Personalized Neural Network Pruning

Abstract

A method for generating a personalization model includes receiving one or more personal data samples from a user. A prototype of personally identifiable information is generated based on the personal data samples. The prototype of personally identifiable information is trained to reflect personal characteristics of the user. A network graph is generated based on the prototype of personally identifiable information. One or more channels of a global network are pruned based on the network graph to generate the personalization model.

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Patent Application No. 17 / 506,646, entitled "PERSONALIZED NEURAL NETWORK PRUNING," filed October 20, 2021, which claims the benefit of U.S. Provisional Patent Application No. 63 / 094,767, entitled "PERSONALIZED NEURAL NETWORK PRUNING," filed October 21, 2020, the disclosures of which are expressly incorporated herein by reference in their entireties.

[0002] Aspects of the present disclosure generally relate to generating personalized neural networks. [Background technology]

[0003] An artificial neural network may comprise an interconnected group of artificial neurons (e.g., neuron models). An artificial neural network may be a computational device or may be represented as a method to be executed by a computational device. A convolutional neural network is a type of feedforward artificial neural network. A convolutional neural network may include a collection of neurons, each with a receptive field, that collectively tile the input space. Convolutional neural networks (CNNs), such as deep convolutional neural networks (DCNs), have numerous applications. Specifically, these neural network architectures are used in various technologies such as image recognition, speech recognition, acoustic scene classification, keyword spotting, autonomous driving, and other classification tasks.

[0004] Edge devices such as smartphones are widely used. Given the many useful applications of neural networks, there is an increasing demand for their use on edge devices and personalized services for such edge devices. However, edge devices have limited computational resources, and generalized models may require more complex networks and more computation. Therefore, conventional personalization methods may not be suitable for edge devices because they use retraining or fine-tuning with limited personal data. Unfortunately, these multiple trainings typically have significant computational costs during on-device learning, which can be excessively burdensome for edge devices. Summary of the Invention [Means for solving the problem]

[0005] The present disclosure is set out in respective independent claims. Some aspects of the disclosure are set out in dependent claims.

[0006] In one aspect of the present disclosure, a computer-implemented method is provided. The method includes receiving one or more personal data samples. The method also includes generating a prototype of personally identifiable information based on the personal data samples. The prototype of personally identifiable information is trained to reflect personal characteristics of a user. In addition, the method includes generating a network graph based on the prototype of personally identifiable information. Furthermore, the method includes pruning one or more channels of a global network based on the network graph to generate a personalized model.

[0007] In another aspect of the present disclosure, an apparatus is provided. The apparatus includes a memory and one or more processors coupled to the memory. The processor is configured to receive one or more personal data samples. The processor is also configured to generate a prototype of personally identifiable information based on the personal data samples. The prototype of personally identifiable information is trained to reflect personal characteristics of a user. Additionally, the processor is configured to generate a network graph based on the prototype of personally identifiable information. Furthermore, the processor is configured to prune one or more channels of a global network based on the network graph to generate a personalized model.

[0008] In another aspect of the present disclosure, an apparatus is provided. The apparatus includes means for receiving one or more personal data samples. The apparatus also includes means for generating a prototype of personally identifiable information based on the personal data samples. The prototype of personally identifiable information is trained to reflect personal characteristics of a user. In addition, the apparatus includes means for generating a network graph based on the prototype of personally identifiable information. Furthermore, the apparatus includes means for pruning one or more channels of a global network based on the network graph to generate a personalized model.

[0009] In a further aspect of the present disclosure, a non-transitory computer-readable medium is provided. The computer-readable medium encodes program code. The program code, when executed by a processor, includes code for receiving one or more personal data samples. The program code also includes code for generating a personal identification information prototype based on the personal data samples. The personal identification information prototype is trained to reflect personal characteristics of a user. In addition, the program code includes code for generating a network graph based on the personal identification information prototype. Furthermore, the program code includes code for pruning one or more channels of a global network based on the network graph to generate a personalization model.

