Quantizing learning model parameters

Abstract

Various aspects of the present disclosure relate to quantizing learning model parameters. The parameters of a trained learning model are quantized taking into account the sensitivity or the tolerance range of the parameters of the learning model so that the prediction or inference performance of the learning model after quantizing its parameters does not degrade beyond a given threshold from its intended (e.g., as trained) prediction or inference performance. The sensitivity or tolerance range of the parameters of the learning model can be determined based on a set of unlabeled data samples, and without having access to the training data or labeled data used to train the learning model. Forward error protection can also be used to provide error protection at different levels based on the sensitivity of parameters of the learning model when communicating the quantized model parameters (e.g., to another device or entity).

Description

Lenovo Ref. No. SMM920240010-WO-PCT 1 QUANTIZING LEARNING MODEL PARAMETERS RELATED APPLICATION

[0001] This application claims priority to U.S. Patent Application Serial No. 63 / 641,765 filed May 2, 2024 entitled “QUANTIZING LEARNING MODEL PARAMETERS,” the disclosure of which is incorporated by reference herein in its entirety. TECHNICAL FIELD

[0002] The present disclosure relates to wireless communications, and more specifically to quantizing learning model parameters. BACKGROUND

[0003] A wireless communications system may include one or multiple network communication devices, which may be otherwise known as network equipment (NE), supporting wireless communications for one or multiple user communication devices, which may be otherwise known as user equipment (UE), or other suitable terminology. The wireless communications system may support wireless communications with one or multiple user communication devices by utilizing resources of the wireless communication system (e.g., time resources (e.g., symbols, slots, subframes, frames, or the like) or frequency resources (e.g., subcarriers, carriers, or the like)). Additionally, the wireless communications system may support wireless communications across various radio access technologies including third generation (3G) radio access technology, fourth generation (4G) radio access technology, fifth generation (5G) radio access technology, among other suitable radio access technologies beyond 5G (e.g., sixth generation (6G)). SUMMARY

[0004] An article “a” before an element is unrestricted and understood to refer to “at least one” of those elements or “one or more” of those elements. The terms “a,” “at least one,” “one or more,” and “at least one of one or more” may be interchangeable. As used herein, including in the claims, “or” as used in a list of items (e.g., a list of items prefaced by a phrase such as “at least one of” or “one or more of” or “one or both of”) indicates an inclusive list such that, for example, a list of at least one of A, B, or C means A or B or C or AB or AC or BC or ABC (i.e., A and B and C). By Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 2 way of another example, a list of at least one of B; or C means A or B or C or AB or AC or BC or ABC (i.e., A and B and C). Also, as used herein, the phrase “based on” shall not be construed as a reference to a closed set of conditions. For example, an example step that is described as “based on condition A” may be based on both a condition A and a condition B without departing from the scope of the present disclosure. In other words, as used herein, the phrase “based on” shall be construed in the same manner as the phrase “based at least in part on”. Further, as used herein, including in the claims, a “set” may include one or more elements.

[0005] A UE for wireless communication is described. The UE may be configured to, capable of, or operable to perform one or more operations as described herein. For example, the UE may be configured to, capable of, or operable to receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0006] A processor (e.g., a standalone processor chipset, or a component of a UE) for wireless communication is described. The processor may be configured to, capable of, or operable to perform one or more operations as described herein. For example, the processor may be configured to, capable of, or operable to receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0007] A method performed or performable by a UE for wireless communication is described. The method may include receiving a trained learning model that includes a set of model parameters; determining, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantizing parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 3

[0008] In some implementations of the UE, and method described herein, the UE, processor, and method may further be configured to, capable of, performed, performable, or operable to determine a sensitivity value for each parameter in the set of model parameters; and communicate, to a device, the set of quantized model parameters, where to communicate the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, to provide error protection to the parameter based at least in part on the sensitivity value for the parameter.

[0009] In some implementations of the UE, processor, and method described herein, to determine the tolerance range for the set of model parameters, the UE, processor, and method may further be configured to, capable of, performed, performable, or operable to determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters.

[0010] In some implementations of the UE, processor, and method described herein, a norm of the error vector is equal to a step size parameter, and the UE, processor, and method may further be configured to, capable of, performed, performable, or operable to repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value.

[0011] In some implementations of the UE, processor, and method described herein, to determine the sensitivity value for each parameter in the set of model parameters, the UE, processor, and method may further be configured to, capable of, performed, performable, or operable to determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 4 error vector to change one or more parameters the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter.

[0012] In some implementations of the UE, processor, and method described herein, a norm of the error vector is equal to a step size parameter, and the UE, processor, and method may further be configured to, capable of, performed, performable, or operable to repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value.

[0013] In some implementations of the UE, processor, and method described herein, to provide error protection to each parameter in the set of quantized model parameters, the UE, processor, and method may further be configured to, capable of, performed, performable, or operable to apply an error control coding to each parameter in the set of quantized model parameters, where for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter.

[0014] In some implementations of the UE, processor, and method described herein, to apply the error control coding includes applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value. In some implementations of the UE, processor, and method described herein, the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network. In some implementations of the UE, processor, and method described herein, the trained learning model comprises an artificial intelligence (AI) / machine learning (ML) model. In some implementations of the UE, processor, and method described herein, the trained learning model comprises a classification model. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 5

[0015] A NE for wireless communication is The NE may be configured to, capable of, or operable to perform one or more operations as described herein. For example, the NE may be configured to, capable of, or operable to receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0016] A processor (e.g., a standalone processor chipset, or a component of a NE) for wireless communication is described. The processor may be configured to, capable of, or operable to perform one or more operations as described herein. For example, the processor may be configured to, capable of, or operable to receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0017] A method performed or performable by a NE for wireless communication is described. The method may include receiving a trained learning model that includes a set of model parameters; determining, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantizing parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0018] In some implementations of the NE, processor, and method described herein, the NE, processor, and method may further be configured to, capable of, performed, performable, or operable to determine a sensitivity value for each parameter in the set of model parameters; and communicate, to a device, the set of quantized model parameters, where to communicate the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, to provide error protection to the parameter based at least in part on the sensitivity value for the parameter. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 6

[0019] In some implementations of the NE, and method described herein, to determine the tolerance range for the set of model parameters, the NE, processor, and method may further be configured to, capable of, performed, performable, or operable to determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters.

[0020] In some implementations of the NE, processor, and method described herein, a norm of the error vector is equal to a step size parameter, and the NE, processor, and method may further be configured to, capable of, performed, performable, or operable to repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value.

[0021] In some implementations of the NE, processor, and method described herein, to determine the sensitivity value for each parameter in the set of model parameters, the NE, processor, and method may further be configured to, capable of, performed, performable, or operable to determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 7

[0022] In some implementations of the NE, and method described herein, a norm of the error vector is equal to a step size parameter, and the NE, processor, and method may further be configured to, capable of, performed, performable, or operable to repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value.

[0023] In some implementations of the NE, processor, and method described herein, to provide error protection to each parameter in the set of quantized model parameters, the NE, processor, and method may further be configured to, capable of, performed, performable, or operable to apply an error control coding to each parameter in the set of quantized model parameters, where for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter.

[0024] In some implementations of the NE, processor, and method described herein, to apply the error control coding includes applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value. In some implementations of the NE, processor, and method described herein, the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network. In some implementations of the NE, processor, and method described herein, the trained learning model comprises an AI / ML model. In some implementations of the NE, processor, and method described herein, the trained learning model comprises a classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 illustrates an example of a wireless communications system in accordance with aspects of the present disclosure.

[0026] Figure 2 illustrates an example of a device in accordance with aspects of the present disclosure.

[0027] Figure 3 illustrates an example of a UE in accordance with aspects of the present disclosure. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 8

[0028] Figure 4 illustrates an example of a in accordance with aspects of the present disclosure.

[0029] Figure 5 illustrates an example of a network equipment (NE) in accordance with aspects of the present disclosure.

[0030] Figure 6 illustrates a flowchart of a method in accordance with aspects of the present disclosure. DETAILED DESCRIPTION

[0031] A learning model may be an artificial intelligence (AI) model, a machine learning (ML) model, an artificial neural network (ANN) model, or a combination thereof. Additionally, a learning model may be defined by one or more algorithms. The learning model may include representations or capabilities for generating inferences from input data based on patterns or relationships identified in the input data. The term inferences can include one or more of determinations, predictions, or values, which may represent outputs of the learning model. The capabilities of the learning model may be defined in terms of one or more parameters of the learning model, such as constants, degrees, weights, biases, among others. In some examples, one or more parameters of a support vector machine (SVM) algorithm can include regularization constant, degree, and gamma. In some examples, one or more parameters of a decision tree algorithm can include a number of tree nodes, attributes (e.g., features) of tree nodes, and threshold values of tree node. In some examples, one or more parameters of a neural network (e.g., a deep neural network) can include neuron weights, bias values and other affine parameters. The device may train a learning model by determining or selecting a set of one or more parameters for the learning model. For example, the device may select a set of one or more parameters for the learning model that minimizes a loss function for a training dataset.

