Error correction and validation in training artificial intelligence / machine learning models

Abstract

Error correction and validation related techniques are described for training robust artificial intelligence / machine learning (AI / ML) models in wireless communications. The AI / ML models are trained with intentional errors. The trained AI / ML models are then utilized in user equipment (UE) or network nodes of a wireless network.

Description

[0001] Cross-referencing

[0002] This disclosure claims priority to U.S. Provisional Patent Applications Nos. 63 / 598,156 and 63 / 598,158, filed November 13, 2023, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure generally relates to wireless communications, and more specifically, to error correction and verification of training robust artificial intelligence / machine learning (AI / ML) models in wireless communications. Background Technology

[0004] Unless otherwise stated, the methods described in this section are not prior art to the following claims and are not considered prior art because they are included in this section.

[0005] In communication systems, such as wireless communications conforming to the Generation Partnership Project (3GPP) standards, many functions on the user equipment (UE) side often have corresponding dual functions on the network side, and vice versa. In the context of artificial intelligence / machine learning (AI / ML), this can be called a bilateral AI / ML model, also known as an autoencoder. For example, the modulation function of the UE / network has a demodulation function on the network / UE; the quantization function of the UE / network has an inverse quantization function on the network / UE; the forward error correction (FEC) encoder of the UE / network has a decoder on the network / UE; the signal shaping function of the UE / network has an inverse shaping function on the network / UE, and vice versa. Some functions / applications also require complementary modules on both the UE and network sides, such as channel state information (CSI) compression, denoising (or noise reduction), quantization, coding, error correction codes, modulation, peak-to-average power ratio (PAPR) reduction, and image compression. In short, the ideal approach in a two-sided AI / machine learning model is to train both ends together so that the functionality of one end is compatible with the corresponding functionality of the other end.

[0006] In the context of artificial intelligence / machine learning, the receptive field (RF) expands rapidly or exponentially across the entire input, depending on the number of layers. In convolutional neural network (CNN) models, the receptive field can rapidly expand from the input layer to an intermediate layer, and then to subsequent intermediate layers, using the output of the previous layer as input. Therefore, capturing the local dependencies of a CNN model is helpful. In transformer models with two intermediate layers, even a single layer can have a global receptive field. That is, each element in a subsequent layer may be influenced by all elements in the previous layer, and elements in subsequent layers often share some common information from the previous layer. Therefore, capturing the local and long-range dependencies of a transformer model is beneficial. Similarly, in deep neural network (DNN) models, even a single layer can have a global receptive field. Therefore, capturing the local and long-range dependencies of a DNN model is also beneficial.

[0007] In many real-world scenarios, interdependencies exist between input elements. That is, some input elements may reveal information about other elements. Some real-world examples include translation (e.g., "It's cloudy today, it's going to rain soon"), autocorrect, and annotation. Therefore, information from missing or corrupted elements can be corrected using information from uncorrupted elements. In applications within the context of wireless communications, such as CSI compression and CSI prediction, a global receptive field is likely to emerge when the input passes through a two-layer grid learning model (e.g., a transformer model or CNN model followed by a DNN model). All latent variable / output elements are influenced by the entire input element; each latent variable element in a bilateral AI / machine learning model or each output element in a one-sided AI / machine learning model may provide information about other elements in the same space. Therefore, any error in the latent variable space or input space will negatively impact the output. For example, an error in the latent variable space (e.g., an error in the transmission medium) occurring during CSI feedback will cause the network decoder's input to no longer be equivalent to the user equipment (UE) encoder's output. The error in the latent variable space alters the decoder's expected input and inevitably affects the decoder's task. Another example is an error occurring in the input space before the data is fed into the AI / machine learning model (e.g., errors in the transmission medium and / or storage). This error alters the nominal input to the AI / machine learning model and inevitably affects the AI / machine learning task. Therefore, an error correction and verification solution is needed when training robust AI / machine learning models in wireless communications. Summary of the Invention

[0008] The following abstract is for illustrative purposes only and is not intended to be limiting in any way. That is, the following abstract aims to introduce the concepts, key points, benefits, and advantages of the novel and non-obvious techniques described herein. Detailed descriptions of specific embodiments will follow. Therefore, the following abstract is not intended to identify the essential characteristics of the claimed subject matter, nor is it intended to define the scope of the claimed subject matter.