[0010] Additional features and advantages of the present disclosure are described below. Those skilled in the art will appreciate that the present disclosure may be readily utilized as a basis for modifying or designing other structures for carrying out the same purposes of the present disclosure. Those skilled in the art will also appreciate that such equivalent constructions do not depart from the teachings of the present disclosure as set forth in the appended claims. The novel features believed characteristic of the present disclosure, both as to its organization and method of operation, together with further objects and advantages, will be better understood from the following description when considered in connection with the accompanying figures. It is to be expressly understood, however, that each of the figures is provided for the purpose of illustration and description only and is not intended to define the scope of the present disclosure.

[0011] The features, nature, and advantages of the present disclosure will become more apparent from the detailed description set forth below when taken in conjunction with the drawings in which like reference characters identify correspondingly throughout. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 illustrates an example implementation of a neural network using a system-on-chip (SOC) including a general-purpose processor, in accordance with some aspects of the present disclosure. [Figure 2A]FIG. 1 illustrates a neural network according to an aspect of the present disclosure. [Figure 2B] FIG. 1 illustrates a neural network according to an aspect of the present disclosure. [Figure 2C] FIG. 1 illustrates a neural network according to an aspect of the present disclosure. [Figure 2D] FIG. 1 illustrates an exemplary deep convolutional network (DCN), according to aspects of the present disclosure. [Figure 3] FIG. 1 is a block diagram illustrating an example deep convolutional network (DCN), according to aspects of the present disclosure. [Figure 4] FIG. 10 is a block diagram illustrating an example pruning of a graph of a full or global model to generate a personalized model, according to aspects of the present disclosure. [Figure 5] FIG. 1 is a block diagram illustrating an example implementation of a personalization model, according to aspects of the present disclosure. [Figure 6A] FIG. 1 is a block diagram illustrating a prototype-based personalized pruning (PPP) model, according to aspects of the present disclosure. [Figure 6B] FIG. 1 is a block diagram illustrating a prototype-based personalized pruning (PPP) model, according to aspects of the present disclosure. [Figure 6C] FIG. 2 is a block diagram illustrating an example residual block, according to aspects of the present disclosure. [Figure 7] 1 is a flowchart illustrating an exemplary method for generating a personalized neural network model, according to aspects of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] The detailed description set forth below with reference to the accompanying drawings is intended as a description of various configurations and does not represent the only configuration in which the concepts described herein may be practiced. The detailed description includes specific details to provide a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form to avoid obscuring such concepts.

[0014] Based on the teachings, those skilled in the art will appreciate that the scope of the present disclosure encompasses any aspect of the present disclosure, whether implemented independently of or in combination with any other aspect of the present disclosure. For example, an apparatus may be implemented or a method may be practiced using any number of the described aspects. Additionally, the scope of the present disclosure is intended to encompass such apparatuses or methods practiced using other structures, functions, or structures and functions in addition to or other than the various aspects of the present disclosure described. It should be understood that any aspect of the present disclosure that is disclosed may be embodied by one or more elements of a claim.

[0015] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any aspect described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects.

[0016] While particular aspects are described herein, numerous variations and permutations of these aspects fall within the scope of the present disclosure. While certain benefits and advantages of the preferred aspects are described, the scope of the present disclosure is not limited to particular benefits, applications, or objectives. Rather, aspects of the present disclosure are intended to be broadly applicable to different technologies, system configurations, networks, and protocols, some of which are illustrated by way of example in the figures and the following description of the preferred aspects. The detailed description and drawings are merely illustrative of the present disclosure, rather than limiting, the scope of which is defined by the appended claims and their equivalents.

[0017] Personalization of deep neural networks is an important topic in computer vision and signal processing, with numerous applications such as recommender systems, smart assistance, speaker verification, and keyword spotting. To achieve broad appeal, a generalized model may be distributed to numerous users. Each user may have locally generated and / or user-specific data, such as data generated via the user's camera, microphone, or sensors on the user's edge device. Such data is sometimes referred to as personal data. Personalization of a global model may have the conflicting objectives of performing well on personal data with relatively few samples while continuing to perform well on general data. Training a global model to perform well in numerous applications relies on training data to encompass all situations of generalization. However, this assumption deviates from real-world scenarios.

