Adaptive real-time adjustment in deep neural networks

Abstract

In some embodiments, the present disclosure can provide a method for adaptive real-time adjustment in a deep neural network, the method comprising: receiving an indication to perform a single-class inference task using a machine learning model trained to perform a multi-class inference task; in response to the indication, applying at least one mask mapped to the single-class inference task during a time frame of execution of the machine learning model hosted by a user device; and reconfiguring the machine learning model to a state in which a first set of nodes and a second set of nodes perform the multi-class inference task. Related systems, methods, and articles of manufacture are also disclosed.

Description

Technical Field

[0001] The topics described in this article involve machine learning. Background Technology

[0002] Machine learning (ML) models can be trained and learned. ML models can take many forms, such as artificial neural networks (or simply neural networks), decision trees, and / or others. Some neural networks can be considered "deep neural networks," which refer to neural networks that include at least two hidden layers. Training of ML models can be supervised (using labeled training data), semi-supervised, or unsupervised. Once trained, ML models can be used to perform inference tasks. Summary of the Invention

[0003] In some embodiments, a method may be provided, comprising: receiving, by a user equipment, an instruction to perform a single-class inference task using a machine learning model trained to perform a multi-class inference task; in response to the instruction, by the user equipment and during a time frame of execution of the machine learning model hosted by the user equipment, applying at least one mask mapped to the single-class inference task, wherein the at least one mask defines, for the single-class inference task,: a first set of nodes at the machine learning model activated to perform the single-class inference task, and a second set of nodes at the machine learning model deactivated while performing the single-class inference task; and by the user equipment reconfiguring the machine learning model to a state in which the first set of nodes and the second set of nodes perform the multi-class inference task.

[0004] In some variations, one or more features disclosed herein (including the following features) can be optionally included in any feasible combination. The machine learning model can be at least partially trained to perform multi-class inference tasks before being deployed to a user device. An instruction to perform a single-class inference task using the machine learning model can be received from at least one other user device requesting the single-class inference task. The instruction to perform a single-class inference task using the machine learning model can be received in response to the single-class inference task being detected by the user device. This time frame can be a predetermined time period, or a time when the single-class inference task is no longer an inference task inferred by the user device. At least one mask can be learned by providing the machine learning model with samples for classification as a single-class inference task and profiling multiple nodes of the machine learning model to determine a second set of nodes not associated with the classification of the machine learning model for the single-class inference task. The weight of the second set of nodes can be below a threshold, indicating that the second set of nodes is not associated with the classification of the single-class inference task. In response to reconfiguration, the machine learning model can perform multi-class inference tasks. In response to the application of at least one mask, a machine learning model can be executed to perform a single-class inference task, wherein a first set of nodes is executed to perform the single-class inference task, and a second set of nodes is deactivated and not executed while the single-class inference task is being performed.

[0005] Depending on the desired configuration, the above aspects and features can be implemented in systems, apparatus, methods, and / or articles. Details of one or more variations of the subject matter described herein are set forth in the accompanying drawings and the following description. The features and advantages of the subject matter described herein will be apparent from the specification, the accompanying drawings, and the claims. Attached Figure Description

[0006] In the attached diagram,

[0007] Figure 1 Example scenarios in which machine learning models, according to some embodiments, can be optimized are described.

[0008] Figure 2 Examples of apparatuses including at least one machine learning model according to some embodiments are described;

[0009] Figure 3A Examples of machine learning models according to some embodiments are described;

[0010] Figure 3B Depicting according to some embodiments Figure 3A An example of a machine learning model where some nodes are temporarily deactivated based on a mask;

[0011] Figure 4Examples of processes for optimizing machine learning models according to some embodiments are described;

[0012] Figure 5 Another example of a process for optimizing a machine learning model, according to some embodiments, is described; and

[0013] Figure 6 Examples of computing systems according to some embodiments are described.

[0014] The same reference numerals in the figure are used to refer to the same or similar items. Detailed Implementation

[0015] Most machine learning (ML) models, such as neural networks, deep neural networks, and / or similar systems, have a fixed architecture relative to the ML model (e.g., each layer and node is fixed). This fixed ML model structure is executed across all layers and nodes to perform inference tasks, such as classifying inputs. When used for classification tasks, the fixed structure of the ML model cannot ignore uninteresting classifications during execution. In other words, even if only a single classification will be performed by the ML model, given the need to infer only that single classification, the ML model cannot adjust its structure, and therefore the entire ML model is still executed.

[0016] Figure 1 Two examples of scenarios according to some embodiments are described, in which the ML model operates under suboptimal conditions and therefore uses excessive processing resources, wastes energy, and causes inference latency.

[0017] refer to Figure 1 At system 100, multiple devices 102A-D (which in this example represent cameras, although other types of sensors or data sources may also be implemented) collaborate in a distributed environment. Each of these 102A-D is equipped with at least one machine learning (ML) model 103A-D, such as a neural network, a deep neural network, and / or the like, which is trained to perform tasks such as object classification, automatic object detection, and / or other types of reasoning tasks.

[0018] In configuration 100, devices 102A-D may experience spatial locality in the detected classifications, where, within a certain time period, the same object (or multiple objects), such as car 104, is likely to be detected by all devices 102A-D at some point in time when car 104 passes through the area (in this example, these devices are cameras, each with a corresponding ML model for automatic object detection and / or classification). In this example, devices 102A-D can be configured collaboratively, where devices 102A-D can signal to each other to, for example, assist in inference tasks. To illustrate this, consider a classification task where devices 102A-D (each device includes a camera and a corresponding ML model 103A-D) are configured to detect objects of a certain type or category, such as the following four categories: trucks, cars, bicycles, and pedestrians (or people). In this example, the first device 102D (including ML model 103D) identifies car 104 and wants to continue tracking it. Therefore, the first device 102D sends signals to one or more of the other devices 102A-C to cooperate, thereby assisting in the detection or identification of the classification of the car 104. Since only the classification "car" is being considered, there is no need to consume resources at the other devices 102A-C to detect other classifications, which in this example are trucks, bicycles, and pedestrians.

[0019] Another use case that might arise from scenario 100 is when the first device, including device 102D, classifies car 104 as "car," but the ML model 103D of the first device is uncertain about the classification of the car (e.g., the confidence in the car classification is below a threshold confidence metric). The low confidence classification of car 104 could be caused by a variety of factors, such as noise in the image, camera orientation or rotation, lighting, and / or other reasons. In this example, other devices 102A-C have different perspectives on car 104; therefore, a distributed collaborative system of other devices can be able to assist in further classification of car 104, thereby improving the overall classification confidence of car 104 (e.g., making the confidence for the car classification above a threshold confidence metric). As in the previous example, the ML models 103A-C of other devices 102A-C only need to perform the task of classifying cars, without wasting computational resources on classifying objects such as trucks, bicycles, or pedestrians.

