antenna arrangement 805-1 may be arranged internally or externally to the terminal device UE. The antenna arrangement 805-1 includes multiple antennas, for instance to be able to perform beam forming. For instance, some UE's may include twelve antenna elements, e.g. four panels each having four cross-polarized antenna elements. The terminal device UE comprises a controller / control (or processing) apparatus 80 which is operable to control the other components of the terminal device UE in addition to performing any suitable combinations of the operations described in connection with terminal device UE with reference to the preceding FIGS. The control apparatus 80 may comprise processing apparatus 801 and memory 802. Computer-readable code 802-2A may be stored on the memory 802, which when executed by the processing apparatus 801, causes the control apparatus 80 to perform any of the operations described herein in relation to the terminal device UE. Although not illustrated in FIG. 13, the memory 802 stores the compression model, and include the delay buffer and, in some examples, a training buffer. Example configurations of the memory 802 and processing apparatus 801 will be discussed in more detail below. The terminal device UE may be, for example, a device that does not need human interaction, such as an entity that is involved in Machine Type Communications (MTC). Alternatively, the terminal device UE may be a device designed for tasks involving human interaction such as making and receiving phone calls between users and streaming multimedia or providing other digital content to a user. Non-limiting examples for the terminal device UE include a smart phone, a laptop, a smartwatch, a tablet computer, an e-reader, a vehicle-based terminal device, such as those mounted on cars, buses, uncrewed aerial vehicles (UAVs), aeroplanes, trains, or boats, or any type of terminal device that may be carried by a user, or worn on their person. Where the terminal device UE is a device designed for human interaction, the user may control the operation of the terminal device UE by means of a suitable user input interface UII 804 such as keypad, voice commands, touch sensitive screen or pad, combinations thereof or the like. A display 803, a speaker and a microphone may also be provided. Furthermore, the terminal device UE may comprise appropriate connectors (either wired or wireless) to other devices and / or for connecting external accessories, for example hands-free equipment, thereto. The terminal device UE may additionally be associated with (e.g., comprises or is in short range wired or wireless communication with) one or a plurality of motion sensors 806 for sensing motion of the mobile device. The terminal device may additionally include other sensors such as a GNSS unit. FIG. 14 is a schematic illustration of an example configuration of a network node NN. As described above, the network node may be a base station, such as a gNodeB or gNB. In some examples, the network node may include a wireless interface via which to communicate with terminal devices UEs. The network node may therefore also be TRP. In such examples, the network node may, as depicted in FIG. 14, comprise a radio frequency antenna array 901 configured to receive and transmit radio frequency signals. Although the network node NN is shown as having an array 901 of four antennas, this is illustrative only. The number of antennas may vary from two to many hundreds. The network node NN further comprises radio frequency interface circuitry 903 configured to interface between the antenna 901 and a control apparatus 90. The radio frequency interface circuitry 903 may also be known as a transceiver. The network node NN also comprises one or more interfaces 909 via which it can communicate with TRPs, other base stations, and other network entities such as those of the core network. The network node control apparatus 90 may be configured to cause the exchange of information with other network elements via the interface 909. In addition, the network node control apparatus 90 may be configured to process signals from the radio frequency interface circuitry 903, control the radio frequency interface circuitry 903 to generate suitable RF signals to communicate information to the terminal devices UEs via the wireless communications link. The network node control apparatus 90 may comprise processing apparatus 902 and memory 904. Computer-readable code 904-2A may be stored on the memory 904, which when executed by the processing apparatus 902, causes the control apparatus 90 to perform any of the operations assigned to the base station TRP1 described above. In some examples the network node NN may be distributed over more than one network / virtual entities (e.g. centralized units, CU, distributed units, DU, and remote radio heads, RRH) which host different network functions or protocol layers. In addition, as should of course be appreciated, the entities UE, NN shown in each of FIGS. 13 and 14 described above may comprise further elements which are not directly involved with processes and operations in respect of which this application is focussed. Memory 904 of the network node stores the compression model and includes the delay buffer. It may additionally include a training buffer. Some further details of components and features of the above-described apparatus / entities / apparatuses and alternatives forthem will now be described. The control apparatuses 80, 90 may comprise processing apparatus 801, 902 communicatively coupled with memory 802, 904. The memory 802, 904 has computer readable instructions 802-2A, 904-2A stored thereon, which when executed by the processing apparatus 801, 902 causes the control apparatus 80, 90 to cause performance of various ones of the operations described herein. The control apparatus 80, 90 may in some instances be referred to, in