[0032] A device, such as a UE or a NE (e.g., a base station or other network entity) may be configured to, capable of, or operable to support one or more learning models, and the one or more learning models may be configured to enhance various aspects of a wireless communication system, including the device. The one or more learning models can perform tasks (e.g., actions, operations) related to signaling (e.g., receiving, transmitting, obtaining, outputting) or other operations (e.g., monitoring, tracking, measuring, detecting, determining) in the wireless communications system. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 9 For example, the device can implement the one more learning models for channel state information (CSI) compression (e.g., minimizing a complexity of CSI and / or encoding the CSI to a more compact representation), for beam prediction, and for positioning procedures, among other tasks. Learning models deployed in wireless communication systems for different purposes, such as CSI estimation, beam prediction, and so forth, may be developed (e.g., generated, trained) at the device or another device or entity (e.g., network entity, server) in the wireless communications system. For example, one or more learning models can be developed at a network node of the wireless communications system such as a base station, at a device at the edge of the wireless communications system such as a UE, in a laboratory or research and development facility of a network vendor or a UE vendor, by a company that develops and sells learning models to the wireless network industry, and so forth.

[0033] A learning model may be transferred over a wireless network from one device or entity to another device or entity, also referred to as model transfer. Model transfer, in the case when the learning model is a deep neural network (DNN), corresponds to transferring parameters or weights of the DNN, including parameters (e.g., weight and / or bias) of each neuron in the DNN and values of affine parameters of its normalization layers (if any) and other information, such as activation functions and architectural details of the DNN. The parameters of a learning model can be high- precision values, such as real 32-bit floating point values, that are determined during training of the learning model. Such high-precision values, such as real 32-bit floating point values, take up a larger amount of storage at devices and consume a larger amount of signaling resources than lower precision values, such as integers. To reduce the storage space and signaling resources consumed for parameters of a learning model, the parameters can be quantized to lower precision numbers (e.g., integers). However, such quantization results in the values of the parameters of the learning model being changed, which degrades the performance of the learning model.

[0034] The techniques discussed herein quantize the parameters of a learning model taking into account the sensitivity or the tolerance range of the parameters of the learning model so that the prediction or inference performance of the learning model after quantizing its parameters does not degrade beyond a given threshold from its intended (e.g., as trained) prediction or inference performance. The sensitivity or tolerance range of the parameters of the learning model is determined by applying an error vector to the parameters that results in a loss value from a loss Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 10 function that does not exceed a threshold from a loss value generated from the loss function using the intended (e.g., as trained) parameters. Accordingly, the sensitivity or tolerance range of the parameters of the learning model can be determined based on a set of unlabeled data samples, and without having access to the training data or labeled data used to train the learning model.

[0035] Forward error correction (FEC) can also be used to reduce errors during model transfer. FEC may include adding redundant bits to the information bits (e.g., values of parameters of a learning model) in such a way that the information bits can be reliably decoded at the receiving end even when some of the transmitted bits are flipped or corrupted during the transmission. The techniques discussed herein provide error protection at different levels based on the sensitivity of parameters of the learning model. A higher level of error protection is provided to the highly sensitive parameters and a lower level of error protection is provided to the parameters having lesser sensitivity (or lesser influence on the learning model prediction).

[0036] Accordingly, the techniques discussed herein use quantization to reduce the amount of memory or storage used to store the learning model and reduce the resources (e.g., time and / or bandwidth) consumed to communicate the learning model (e.g., to another device or entity). The quantization further allows the inference or prediction performed using the learning model to be faster because the computations involve lower-precision (e.g., finite-precision) arithmetic rather than higher-precision (e.g., infinite-precision) arithmetic. The quantization values to use are determined based on a loss function that does not need, and typically does not have, access to the training data or labeled data used to train the learning model. Furthermore, different levels of error protection are provided to different parameters based on the sensitivity of the parameters, reducing communication overhead communicating the learning model (e.g., to another device or entity) by using a lower level of error protection for less sensitive parameters.

[0037] Reference is made herein to communicating data or information, such as communicating a learning model or model parameters. It is to be appreciated that other terms may be used interchangeably with communicating, such as transmitting, receiving, outputting, forwarding, retrieving, obtaining, and so forth. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 11

[0038] Aspects of the present disclosure in the context of a wireless communications system.

[0039] Figure 1 illustrates an example of a wireless communications system 100 in accordance with aspects of the present disclosure. The wireless communications system 100 may include one or more NE 102, one or more UE 104, and a core network (CN) 106. The wireless communications system 100 may support various radio access technologies. In some implementations, the wireless communications system 100 may be a 4G network, such as an LTE network or an LTE-Advanced (LTE-A) network. In some other implementations, the wireless communications system 100 may be a new radio (NR) network, such as a 5G network, a 5G-Advanced (5G-A) network, or a 5G ultrawideband (5G-UWB) network. In other implementations, the wireless communications system 100 may be a combination of a 4G network and a 5G network, or other suitable radio access technology including Institute of Electrical and Electronics Engineers (IEEE) 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802.20. The wireless communications system 100 may support radio access technologies beyond 5G, for example, 6G. Additionally, the wireless communications system 100 may support technologies, such as time division multiple access (TDMA), frequency division multiple access (FDMA), or code division multiple access (CDMA), etc.

[0040] The one or more NE 102 may be dispersed throughout a geographic region to form the wireless communications system 100. One or more of the NE 102 described herein may be or include or may be referred to as a network node, a base station, a network element, a network function, a network entity, a radio access network (RAN), a NodeB, an eNodeB (eNB), a next- generation NodeB (gNB), or other suitable terminology. An NE 102 and a UE 104 may communicate via a communication link, which may be a wireless or wired connection. For example, an NE 102 and a UE 104 may perform wireless communication (e.g., receive signaling, transmit signaling) over a Uu interface.

[0041] An NE 102 may provide a geographic coverage area for which the NE 102 may support services for one or more UEs 104 within the geographic coverage area. For example, an NE 102 and a UE 104 may support wireless communication of signals related to services (e.g., voice, video, packet data, messaging, broadcast, etc.) according to one or multiple radio access technologies. In some implementations, an NE 102 may be moveable, for example, a satellite associated with a non- terrestrial network (NTN). In some implementations, different geographic coverage areas associated Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 12 with the same or different radio access may overlap, but the different geographic coverage areas may be associated with different NE 102.

[0042] The one or more UE 104 may be dispersed throughout a geographic region of the wireless communications system 100. A UE 104 may include or may be referred to as a remote unit, a mobile device, a wireless device, a remote device, a subscriber device, a transmitter device, a receiver device, or some other suitable terminology. In some implementations, the UE 104 may be referred to as a unit, a station, a terminal, or a client, among other examples. Additionally, or alternatively, the UE 104 may be referred to as an Internet-of-Things (IoT) device, an Internet-of- Everything (IoE) device, or machine-type communication (MTC) device, among other examples.

[0043] A UE 104 may be able to support wireless communication directly with other UEs 104 over a communication link. For example, a UE 104 may support wireless communication directly with another UE 104 over a device-to-device (D2D) communication link. In some implementations, such as vehicle-to-vehicle (V2V) deployments, vehicle-to-everything (V2X) deployments, or cellular-V2X deployments, the communication link may be referred to as a sidelink. For example, a UE 104 may support wireless communication directly with another UE 104 over a PC5 interface.

[0044] An NE 102 may support communications with the CN 106, or with another NE 102, or both. For example, an NE 102 may interface with other NE 102 or the CN 106 through one or more backhaul links (e.g., S1, N2, N6, or other network interface). In some implementations, the NE 102 may communicate with each other directly. In some other implementations, the NE 102 may communicate with each other indirectly (e.g., via the CN 106). In some implementations, one or more NE 102 may include subcomponents, such as an access network entity, which may be an example of an access node controller (ANC). An ANC may communicate with the one or more UEs 104 through one or more other access network transmission entities, which may be referred to as a radio heads, smart radio heads, or transmission-reception points (TRPs).

[0045] The CN 106 may support user authentication, access authorization, tracking, connectivity, and other access, routing, or mobility functions. The CN 106 may be an evolved packet core (EPC), or a 5G core (5GC), which may include a control plane entity that manages access and mobility (e.g., a mobility management entity (MME), an access and mobility management functions (AMF)) and a user plane entity that routes packets or interconnects to Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 13 external networks (e.g., a serving gateway (S- , a packet data network (PDN) gateway (P-GW), or a user plane function (UPF)). In some implementations, the control plane entity may manage non-access stratum (NAS) functions, such as mobility, authentication, and bearer management (e.g., data bearers, signal bearers, etc.) for the one or more UEs 104 served by the one or more NE 102 associated with the CN 106.

[0046] The CN 106 may communicate with a packet data network over one or more backhaul links (e.g., via an S1, N2, N6, or other network interface). The packet data network may include an application server. In some implementations, one or more UEs 104 may communicate with the application server. A UE 104 may establish a session (e.g., a protocol data unit (PDU) session, or the like) with the CN 106 via an NE 102. The CN 106 may route traffic (e.g., control information, data, and the like) between the UE 104 and the application server using the established session (e.g., the established PDU session). The PDU session may be an example of a logical connection between the UE 104 and the CN 106 (e.g., one or more network functions of the CN 106).

[0047] In the wireless communications system 100, the NEs 102 and the UEs 104 may use resources of the wireless communications system 100 (e.g., time resources (e.g., symbols, slots, subframes, frames, or the like) or frequency resources (e.g., subcarriers, carriers)) to perform various operations (e.g., wireless communications). In some implementations, the NEs 102 and the UEs 104 may support different resource structures. For example, the NEs 102 and the UEs 104 may support different frame structures. In some implementations, such as in 4G, the NEs 102 and the UEs 104 may support a single frame structure. In some other implementations, such as in 5G and among other suitable radio access technologies, the NEs 102 and the UEs 104 may support various frame structures (i.e., multiple frame structures). The NEs 102 and the UEs 104 may support various frame structures based on one or more numerologies.