[0020] In system 199 ( Figure 1 In another scenario, a single device 106 (such as a smartphone with one or more sensors) may further include at least one ML model 107. Figure 1In the example, device 106 is equipped with ML model 107, which uses sensor data (obtained from device 106 or other sources) to predict, for example, human activities. In this example, temporal locality exists, meaning that when a user begins an activity, the user is likely to continue that activity for a period of time. For example, if a user starts cycling 108 and the cycling is for a certain amount of time (e.g., 2 minutes or other time intervals), the probability that the user will continue the cycling activity in the next (multiple) time frames (e.g., the next 2 minutes, 4 minutes, or other time intervals) is high. To avoid wasting computation time and resources, ML model 107 can be optimized to detect only the "cycling" category, so that when ML model 107 stops detecting "cycling" activities, it can recover multi-class classifications (e.g., exercise, cycling, running, walking, etc.).

[0021] In some embodiments, a method using spatial locality and / or temporal locality is provided to allow ML models to reduce their resource utilization by at least avoiding, for example, classification computations for unlikely classifications. This can reduce inference latency (e.g., the time spent by the ML model performing inference tasks), reduce energy consumption, and reduce device resource utilization (e.g., in terms of processing and / or storage resources at the device).

[0022] As mentioned earlier, ML models (such as deep neural networks) can be viewed as a monolithic structure that requires end-to-end execution (e.g., all its layers and nodes) to perform inference tasks. Therefore, given an input, the ML model executes completely across all nodes to generate an output for the inference task. Consequently, resources can be unnecessarily consumed by the ML model for each and every inference computation.

[0023] In some embodiments, a method is provided, for example, to reduce inference latency and decrease the energy and resources consumed by the ML model when not all classifications are required (e.g., when the inference task is a classification task). Figure 1 The example illustration is shown below. In some embodiments, the machine learning model is adaptively configured such that one or more nodes of the ML model are deactivated based on a mask for execution, such that activated nodes are needed to perform the classification task, while deactivated nodes are not needed to perform the classification task. This adaptation is temporary to account for the aforementioned spatial and / or temporal locality, allowing the ML model to revert to its original unadapted state (e.g., deactivated nodes are activated) to enable the ML to perform more comprehensive multi-class classification. Additionally, this adaptation can be real-time, in a sense, triggered as needed, such as by... Figure 1 As shown in the example. In the adapted state (e.g., with one or more nodes deactivated), the ML model temporarily only computes certain categories (e.g., such as...). Figure 1 The example shows a single category, such as cars), without performing other classifications (e.g., Figure 1 (The categories of trucks, bicycles, and pedestrians).

[0024] In some embodiments, the ML model is configured such that computations related to the inference task are optimized when not all categories are needed. This optimization is provided by at least one or more masks for each category. The mask for a given category can be activated (or, for example, deactivated) during the runtime execution of the ML model to ensure that only certain specific categories are computed by the ML model. This process ensures optimization of the ML model regarding both inference latency, computational resources, and / or energy consumption. In other words, this is a novel approach to saving runtime resources by running only a portion of the ML model responsible for inference of a specific category (or multiple categories).

[0025] To determine the mask used for classification, analytical techniques can be performed on ML models 103A-D (and / or ML model 107) at position 100, using only a small sample of input data. For example, the mask is determined by learning which activations (e.g., nodes of a neural network) are used or necessary for performing the desired specific classification. Referring to the "car" classification example above, the mask could indicate one or more ML model nodes that can be temporarily deactivated while the ML model is performing the car classification task (because deactivated nodes are not needed for that task).

[0026] Before providing additional descriptions of the optimization and adaptability of the ML model disclosed herein, an example device 200 (also referred to herein as a user device) including ML model 299 is referenced. Figure 2 Described. Reference Figure 2 Device 200 may be included in devices 102A-D, device 106, and / or other processor-based devices capable of hosting ML models.

[0027] The device 200 may include one or more of the following: at least one processor 210 (such as a central processing unit and / or etc.), at least one memory 220, at least one storage device 230, at least one input and / or output device 242, at least one sensor 245, at least one graphics processing unit 247, and at least one camera 249, all of which may be coupled via bus 250.

[0028] like Figure 2As shown, processor 210 is capable of processing instructions for execution within device 200. In some implementations of the present subject, processor 210 may be a single-threaded processor. Alternatively, processor 210 may be a multi-threaded processor. The processor may be a multi-core processor with multiple processors, or a single-core processor. Processor 210 is capable of processing instructions stored in memory 220 and / or storage device 230, and / or generating displays of information for a user interface provided via input / output device 242. Memory 220 is a computer-readable medium, such as volatile and / or non-volatile media, storing information within device 200. Memory 220 is capable of storing data structures representing nodes of an ML model (e.g., parameters for at least one ML model, such as weights and / or other configuration information). Storage device 230 may be capable of providing persistent storage for device 200. Storage device 230 may be any type of storage device (e.g., hard disk drive, optical disk drive, and / or other suitable persistent storage or memory mechanism). Input / output device 242 provides input / output operations for device 200. In some implementations of the present subject, input / output device 242 includes a keyboard and / or pointing device. In various implementations, input / output device 242 includes a display unit for displaying a graphical user interface. According to some implementations of the present subject, input / output device 242 can provide input / output operations for network devices (e.g., coupled to a network, bus, and / or the like). For example, input / output device 242 may include an Ethernet or WiFi port or other network port for communication with one or more wired and / or wireless networks (e.g., a local area network (LAN), a wide area network (WAN), or the Internet).

[0029] Device 200 may include at least one graphics processing unit 247. This GPU may be used to execute ML models 299. Alternatively or additionally, the at least one graphics processing unit 247 may include an AI chip and / or the like.

[0030] exist Figure 2 In the example, device 200 may include one or more sensors. For example, the sensor may include an accelerometer, gyroscope, ECG sensor, heart rate sensor, blood glucose sensor, EMG sensor, and / or other types of sensors. Figure 2 In this example, the device also includes a camera 249, such as a video camera, a still camera, an infrared camera, and / or the like. The sensor can provide input 297A to the ML model 299, although this input can also be obtained from other sources.

[0031] Figure 2 An example of an ML model 299 hosted at device 200 is also depicted. For example, ML model 299 may be included... Figure 1In each of the ML models 103A-D and / or 107. The structure and type of ML model 299 are for illustrative purposes only; other types and structures may also be used.

[0032] The ML model 299 can be trained to perform inference tasks, such as performing multiple classifications on the data provided at input 297A. As previously mentioned, the data provided at the input can include various types of data based on the inference task performed by the ML model. For example, the input data can include images, sensor data, and / or others. The classification task refers to classifying objects (presented at input 297A) into members of at least one category, such as cars, pedestrians, and / or other categories. The classification can be indicated by the output 297B of the ML model. (See reference...) Figure 1 As described in the example, ML model 299 can output a classification indication of input 297A (297B). For example, given an image of a car, ML model 299 can output an indication that the image is classified as "car". Figure 2 Examples of category 295A are described, such as trucks, cars, bicycles, people, exercise, cycling, running, and walking, as well as examples of category 295B, such as exercise, cycling, running, or walking. These categories are merely examples; other categories can be classified by the ML model. Furthermore, although the number of items in a particular category may vary... Figure 2 The categories are described for classification, but ML models can also be trained to infer more or fewer categories.