general terms, as "apparatus". The processing apparatus 801, 902 may be of any suitable composition and may include one or more processors 801A, 902A of any suitable type or suitable combination of types. For example, the processing apparatus 801, 902 may be a programmable processor that interprets computer program instructions 802-2A, 904-2A and processes data. The processing apparatus 801, 902 may include plural programmable processors. Alternatively, the processing apparatus 801, 902 may be, for example, programmable hardware with embedded firmware. The processing apparatus 801, 902 may be termed processing means. The processing apparatus 801, 902 may alternatively or additionally include one or more Application Specific Integrated Circuits (ASICs). In some instances, processing apparatus 801, 902 may be referred to as computing apparatus. The processing apparatus 801, 902 is coupled to the memory (which may be referred to as one or more storage devices) 802, 904 and is operable to read / write data to / from the memory 802, 904. The memory 802, 904 may comprise a single memory unit or a plurality of memory units, upon which the computer readable instructions (or code) 802-2A, 904-2A is stored. For example, the memory 802, 904 may comprise both volatile memory 802-1 and non-volatile memory 802-2. For example, the computer readable instructions / program code 802-2A, 904-2A may be stored in the non-volatile memory 802-2, 904-2 and may be executed by the processing apparatus 801, 902 using the volatile memory 802-1, 904-1 for temporary storage of data or data and instructions. In some examples, a transmission buffer 802-1B of the terminal device UE may be constituted by volatile memory 802-1 of the UE control apparatus 80. Examples of volatile memory include RAM, DRAM, and SDRAM etc. Examples of non-volatile memory include ROM, PROM, EEPROM, flash memory, optical storage, magnetic storage, etc. The memories in general may be referred to as non-transitory computer readable memory media. The term 'memory', in addition to covering memory comprising both non-volatile memory and volatile memory, may also cover one or more volatile memories only, one or more non-volatile memories only, or one or more volatile memories and one or more non-volatile memories. The computer readable instructions / program code 802-2A, 904-2A may be preprogrammed into the control apparatus 80, 90. Alternatively, the computer readable instructions 802-2A, 904-2A may arrive at the control apparatus 80, 90 via an electromagnetic carrier signal or may be copied from a physical entity 1000 such as a computer program product, a memory device or a record medium such as a CD-ROM or DVD an example of which is illustrated in Fig. 15. The computer readable instructions 802-2A, 904-2A may provide the logic and routines that enables the entities devices / apparatuses to perform the functionality described above. The combination of computer-readable instructions stored on memory (of any of the types described above) may be referred to as a computer program product. Embodiments of the technology described herein may be implemented in software, hardware, application logic ora combination of software, hardware and application logic. The software, application logic and / or hardware may reside on memory, or any computer media. In an example embodiment, the application logic, software or an instruction set is maintained on any one of various conventional computer-readable media. In the context of this document, a "memory" or "computer-readable medium" may be any media or means that can contain, store, communicate, propagate or transport the instructions for use by or in connection with an instruction execution system, apparatus, or device, such as a computer. Reference to, where relevant, "computer-readable storage medium", "computer program product", "tangibly embodied computer program" etc., or a "processor" or "processing apparatus" etc. should be understood to encompass not only computers having differing architectures such as single / multi-processor architectures and sequencers / parallel architectures, but also specialised circuits such as field programmable gate arrays FPGA, application specify circuits ASIC, signal processing devices and other devices. References to computer program, instructions, code etc. should be understood to express software for a programmable processor firmware such as the programmable content of a hardware device as instructions for a processor or configured or configuration settings for a fixed function device, gate array, programmable logic device, etc. If desired, the different functions discussed herein may be performed in a different order and / or concurrently with each other. Furthermore, if desired, one or more of the above-described functions may be optional or may be combined. Similarly, it will also be appreciated that flow diagrams described herein are examples only and that various operations depicted therein may be omitted, reordered and or combined. Although the methods and apparatuses have been described in connection with a New Radio (NR) network, it will be appreciated that they are not limited to such networks and are applicable to radio networks of various different types. Although various aspects of the methods and apparatuses described herein are set out in the independent claims, other aspects may comprise other combinations of features from the described embodiments and / or the dependent claims with the features of the independent claims, and not solely the combinations explicitly set out in the claims. It is also noted herein that while various examples are described above, these descriptions should not be viewed in a limiting sense. Rather, there are several variations and modifications which may be made without departing from the scope of the present invention as defined in the appended claims.