[0048] One or more numerologies may be supported in the wireless communications system 100, and a numerology may include a subcarrier spacing and a cyclic prefix. A first numerology (e.g., ^=0) may be associated with a first subcarrier spacing (e.g., 15 kHz) and a normal cyclic prefix. In some implementations, the first numerology (e.g., ^=0) associated with the first subcarrier spacing (e.g., 15 kHz) may utilize one slot per subframe. A second numerology (e.g., ^=1) may be associated with a second subcarrier spacing (e.g., 30 kHz) and a normal cyclic prefix. A third numerology (e.g., ^=2) may be associated with a third subcarrier spacing (e.g., 60 kHz) and a Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 14 normal cyclic prefix or an extended cyclic A fourth numerology (e.g., ^=3) may be associated with a fourth subcarrier spacing (e.g., 120 kHz) and a normal cyclic prefix. A fifth numerology (e.g., ^=4) may be associated with a fifth subcarrier spacing (e.g., 240 kHz) and a normal cyclic prefix.

[0049] A time interval of a resource (e.g., a communication resource) may be organized according to frames (also referred to as radio frames). Each frame may have a duration, for example, a 10 millisecond (ms) duration. In some implementations, each frame may include multiple subframes. For example, each frame may include 10 subframes, and each subframe may have a duration, for example, a 1 ms duration. In some implementations, each frame may have the same duration. In some implementations, each subframe of a frame may have the same duration.

[0050] Additionally or alternatively, a time interval of a resource (e.g., a communication resource) may be organized according to slots. For example, a subframe may include a number (e.g., quantity) of slots. The number of slots in each subframe may also depend on the one or more numerologies supported in the wireless communications system 100. For instance, the first, second, third, fourth, and fifth numerologies (i.e., ^=0, ^=1, ^=2, ^=3, ^=4) associated with respective subcarrier spacings of 15 kHz, 30 kHz, 60 kHz, 120 kHz, and 240 kHz may utilize a single slot per subframe, two slots per subframe, four slots per subframe, eight slots per subframe, and 16 slots per subframe, respectively. Each slot may include a number (e.g., quantity) of symbols (e.g., OFDM symbols). In some implementations, the number (e.g., quantity) of slots for a subframe may depend on a numerology. For a normal cyclic prefix, a slot may include 14 symbols. For an extended cyclic prefix (e.g., applicable for 60 kHz subcarrier spacing), a slot may include 12 symbols. The relationship between the number of symbols per slot, the number of slots per subframe, and the number of slots per frame for a normal cyclic prefix and an extended cyclic prefix may depend on a numerology. It should be understood that reference to a first numerology (e.g., ^=0) associated with a first subcarrier spacing (e.g., 15 kHz) may be used interchangeably between subframes and slots.

[0051] In the wireless communications system 100, an electromagnetic (EM) spectrum may be split, based on frequency or wavelength, into various classes, frequency bands, frequency channels, etc. By way of example, the wireless communications system 100 may support one or multiple operating frequency bands, such as frequency range designations FR1 (410 MHz – 7.125 GHz), FR2 (24.25 GHz – 52.6 GHz), FR3 (7.125 GHz – 24.25 GHz), FR4 (52.6 GHz – 114.25 GHz), Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 15 FR4a or FR4-1 (52.6 GHz – 71 GHz), and FR5 GHz – 300 GHz). In some implementations, the NEs 102 and the UEs 104 may perform wireless communications over one or more of the operating frequency bands. In some implementations, FR1 may be used by the NEs 102 and the UEs 104, among other equipment or devices for cellular communications traffic (e.g., control information, data). In some implementations, FR2 may be used by the NEs 102 and the UEs 104, among other equipment or devices for short-range, high data rate capabilities.

[0052] FR1 may be associated with one or multiple numerologies (e.g., at least three numerologies). For example, FR1 may be associated with a first numerology (e.g., ^=0), which includes 15 kHz subcarrier spacing; a second numerology (e.g., ^=1), which includes 30 kHz subcarrier spacing; and a third numerology (e.g., ^=2), which includes 60 kHz subcarrier spacing. FR2 may be associated with one or multiple numerologies (e.g., at least 2 numerologies). For example, FR2 may be associated with a third numerology (e.g., ^=2), which includes 60 kHz subcarrier spacing; and a fourth numerology (e.g., ^=3), which includes 120 kHz subcarrier spacing.

[0053] A UE 104 or a NE 102 can use a learning model. For example, a NE 102 or component of the CN 106 may train a learning model for use by multiple UEs 104. The learning model is communicated to the UEs 104, each of which quantizes the model parameters of the learning model into quantized model parameters, and the quantized model parameters are transmitted to another device (e.g., a NE 102 or a component of the CN 106).

[0054] Figure 2 illustrates an example of a device 200 in accordance with aspects of the present disclosure. The device 200 can be, for example, a NE 102, a UE 104, or a component of the CN 106 of Figure 1. The device receives a trained learning model 202 that includes a set of model parameters. The trained learning model 202 can have been trained at various locations, such as a base station (e.g., gNB), or at a device at the edge of the wireless communications system such as a UE. A quantizer 204 determines, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters of the trained learning model 202, and quantizes parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters 206 for the trained learning model. The device 200 can communicate (e.g., transmit, output, forward) the quantized model parameters 206 to another device (e.g., a NE 102, a UE 104, or a component of the CN 106 of Figure 1). This communication can Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 16 include providing error protection to each in the set of quantized model parameters based at least in part on a sensitivity value for the parameter. The device 200 also uses the set of quantized model parameters 206 to generate a quantized learning model 208. The quantized learning model 208 is, for example, the trained learning model 202 with the set of parameters of the trained learning model 202 being replaced with the set of quantized model parameters 206.

[0055] A learning model (e.g., AI / ML model) can perform various functions, including replacing a conventional non-AI / ML module in a transceiver chain, for example, of the device 200 with the learning model developed through data driven methods. This includes use cases of AI / ML models for CSI compression, beam prediction, positioning and so forth in the transceiver chain.

[0056] As an example, let x^ ∈ ^ and y^ ∈ ^ denote an input sample and the correspondinglabel (also referred to as the expected output sample, or the expected prediction or inference from the learning model for input x^). Here, x^can be a scalar, a vector, a matrix, or a tensor (e.g., x^is a scalar or a one or multi-dimensional vector) and, similarly, y^is a scalar, a vector, a matrix, or a tensor. ^ and ^ denote the input sample space and the output sample space, respectively. When y^assumes discrete and finitely many values, then the learning model is called a classifier model. When y^ assumes continuous values with ^ = ℝ (or ^ = ℂ) then the learning model is referred toas a model. A learning model may be defined by a mapping or a function ^^: ^ → ^,where ^ = {^^}^^^^ denotes the set of parameters (also referred to as a set of model parameters) thattraining the learning model.

[0057] During a model transfer, the set of parameters (e.g., weights, biases of neurons along with any affine parameters of the normalization layers) that are represented by w^, … , w^,constitute a major fraction of information to be transferred. Note that ^^ ∈ ℝ, ^ = 1, … , ^, whichmeans that parameter values can assume any value over the real line. To transfer (e.g., send, transmit) such parameter values over a wireless channel, the parameter values are quantized and expressed in binary using a finite number of bits. As the wireless channel is noisy and transmissions are prone to distortions and errors due to multiple imperfections in transmit-receive chain hardware units, for example, of the device 200, and distortions introduced by the wireless channel, the bits representing each parameter are protected using FEC. According to FEC, the device 200 may add redundant bits to information bits (e.g., values of parameters of a learning model) in such a way that Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 17 the information bits can be reliably decoded when some of the transmitted bits are flipped or corrupted during the transmission. Thus, transferring a learning model, such as a DNN model,including parameters of the DNN model involves quantization of the parameters ^^, … , ^^, andprotecting the quantized parameter values by adding redundancy through FEC.

[0058] For a given learning model, the task of model transfer may be performed multiple times. In many situations, a network node (e.g., the device 200) may update a model and transfer (e.g., send, transmit) the updated model parameters to another node over the wireless network. Thus, model transfer can consume a significant amount of wireless network resources.

[0059] Another situation where quantization of the parameters of a trained or learned learning model is valuable is the following. It can be understood that, performing computations involving all integer valued parameters (e.g., having a finite precision) would be easier than performing computations involving real numbers that require an infinite, or very long, decimal expansion. While implementing algorithms, methods, procedures, learning models involving real numbers (e.g., numbers that have an infinite decimal expansion), the real numbers can be represented with floating point numbers. For example, quantizing real numbers to 32-bit floating-point numbers is a practical implementation of algorithms, methods, procedures, or learning models whose parameter values are given by real numbers. Quantizing real numbered parameters to a finite-precision may result in losing the precision or accuracy in the output, but often a close approximation to the actual output suffices.

[0060] With regard to implementation of a learning model, quantizing real numbered parameters to a finite-precision has at least two advantages: it takes lower storage space or memory to store the learning model and reduces the signaling overhead in transferring (e.g., sending, transmitting) the learning model over a wireless communications system; and the inference or prediction using the learning would be faster as the computations involves finite-precision arithmetic rather than infinite-precision arithmetic.