[0033] While some examples in the document refer to the ML models used for training and inference as deep neural networks, other types of ML models can also be used. For example, ML models can include a variety of ML model types, including one or more of the following: neural networks, convolutional neural networks, decision trees (multiples), graph neural networks (multiples), and / or other types of machine learning models (which include nodes or paths that can be activated and deactivated, as disclosed herein), as well as other inference types such as clustering and regression.

[0034] In the case of ML models, there may be a training phase (during which the ML model is trained to perform a task, such as an inference task) and an inference phase (during which the ML model performs an inference task). An example of an inference task is classifying objects into one or more categories. For example, a classification task might categorize input data representing objects into trucks, cars, monkeys, and / or any other type of object.

[0035] For example, in the case of supervised learning, the ML model provides ML model 299 with examples of what (and / or what not) objects belong to a category at input 297A. In the example of car classification, the ML model can receive input from the training set, such as examples of images including cars (and other types of images), until the output 297B of the ML model learns to recognize the "car" category of the object. Once trained, the ML model is then able to advance to perform inference tasks on, for example, other images outside the training set, to recognize, for example, the "car" category presented to ML model 299 as input 297A.

[0036] In the case of training ML model 299, the training process includes at least: (1) deploying the ML model (such as ML model 299) to a device (such as device 102A-D and / or 106); and (2) analyzing the ML model at the corresponding device hosting the ML model.

[0037] In some example embodiments, as previously described, the ML model is provided to one or more devices that host the ML model. For example, an ML model (which can be trained to perform a multi-class classification reasoning task) is provided to the device. (See reference) Figures 1 to 2 For example, devices 102A-D may each be provided with an ML model (such as ML models 103A-D) that is trained to classify multiple categories, as shown in category 295A. Similarly, device 106 may be provided with an ML model 107 that is trained to classify multiple categories, as shown in category 295B.

[0038] In some embodiments, ML models 103A-D and / or 107 are pre-trained and then deployed to devices 102A-D and / or 106, although some training and / or retraining may occur after deployment to the device.

[0039] In some embodiments, some (if not all) ML models 103A-D operate in a distributed or collaborative manner, as referenced Figure 1 The system 100 described herein. Additionally, some (if not all) of the ML models 103A-D may be identical with respect to structure (e.g., number of nodes, number of layers, interconnections, etc.) and / or configuration (e.g., weights of each node in the nodes, etc.). Alternatively or additionally, some (if not all) of the ML models 103A-D may be different with respect to structure and / or configuration. Furthermore, some (if not all) of the ML models may be similarly trained to classify some (if not all) of objects of the same type. References Figure 2For example, ML models 103A-D can each be trained to classify input images into categories such as trucks, cars, bicycles, or people (see, for example, category 295A). ML model 107 can be trained to classify input images into categories such as cycling, running, walking, etc. (see, for example, category 295B).

[0040] In some embodiments, analyzer 240 may perform analysis of the ML model to determine which nodes of the ML model are needed (or not needed) for classification. For example, analyzer 240 may use a relatively small set of examples for each category in the ML model.

[0041] refer to Figure 2 For example, analyzer 240 can provide a sample input set (as input 297A) for each of multiple categories, where ML model 299 is trained to classify those multiple categories. For instance, given categories 295A 1-4, the analyzer can provide samples at input 297A for category 1 (e.g., 100 truck images or other quantities), category 2 (e.g., 100 car images or other quantities), category 3 (e.g., 100 bicycle images or other quantities), and category 4 (e.g., 100 people images or other quantities). These samples allow the analyzer to understand which nodes of the ML model are (or are not) required to perform each classification. In other words, the weights of a given node in the neural network can indicate whether certain nodes share some classification for a given category.

[0042] For each category that can be classified by the ML model, analyzer 240 determines (e.g., learns) a mask (also referred to as a "node mask"). This mask represents the path (e.g., which(nodes)(s)) activated in the ML model 299 for a specific category. Referring to the previous example of four categories, analyzer 240 determines a mask for the first category, a mask for the second category, a mask for the third category, and a mask for the fourth category.

[0043] To further illustrate, the analyzer acts as a selective gating mechanism within ML models, such as deep neural networks. Therefore, the analyzer systematically evaluates the contribution of each node to the performance of the ML model's inference task, such as classification. For example, given an input sample, the analyzer analyzes the impact of each node's activation on the output logistic value by adjusting the activation of each node, thus allowing the analyzer to measure the impact of nodes on classification decisions. This evaluation can be performed by calculating the impact of nodes on the model's loss function with respect to the target class, thereby identifying nodes that are crucial for minimizing classification error.

[0044] Through iterative adjustments, the analyzer updates the mask, which dynamically adjusts the activation of each node. Nodes that consistently have a significant impact on minimizing the loss are strengthened, while those showing minimal impact are selectively suppressed. Instead of applying a hard threshold, this process leverages gradient-based optimization to fine-tune the contribution of each node, thereby continuously optimizing the mask value. The result is a sparse representation where only the most informative nodes remain active, effectively pruning the network while preserving its prediction accuracy.

[0045] By constructing task-specific masks, the analyzer enables more compact and efficient models, where computational resources are concentrated on the nodes that contribute most to the target classification. This selective activation aligns with principles of model compression and interpretability, promoting efficient and robust networks tailored for accurate classification.

[0046] The determined node mask can then be stored, for example, at the device, such as in storage device 230 or memory 220 (e.g., in a structured manner, such as a database, or in an unstructured manner, such as object storage), although the node mask can also be stored in other locations.

[0047] Table 1 below depicts an example of a node mask for two categories (Category 1 and Category 2). Although, based on the ML model structure, the node mask may include additional categories and layers. Figure 3A The corresponding ML model 300 is described, and the node masks in Table 1 can be applied to this ML model 300. For example, Figure 3A An ML model 300 configured as a deep neural network is depicted. This ML model comprises four hidden layers 304A-B, and each layer includes four nodes (labeled 1-4), although other ML model architectures can be implemented. Input is provided at input layer 302A, and output is provided by output layer 302B.

[0048] Refer to Table 1 and Figures 3A-3B And assuming ML model 300 classifies two categories, for example, for category 1, layer 1 node mask values ​​of 0.2, 0.3, 0.0, 0.4 can indicate that node 3 at layer 1 can be deactivated during execution. In this example, mask values ​​below a threshold (such as 0.1) indicate that the corresponding node can be deactivated, although other thresholds or deactivation methods can also be used. For layer 2, mask values ​​of 0.5, 0.4, 0.2, 0.3 can indicate that no node should be deactivated during execution. For layer 3, mask values ​​of 0.0, 0.2, 0.0, 0.3 indicate that node 1 should be deactivated for execution, and so on for all categories and layers. Although the previous example used 2 categories and 4 nodes, this is just an example, as other categories, layers, and nodes can also be used.