Claims

1. A terminal device comprising:means for transmitting, to a network, capability information which indicates that the terminal device is capable of handling compressed downlink control information, DCI;means for receiving, from a network node of the network and subsequent to transmitting the capability information, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI;means for storing the trained model at the terminal device;means for receiving a first instance of compressed DCI from the network node; andmeans for using the stored trained model to decompress the first compressed DCI.

2. The terminal device of claim 1 comprising:means for, subsequent to storing the trained model at the terminal device and prior to receiving the first compressed DCI, transmitting, to the network node, an indication that the network node can begin transmitting compressed DCI to the terminal device.

3. The terminal device of any preceding claim, wherein the first information includes an identifier associated with the trained model.

4. The terminal device of claim 3, comprising:means for downloading the trained model from a remote storage system based on the identifier.

5. The terminal device of claim 1 or claim 2, wherein the first information includes the trained model.

6. The terminal device of claim 5 comprising:means for, prior to receiving the first information including the trained model, receiving, from the network node, an indication that the trained model is available for delivery to the terminal device.

7. The terminal device of any preceding claim, comprising:means for receiving second information identifying which previously-received instances of DCI are to be used by the terminal device when decompressing the first compressed DCI using the stored trained model.

8. The terminal device of any preceding claim, comprising:means for receiving an indication that the stored trained model is to be updated.

9. The terminal device of claim 9, comprising:means for, subsequent to receiving the indication that the stored trained model is to be updated, receiving, from the network node, multiple instances of uncompressed DCI.

10. The terminal device of claim 9, comprising:means for storing the multiple instances of uncompressed DCI in a training buffer at the terminal device; andmeans for updating the trained model using the multiple instances of uncompressed DCI stored in the training buffer.

11. The terminal device of claim 10, comprising:means for transmitting, to the network, capability information which indicates that the terminal device is capable of updating a model that is to be used by the terminal device for decompressing compressed DCI; andmeans for receiving, from the network node, an indication that the terminal device should begin updating the trained model, wherein the trained model is updated using the multiple instances of uncompressed DCI stored in the training buffer in response to receiving the indication that the terminal device should begin updating the trained model.

12. The terminal device of claim 8 or 9, comprising:means for, subsequent to receiving the indication that the trained model is to be updated, receiving, from the network node, third information relating to an updated model that is to be used by the terminal device for decompressing compressed DCI;means for storing the updated model at the terminal device;means for receiving a further instance of compressed DCI from the network node; andmeans for using the stored updated model to decompress the further instance of compressed DCI.

13. The terminal device of any preceding claim, wherein the first instance of compressed DCI includes a first data portion of one or more consecutive bits, which indicate that the first instance of compressed DCI is compressed DCI.

14. The terminal device of claim 13, wherein the first instance of compressed DCI includes a second data portion of one or more consecutive bits, which indicate a length of the payload of the first instance of compressed DCI.

15. The terminal device of claim 14, wherein the first instance of compressed DCI includes a third data portion which carries the payload of the first instance of compressed DCI, wherein the third data portion has the length indicated by the second data portion.

16. The terminal device of any preceding claim, comprising:means for storing the decompressed first instance of DCI in a memory at the terminal device for use in decompressing a subsequently-received instance of compressed DCI.

17. A network node comprising:means for receiving capability information which indicates that a terminal device served by the network node is capable of handling compressed downlink control information, DCI;means for transmitting, to the terminal device, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI; andmeans for transmitting a first instance of compressed DCI to the terminal device, the first instance of compressed DCI having been compressed using the trained model.

18. The network node of claim 17 comprising:means for, prior to transmitting the first instance of compressed DCI to the terminal device, receiving, from the terminal device, an indication that the node can begin transmitting compressed DCI to the terminal device.

19. The network node of claim 17 or 18, wherein the first information includes an identifier associated with the trained model, the identifier associated with the trained model being for use by the terminal device in retrieving a copy of the trained model from a remote storage system20. The network node of claim 17 or claim 18, wherein the first information includes a copy of the trained model.

21. The network node of any of claims 17 to 20, comprising:means for providing, to the terminal device, second information which identifies which previously-received instances of DCI are to be used by the terminal device when decompressing the first instance of compressed DCI using the copy of the trained model.

22. The network node of any of claims 17 to 21, comprising:means for transmitting, to the terminal device, an indication that the trained model is to be updated.

23. The network node of claim 22, comprising:means for, subsequent to transmitting, to the terminal device, the indication that the trained model is to be updated, transmitting, to the terminal device, multiple instances of uncompressed DCI; andmeans for storing the multiple instances of uncompressed DCI in a training buffer at the network node, wherein the trained model is updated by the network node using the multiple instances of uncompressed DCI stored in the training buffer.

24. The network node of claim 23 comprising:means for receiving capability information which indicates that the terminal device is capable of updating a trained model that is to be used by the terminal device for decompressing compressed DCI; andmeans for transmitting, to the terminal device, an indication that the terminal device should begin updating the copy of the trained model that is stored at the terminal device.

25. A method comprising:transmitting, by a terminal device and to a network, capability information which indicates that the terminal device is capable of handling compressed downlink control information, DCI;subsequent to transmitting the capability information, receiving, by the terminal device and from a network node of the network, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI;storing, by the terminal device, the trained model at the terminal device;receiving, by the terminal device, a first instance of compressed DCI from the network node; andusing, by the terminal device, the stored trained model to decompress the first compressed DCI.

26. A method comprising:receiving, by a network node, capability information which indicates that a terminal device served by the network node is capable of handling compressed downlink control information, DCI;transmitting, by the network node and to the terminal device, first information relating to a trained model that is to be used by the terminal device for decompressing compressed DCI; andtransmitting, by the network node and to the terminal device, a first instance of compressed DCI, the first instance of compressed DCI having been compressed using the trained model.

Images