[0009] One objective of this disclosure is to propose solutions or strategies for addressing the problems described herein. More specifically, the various solutions proposed herein relate to error correction and verification for training robust artificial intelligence / machine learning (AI / ML) models in wireless communications. It is believed that implementations of the proposed solutions can solve or mitigate the aforementioned problems. The various solutions proposed herein can be applied to a variety of applications and scenarios, such as, but not limited to, channel state information (CSI) compression, denoising (or noise reduction), quantization, coding, error correction codes, modulation, peak-to-average power ratio (PAPR) reduction, and image compression.

[0010] In one aspect, a method may include training an artificial intelligence / machine learning model using intentional errors. The method may also include a processor utilizing the trained artificial intelligence / machine learning model in a network node of a user equipment (UE) or wireless network.

[0011] It is worth noting that while the content described herein may be set against the backdrop of certain wireless access technologies, networks, and wireless communication network topologies, such as fifth-generation (5G) / new radio (NR) / sixth-generation (6G) mobile communications, the proposed concepts, schemes, and any variations / derivatives thereof can also be implemented, used, and realized in other types of wireless access technologies, networks, and network topologies, such as, but not limited to, Evolved Packet System (EPS), Long-Term Evolution (LTE), LTE-Advanced, LTE-Advanced Pro, Internet of Things (IoT), Narrow Band Internet of Things (NB-IoT), Industrial Internet of Things (IIoT), Vehicle-to-Everything (V2X), and non-terrestrial network (NTN) communications. Therefore, the scope of this disclosure is not limited to the examples described herein. Attached Figure Description

[0012] The accompanying drawings are included in this specification to further understand this disclosure and form part of this disclosure. The drawings illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure. It will be understood that the drawings are not necessarily drawn to scale, as some components may be shown out of proportion to their actual dimensions in order to clearly illustrate the concepts of this disclosure.

[0013] Figure 1 This is a schematic diagram of an example network environment in which various proposed solutions of this disclosure can be implemented.

[0014] Figure 2 This is an example design schematic diagram based on the proposed solution in this disclosure.

[0015] Figure 3 This is an example design schematic diagram based on the proposed solution in this disclosure.

[0016] Figure 4 This is an example design schematic diagram based on the proposed solution in this disclosure.

[0017] Figure 5 This is an example design schematic diagram based on the proposed solution in this disclosure.

[0018] Figure 6 This is an example design schematic diagram based on the proposed solution in this disclosure.

[0019] Figure 7 This is an example design schematic diagram based on the proposed solution in this disclosure.

[0020] Figure 8 This is an example communication system block diagram based on the proposed solution of this disclosure.

[0021] Figure 9 This is an example flowchart based on the proposed solution in this disclosure. Detailed Implementation

[0022] Detailed embodiments and implementations of the subject matter claimed in this application are disclosed herein. However, it should be understood that the disclosed embodiments and implementations are merely illustrative of the claimed subject matter, which can be implemented in various forms. This disclosure can be implemented in many different forms and should not be construed as being limited to the exemplary embodiments and implementations set forth herein. Rather, these exemplary embodiments and implementations are provided to make the description of this disclosure more detailed and complete, and to fully communicate the scope of this disclosure to those skilled in the art. In the following description, details of well-known features and techniques may be omitted to avoid unnecessarily obscuring the presented embodiments and implementations.

[0023] Overview

[0024] According to the implementation of this disclosure, various techniques, methods, schemes, and / or solutions are involved in error correction and verification for training robust artificial intelligence / machine learning models in wireless communications. According to this disclosure, multiple possible solutions can be implemented individually or in combination. That is, although these possible solutions are described separately below, two or more possible solutions can be implemented in some combination.

[0025] Figure 1 An example network environment 100 is shown in which the various solutions and schemes disclosed herein can be implemented. Figures 2 to 9 Examples of implementing various proposed schemes in a network environment 100 according to this disclosure are shown. The following description of the various proposed schemes is in conjunction with... Figures 1 to 9 It was carried out.