[0061] Accordingly, quantization of the set of parameters of a learning model and the amount of error protection given to each parameter during model transfer (e.g., sending, transmitting) over wireless channel would affect the amount of information or number of bits that are transmitted over the wireless channel for model transfer. Further, quantization of the set of parameters of a learning Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 18 model is an important aspect in practically them to use in wireless networks, as it will affect the amount of storage space used to store the learning model and it will affect the speed of computations performed by the learning model during inference.

[0062] Quantization of the set of parameters will affect the prediction or inference performance of the learning model. One problem is to quantize the parameters while ensuring that the model performance does not degrade beyond a threshold value. This problem can be addressed by determining the sensitivity or tolerance range of the set of learning model parameters and quantizing the set of parameters in accordance with the determined sensitivity or tolerance range. Further, if the sensitivity of different parameters of the learning model is known, a higher level of error protection can be provided to the highly sensitive parameters and a lower level of error protection to the parameters having lesser sensitivity (or lesser influence on the model prediction). As the amount of redundancy and, hence, the number of bits to be transmitted will increase with the amount of error protection, this kind of unequal error protection can potentially achieve an optimal or desired trade-off inference performance of the learning model and the amount of error protection that is added.

[0063] The techniques discussed herein determine the sensitivity, or the tolerance range, of the set of parameters of a given learning model so that the prediction or inference performance of the learning model after quantizing its parameters does not degrade beyond a given threshold from its trained (e.g., optimal or desired) prediction or inference performance. These techniques address the practically important and challenging scenario of determining the sensitivity or tolerance range of the set of parameters when a set of unlabeled data samples is available but access to the training data or labeled data is not available.

[0064] Some methods of quantizing learning model parameters require the training data set ^^or, at least a labeled data set to determine the sensitivity of, or tolerance range for, the learned or trained parameters of the learning model. Knowing the training data set is very difficult in some situations. For example, a parametric server (PS) performing model aggregation in a federated learning setting will not have access to the training data, but the PS has to transfer the aggregated model parameters back to the users participating in the federated learning process. By way of another example, consider a situation where a learning model is developed by a third party (e.g., a company specialized in developing learning models that is not a UE or a network vendor) and Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 19 supplied to a network operator to deploy the at a network node (e.g., gNB) or at one or more edge devices (e.g., UEs). Practically, it is difficult for the third party to supply the training data used to develop the model to its customer purchasing the learning model due to privacy concerns and due to the cost incurred in sharing training data sets. At the same time, the network operator purchasing the learning model might want to quantize the parameters of the learning model before deploying it on resource (e.g., memory, computational complexity) constrained devices, such as edge devices, while ensuring that inference or prediction performance of the learning model does not degrade beyond a certain limit due to quantization of its parameters.

[0065] Such scenarios benefit from a technique that allows determining the sensitivity, or tolerance range, of the learning model parameters without having access to the training data set or, in general, a labeled data set. The proposed techniques discussed herein use a set of unlabeled data samples. It is to be noted that the techniques discussed herein can be used on classification learning models, e.g., learning models trained for a classification task.

[0066] A learning model (e.g., an AI / ML model) is developed by training the model over one or more data sets. The data sets can be labeled or unlabeled, leading to supervised training / learning or unsupervised / self-supervised training / learning, respectively. In the case of learning models for wireless communications, the models are typically developed based on training and testing with data sets constructed from simulated data or from the real-world data collected from functioning wireless networks or a combination of both simulated and real-world data.

[0067] Without loss of generality, consider developing a learning method through supervised learning. However, it should be noted that the techniques discussed herein are applicable to learning models developed through unsupervised learning and self-supervised methods as well.

[0068] Let data samples ^^ = {!x^, y^)}^"#^^^ denote the set of labeled training data samples,where x^ ∈ ^ and y^ ∈ ^ denote an input sample and the corresponding label, or equivalently theexpected output sample from a learning model for input x^. It should be noted that y^is also known as the prediction for input x^. In unsupervised or self-supervised learning, the training data set is anunlabeled data set ^^ = {x^}^"#^^^ .Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 20

[0069] A learning model is generally a or a function, ^^, where ^^: ^ → ^. Here,^ = {^^}^^^^ denotes the set of model parameters that are learned during the process of training thethe values of model parameters, e.g., determining ^ using the data set ^^is the “training the model” or “learning the model”. The procedure of supervised learning or training of a learning model consists of reducing (e.g., minimizing) a loss function $. More precisely, the set of optimal parameters (^) is determined by solving the following optimization problem: ^=^ %'∈^ℝ&( $ !^), ^^ )

[0070] The learned or trained mapping) ^^, especially when theunderlying learning model is a DNN, generates a probability distribution *!y|x; ^) over thepredictions y, conditioned on the input x and parameterized by ^, and the conditional distribution is differentiable in ^. Thus, the trained DNN, after training on a data set ^^, generates aprobability distribution P!y|x; ^) that is differentiable in ^.

[0071] When the task is classification, output space contains finitely many values, and thecardinality of the set ^ is finite. Hence, y is a discrete variable with y ∈ {y^, … , y.} and *!y|x; ^)is the probability that y is the predicted label for x under the model parameters ^. In other words,*!y / |x^; ^) is the probability that y / is the label, prediction, or class (with c∈ {1, … , 0}) for thegiven input x^, as per the prediction or inference made by the DNN with ^ as its model parameters.Note that ∑. / ^^ *!y / |x^; ^) = 1, or, equivalently, ∑23∈{24,…,25} *!y / |x^; ^) = 1.

[0072] hand, when the expected output, or the expected prediction or the expectedinference y^ for an input x^ is a continuous valued real number, e.g., when y^ ∈ ℝ, the task beingperformed is referred to as regression and the learning model is referred to as a regression model.

[0073] In a trained DNN, with the trained (e.g., optimal) model parameters given by ^ =6^^, … , ^^78 ∈ ℝ^, the prediction or inference performance of the DNN is more sensitive to someof the parameters in ^^, … , ^^ than the other parameters. In other words, some of the parametersplay a more important role in deciding the output of the DNN. When the value of a sensitive parameter in a DNN is changed even by a small amount from its leaned or trained (e.g., optimal) value, there will be a significant change in the DNN’s prediction or inferences. At the same time, Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 21 some of the parameters of the DNN are not meaning that the DNN’s prediction or inference performance is relatively robust to changes in the values those parameters.

[0074] The trained (e.g., optimal) parameter vector ^ can be modified to produce a parametervector ^9 , where ^9 = ^ + ;, and where E is an error or a perturbation vector of the same size asthat of ^. The prediction or inference performance of a learning model with ^9 as its parameterswould be inferior compared to the learning model with ^ as its parameters. In other words, the performance of ^^9would be inferior to the performance of ^^. If the performance of ^^9is still acceptable, e.g., if the performance of ^^9is not lower than a certain level or threshold when compared with the performance of ^^, then the vector ; represents an acceptable or tolerable change in the parameter values of the DNN under consideration, from its optimal values.

[0075] Accordingly, if the sensitive parameters that are highly influential in determining the prediction or inference performance of the AI / ML model are identified, then quantization of the learning model parameters can be performed such that the sensitive parameters are quantized with a high precision or high resolution quantizer, and the remaining parameters are quantized with a low precision or low resolution quantizer. In other words, sensitive parameters are represented by a larger number of bits and the other (not-so-sensitive) parameters are represented with a lower number of bits. This would potentially reduce the total number of bits needs to represent the DNN parameters, while ensuring desired level of inference performance. Furthermore, the sensitive parameters can be provided with more protection against channel corruption and noise by employing low-rate error control codes, while other parameters can be sent over the channel with relatively lower levels of protection by employing higher rate error control codes. This kind of unequal error protection (UEP) would improve usage of the wireless network resources for transferring the model parameters. A rate of an error control code can be defined as code rate =^DEFGH IJ DKGJDL MNJIHEOPMIN FMPKQIPOL NDEFGH IJ FMPK OJPGH RISMNT.

[0076] With respect to training-based quantization, quantization of the DNN parameters is considered during the training of the network itself. Thus, during training, the parameter values are restricted to discrete values having a finite precision, instead of continuous values over the real line. However, such methods are not applicable when an end user, who receives a DNN from a vendor, would like to quantize its weights to improve resource usage (such as storage, speed of Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 22 computation, network resources required to the learning model to another node or device), while willing to compromise on the inference performance of the AI / ML model.

[0077] With respect to quantization based on parameter sensitivity or permissible loss in inference performance, methods that first determine the sensitivity of the DNN parameters and then perform quantization use training data, or at least labeled data to determine the sensitivity. However, having access to the training data, or labeled data, is not a valid assumption in all practical scenarios, as discussed above.

[0078] There exist some practical scenarios, where it is desirable to determine the sensitivity, or tolerance range of the parameters of a trained learning model, such as a trained DNN, without having access to its training data or labeled data. The techniques discussed herein allow for determining the sensitivity, or tolerance range of the learning model parameters, not only the overall tolerance or sensitivity, but tolerance or sensitivity without having access to the training data set or, in general, a labeled data set. The techniques discussed herein use only a set of unlabeled data samples.

[0079] Further, unequal error protection of model parameters during model transfer over wireless networks is discussed herein. Non-uniform error protection for the model parameters, based on how sensitive the model performance is with respect to those parameters, is discussed here. If the individual tolerance ranges are known, unequal error protection can be performed. The concern / focus of existing works, that determine the error vector. The magnitude of the individual elements of the error vector ; are considered using the techniques discussed herein, rather than just determining the tolerable distortion or sensitivity of parameters (based on training data or labeled data) using the total amount of permissible distortion, which is given by a suitably chosen norm of E. By considering the magnitudes of the individual elements of the error vector, unequal error protection of the model parameters during model transfer over the wireless network is provided.