[0049] For example, in the example in Table 1, the mask values ​​are obtained by iteratively adjusting the mask value of each node based on its contribution to reducing the classification loss for a given class. To further illustrate, these values ​​are updated using the gradient of the loss relative to the mask of each node, guided by an optimizer (e.g., stochastic gradient descent (SGD) or other type of optimizer). Therefore, this adjustment process can identify the importance of each node for accurate classification, helping to determine which nodes should remain active or be deactivated. While mask values ​​may resemble “weights,” they do not function like typical model weights. Mask values ​​are more like binary or scaled indicators that selectively enable or disable specific nodes, thereby creating a custom structure for the model being optimized for the target class.

[0050] When ML Model 300 is running, it can deactivate certain nodes based on node mask values. Referring to the previous example and considering a single-class 1 inference task for ML Model 300, the ML model will deactivate (as indicated by "x"). Figure 3B The nodes are shown in Table 1. Therefore, during the execution of ML model 300, in response to the mask in Table 1 (where the ML model only classifies "Category 1"), node 3 at layer 1, as well as nodes 1 and 3 at layer 3, will be deactivated, while other nodes will be active, and thus inference for category 1 based on the given input layer 302A will be performed during runtime. As previously stated, when the ML model returns to multi-class, the ML model will reactivate node 3 at layer 1, as well as nodes 1 and 3 at layer 3. Similarly, if the ML model only needs to infer category 2, the mask in Table 1 will cause node 2 at layer 2, as well as nodes 2 and 3 at layer 4, to be temporarily deactivated during runtime.

[0051] Table 1

[0052] In some implementations, some devices (such as devices 102A-D) may each include a corresponding ML model 103A-D, where each device has the same ML model architecture (such as ML model 299), but ML models 103A-D may have different configurations, such as weights at nodes. Alternatively or additionally, some devices (such as devices 102A-D) may each include a corresponding ML model 103A-D with a different ML model architecture. In either case, analyzer 240 can analyze each ML model at each device separately, so each ML model is provided with a different mask. However, if some of these devices (such as devices 102A-D) include the same ML model 103A-D architecture and the same ML model architecture and configuration, analyzer 240 can perform a single analysis, and the same mask can be used on all ML models 103A-D.

[0053] When ML models are trained and analyzed, and the masks ("node masks") are thus determined, these ML models can be used during inference to adjust which nodes of the ML model are activated or deactivated (e.g., ...). Figure 3B (The location is described) to optimize the operation of the ML model to perform a given classification.

[0054] During inference operations, devices 102A-D can initially perform their inference tasks on all categories for which the ML model has been trained. For example, ML model 103A-D can classify all categories, such as trucks, cars, bicycles, and pedestrians (see, for example, category 295A).

[0055] For example, suppose a device (such as device 102D) wants other collaborating devices (such as devices 102A-C) to check for cars within a certain time frame (this time frame could be, for example, a predetermined time, a configurable time, and / or otherwise determined or set). For example, this might be because device 102D seeks confirmation that car 104 has been classified (by device 102D using low confidence, or for other reasons). In this case, device 102D can send an instruction to signal other devices 102A-C to check for a specific category (i.e., only for car category detection) within that time frame (e.g., the next 60 seconds or other time frames). For example, as previously described, when device 102D wants to check whether a sample (acquired by device 102D) is correctly classified when its predictive entropy is high and the sample classification has low confidence, device 102D can send an instruction to other devices. For example, the ML model 103D at device 102D can classify an object as having the category of car or van with low confidence (e.g., below a threshold number). Regardless, when at least one of the other devices 102A-C confirms within a time frame that the object (such as car 104) is of the car category in response to device 102D, devices 102A-D can return to detecting all of their configured categories. Alternatively or additionally, the time frame may elapse without any response from the other devices 102A-C to device 102D, in which case devices 102A-D can return to detecting all of their configured categories. Furthermore, the signaling and coordination process between devices can be repeated by, for example, any device in the device group 102A-D.

[0056] In use cases with only a single device (such as device 106), the above process can also be used. For example, when device 106 detects a specific category of activity (such as running (see category 295B)), device 106 may not be interested in, for example, classifying the input as a cycling category within a certain time frame. Therefore, only the node associated with the running category can be activated, while other nodes of the ML model can be deactivated using this mask. And, for example, when the time frame ends, device 106, including ML model 107, can return to a state where all categories 295B have been detected.

[0057] During inference, when the device enters an optimized state where a specific category is classified, the device can retrieve a mask for that category. For example, the device can retrieve a specific mask for category 1, as shown in the example at Table 1.

[0058] Figure 4 Examples of processes for adjusting an ML model by deactivating nodes based on a mask, according to some embodiments, are described. Figure 4Includes a first device 102D and a second device 102A, each of which may include, for example, a storage device 230, an analyzer 240, and an ML model 299 (labeled as an inference component). Although Figure 4 A certain number and configuration of devices are described, but this is only an example to illustrate the process, as other numbers and configurations of devices can also be used.

[0059] exist Figure 4 In the example, cloud server 404 can (at point 0) provide (multiple) ML models to first device 102D and second device 102A. Furthermore, cloud server 404 can (at point 1) provide analysis data (e.g., samples for each category) for analysis of the ML models by analyzer 240 at each device. The analysis data can be stored, for example, in storage device 230. Although... Figure 4 The example depicts cloud server 404 providing ML models to devices 102A and 102D, but ML models can also be provided in other ways.

[0060] At two locations, masks are learned for the first device 102D and the second device 102A. For example, analyzer 240 may request data (e.g., samples for each category) from storage device 230 for analyzing ML model 299, and data storage may respond by sending the requested data to be used for analyzing the ML model. For example, the data samples used for analyzing the ML model may include relatively small subsets of samples in each category as described above (e.g., 100 samples for category 1, 100 samples for category 2, etc.). Analyzer 240 uses these data samples to learn a mask (e.g., see Table 1) such that the mask indicates which nodes need to be activated (and / or which nodes can be deactivated) to perform a given classification task. When the mask is learned, the analyzer may save and store the mask in storage device 230. Figure 4 In the example, each device 102A and 102D can perform its own analysis, but if these devices have the same structure and configuration, a single mask can be determined by one of these devices and shared with the other device (e.g., see 2a(ii)).

[0061] As described above, the ML model 299 at devices 102A and 102D is operable such that it performs inference tasks across all categories it has been trained on, with all nodes at each layer being activated. However, when there is an opportunity (e.g., due to spatial or temporal locality) to optimize the ML model 299, it may be made to classify only a single category, rather than all categories it has been trained on. For example, device 102D may request device 102A to focus on a specific category at point 3, as referenced above. Figure 1 The aforementioned "automobile" category. In Figure 4 In the example, ML model 299 at device 102D sends a signal to instruct ML model 299 at device 102A to focus on a specific category instead of inferring all categories (e.g., focus on inferring only the "car" category instead of inferring all four categories 295A). In response, a mask can be requested by ML model 299 from storage device 230, which then sends the mask to ML model 299.