[0026] See Figure 1 Network environment 100 may involve user equipment (UE) 110 communicating wirelessly with radio access network (RAN) 120 (e.g., 5G NR / 6G mobile network, other types of networks such as non-terrestrial network (NTN), or future generation network). UE 110 may communicate wirelessly with RAN 120 via terrestrial network node 125 (e.g., base station, eNB, gNB, or transmit-and-receive point (TRP)) or non-terrestrial network node 128 (e.g., satellite), and UE 110 may be within the coverage area of ​​cell 135 associated with terrestrial network node 125 and / or non-terrestrial network node 128. RAN 120 may be part of wireless network 130. In network environment 100, UE 110 and wireless network 130 (via terrestrial network node 125 and / or non-terrestrial network node 128) can implement various schemes for error correction and verification of training robust artificial intelligence / machine learning models in wireless communication. It is worth noting that although the various proposed schemes, options, and methods are described separately below, in practical applications, these proposed schemes, options, and methods can be implemented individually or in combination. That is, in some cases, each or more proposed schemes, options, and methods can be implemented individually or separately. In other cases, some or all of the proposed schemes, options, and methods can be implemented in combination.

[0027] Figure 2An example design 200 based on the scheme proposed in this disclosure is shown. In the proposed scheme, in a bilateral artificial intelligence / machine learning model, an error corrector or error detector can be used to correct or detect errors before they are sent to subsequent stages or the receiving end for further processing. Here, the term "pluggable" refers to the ability to use an error corrector (which may be interchangeably referred to herein as a "pluggable corrector," "lightweight pluggable corrector," or LPC) without altering the artificial intelligence / machine learning model. Figure 2 As shown in section (A), and / or the error detector (which may be interchangeably referred to herein as a "pluggable corrector", "lightweight pluggable corrector", or LPD), as Figure 2 As shown in Section (B), add or insert artificial intelligence / machine learning models to improve performance. See also Figure 2 In Part (A), AI / machine learning models can map erroneous latent variables to correct latent variables during training by utilizing an error corrector. For example, during training, intentional errors are provided to the LPC along with the input or latent variable input. The LPC outputs the corrected latent variable, which is compared to the true input or true latent variable. The comparison result can be used as feedback for backpropagation to correct errors in training the AI / machine learning model, thereby improving its performance. See also... Figure 2 In part (B), AI / machine learning models can map erroneous latent variables to correct latent variables during training by utilizing error detectors. For example, during training, intentional errors are provided to an LPD along with inputs or latent variable inputs. The LPD outputs a detection result (e.g., an error detection probability or a simple "yes" or "no" answer regarding error detection), which is compared to the intentional error. The comparison result can be used as feedback for backpropagation to train the AI / machine learning model's error detection, thereby improving the model's performance. It is noteworthy that error correctors can not only detect errors but also repair or otherwise correct them, while error detectors can only detect errors and cannot repair or correct them.

[0028] Figure 3 An example design 300 based on the scheme proposed in this disclosure is shown. Design 300 may involve utilizing LPC or LPD in the inference phase of a bilateral artificial intelligence / machine learning model. See also Figure 3LPD / LPC can be inserted into the inference stage between the encoder (e.g., at the UE) and decoder (e.g., at the network) of a bilateral AI / machine learning model. LPD / LPC detects or corrects errors and provides the detection decision or correction result as feedback. The advantage is that the performance of AI / machine learning models trained in this way can be improved, especially through error correction via LPC. Furthermore, error detection via LPD avoids unnecessary execution of erroneous inputs. Additionally, retransmission can be requested when an error is detected by LPD / LPC. Moreover, the use of LPD / LPC can go beyond the capabilities of error correction codes.

[0029] Figure 4 An example design 400 based on the scheme proposed in this disclosure is shown. Design 400 may involve utilizing LPC or LPD in a one-sided artificial intelligence / machine learning model. See also Figure 4 In section (A), LPD or LPC can be applied to the input of a one-sided artificial intelligence / machine learning model to train the model to detect or correct errors. See also Figure 4 In part (B), LPD or LPC can be applied during the inference phase to train the model to detect or correct errors.

[0030] Figure 5 An example design 500 based on the scheme proposed in this disclosure is shown. Design 500 may involve utilizing LPC or LPD in a bilateral artificial intelligence / machine learning model. See also Figure 5 Considering the bilateral model from an end-to-end perspective, it can be viewed as a single, one-sided artificial intelligence / machine learning model. Therefore, LPD / LPC can be like... Figure 4 The above is applied to a bilateral artificial intelligence / machine learning model.