[0080] Consider a learning model (e.g., an AI / ML model), which is essentially a mapping or afunction, ^^, where ^^: ^ → ^, where ^ denotes the set of optimal model parameters that arelearned during the process of training the model, ^ is the input sample space and ^ is the output sample space. The set of model parameters can be denoted as a vector; e.g., instead of treating ^ asa set, consider ^ as a one dimensional vector such that ^ = 6^^, … , ^^7Q ∈ ℝ^ denotes the vectorFirm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 23 of trained (e.g., optimal) model parameters during the training process and ^ is the total number of parameters of the learning model. Superscript U denotes transpose operation. In the case when the learning model is a DNN, the DNN would include certain number of layers, e.g., V layerswith ℓ^X layer including ^ℓ parameters. Thus, for a DNN including ^ number of parameters, ^ =6^ , … , ^ 78 represen8 ^^ ts the vector of all its parameters and the vector ^ℓ = Y^ℓ,^, … , ^ℓ,^ℓZrepresents the parameters of the ℓ^X layer of the DNN. Note that ^ℓ <

[0081] In supervisedparameters ^ are determined by reducing (e.g., minimizing) aloss function $!⋅) over a labeled training data set ^^ = {!x^, y^)}^"#^^^ , where x^ ∈ ^ and y^ ∈ ^denote an input sample and the corresponding label, or, equivalently, expected output sample from the learning model for input x^. Note that y^is also known as the prediction for input x^. Formally,^ = 6^^, … , ^^7 is determined as follows:^ =^ %'∈^ℝ&] $ !^), ^^ )

[0082] The learned or trained model (or the mapping) ^^, e.g., when the underlying learningmodel is a DNN generates a probability distribution *!y|x; ^) over the predictions y, conditionedon the input x and parameterized by ^, and the conditional distribution is differentiable in ^. Thus, the trained DNN, after training on a data set ^^, generates a probability distribution*!y|x; ^) that is differentiable in ^.

[0083] When the task is classification, the model output space contains finitely many values,and the cardinality of the set ^ is finite. Hence, y is a discrete variable with y ∈ {1, … , 0} and*!y|x; ^) is the probability that y is the predicted label for x under the model parameters ^. Inother words, *!y / |x^; ^) is the probability that y / is the label, prediction, or class for the giveninput x^, as per the prediction or inference made by the DNN with ^ as its model parameters. Notethat ∑. / ^^ *!y / |x^; ^) = 1, or, equivalently, ∑23∈^ *!y / |x^; ^) = 1.

[0084] Although some of the discussions herein refer a learning model that is a classifier DNN (e.g., a DNN trained to perform the task of classifying the input data samples), it is to be appreciated that the techniques discussed herein can also be used with other types of learning models (e.g., other classifier learning models). Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 24

[0085] In a trained DNN, with the optimal parameters given by ^ = 6^^, … , ^^78, theprediction or inference performance of the DNN is more sensitive to some of the parameters in^^, … , ^^. In other words, some of the parameters play a more important role in deciding the outputof the DNN.

[0086] The techniques discussed herein measure the sensitivity of DNN Parameters with unlabeled data samples. These techniques can depend only on the availability of a set of unlabeleddata samples, denoted by ^ = {x^,  x`,   … ,  xa}.

[0087] Consider a trained learning model, or a DNN, with optimal parameters ^ =6^^, … , ^^7 ∈ ℝ^ and a set of unlabeled data samples ^ = {x^,  x`,   … ,  xa}. Let $!^, ^) denotevalue of the loss function computed by the learning model with parameters ^ = 6^^, … , ^^78 overthe set of input samples (or unlabeled samples), ^ = {x^,  x`,   … ,  xa}. Further $^X is the acceptableloss in performance of the learning model with quantized or corrupted parameters compared with the learning model performance with learned or trained (e.g., optimal) parameters.

[0088] Without access to the training data set or a labeled data set, consider b^!^; y) as the lossfunction, e.g., consider $!^, ^) = b^!^; y), where b^!^; y) is the mutual information between theunlabeled data samples ^ = {x^,  x`,   … ,  xa} and the corresponding predictions or inferencesy^, … , ya produced by the learning model with parameters ^. Note that the subscript ^ in b^!^; y)emphasizes that the mutual information is computed with ^ = 6^^, … , ^^78 as the parameters ofthe learning model.

[0089] The mutual information b^!^; y) can be computed as follows. Recall that ^ is the set ofall the parameters of the learning model, and that the mapping ^^produces a conditionaldistribution *!y|x; ^). Compute *!y|x^; ^) for x^ ∈ ^. Compute the empirical marginaldistribution of predicted labels: *!y; ^) = 1a c& d *!y|x^; ^)^^^

[0090] Note that computed distribution *c!y; ^) is an approximation of *!y; ^), the truemarginal distribution of y. The superscript tilde is an indication of this approximation. Mutual information, and hence, the loss function, is given by Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 25 j where=−anda a 1 d^) = − 1d ^) klm ^)

[0091] arecarried out with the learning model having ^ + ; as its parameters, rather than ^.

[0092] The sensitivity, or tolerance range of the AI / ML model parameters can be formulated as the following optimization problem: ;pqr = stm %]su 6$!^ + ;, ^) − $!^, ^)7 < $^X ^ Equation (1)o∈ℝwheremodel withparameters ^ + ; = 6^^ + v^, … , ^^+v^78 over the set of unlabeled samples ^ and $^X is theacceptable level of degradation in the prediction or inference performance of the learning model.

[0093] Solving the above optimization problem is equal to computing an ^ length error vector ; such that the difference in the loss value with optimal parameters ^ and perturbed parameters^ + ; is within an acceptable limit, determined by $^X. Such a vector ; gives the tolerance rangeof the parameters of the learning model. If v^, the ^^Xelement of the error vector ; has a higher magnitude, it implies that ^^, the ^^Xof the learning model, has a higher tolerance range or, relatively, less influence on the learning model performance and hence, it has a lower sensitivity. On the other hand, if an element v^of the error vector ; has a lower magnitude, it implies that the ^^Xparameter of the learning model, ^^, has, relatively, a lower tolerance range and it has, relatively, a higher sensitivity. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 26

[0094] The optimization problem of (1) can be solved in different ways. One technique for determining an approximated version of such an error vector ;, based on an iterative numerical approach, is discussed below.

[0095] An iterative numerical approach for determining the sensitivity, or tolerance range, ofthe parameters of a learning model ^^ based on & unlabeled samples ^ = {x^,  x`,   … ,  xa} includesas inputs: trained (e.g., optimal) parameters ^ = 6^^, … , ^^7 ∈ ℝ^ of the learning model; a set ofunlabeled data samples ^ = {x^,  x`,   … ,  xa}; a step size Δ, where Δ > 0 is a small positive value;and a threshold on loss value $^X. The output is a vector ;pqr (an ^ × 1 error vector indicating thetolerance or sensitivity of the trained (e.g., optimal) learning model parameters).

[0096] The iterative approach is as follows.

[0097] (1). Compute *!y|x^; ^) for x^ ∈ ^.

[0098] (2). Compute the empirical marginal distribution of predicted labels: a − 1 j where,*c!y; ^) = 1& d *!y|x^; ^)^^^ ^)) = − d ^) klm ^)

[0099] (3)78 as theparameters of the learning model, where ^9 = ^ + ;, where ; = 6v ^^, … , v^7 ∈ ℝ , ; ∈ ℰrepresents the error vector, and ℰ denotes the set of all possible ^ length vectors with Euclideannorm equal to } > 0; i.e., ℰ = {;: ∥ ; ∥`= }}.Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 27

[0100] Using Taylor expansion, first order of $i^9 , ^j − $!^, ^) is given byi^9 − ^j8∇^$!^, ^) , where ∇^$!^, ^) = ^^!^,^)^^ (i.e., ∇^$!^, ^) is the gradient of $!^, ^)and $i^9 , ^j − $!^, ^) is a scalar value.

[0102] (4). Set } = Δ

[0103] (5). The error vector ;^that results in highest loss, under the constraint that the Euclidean norm of the error vector is equal to } can be computed as follows: ;^ = stm % 8o∈sℰu ; i∇^$!^, ^)jwhere ℰ = {;: ∥ ; ∥ = }}, ℰ ⊂valued vectors withnorm of each vector equal to }.

[0104] (6). Set ^9 = ^ + ;^ and compute $i^9 , ^j − $!^, ^).

[0105] (7). If 6$i^9 , ^j − $!^, ^)7 < $^X, then increment } ← R + Δ , set ;EGE = ;^ , andreturn to step (5) to compute the error vector ;^.

[0106] (8). If 6$i^9 , ^j − $!^, ^)7 = $^X, then OUTPUT ;pqr = ;^ and stop.

[0107] (9). If 6$i^9 , ^j − $!^, ^)7 > $^X, then OUTPUT ;pqr = ;EGE and stop.

[0108] A large (e.g., maximum) amount of change allowed in the parameters of the AI / ML model, after quantization, is given by the norm of the error vector ;pqrproduced by the above algorithm. Thus, maximum tolerance with a given threshold $^Xon the loss in performance, measured in terms of the Euclidean norm of the maximum permissible error vector, denoted by Upqr, is given by, Upqr = ∥ ;pqr ∥`.