[0062] At point 3, each ML model 299 will activate one or more nodes based on a mask. For example, see reference... Figure 3B During the execution of the ML model, in response to the mask in Table 1 (where the ML model only classifies "Category 1"), node 3 at layer 1, as well as nodes 1 and 3 at layer 3, will be deactivated, while the other nodes will be active, and thus perform inference for category 1 given the input during runtime. As previously mentioned, the deactivation of these nodes may occur within a given time frame, although it can also be triggered in other ways to return to a state where all nodes are active. For example, since the deactivation of nodes is temporary, the ML model could return to a state where all nodes are active at point 4 to enable classification of all trained categories (such as categories 1 through 4 at category 295A).

[0063] As mentioned earlier, each processor-based device (such as devices 102A-D, 106, and 200) can include a pre-trained ML model configured to perform a specific classification task, such as a multi-class classifier. Therefore, consider a classification problem where x represents the input and y is a set... element This represents the target category corresponding to one of C categories. Formally, an ML model (such as a neural network) can be represented as follows: , in These are the weights of nodes in a trained ML model deployed on processor-based devices such as user devices, smartphones, cameras, wearables, and / or similar devices.

[0064] To isolate specific categories during the inference phase of the ML model, as previously described, the ML model is analyzed by analyzer 240 to determine a mask for each category (see, for example, Table 1). As mentioned above, since this is a very lightweight learning phase performed using a small subset of data, the process can be performed on the processor-based device itself, although the mask can also be determined externally and provided to the device.

[0065] As mentioned earlier, for analysis and mask learning, the device is provided with a set of examples for each category. These examples can be sent (e.g., from another device or cloud server) to a device with an initial ML model. This allows the device to learn the binary mask for each category ( (For some, if not all, layers of the ML model), where the value learned only for the mask is... Meanwhile, the original pre-trained weights W of the ML model remain unchanged. Therefore, the output is calculated as follows: , in This represents element-wise product. (For learning masks) loss function The difference between the original prediction and the masked prediction is minimized as follows: (1)

[0066] Used for masking The gradient can be calculated as follows: = + (2) in This is the weighting factor.

[0067] Examples of algorithms for learning masks are depicted in Table 2. The mask creation process can be a fine-tuning mechanism that selectively enables or disables nodes in the ML model based on their contribution to accurate predictions. By iteratively adjusting the mask to balance prediction accuracy with node utilization, a mask is determined that maintains classification performance even with potentially fewer active nodes. This allows the ML model to be more efficient while still ensuring good performance on the target class.

[0068] To further illustrate, the mask creation process can begin by initializing the mask, with all nodes active, allowing the full ML model to make predictions. For each sample of the target class, a prediction is made using the entire network, identifying the predicted class. Next, the loss is computed, along with its gradient relative to each mask value, revealing each node's contribution to reducing the loss. Using an optimizer (e.g., stochastic gradient descent (SGD), the mask values ​​are adjusted, preserving nodes that reduce the loss and deactivating less important nodes. After each adjustment, prediction consistency is checked to ensure the class remains correct; if it changes, the mask adjustment is modified or reversed. Once the loss is minimized and the predictions remain stable, the optimized mask is finalized, selectively activating nodes crucial for accurate classification and enhancing model efficiency without compromising performance.

[0069] Table 2

[0070] A learned mask for each category in (multiple) ML models can be used to tune the corresponding ML model, thereby optimizing it. In some embodiments, a threshold... The mask values ​​in Table 1 can be applied to determine the final path for ML model nodes of a specific category. Threshold It can be determined or notified in various ways, such as by value and / or by calculation. In the case of by value, the threshold... This can be indicated by, for example, discarding mask values ​​that are below a threshold and close to, for example, "0.0". This means that the corresponding nodes (those below the threshold and close to "0.0") will not contribute any prediction weight to the category prediction. For example, It can be set to 0.2, in which case it has less than Any node with a value of will be discarded from the ML model computation for the category (i.e., adaptively deactivated). The threshold is calculated... The threshold can be determined by processes such as using energy, entropy, or other sensitivity indicators of whether a node contributes to the prediction of a particular category.

[0071] After the initial mask is created (which may only be executed once for a set of classifications of the ML model), the mask can be used during the execution of inference for a specific category to adjust which nodes are activated and which are deactivated in the ML model (which is trained for multi-class classification).

[0072] Figure 5 Examples of processes for adaptively optimizing ML models (such as neural networks) using masks are described according to some embodiments.

[0073] At 510, the process may include, according to some embodiments, receiving instructions from a user device to perform a single-class inference task using a machine learning model trained to perform multi-class inference tasks. (See reference...) Figure 4 For example, device 102A can receive instructions (from device 102D) (e.g., see [reference]). Figure 4 (Send signal 3). In this case, device 102A may choose to focus on a specific category (e.g., "car" category) among the multiple categories that machine learning model 299 is able to classify.

[0074] At 515, the process may be implemented according to some embodiments, including: in response to an instruction, by a user device and during the execution of a machine learning model hosted by the user device, applying at least one mask mapped to a single-class inference task, wherein the at least one mask defines, for the single-class inference task, a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while the single-class inference task is being performed. (See reference...) Figure 4 For example, when the instruction of 510 is received, device 102A can apply a mask. For instance, the mask (e.g., specific to the "car" inference classification task associated with the instruction) can be retrieved from storage device 230 and used at ML model 299 at device 102A. Additionally, for single-class inference tasks, the mask defines which nodes of the machine learning model are activated during the execution of ML model 299 to perform the single-class inference task. Furthermore, the mask can also define or indicate which nodes of ML model 299 should be deactivated while performing the single-class inference task. Table 1 and... Figure 3B Further examples of masking and dynamic activation and deactivation of nodes are shown. Furthermore, activation and deactivation can occur within a single time frame, after which all nodes of the ML model are active, thus enabling ML model 299 to return to multi-class classification.

[0075] At point 520, the process can be implemented according to some embodiments, including: reconfiguring a machine learning model by a user device into a state where the first set of nodes and the second set of nodes perform multiple types of inference tasks. (See reference) Figure 4 For example, an ML model can return (see, for example, see...) Figure 4 4) Initial state or initial settings, so for example, the nodes of the ML model are active, and the ML model 299 is able to return to multi-class classification.

[0076] Figure 6A block diagram of apparatus 10 according to some embodiments is illustrated. As previously described, the disclosed machine learning model may be included in this apparatus. Apparatus 10 may include or be included in user equipment, such as user equipment (e.g., user entity, PRU, etc.). Generally, various embodiments of user equipment may include cellular phones (such as smartphones), tablet computers, personal digital assistants (PDAs) with wireless communication capabilities, portable computers with wireless communication capabilities, image capture devices (such as digital cameras) with wireless communication capabilities, gaming devices with wireless communication capabilities, music storage and playback devices with wireless communication capabilities, internet devices that allow wireless internet access and browsing, tablet computers with wireless communication capabilities, and portable units or terminals including combinations of such functions; additionally, it may include vehicles (such as cars and / or trucks) and aircraft (such as manned or unmanned aircraft) and portable units or terminals including combinations of such functions. User equipment may include IoT devices, industrial IoT (IIoT) devices and / or the like, or be included in IoT devices, industrial IoT (IIoT) devices and / or the like. For example, in the case of IoT devices or IIoT devices, the UE can be configured to operate with fewer resources (e.g., in terms of power, processing speed, memory, etc.) compared to a smartphone.