[0031] Figure 6 An example design 600 based on the scheme proposed in this disclosure is shown. Design 600 may involve training a robust bilateral artificial intelligence / machine learning model. Under the proposed scheme, the bilateral artificial intelligence / machine learning model can be forced to focus on the correction of latent variables during the training phase, thereby exposing it to intentional errors during the training phase. See also Figure 6Intentional errors can be introduced in the intermediate stage between the encoder and decoder (e.g., simulating errors in transmission) or directly as input to the decoder. The output generated by the erroneous latent variable or input is then compared with the true input / latent variable, and the comparison result can be used as feedback for backpropagation to train the AI / machine learning model for error detection or correction. The trained encoder and decoder can then be used for inference. Design 600 can be easily extended to all training types. Furthermore, the design can be generalized to accommodate any type of error. Moreover, the design can even be extended to one-sided AI / machine learning models with intentional errors in the input.

[0032] Figure 7 An example design 700 based on the proposed scheme is shown. Design 700 may involve training a robust one-sided artificial intelligence / machine learning (AI / ML) model. Under the proposed scheme, the one-sided AI / ML model can be forced to focus on the correction of latent variables during the training phase, thereby exposing the model to intentional errors during the training phase. See also Figure 7 Intentional errors can be introduced into the input, resulting in erroneous input being provided to a one-sided AI / machine learning model during training. The model's output based on this erroneous input can then be compared with the true input / latent variable, and the comparison result can be used as feedback for backpropagation to train the AI / machine learning model for error detection or correction. The trained AI / machine learning model can then be used for inference. Design 700 can be easily extended to all input types, such as words, images, long texts, numerical values, or measurements. Furthermore, this design can be generalized to apply to any type of error.

[0033] Given the foregoing, under some of the proposed solutions, error correction and validation in AI / machine learning models can involve the correlation between the inputs and latent variables of a typical AI / machine learning model. The correlation between latent variables and inputs (LPC) can be used for both bilateral AI / machine learning models and their inputs, as well as for one-sided AI / machine learning models. Similarly, the detection of latent variables and inputs (LPD) can be used for both bilateral AI / machine learning models and their inputs, as well as for one-sided AI / machine learning models. Furthermore, under other proposed solutions, robust bilateral AI / machine learning models and robust one-sided AI / machine learning models can be trained by intentionally introducing errors during the training phase without using LPD or LPC. Robust AI / machine learning models can be implemented in User Equipment (UE) 110, terrestrial network node 125, non-terrestrial network node 128, Radio Access Network (RAN) 120, and Radio Network 130.

[0034] Example Implementation

[0035] Figure 8 An example communication system 800 according to an embodiment of this disclosure is shown, comprising at least one example device 810 and one example device 820. Devices 810 and 820 are capable of performing various functions to implement the schemes, techniques, processes, and methods described herein, involving channel state information (CSI) compression and decompression, including the various schemes proposed above for various designs, concepts, schemes, systems, and methods, including network environment 100, and the processes described below.

[0036] Devices 810 and 820 may be part of an electronic device, which may be a network device or user equipment (UE) (e.g., UE 110), such as a portable or mobile device, wearable device, in-vehicle device or vehicle, wireless communication device, or computing device. For example, devices 810 and 820 may be implemented in a smartphone, smartwatch, personal digital assistant, electronic control unit (ECU) in a vehicle, digital camera, or computing device such as a tablet, laptop, or notebook computer. Devices 810 and 820 may also be part of a machine-type device, such as an Internet of Things (IoT) device, such as a non-movable or fixed device, home appliance, roadside unit (RSU), wired communication device, or computing device. For example, devices 810 and 820 may be implemented in a smart thermostat, smart refrigerator, smart door lock, wireless speaker, or home control center. When implemented as a network device, device 810 and / or device 820 may be implemented in an eNodeB in an LTE, LTE-Advanced, or LTE-Advanced Pro network, or in a gNB or TRP in a 5G / New Radio (NR), 6G, or IoT network.