[0109] Note that the actual algorithm is depicted by the steps (4) through (9) (and steps (1) through (3) are to define or explain the different terms and their computations). Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 28

[0110] The proposed technique and the numerical procedure outlined above are discussed by considering Euclidean norm, or, equivalently, $`norm of the error vector (e.g., ∥; ∥`). It should be noted that the proposed technique and the iterative numerical procedure oralgorithm can be generalized to work with any p-norm, e.g., ∥ ; ∥^, where ^ ≥ 1.

[0111] The numerical iterative method or algorithm is depicted by the steps (4) through (9), and steps (1) through (3) are to define or explain the different terms and their computations.

[0112] The accuracy of the estimated Upqrimproves as the value of Δ becomes smaller.

[0113] For an accurate estimate of the sensitivity or tolerance range of the learning modelparameters, it is desirable that the set of data samples ^ = {x^,  x`,   … ,  xa} are independent andidentically distributed (i.i.d) data samples with very similar statistical characteristics as that of the set of training data, ^^, used to develop the learning model.

[0114] The accuracy of the proposed technique improves with the number of data points & inthe data set ^ = {x^,  x`,   … ,  xa}.

[0115] With respect to quantization based on the tolerance range Upqr, having obtained the permissible or allowed tolerance range Upqr, quantization of the model parameters can beperformed in such a way that ∥ ^^ − ^ ∥ ^`≤ Upqr, where ^ is the quantized version of ^. E.g.,^^ = Y^^ , … , ^^Z8is a ^ ^^ ^ n ^ length vector containing the quantized parameter values {^^ }^^^ ,whereversion of ^^.

[0116] Note that the quantization can be a scalar quantization, quantizing each parameter ^^individually as a scalar, or a vector quantizer that quantizes the vector ^ at once.

[0117] With respect to unequal error protection based on error vector ;pqr, a higher value of ^vpqr,^^ (absolute value of the ^^Xelement in the vector ;pqr) indicates that the parameter ^^can tolerate a higher variation in its value, while keeping the performance of the learning model within the limit dictated by the threshold $^X. Such a parameter may be referred to as “not a very sensitive” parameter. On the other hand, a low value of ^vpqr,^^ indicates that the parameter ^^is to be safeguarded so that its value does not change much due to any kind of noisy or corrupt conditions (e.g., due to hardware imperfections or due to the adverse channel) encountered during the Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 29 transmission of these parameters over a Such a parameter might be referred to as a “sensitive” parameter. Thus, during the model transfer, the error protection provided to the parameters through FEC can be optimized by providing higher protection to the highly sensitive parameters (e.g., those having a lower absolute value of vpqr,^) by employing a lower rate code and relatively lower protection to the non-very sensitive parameters by employing a higher rate codes.

[0118] The proposed techniques enable reducing the signaling overhead through informed quantization of, and unequal error protection for, classification learning model parameters without the need for training data and labeled data. Determining the sensitivity or tolerance range of the model parameters without training data and without labeled data by considering the mutual information between the learning model input and output as the cost function is discussed herein. Unequal error protection for the learning model parameters during model transfer over the wireless channels, by determining the tolerance range or sensitivity of each individual parameter of the learning model is also discussed herein.

[0119] Figure 3 illustrates an example of a UE 300 in accordance with aspects of the present disclosure. The UE 300 may include a processor 302, a memory 304, a controller 306, and a transceiver 308. The processor 302, the memory 304, the controller 306, or the transceiver 308, or various combinations thereof or various components thereof may be examples of means for performing various aspects of the present disclosure as described herein. These components may be coupled (e.g., operatively, communicatively, functionally, electronically, electrically) via one or more interfaces.

[0120] The processor 302, the memory 304, the controller 306, or the transceiver 308, or various combinations or components thereof may be implemented in hardware (e.g., circuitry). The hardware may include a processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or other programmable logic device, or any combination thereof configured as or otherwise supporting a means for performing the functions described in the present disclosure.

[0121] The processor 302 may include an intelligent hardware device (e.g., a general-purpose processor, a DSP, a CPU, an ASIC, an FPGA, or any combination thereof). In some implementations, the processor 302 may be configured to operate the memory 304. In some other Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 30 implementations, the memory 304 may be into the processor 302. The processor 302 may be configured to execute computer-readable instructions stored in the memory 304 to cause the UE 300 to perform various functions of the present disclosure.

[0122] The memory 304 may include volatile or non-volatile memory. The memory 304 may store computer-readable, computer-executable code including instructions when executed by the processor 302 cause the UE 300 to perform various functions described herein. The code may be stored in a non-transitory computer-readable medium such as the memory 304 or another type of memory. Computer-readable media includes both non-transitory computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A non-transitory storage medium may be any available medium that may be accessed by a general-purpose or special-purpose computer.

[0123] In some implementations, the processor 302 and the memory 304 coupled with the processor 302 may be configured to cause the UE 300 to perform one or more of the functions described herein (e.g., executing, by the processor 302, instructions stored in the memory 304). For example, the processor 302 may support wireless communication at the UE 300 in accordance with examples as disclosed herein. The UE 300 may be configured to or operable to support a means for receiving a trained learning model that includes a set of model parameters; determining, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantizing parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0124] Additionally, the UE 300 may be configured to support any one or combination of determining a sensitivity value for each parameter in the set of model parameters; and communicating to a device, the set of quantized model parameters, where communicating the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, providing error protection to the parameter based at least in part on the sensitivity value for the parameter; where determining the tolerance range for the set of model parameters further comprises: determining a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generating a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generating a modified trained learning model by replacing the set of model parameters with the set of modified model Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 31 parameters; determining a second loss value of loss function using the set of unlabeled data samples as an input to the modified trained learning model; and using, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters; where a norm of the error vector is equal to a step size parameter, and further including: repeatedly increasing a value of the step size parameter, generating the set of modified model parameters, generating the modified trained learning model, determining the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where determining the sensitivity value for each parameter in the set of model parameters further comprises: determining a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generating a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generating a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determining a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and using, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter; where a norm of the error vector is equal to a step size parameter, and further including: repeatedly increasing a value of the step size parameter, generating the set of modified model parameters, generating the modified trained learning model, determining the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to provide error protection to each parameter in the set of quantized model parameters further comprises: applying an error control coding to each parameter in the set of quantized model parameters, where for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter; where to apply the error control coding further comprises applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value; where the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network; Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 32 where the trained learning model comprises an model; where the trained learning model comprises a classification model; where the method is implemented in a UE; where the method is implemented in a network equipment.

[0125] Additionally, or alternatively, the UE 300 may support at least one memory (e.g., the memory 304) and at least one processor (e.g., the processor 302) coupled with the at least one memory and configured to or operable to cause the UE to: receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0126] Additionally, the UE 300 may be configured to support any one or combination of the at least one processor is configured to or operable to determine a sensitivity value for each parameter in the set of model parameters; and communicate, to a device, the set of quantized model parameters, where to communicate the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, to provide error protection to the parameter based at least in part on the sensitivity value for the parameter; where to determine the tolerance range for the set of model parameters, the at least one processor is further configured to cause the UE to: determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters; where a norm of the error vector is equal to a step size parameter, and the at least one processor is further configured to cause the UE to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 33 threshold value; where to determine the value for each parameter in the set of model parameters, the at least one processor is further configured to cause the UE to: determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter; where a norm of the error vector is equal to a step size parameter, and the at least one processor is further configured to cause the UE to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to provide error protection to each parameter in the set of quantized model parameters, the at least one processor is further configured to cause the UE to: apply an error control coding to each parameter in the set of quantized model parameters, where for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter; where to apply the error control coding includes applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value; where the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network; where the trained learning model comprises an AI / ML model; where the trained learning model comprises a classification model; where the apparatus comprises a UE; where the apparatus comprises a network equipment.

[0127] The controller 306 may manage input and output signals for the UE 300. The controller 306 may also manage peripherals not integrated into the UE 300. In some implementations, the controller 306 may utilize an operating system such as iOS®, ANDROID®, WINDOWS®, or other Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 34 operating systems. In some implementations, 306 may be implemented as part of the processor 302.

[0128] In some implementations, the UE 300 may include at least one transceiver 308. In some other implementations, the UE 300 may have more than one transceiver 308. The transceiver 308 may represent a wireless transceiver. The transceiver 308 may include one or more receiver chains 310, one or more transmitter chains 312, or a combination thereof.

[0129] A receiver chain 310 may be configured to receive signals (e.g., control information, data, packets) over a wireless medium. For example, the receiver chain 310 may include one or more antennas to receive a signal over the air or wireless medium. The receiver chain 310 may include at least one amplifier (e.g., a low-noise amplifier (LNA)) configured to amplify the received signal. The receiver chain 310 may include at least one demodulator configured to demodulate the receive signal and obtain the transmitted data by reversing the modulation technique applied during transmission of the signal. The receiver chain 310 may include at least one decoder for decoding the demodulated signal to receive the transmitted data.

[0130] A transmitter chain 312 may be configured to generate and transmit signals (e.g., control information, data, packets). The transmitter chain 312 may include at least one modulator for modulating data onto a carrier signal, preparing the signal for transmission over a wireless medium. The at least one modulator may be configured to support one or more techniques such as amplitude modulation (AM), frequency modulation (FM), or digital modulation schemes like phase-shift keying (PSK) or quadrature amplitude modulation (QAM). The transmitter chain 312 may also include at least one power amplifier configured to amplify the modulated signal to an appropriate power level suitable for transmission over the wireless medium. The transmitter chain 312 may also include one or more antennas for transmitting the amplified signal into the air or wireless medium.