[0077] Device 10 may include at least one antenna 12 communicating with transmitter 14 and receiver 16. Alternatively, the transmitting and receiving antennas may be separate. Device 10 may also include a processor 20 configured to provide signals to the transmitter and receive signals from the receiver, respectively, and to control the functions of the device. Processor 20 may be configured to implement control signaling via electronic leads to the transmitter and receiver to control the functions of the transmitter and receiver. Similarly, processor 20 may be configured to implement control signaling via electronic leads connecting processor 20 to other elements such as displays or memory to control other elements of device 10. For example, processor 20 may be implemented in a variety of ways, including circuit systems, at least one processing core, one or more microprocessors having accompanying digital signal processors (DSPs), one or more processors without accompanying DSPs, one or more coprocessors, one or more multi-core processors, one or more controllers, processing circuit systems, one or more computers, various other processing elements, including integrated circuits (e.g., application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and / or combinations thereof). Therefore, although Figure 6While shown as a single processor, in some embodiments, processor 20 may include multiple processors or processing cores. Alternatively or additionally, for example, device 10 may include a GPU, AI chip, and / or other aspects to at least partially perform the functions described above. Figure 2 The ML model is described above. Alternatively or additionally, for example, device 10 may be coupled to or include the above-mentioned reference. Figure 2 The aforementioned sensors, cameras, and / or etc.

[0078] Device 10 may be capable of operating with one or more air interface standards, communication protocols, modulation types, access types, and / or the like. Signals transmitted and received by processor 20 may include signaling information according to the applicable cellular system's air interface standard, and / or any number of different wired or wireless network technologies, including but not limited to Wi-Fi, Wireless Local Access Network (WLAN) technologies (such as IEEE 802.11, 802.16, 802.3, ADSL, DOCSIS, and / or the like). Additionally, these signals may include voice data, user-generated data, user-requested data, and / or the like.

[0079] For example, the cellular modem in device 10 and / or therein may be capable of operating according to various first-generation (1G) communication protocols, second-generation (2G or 2.5G) communication protocols, third-generation (3G) communication protocols, fourth-generation (4G) communication protocols, fifth-generation (5G) communication protocols, sixth-generation (6G) communication protocols, Internet Protocol Multimedia Subsystem (IMS) communication protocols (e.g., Session Initiation Protocol (SIP) and / or etc.). For example, device 10 may be capable of operating according to 2G wireless communication protocols IS-136, Time Division Multiple Access (TDMA), Global System for Mobile Communications (GSM), IS-95, Code Division Multiple Access (CDMA) and / or etc. Additionally, for example, device 10 may be capable of operating according to 2.5G wireless communication protocols General Packet Radio Service (GPRS), Enhanced Data GSM Environment (EDGE) and / or etc. Furthermore, for example, device 10 may be able to operate according to 3G wireless communication protocols, such as Universal Mobile Telecommunications System (UMTS), Code Division Multiple Access 2000 (CDMA2000), Wideband Code Division Multiple Access (WCDMA), Time Division Synchronous Code Division Multiple Access (TD-SCDMA), and / or the like. Device 10 may also be able to operate according to 3.9G wireless communication protocols, such as Long Term Evolution (LTE), Evolved Universal Terrestrial Radio Access Network (E-UTRAN), and / or the like. Additionally, for example, device 10 may be able to operate according to 4G wireless communication protocols, such as Advanced LTE, 5G, 6G, and / or the like, as well as similar wireless communication protocols that may be developed subsequently.

[0080] It should be understood that processor 20 may include circuitry for implementing the audio / video and logic functions of device 10. For example, processor 20 may include digital signal processor devices, microprocessor devices, analog-to-digital converters, digital-to-analog converters, and / or the like. The control and signal processing functions of device 10 may be distributed among these devices according to their respective capabilities. Processor 20 may also additionally include an internal voice encoder (VC) 20a, an internal data modem (DM) 20b, and / or the like. Furthermore, processor 20 may include the ability to operate one or more software programs, which may be stored in memory. Typically, processor 20 and the stored software instructions may be configured to cause device 10 to perform actions. For example, processor 20 may be able to operate a connectivity program, such as a web browser. The connectivity program may allow device 10 to send and receive network content (such as location-based content) according to protocols such as Wireless Application Protocol (WAP), Hypertext Transfer Protocol (HTTP), and / or similar protocols.

[0081] Device 10 may also include a user interface, including, for example, headphones or speaker 24, ringer 22, microphone 26, display 28, user input interface, and / or the like, which are operatively coupled to processor 20. As described above, display 28 may include a touch-sensitive display, allowing a user to make selections, input values, and / or the like through touch and / or gestures. Processor 20 may also include a user interface circuitry configured to control at least some functions of one or more elements in the user interface, such as speaker 24, ringer 22, microphone 26, display 28, and / or the like. Processor 20 and / or the user interface circuitry including processor 20 may be configured to control one or more functions of one or more elements in the user interface via computer program instructions (e.g., software and / or firmware), which are stored in a memory accessible to processor 20, such as volatile memory 40, non-volatile memory 42, and / or the like. Device 10 may include a battery for powering various circuits associated with the mobile terminal, such as circuitry for providing mechanical vibration as a detectable output. The user input interface may include devices that allow device 20 to receive data, such as keyboard 30 (which may be a virtual keyboard displayed on display 28 or an externally coupled keyboard) and / or other input devices.

[0082] like Figure 6As shown, device 10 may also include one or more mechanisms for sharing and / or acquiring data. For example, device 10 may include a short-range radio frequency (RF) transceiver and / or interrogator 64, so that data can be shared with and / or acquired from electronic devices according to RF technology. Device 10 may include other short-range transceivers, such as infrared (IR) transceiver 66, Bluetooth™ transceiver 68 operating using Bluetooth™ (BT) wireless technology, wireless universal serial bus (USB) transceiver 70, Bluetooth™ Low Energy transceiver, ZigBee transceiver, ANT transceiver, cellular device-to-device transceiver, wireless LAN link transceiver, and / or any other short-range radio technology. For example, device 10, particularly a short-range transceiver, may be able to send data to and / or receive data from electronic devices in its vicinity (e.g., within 10 meters). The device 10, including a Wi-Fi or wireless LAN modem, can also send and / or receive data from electronic devices according to various wireless network technologies, including 6LoWPAN, Wi-Fi, Wi-Fi Low Energy, WLAN technologies such as IEEE 802.11, IEEE 802.15, IEEE 802.16 and / or others.