[0037] In some embodiments, devices 810 and 820 may take the form of one or more integrated circuit (IC) chips, such as, but not limited to, one or more single-core processors, one or more multi-core processors, one or more complex-instruction-set-computing (CISC) processors, or one or more reduced-instruction-set-computing (RISC) processors. In all the above embodiments, devices 810 and 820 may be implemented as network devices or user equipment. Devices 810 and 820 may at least include... Figure 8 The components shown include, for example, processors 812 and 822. Devices 810 and 820 may also include one or more other components unrelated to the present disclosure (e.g., internal power supply, display device, and / or user interface device); therefore, for the sake of brevity, Figure 8 Such components of devices 810 and 820 are not shown and are not described below.

[0038] In one aspect, processors 812 and 822 may take the form of one or more single-core processors, one or more multi-core processors, or one or more CISC or RISC processors. That is, although the singular term "processor" is used herein to refer to processors 812 and 822, in some embodiments processors 812 and 822 may include multiple processors, and in other embodiments may be a single processor, all in accordance with this disclosure. In another aspect, processors 812 and 822 may take the form of hardware (and optionally firmware) and include electronic components, such as, but not limited to, one or more transistors, one or more diodes, one or more capacitors, one or more resistors, one or more inductors, one or more memristors, and / or one or more transformers, which are configured and arranged according to the specific purposes of this disclosure. In other words, in at least some embodiments, processors 812 and 822 are special-purpose machines specifically designed, arranged, and configured to perform specific tasks, including error correction and verification for training robust artificial intelligence / machine learning models in wireless communications, in accordance with various embodiments of this disclosure.

[0039] In some embodiments, device 810 may further include a transceiver 816 connected to processor 812. Transceiver 816 enables the transmission and reception of wireless data. In some embodiments, transceiver 816 can wirelessly communicate with different types of wireless networks and different radio access technologies (RATs). In some embodiments, transceiver 816 may be equipped with multiple antenna ports (not shown), such as four antenna ports. That is, transceiver 816 may be equipped with multiple transmit antennas and multiple receive antennas for multiple-input multiple-output (MIMO) wireless communication. In some embodiments, device 820 may further include a transceiver 826 connected to processor 822. Transceiver 826 may be a transceiver capable of wirelessly transmitting and receiving data. In some embodiments, transceiver 826 can wirelessly communicate with different types of user equipment / wireless networks and different RATs. In some embodiments, transceiver 826 may be equipped with multiple antenna ports (not shown), such as four antenna ports. That is, transceiver 826 may be equipped with multiple transmit antennas and multiple receive antennas for MIMO wireless communication.

[0040] In some embodiments, device 810 may further include a memory 814 connected to processor 812 and accessible and able to store data by processor 812. In some embodiments, device 820 may further include a memory 824 connected to processor 822 and accessible and able to store data by processor 822. Memory 814 and memory 824 may include a random access memory (RAM), such as dynamic RAM (DRAM), static RAM (SRAM), thyristor RAM (T-RAM), and / or zero-capacitor RAM (Z-RAM). Alternatively, or furthermore, memory 814 and memory 824 may include a read-only memory (ROM), such as a mask ROM, programmable ROM (PROM), erasable programmable ROM (EPROM), and / or electrically erasable programmable ROM (EEPROM). Alternatively, or further, memory 814 and memory 824 may include a non-volatile random access memory (NVRAM), such as flash memory, solid-state memory, ferroelectric RAM (FeRAM), magnetoresistive RAM (MRAM), and / or phase-change memory.

[0041] Devices 810 and 820 may be communication entities capable of communicating with each other according to various proposed schemes of this disclosure. For ease of illustration and without limitation, the following describes the capabilities of device 810 as a user equipment (e.g., UE 110) and device 820 as a network node (e.g., network node 125) in a network (e.g., network 130 as a 5G / New Radio (NR) or 6G mobile network) within the context of example flow 900.

[0042] Example Process

[0043] Figure 9An example flow 900 according to an embodiment of this disclosure is illustrated. Flow 900 may represent one aspect, or in part or in whole, of the various proposed designs, concepts, schemes, systems, and methods described above regarding error correction and verification of training robust artificial intelligence / machine learning models in wireless communication, including the aforementioned related content. Flow 900 may include one or more operations, actions, or functions, as shown in one or more modules. Although shown as discrete modules, the individual modules of each flow may be divided into more modules, merged into fewer modules, or omitted depending on the desired implementation. Furthermore, the modules / submodules of each flow may be executed in the order shown in the figures, or in a different order. Additionally, one or more modules / submodules of each flow may be executed iteratively. Flow 900 may be implemented by or in device 810 and / or device 820 and any variations thereof. For illustrative purposes only and without limitation, each process is described below as an implementation in a network (e.g., a 5G / New Radio (NR) or 6G mobile network) where device 810 is a user equipment (e.g., user equipment 110) and device 820 is a network node or base station (e.g., terrestrial network node 120). Process 900 may begin at module 910.