[0131] Figure 4 illustrates an example of a processor 400 in accordance with aspects of the present disclosure. The processor 400 may be an example of a processor configured to perform various operations in accordance with examples as described herein. The processor 400 may include a controller 402 configured to perform various operations in accordance with examples as described herein. The processor 400 may optionally include at least one memory 404, which may be, for example, an L1 / L2 / L3 cache. Additionally, or alternatively, the processor 400 may optionally Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 35 include one or more arithmetic-logic units 406. One or more of these components may be in electronic communication or otherwise coupled (e.g., operatively, communicatively, functionally, electronically, electrically) via one or more interfaces (e.g., buses).

[0132] The processor 400 may be a processor chipset and include a protocol stack (e.g., a software stack) executed by the processor chipset to perform various operations (e.g., receiving, obtaining, retrieving, transmitting, outputting, forwarding, storing, determining, identifying, accessing, writing, reading) in accordance with examples as described herein. The processor chipset may include one or more cores, one or more caches (e.g., memory local to or included in the processor chipset (e.g., the processor 400) or other memory (e.g., random access memory (RAM), read-only memory (ROM), dynamic RAM (DRAM), synchronous dynamic RAM (SDRAM), static RAM (SRAM), ferroelectric RAM (FeRAM), magnetic RAM (MRAM), resistive RAM (RRAM), flash memory, phase change memory (PCM), and others).

[0133] The controller 402 may be configured to manage and coordinate various operations (e.g., signaling, receiving, obtaining, retrieving, transmitting, outputting, forwarding, storing, determining, identifying, accessing, writing, reading) of the processor 400 to cause the processor 400 to support various operations in accordance with examples as described herein. For example, the controller 402 may operate as a control unit of the processor 400, generating control signals that manage the operation of various components of the processor 400. These control signals include enabling or disabling functional units, selecting data paths, initiating memory access, and coordinating timing of operations.

[0134] The controller 402 may be configured to fetch (e.g., obtain, retrieve, receive) instructions from the memory 404 and determine subsequent instruction(s) to be executed to cause the processor 400 to support various operations in accordance with examples as described herein. The controller 402 may be configured to track memory addresses of instructions associated with the memory 404. The controller 402 may be configured to decode instructions to determine the operation to be performed and the operands involved. For example, the controller 402 may be configured to interpret the instruction and determine control signals to be output to other components of the processor 400 to cause the processor 400 to support various operations in accordance with examples as described herein. Additionally, or alternatively, the controller 402 may be configured to manage Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 36 flow of data within the processor 400. The 402 may be configured to control transfer of data between registers, ALUs 406, and other functional units of the processor 400.

[0135] The memory 404 may include one or more caches (e.g., memory local to or included in the processor 400 or other memory, such as RAM, ROM, DRAM, SDRAM, SRAM, MRAM, flash memory, etc. In some implementations, the memory 404 may reside within or on a processor chipset (e.g., local to the processor 400). In some other implementations, the memory 404 may reside external to the processor chipset (e.g., remote to the processor 400).

[0136] The memory 404 may store computer-readable, computer-executable code including instructions that, when executed by the processor 400, cause the processor 400 to perform various functions described herein. The code may be stored in a non-transitory computer-readable medium such as system memory or another type of memory. The controller 402 and / or the processor 400 may be configured to execute computer-readable instructions stored in the memory 404 to cause the processor 400 to perform various functions. For example, the processor 400 and / or the controller 402 may be coupled with or to the memory 404, the processor 400, and the controller 402, and may be configured to perform various functions described herein. In some examples, the processor 400 may include multiple processors and the memory 404 may include multiple memories. One or more of the multiple processors may be coupled with one or more of the multiple memories, which may, individually or collectively, be configured to perform various functions herein.

[0137] The one or more ALUs 406 may be configured to support various operations in accordance with examples as described herein. In some implementations, the one or more ALUs 406 may reside within or on a processor chipset (e.g., the processor 400). In some other implementations, the one or more ALUs 406 may reside external to the processor chipset (e.g., the processor 400). One or more ALUs 406 may perform one or more computations such as addition, subtraction, multiplication, and division on data. For example, one or more ALUs 406 may receive input operands and an operation code, which determines an operation to be executed. One or more ALUs 406 may be configured with a variety of logical and arithmetic circuits, including adders, subtractors, shifters, and logic gates, to process and manipulate the data according to the operation. Additionally, or alternatively, the one or more ALUs 406 may support logical operations such as AND, OR, exclusive-OR (XOR), not-OR (NOR), and not-AND (NAND), enabling the one or more ALUs 406 to handle conditional operations, comparisons, and bitwise operations. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 37

[0138] The processor 400 may support communication in accordance with examples as disclosed herein. The processor 400 may be configured to or operable to support at least one controller (e.g., the controller 402) coupled with at least one memory (e.g., the memory 404) and configured to cause the processor to: receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0139] Additionally, the processor 400 may be configured to or operable to support any one or combination of the at least one controller is configured to cause the processor to determine a sensitivity value for each parameter in the set of model parameters; and communicate, to a device, the set of quantized model parameters, where to communicate the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, to provide error protection to the parameter based at least in part on the sensitivity value for the parameter; where to determine the tolerance range for the set of model parameters, the at least one controller is further configured to cause the processor to: determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters; where a norm of the error vector is equal to a step size parameter, and the at least one controller is further configured to cause the processor to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to determine the sensitivity value for each parameter in the set of model parameters, the at least one processor is further configured to cause the processor Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 38 to: determine a first loss value of a loss the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter; where a norm of the error vector is equal to a step size parameter, and the at least one controller is further configured to cause the processor to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to provide error protection to each parameter in the set of quantized model parameters, the at least one controller is further configured to cause the processor to: apply an error control coding to each parameter in the set of quantized model parameters, where for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter; where to apply the error control coding includes applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value; where the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network; where the trained learning model comprises an AI / ML model; where the trained learning model comprises a classification model; where the processor is included in a UE; where the processor is included in a network equipment.

[0140] Figure 5 illustrates an example of a NE 500 in accordance with aspects of the present disclosure. The NE 500 may include a processor 502, a memory 504, a controller 506, and a transceiver 508. The processor 502, the memory 504, the controller 506, or the transceiver 508, or various combinations thereof or various components thereof may be examples of means for performing various aspects of the present disclosure as described herein. These components may be Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 39 coupled (e.g., operatively, communicatively, electronically, electrically) via one or more interfaces.

[0141] The processor 502, the memory 504, the controller 506, or the transceiver 508, or various combinations or components thereof may be implemented in hardware (e.g., circuitry). The hardware may include a processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or other programmable logic device, or any combination thereof configured as or otherwise supporting a means for performing the functions described in the present disclosure.

[0142] The processor 502 may include an intelligent hardware device (e.g., a general-purpose processor, a DSP, a CPU, an ASIC, an FPGA, or any combination thereof). In some implementations, the processor 502 may be configured to operate the memory 504. In some other implementations, the memory 504 may be integrated into the processor 502. The processor 502 may be configured to execute computer-readable instructions stored in the memory 504 to cause the NE 500 to perform various functions of the present disclosure.

[0143] The memory 504 may include volatile or non-volatile memory. The memory 504 may store computer-readable, computer-executable code including instructions when executed by the processor 502 cause the NE 500 to perform various functions described herein. The code may be stored in a non-transitory computer-readable medium such as the memory 504 or another type of memory. Computer-readable media includes both non-transitory computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A non-transitory storage medium may be any available medium that may be accessed by a general-purpose or special-purpose computer.

[0144] In some implementations, the processor 502 and the memory 504 coupled with the processor 502 may be configured to cause the NE 500 to perform one or more of the functions described herein (e.g., executing, by the processor 502, instructions stored in the memory 504). For example, the processor 502 may support wireless communication at the NE 500 in accordance with examples as disclosed herein. The NE 500 may be configured to support a means for receiving a trained learning model that includes a set of model parameters; determining, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantizing Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 40 parameters in the set of model parameters least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0145] Additionally, the NE 500 may be configured to support any one or combination of determining a sensitivity value for each parameter in the set of model parameters; and communicating to a device, the set of quantized model parameters, where communicating the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, providing error protection to the parameter based at least in part on the sensitivity value for the parameter; where determining the tolerance range for the set of model parameters further comprises: determining a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generating a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generating a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determining a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and using, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters; where a norm of the error vector is equal to a step size parameter, and further including: repeatedly increasing a value of the step size parameter, generating the set of modified model parameters, generating the modified trained learning model, determining the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where determining the sensitivity value for each parameter in the set of model parameters further comprises: determining a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generating a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generating a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determining a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and using, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 41 vector corresponding to the parameter; where a of the error vector is equal to a step size parameter, and further including: repeatedly increasing a value of the step size parameter, generating the set of modified model parameters, generating the modified trained learning model, determining the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to provide error protection to each parameter in the set of quantized model parameters further comprises: applying an error control coding to each parameter in the set of quantized model parameters, where for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter; where to apply the error control coding further comprises applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value; where the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network; where the trained learning model comprises an AI / ML model; where the trained learning model comprises a classification model; where the method is implemented in a UE; where the method is implemented in a network equipment.