[0083] Device 10 may include memory such as a Subscriber Identity Module (SIM) 38, a Removable Subscriber Identity Module (R-UIM), eUICC, UICC, U-SIM, and / or the like, which may store information elements related to mobile subscribers. In addition to the SIM, device 10 may also include other removable and / or fixed memory. Device 10 may include volatile memory 40 and / or non-volatile memory 42. For example, volatile memory 40 may include random access memory (RAM), including dynamic RAM and / or static RAM, on-chip or off-chip cache, and / or the like. Non-volatile memory 42 may be embedded and / or removable, for example, it may include read-only memory, flash memory, magnetic storage devices (e.g., hard disks, floppy disk drives, magnetic tape, optical disk drives, and / or media), non-volatile random access memory (NVRAM), and / or the like. Similar to volatile memory 40, non-volatile memory 42 may include a cache for temporary data storage. At least a portion of the volatile memory and / or non-volatile memory may be implemented in processor 20. These memories can store one or more software programs, instructions, information fragments, data, and / or the like, which can be used by the apparatus to perform the operations disclosed herein.

[0084] The memory may include identifiers, such as an International Mobile Equipment Identity (IMEI) code, which can uniquely identify device 10. In an example embodiment, processor 20 may be configured using computer code stored in memory 40 and / or 42 to provide the operations disclosed herein concerning a UE (such as a user entity).

[0085] Some embodiments disclosed herein can be implemented as software, hardware, application logic, or a combination of software, hardware, and application logic. For example, the software, application logic, and / or hardware may reside on memory 40, control device 20, or electronic components. In some embodiments, application logic, software, or instruction sets are held on any of a variety of conventional computer-readable media. In the context of this document, a “computer-readable storage medium” can be any non-transitory medium capable of containing, storing, communicating, propagating, or transmitting instructions for use by or in connection with an instruction execution system, apparatus, or device, such as a computer or data processor circuit system; a computer-readable medium may include a non-transitory computer-readable storage medium, which can be any medium capable of containing or storing instructions for use by or in connection with an instruction execution system, apparatus, or device, such as a computer.

[0086] Without limiting the scope, interpretation, or application of the foregoing claims in any way, the technical effects of one or more example embodiments disclosed herein may include reduced inference latency (e.g., the time spent by the ML model performing inference tasks), reduced energy consumption, and / or reduced device resource utilization (e.g., in terms of processing and / or memory resources at the device). Additionally, the ML model may be specifically tailored for a particular inference category, thereby enhancing the available resources for performing that particular category.

[0087] In view of the above implementation of the subject matter, this application discloses the following list of examples, wherein a feature of a single example, or a combination of more than one feature of the examples, and optionally a combination with one or more features of one or more other examples, also belongs to the further examples within the disclosure of this application. Example 1. A method that includes: The user equipment receives instructions to perform a single-class reasoning task using a machine learning model, which is trained to perform multi-class reasoning tasks. In response to an instruction, during a time frame of execution of a machine learning model hosted by the user device, the application is mapped to at least one mask for a single-class inference task, wherein the at least one mask defines, for the single-class inference task: a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while performing the single-class inference task; and The user device reconfigures the machine learning model into a state where the first set of nodes and the second set of nodes perform multiple types of inference tasks. Example 2. Following the approach of Example 1, the machine learning model is at least partially trained to perform multi-class inference tasks before being deployed to the user device. Example 3. The method according to any one of Examples 1 to 2, wherein the instruction to perform a single-class inference task using a machine learning model is received from at least one other user device that requests the single-class inference task. Example 4. A method according to any of Examples 1 to 3, wherein an instruction to perform a single-class inference task using a machine learning model is received in response to the single-class inference task being detected by the user device. Example 5. The method of any one of Examples 1 to 4, wherein the time frame is a predetermined time period or the time when a single inference task is no longer an inference task inferred by the user device. Example 6. According to the method of any one of Examples 1 to 5, wherein at least one mask is learned by: providing a machine learning model with samples for classification as a single-class inference task, and analyzing multiple nodes of the machine learning model to determine a second set of nodes that are not associated with the classification of the machine learning model for the single-class inference task. Example 7. Using the method of any of Examples 1 to 6, where the weight of the second node set is below a threshold, indicating that the second node set is not associated with the classification of a single-class reasoning task. Example 8. Based on any of Examples 1 through 7, where the machine learning model performs a multi-class inference task in response to reconfiguration. Example 9. A method according to any one of Examples 1 to 8, wherein in response to applying at least one mask, a machine learning model is executed to perform a single-class inference task, wherein a first set of nodes is executed to perform the single-class inference task, and a second set of nodes is deactivated and not executed while the single-class inference task is being performed. Example 10. An apparatus comprising: At least one processor; and At least one memory, the at least one memory including instructions, which, when executed by at least one processor, cause an operation including: Receive instructions to perform a single-class reasoning task using a machine learning model, which is trained to perform multi-class reasoning tasks; In response to an instruction, during a time frame of execution of the machine learning model hosted by the device, the application is mapped to at least one mask for a single-class inference task, wherein the at least one mask defines, for the single-class inference task: a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while performing the single-class inference task; and Reconfigure the machine learning model into a state where the first set of nodes and the second set of nodes perform multiple types of inference tasks. Example 11. The apparatus according to Example 10, wherein the machine learning model is at least partially trained to perform multi-class inference tasks before being deployed to the apparatus. Example 12. An apparatus according to any one of Examples 10 to 11, wherein an instruction to perform a single-class reasoning task using a machine learning model is received from at least one other apparatus that requests the single-class reasoning task. Example 13. An apparatus according to any one of Examples 10 to 12, wherein an instruction to perform a single-class reasoning task using a machine learning model is received in response to the single-class reasoning task being detected by the apparatus. Example 14. An apparatus according to any one of Examples 10 to 13, wherein the time frame is a predetermined time period, or the time when a single-class reasoning task is no longer a reasoning task inferred by the apparatus. Example 15. An apparatus according to any one of Examples 10 to 14, wherein at least one mask is learned by: providing a machine learning model with samples for classification as a single-class reasoning task, and analyzing multiple nodes of the machine learning model to determine a second set of nodes that are not associated with the classification of the machine learning model for the single-class reasoning task. Example 16. A device according to any of Examples 10 to 15, wherein the weight of the second set of nodes is below a threshold, indicating that the second set of nodes is not associated with the classification of a single-class reasoning task. Example 17. An apparatus according to any of Examples 10 to 16, wherein a machine learning model performs a multi-class inference task in response to reconfiguration. Example 18. An apparatus according to any one of Examples 10 to 17, wherein in response to the application of at least one mask, a machine learning model is executed to perform a single-class inference task, wherein a first set of nodes is executed to perform the single-class inference task, and a second set of nodes is deactivated and not executed while the single-class inference task is being performed. Example 19. An apparatus according to any one of Examples 10 to 18, wherein the apparatus includes or is included in a user equipment. Example 20. A non-transitory computer-readable storage medium comprising instructions that, when executed by at least one processor, cause an operation including: Receive instructions to perform a single-class reasoning task using a machine learning model, which is trained to perform multi-class reasoning tasks; In response to an instruction, during a time frame of execution of a machine learning model hosted by the user device, the application is mapped to at least one mask for a single-class inference task, wherein the at least one mask defines, for the single-class inference task: a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while performing the single-class inference task; and Reconfigure the machine learning model into a state where the first set of nodes and the second set of nodes perform multiple types of inference tasks. Example 21. An apparatus comprising: A component for receiving instructions to perform a single-class reasoning task using a machine learning model, which is trained to perform multi-class reasoning tasks. A component for responding to an instruction to apply at least one mask mapped to a single-class inference task during a time frame of execution of a machine learning model hosted by the user device, wherein the at least one mask defines, for the single-class inference task: a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while performing the single-class inference task; and A component used to reconfigure a machine learning model into a state where the first set of nodes and the second set of nodes perform multiple types of inference tasks. Example 22. The apparatus of Example 21 further includes a component for performing at least one function of any of Examples 2 to 9.