[0044] At 910, process 900 may involve the processor 812 of device 810 (e.g., as user equipment 110) using intentionally erroneous training of an artificial intelligence / machine learning model. Alternatively, or additionally, process 900 may involve the processor 822 of device 820 (e.g., as a terrestrial network node 125 or non-terrestrial network node 128 of wireless network 130) using the intentionally erroneous training of an artificial intelligence / machine learning model. Process 900 may continue from 910 to 920.

[0045] At 920, process 900 may involve processor 812 utilizing the trained artificial intelligence / machine learning model in wireless communication (e.g., with device 820). Alternatively, or additionally, process 900 may involve processor 822 utilizing the trained artificial intelligence / machine learning model in wireless communication (e.g., with device 810).

[0046] In some implementations, during the training of an AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error corrector into the AI / machine learning model to detect and correct intentional errors and inputs or latent variables, thereby generating corrected inputs or latent variables. Furthermore, the corrected inputs or latent variables may be compared with the true inputs or latent variables to provide backpropagation feedback to the AI / machine learning model.

[0047] In some implementations, during the training of the artificial intelligence / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error detector into the artificial intelligence / machine learning model to detect the intentional error and the input or latent variable, thereby generating a detection result. Furthermore, the detection result may be compared with the intentional error to provide backpropagation feedback to the artificial intelligence / machine learning model.

[0048] In some implementations, during the training of an artificial intelligence / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error corrector during the inference phase between the encoder and decoder of the artificial intelligence / machine learning model to generate corrected latent variables as feedback to the artificial intelligence / machine learning model.

[0049] In some implementations, during the training of an artificial intelligence / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error detector during the inference phase between the encoder and decoder of the artificial intelligence / machine learning model to generate detection decisions as feedback to the artificial intelligence / machine learning model.

[0050] In some implementations, during the training of an AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error corrector into the one-sided AI / machine learning model to detect and correct the intentional error and input, thereby generating a corrected input and comparing it with the real input to provide backpropagation feedback for the one-sided AI / machine learning model.

[0051] In some implementations, during the training of an AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error corrector during the inference phase of the one-sided AI / machine learning model to generate corrected inputs for the one-sided AI / machine learning model.

[0052] In some implementations, when training an AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error detector into the one-sided AI / machine learning model to detect the intentional error and input, thereby generating a detection decision and comparing it with the real input to provide backpropagation feedback to the one-sided AI / machine learning model.

[0053] In some implementations, during the training of an AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error detector during the inference phase of the one-sided AI / machine learning model to generate erroneous decisions for the one-sided AI / machine learning model.

[0054] In some implementations, during the training of the AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error corrector into the bilateral AI / machine learning model to detect and correct the intentional error and input, thereby generating a corrected input and comparing it with the true input to provide backpropagation feedback to the bilateral AI / machine learning model. In some implementations, during the training of the AI / machine learning model, process 900 may also involve processor 812 or processor 822 inserting the error corrector during the inference phase of the bilateral AI / machine learning model to generate a corrected result for the bilateral AI / machine learning model.

[0055] In some implementations, during the training of the AI / machine learning model, process 900 may involve processor 812 or processor 822 inserting an error detector into the bilateral AI / machine learning model to detect the intentional error and input, thereby generating a detection decision and comparing it with the real input to provide backpropagation feedback to the bilateral AI / machine learning model. In some implementations, during the training of the AI / machine learning model, process 900 may also involve processor 812 or processor 822 inserting the error detector during the inference phase of the bilateral AI / machine learning model to generate a decision for the bilateral AI / machine learning model.

[0056] In some implementations, during the training of the AI / machine learning model, process 900 may involve processor 812 or processor 822 introducing the intentional error into the latent variables during the training phase between the encoder and decoder of the bilateral AI / machine learning model, to generate an output from the decoder based on the erroneous latent variables. In some implementations, during the training of the AI / machine learning model, process 900 may also involve processor 812 or processor 822 comparing the decoder output with the true value to provide backpropagation feedback to the bilateral AI / machine learning model.