[0146] Additionally, or alternatively, the NE 500 may support at least one memory (e.g., the memory 504) and at least one processor (e.g., the processor 502) coupled with the at least one memory and configured to or operable to cause the NE to: receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

[0147] Additionally, the NE 500 may be configured to or operable to support any one or combination of the at least one processor is configured to cause the NE to determine a sensitivity value for each parameter in the set of model parameters; and communicate, to a device, the set of quantized model parameters, where to communicate the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, to provide error protection to the parameter based at least in part on the sensitivity value for the parameter; where to determine the tolerance range for the set of model parameters, the at least one processor is further configured to Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 42 cause the NE to: determine a first loss value of function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters; where a norm of the error vector is equal to a step size parameter, and the at least one processor is further configured to cause the NE to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to determine the sensitivity value for each parameter in the set of model parameters, the at least one processor is further configured to cause the NE to: determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter; where a norm of the error vector is equal to a step size parameter, and the at least one processor is further configured to cause the NE to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value; where to provide error protection to each parameter in the set of quantized model parameters, the at least one processor is further configured to cause the NE to: apply an error control coding to each parameter in the set of quantized model parameters, where for each Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 43 parameter in the set of quantized model a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter; where to apply the error control coding includes applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value; where the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network; where the trained learning model comprises an AI / ML model; where the trained learning model comprises a classification model; where the apparatus comprises a UE; where the apparatus comprises a network equipment.

[0148] The controller 506 may manage input and output signals for the NE 500. The controller 506 may also manage peripherals not integrated into the NE 500. In some implementations, the controller 506 may utilize an operating system such as iOS®, ANDROID®, WINDOWS®, or other operating systems. In some implementations, the controller 506 may be implemented as part of the processor 502.

[0149] In some implementations, the NE 500 may include at least one transceiver 508. In some other implementations, the NE 500 may have more than one transceiver 508. The transceiver 508 may represent a wireless transceiver. The transceiver 508 may include one or more receiver chains 510, one or more transmitter chains 512, or a combination thereof.

[0150] A receiver chain 510 may be configured to receive signals (e.g., control information, data, packets) over a wireless medium. For example, the receiver chain 510 may include one or more antennas to receive a signal over the air or wireless medium. The receiver chain 510 may include at least one amplifier (e.g., a low-noise amplifier (LNA)) configured to amplify the received signal. The receiver chain 510 may include at least one demodulator configured to demodulate the receive signal and obtain the transmitted data by reversing the modulation technique applied during transmission of the signal. The receiver chain 510 may include at least one decoder for decoding the demodulated signal to receive the transmitted data.

[0151] A transmitter chain 512 may be configured to generate and transmit signals (e.g., control information, data, packets). The transmitter chain 512 may include at least one modulator for modulating data onto a carrier signal, preparing the signal for transmission over a wireless medium. The at least one modulator may be configured to support one or more techniques such as amplitude Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 44 modulation (AM), frequency modulation (FM) digital modulation schemes like phase-shift keying (PSK) or quadrature amplitude modulation (QAM). The transmitter chain 512 may also include at least one power amplifier configured to amplify the modulated signal to an appropriate power level suitable for transmission over the wireless medium. The transmitter chain 512 may also include one or more antennas for transmitting the amplified signal into the air or wireless medium.

[0152] Figure 6 illustrates a flowchart of a method in accordance with aspects of the present disclosure. The operations of the method may be implemented by a device as described herein, such as a UE or a NE. In some implementations, the UE may execute a set of instructions to control the function elements of the UE to perform the described functions. Additionally or alternatively, the NE may execute a set of instructions to control the function elements of the NE to perform the described functions.

[0153] At 602, the method may include receiving a trained learning model that includes a set of model parameters. The operations of 602 may be performed in accordance with examples as described herein. In some implementations, aspects of the operations of 602 may be performed by a UE as described with reference to Figure 3. Additionally or alternatively, aspects of the operations of 602 may be performed by a NE as described with reference to Figure 5.

[0154] At 604, the method may include determining, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters. The operations of 604 may be performed in accordance with examples as described herein. In some implementations, aspects of the operations of 604 may be performed by a UE as described with reference to Figure 3. Additionally or alternatively, aspects of the operations of 604 may be performed by a NE as described with reference to Figure 5.

[0155] At 606, the method may include quantizing parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model. The operations of 606 may be performed in accordance with examples as described herein. In some implementations, aspects of the operations of 606 may be performed a UE as described with reference to Figure 3. Additionally or alternatively, aspects of the operations of 604 may be performed by a NE as described with reference to Figure 5. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 45

[0156] It should be noted that the method herein describes a possible implementation, and that the operations and the steps may be rearranged or otherwise modified and that other implementations are possible.

[0157] It should be noted that the method described herein describes a possible implementation, and that the operations and the steps may be rearranged or otherwise modified and that other implementations are possible.

[0158] The description herein is provided to enable a person having ordinary skill in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to a person having ordinary skill in the art, and the generic principles defined herein may be applied to other variations without departing from the scope of the disclosure. Thus, the disclosure is not limited to the examples and designs described herein but is to be accorded the broadest scope consistent with the principles and novel features disclosed herein. Firm Ref. No. SMM920240010- WO-PCT

Claims

Lenovo Ref. No. SMM920240010-WO-PCT 46 What is claimed is:

1. An apparatus, comprising: at least one memory; and at least one processor coupled with the at least one memory and operable to cause the apparatus to: receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

2. The apparatus of claim 1, wherein the at least one processor is further operable to cause the apparatus to: determine a sensitivity value for each parameter in the set of model parameters; and communicate, to a device, the set of quantized model parameters, wherein to communicate the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, to provide error protection to the parameter based at least in part on the sensitivity value for the parameter.

3. The apparatus of claim 1, wherein to determine the tolerance range for the set of model parameters, the at least one processor is further operable to cause the apparatus to: determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 47 use, based at least in part on whether a between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters.

4. The apparatus of claim 3, wherein a norm of the error vector is equal to a step size parameter, and the at least one processor is further operable to cause the apparatus to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value.

5. The apparatus of claim 2, wherein to determine the sensitivity value for each parameter in the set of model parameters, the at least one processor is further operable to cause the apparatus to: determine a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generate a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generate a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determine a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and use, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, as the sensitivity value for each parameter in the set of model parameters, a value in the error vector corresponding to the parameter.

6. The apparatus of claim 5, wherein a norm of the error vector is equal to a step size parameter, and the at least one processor is further operable to cause the apparatus to: repeatedly increase a value of the step size parameter, generate the set of modified model parameters, generate the modified trained learning model, determine the second loss value using the set of unlabeled data samples as input to the modified trained learning model, until the difference between the first loss value and the second loss value is greater than the threshold value. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 48 7. The apparatus of claim 5, wherein to error protection to each parameter in the set of quantized model parameters, the at least one processor is further operable to cause the apparatus to: apply an error control coding to each parameter in the set of quantized model parameters, wherein for each parameter in the set of quantized model parameters, a rate of the error control coding applied to the parameter is based at least in part on the sensitivity value for the parameter.

8. The apparatus of claim 7, wherein to apply the error control coding includes applying a higher rate of error control coding to parameters having a higher sensitivity value than to parameters having a lower sensitivity value.

9. The apparatus of claim 1, wherein the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network.

10. The apparatus of claim 1, wherein the trained learning model comprises an artificial intelligence (AI) / machine learning (ML) model.

11. The apparatus of claim 1, wherein the trained learning model comprises a classification model.

12. The apparatus of claim 1, wherein the apparatus comprises a user equipment (UE).

13. The apparatus of claim 1, wherein the apparatus comprises a network equipment.

14. A processor, comprising: at least one controller coupled with at least one memory and operable to cause the processor to: receive a trained learning model that includes a set of model parameters; determine, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantize parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

15. A method, comprising: Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 49 receiving a trained learning model that a set of model parameters; determining, based at least in part on a set of unlabeled data samples, a tolerance range for the set of model parameters; and quantizing parameters in the set of model parameters based at least in part on the tolerance range to generate a set of quantized model parameters for the trained learning model.

16. The method of claim 15, further comprising: determining a sensitivity value for each parameter in the set of model parameters; and communicating to a device, the set of quantized model parameters, wherein communicating the set of quantized model parameters includes, for each parameter in the set of quantized model parameters, providing error protection to the parameter based at least in part on the sensitivity value for the parameter.

17. The method of claim 15, wherein determining the tolerance range for the set of model parameters further comprises: determining a first loss value of a loss function using the set of unlabeled data samples as an input to the trained learning model; generating a set of modified model parameters by using an error vector to change one or more parameters of the set of model parameters; generating a modified trained learning model by replacing the set of model parameters with the set of modified model parameters; determining a second loss value of the loss function using the set of unlabeled data samples as an input to the modified trained learning model; and using, based at least in part on whether a difference between the first loss value and the second loss value is less than a threshold value, a norm of the error vector as the tolerance range for the set of model parameters.

18. The method of claim 15, wherein the trained learning model comprises a neural network and the model parameters comprise weights of multiple neurons in the neural network.

19. The method of claim 15, wherein the trained learning model comprises an artificial intelligence (AI) / machine learning (ML) model. Firm Ref. No. SMM920240010- WO-PCTLenovo Ref. No. SMM920240010-WO-PCT 50 20. The method of claim 15, wherein the is implemented in a user equipment (UE) or in a network equipment. Firm Ref. No. SMM920240010- WO-PCT

Images