[0088] Depending on the required configuration, the topics described herein can be embodied in systems, apparatuses, methods, and / or articles. For example, the base stations and user equipment (or one or more components thereof) and / or processes described herein can be implemented using one or more of the following: processors that execute program code, application-specific integrated circuits (ASICs), digital signal processors (DSPs), embedded processors, field-programmable gate arrays (FPGAs), and / or combinations thereof. These various implementations can include implementations in one or more computer programs that are executable and / or interpretable on a programmable system comprising: at least one programmable processor (which may be a dedicated processor or a general-purpose processor), at least one input device, and at least one output device, wherein the at least one programmable processor is coupled to receive data and instructions from and to a storage system. These computer programs (also referred to as programs, software, software applications, applications, components, program code, or code) include machine instructions for the programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages ​​and / or assembly / machine languages. As used herein, "computer-readable medium" means any computer program product, machine-readable medium, computer-readable storage medium, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions. Similarly, systems that may include a processor and memory coupled to the processor are also described herein. The memory may include one or more programs that cause the processor to perform one or more of the operations described herein.

[0089] Although several variations have been described in detail above, other modifications or additions are still possible. In particular, additional features and / or variations may be provided in addition to the features and / or variations described herein. Furthermore, the above implementations may refer to various combinations and sub-combinations of the disclosed features, and / or combinations and sub-combinations of the aforementioned additional features. Other embodiments are within the scope of the foregoing claims.

[0090] If desired, the different functions discussed herein may be performed in different orders and / or in parallel with each other. Furthermore, one or more of the above functions may be optional or may be combined, if desired. Although aspects of some embodiments in the examples are set forth in the independent claims, other aspects of some embodiments in the examples include other combinations of features from the said embodiments and / or dependent claims with features of the independent claims, not just those expressly listed in the claims. It should also be noted herein that while exemplary embodiments have been described above, these descriptions should not be considered limiting. Rather, several variations and modifications may be made without departing from the scope of some embodiments as defined in the appended claims. Other embodiments may be within the scope of the foregoing claims. The term “based on” includes “at least based on”. Unless otherwise stated, the phrase “such as” is used to mean “such as, for example”.

Claims

1. A method comprising: The user equipment receives an instruction to perform a single-class reasoning task using a machine learning model, which is trained to perform multi-class reasoning tasks; In response to the instruction, the user equipment applies at least one mask mapped to the single-class inference task during a time frame of the execution of the machine learning model hosted by the user equipment, wherein the at least one mask defines for the single-class inference task: a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while the single-class inference task is being performed. as well as The user equipment reconfigures the machine learning model to a state in which the first set of nodes and the second set of nodes perform the multi-class inference task.

2. The method of claim 1, wherein the machine learning model is at least partially trained to perform the multi-class inference task before being deployed to the user device.

3. The method of claim 1, wherein the instruction to perform the single-class inference task using the machine learning model is received from at least one other user device that requested the single-class inference task.

4. The method of claim 1, wherein the instruction to perform the single-class inference task using the machine learning model is received in response to the single-class inference task being detected by the user equipment.

5. The method according to claim 1, wherein the time frame is a predetermined time period, or the time when the single-type inference task is no longer an inference task inferred by the user device.

6. The method of claim 1, wherein the at least one mask is learned by: providing the machine learning model with samples for classification as the single-class inference task, and analyzing a plurality of nodes of the machine learning model to determine a second set of nodes not associated with the classification of the machine learning model by the single-class inference task.

7. The method of claim 6, wherein the weight of the second node set is below a threshold, indicating that the second node set is not associated with the classification of the single-class reasoning task.

8. The method of claim 1, wherein in response to the reconfiguration, the machine learning model performs the multi-class inference task.

9. The method of claim 1, wherein in response to applying the at least one mask, the machine learning model is executed to perform the single-class inference task, wherein the first node set is executed to perform the single-class inference task, and the second node set is deactivated and not executed while the single-class inference task is being performed.

10. An apparatus comprising: At least one processor; as well as At least one memory, the at least one memory including instructions that, when executed by the at least one processor, cause an operation including: Receive an instruction to perform a single-class reasoning task using a machine learning model, which is trained to perform multi-class reasoning tasks; In response to the instruction, during a time frame of execution of the machine learning model hosted by the device, the application is mapped to at least one mask of the single-class inference task, wherein the at least one mask defines, for the single-class inference task: a first set of nodes activated at the machine learning model to perform the single-class inference task, and a second set of nodes deactivated at the machine learning model while performing the single-class inference task; and The machine learning model is reconfigured to a state in which the first set of nodes and the second set of nodes perform the multi-class inference task.

11. The apparatus of claim 10, wherein the machine learning model is at least partially trained to perform the multi-class inference task before being deployed to the apparatus.

12. The apparatus of claim 10, wherein the instruction to perform the single-class reasoning task using the machine learning model is received from at least one other apparatus that requested the single-class reasoning task.

13. The apparatus of claim 10, wherein the instruction to perform the single-class reasoning task using the machine learning model is received in response to the single-class reasoning task being detected by the apparatus.

14. The apparatus of claim 10, wherein the time frame is a predetermined time period or the time when the single-type reasoning task is no longer a reasoning task inferred by the apparatus.

15. The apparatus of claim 10, wherein the at least one mask is learned by: providing the machine learning model with samples for classification as the single-class inference task, and analyzing a plurality of nodes of the machine learning model to determine a second set of nodes not associated with the classification of the machine learning model by the single-class inference task.

16. The apparatus of claim 15, wherein the weight of the second node set is below a threshold, indicating that the second node set is not associated with the classification of the single-class reasoning task.

17. The apparatus of claim 10, wherein the machine learning model performs the multi-class inference task in response to the reconfiguration.

18. The apparatus of claim 10, wherein in response to applying the at least one mask, the machine learning model is executed to perform the single-class inference task, wherein the first node set is executed to perform the single-class inference task, and the second node set is deactivated and not executed while the single-class inference task is being performed.

19. The apparatus of claim 10, wherein the apparatus comprises or is included in a user equipment.

Images