[0057] In some implementations, during the training of an AI / machine learning model, process 900 may involve processor 812 or processor 822 introducing the intentional error along with the input as erroneous input to the one-sided AI / machine learning model to generate an output from the erroneous input. In some implementations, during the training of the AI / machine learning model, process 900 may also involve processor 812 or processor 822 comparing the one-sided AI / machine learning model output with the true value to provide backpropagation feedback to the one-sided AI / machine learning model.

[0058] In some implementations, when utilizing a trained artificial intelligence / machine learning model, process 900 may involve processor 812 or processor 822 utilizing the trained artificial intelligence / machine learning model when performing channel state information (CSI) compression, noise reduction, quantization, encoding, error correction codes, modulation, peak-to-average power ratio (PAPR) reduction, or image compression.

[0059] Additional notes

[0060] The topics described herein sometimes demonstrate that different components are contained within or connected to other different components. It should be understood that such architectures are merely examples, and many other architectures can actually be implemented to achieve the same functionality. Conceptually, any arrangement of components to achieve the same functionality is effectively “associated” to achieve the desired function. Therefore, any two components combined in this document to achieve a specific function can be considered “associated” to achieve the desired function, regardless of the architecture or intermediate components. Similarly, any two components so associated can also be considered “operably connected” or “operably coupled” to achieve the desired function, and any two components that can be so associated can also be considered “operably coupled” to achieve the desired function. Specific examples of operable coupling include, but are not limited to, physically matable and / or physically interactive components, wirelessly interactive and / or wirelessly interactive components, and logically interactive and / or logically interactive components.

[0061] Furthermore, regarding the use of almost all plural and / or singular terms in this document, those skilled in the art can appropriately convert plural to singular and / or singular to plural depending on the context and / or application. For clarity, various permutations and combinations of singular / plural are explicitly listed herein.

[0062] Furthermore, those skilled in the art will understand that the terms used herein, particularly in appended claims, such as the body portion of appended claims, are generally considered "open-ended" terms. For example, the word "comprising" should be interpreted as "comprising but not limited to," the word "having" should be interpreted as "having at least," and the word "including" should be interpreted as "including but not limited to," etc. Those skilled in the art will also further understand that if a particular quantity introduced in a claim is intentional, that intention will be explicitly stated in the claim; if no such statement is made, then that intention does not exist. For example, for ease of understanding, the following appended claims may contain the use of the introductory phrases "at least one" and "one or more" to introduce the content of the claim. However, the use of such phrases should not be interpreted as limiting any particular claim containing that content to containing only one instance of that content, even if the same claim contains the introductory phrases "one or more" or "at least one" and indefinite articles such as "one" or "a," for example, "one" and / or "a" should be interpreted as "at least one" or "one or more"; the same applies to the use of definite articles used to introduce the content of the claim. Furthermore, even if a specific quantity is explicitly stated in the claims, those skilled in the art will recognize that such statement should be interpreted as at least the stated quantity. For example, the statement "two items" alone, without any other modification, means at least two items, or two or more items.

[0063] Furthermore, when using conventions such as "at least one A, B, and C," such structures are generally intended to make the meaning of the convention clear to those skilled in the art. For example, "a system having at least one A, B, and C" includes, but is not limited to, systems with only A, only B, only C, A and B, A and C, B and C, and systems where A, B, and C coexist. Similarly, when using conventions such as "at least one A, B, or C," such structures are generally intended to make the meaning of the convention clear to those skilled in the art. For example, "a system having at least one A, B, or C" includes, but is not limited to, systems with only A, only B, only C, A and B, A and C, B and C, and systems where A, B, and C coexist. Those skilled in the art will further understand that almost all extractive terms and / or phrases presenting two or more alternative terms appearing in descriptions, claims, or drawings should be understood to include one term, any term, or both terms. For example, the phrase "A or B" should be understood to include the possibility of "A" or "B" or "A and B."

[0064] As can be seen from the foregoing, this document describes various embodiments of the present disclosure for illustrative purposes, and various modifications can be made without departing from the scope and spirit of the present disclosure. Therefore, the various embodiments disclosed herein are not intended to be limiting, and the true scope and spirit are defined by the following claims.

Claims

1. A method comprising: Train an artificial intelligence / machine learning model using an intentional error; as well as The trained artificial intelligence / machine learning model is utilized in a user device or a network node of a wireless network.

2. The method of claim 1, wherein training the artificial intelligence / machine learning model includes inserting an error corrector into the artificial intelligence / machine learning model to detect and correct the intentional error and an input or latent variable, thereby generating a corrected input or latent variable.

3. The method of claim 2, wherein the modified input or latent variable is compared with a real input or latent variable to provide a backpropagation as feedback for the artificial intelligence / machine learning model.

4. The method of claim 1, wherein training the artificial intelligence / machine learning model includes inserting an error detector into the artificial intelligence / machine learning model to detect the intentional error and an input or latent variable, thereby generating a detection result.

5. The method of claim 4, wherein the detection result is compared with the intentional error to provide a backpropagation as feedback for the artificial intelligence / machine learning model.

6. The method of claim 1, wherein training the artificial intelligence / machine learning model includes inserting an error corrector into an inference stage between an encoder and a decoder of the artificial intelligence / machine learning model to generate a corrected latent variable as feedback to the artificial intelligence / machine learning model.

7. The method of claim 1, wherein training the artificial intelligence / machine learning model includes inserting an error detector into an inference stage between an encoder and a decoder of the artificial intelligence / machine learning model to generate a detection decision as feedback to the artificial intelligence / machine learning model.

8. The method of claim 1, wherein training the AI / machine learning model includes inserting an error corrector into a one-sided AI / machine learning model to detect and correct the intentional error and an input, thereby generating a corrected input and comparing it with a real input to provide a backpropagation as feedback for the one-sided AI / machine learning model.

9. The method of claim 1, wherein training the artificial intelligence / machine learning model includes inserting an error corrector during an inference phase of the one-sided artificial intelligence / machine learning model to generate a corrected input to the one-sided artificial intelligence / machine learning model.

10. The method of claim 1, wherein training the AI / machine learning model includes inserting an error detector into the one-sided AI / machine learning model to detect the intentional error and an input, thereby generating a detection decision and comparing it with a real input to provide a backpropagation as feedback for the one-sided AI / machine learning model.

11. The method of claim 1, wherein training the AI / machine learning model includes inserting an error detector into an inference phase of the one-sided AI / machine learning model to generate an incorrect decision for the one-sided AI / machine learning model.

12. The method of claim 1, wherein training the AI / machine learning model includes inserting an error corrector into the bilateral AI / machine learning model to detect and correct the intentional error and an input, thereby generating a corrected input and comparing it with a real input to provide a backpropagation as feedback for the bilateral AI / machine learning model.

13. The method of claim 12, wherein training the AI / machine learning model further includes inserting the error corrector during an inference phase of the bilateral AI / machine learning model to generate a correction result for the bilateral AI / machine learning model.

14. The method of claim 1, wherein training the AI / machine learning model includes inserting an error detector into the bilateral AI / machine learning model to detect the intentional error and input, thereby generating a detection decision and comparing it with a real input to provide a backpropagation as feedback for the bilateral AI / machine learning model.

15. The method of claim 14, wherein training the AI / machine learning model further comprises inserting the error detector into an inference phase of the bilateral AI / machine learning model to generate a decision for the bilateral AI / machine learning model.

16. The method of claim 1, wherein training the artificial intelligence / machine learning model comprises introducing the intentional error into a latent variable through a training phase between an encoder and a decoder of the bilateral artificial intelligence / machine learning model to generate an output from the decoder based on an error latent variable.

17. The method of claim 16, wherein training further comprises comparing the output of the decoder with a true value to provide a backpropagation as feedback for the bilateral artificial intelligence / machine learning model.

18. The method of claim 1, wherein training the artificial intelligence / machine learning model comprises introducing the intentional error along with an input as an error input into the one-sided artificial intelligence / machine learning model to generate an output from the error input.

19. The method of claim 18, wherein training further comprises comparing the output of the one-sided AI / machine learning model with a true value to provide a backpropagation as feedback to the one-sided AI / machine learning model.

20. The method of claim 1, wherein utilizing the trained artificial intelligence / machine learning model includes utilizing the trained artificial intelligence / machine learning model when performing channel state information compression, noise reduction, quantization, encoding, error correction codes, modulation, peak-to-average power ratio reduction, or image compression.

Images