[0018] One approach to personalizing deep neural networks is to fine-tune a global model using limited personal data. However, storing individual global deep neural network (DNN) models on memory-limited edge devices is impractical, and optimizing DNN models using limited personal data is difficult. Deploying DNNs requires powerful computational resources, and deploying DNNs on edge devices is therefore a challenging problem due to memory constraints and computational load. To solve this problem, numerous model compression methods, such as knowledge distillation, model quantization, and model pruning, have been considered. However, such conventional methods require additional training stages or retraining. Therefore, additional computational costs may be unavoidable.

[0019] Aspects of the present disclosure are directed to prototype-based personalization pruning (PPP), which uses prototype representations to learn a metric space. PPP considers model complexity as well as personalization. PPP selects a subset graph using a prototype module that uses prototypes of personally identifiable information derived from limited personalization data. For this reason, the model structure of PPP may be dynamically customized to enhance personalization and reduce model complexity. According to aspects of the present disclosure, a prototype of personally identifiable information for each group of personal data is provided. The prototype may represent characteristics of the personal data. A dynamic path network may be trained for personalization pruning to generate a personalized model, which is a pruned version of the generalized model or full model without fine-tuning using personal data.

[0020] 1 shows an example implementation of a system-on-chip (SOC) 100, which may include a central processing unit (CPU) 102 or multi-core CPU configured for personalized neural network pruning. Variables (e.g., neural signals and synaptic weights), system parameters associated with a computational device (e.g., a neural network with weights), delays, frequency bin information, and task information may be stored in memory blocks associated with a neural processing unit (NPU) 108, a memory block associated with the CPU 102, a memory block associated with a graphics processing unit (GPU) 104, a memory block associated with a digital signal processor (DSP) 106, a memory block 118, or may be distributed across multiple blocks. Instructions executed in the CPU 102 may be loaded from a program memory associated with the CPU 102 or from the memory block 118.

[0021] SOC 100 may also include a connectivity block 110 that may include a GPU 104, a DSP 106, fifth-generation (5G) connectivity, fourth-generation long-term evolution (4G LTE) connectivity, Wi-Fi connectivity, USB connectivity, Bluetooth connectivity, etc., as well as additional processing blocks adapted for specific functions, such as a multimedia processor 112 that may detect and recognize gestures, for example. In one implementation, NPU 108 is implemented in CPU 102, DSP 106, and / or GPU 104. SOC 100 may also include a sensor processor 114, an image signal processor (ISP) 116, and / or a navigation module 120 that may include a global positioning system.

[0022] The SOC 100 may be based on the ARM instruction set. In one aspect of the present disclosure, the instructions loaded into the general-purpose processor 102 may include code for receiving one or more personal data samples. The general-purpose processor 102 may also include code for generating a personal identity prototype based on the personal data samples. The personal identity prototype is trained to reflect the user's personal characteristics. The general-purpose processor 102 may also include code for generating a network graph based on the personal identity prototype. The general-purpose processor 102 may further include code for pruning one or more channels of a global network based on the network graph to generate a personalization model.

[0023] Deep learning architectures may perform object recognition tasks by learning to represent inputs at successively higher levels of abstraction within each layer, thereby establishing useful feature representations of the input data. In this way, deep learning addresses a major bottleneck of traditional machine learning. Prior to the advent of deep learning, machine learning approaches to object recognition problems may have relied heavily on human-designed features, possibly in combination with shallow classifiers. A shallow classifier may, for example, be a two-class linear classifier in which a weighted sum of feature vector components may be compared to a threshold to predict which class an input belongs to. Human-designed features may be templates or kernels adapted to a particular problem domain by an engineer with domain expertise. In contrast, deep learning architectures may learn, but through training, to represent features similar to those a human engineer might design. Furthermore, deep networks may learn to represent and recognize new types of features that humans may not have considered.

[0024] Deep learning architectures may learn a hierarchy of features. When presented with visual data, for example, a first layer may learn to recognize relatively simple features, such as edges, in the input stream. In another example, when presented with auditory data, the first layer may learn to recognize spectral power at specific frequencies. A second layer, taking the output of the first layer as input, may learn to recognize combinations of features, such as simple shapes in the case of visual data, or combinations of sounds in the case of auditory data. For example, higher layers may learn to represent complex shapes in visual data or words in auditory data. Even higher layers may learn to recognize common visual objects or spoken phrases.

[0025] Deep learning architectures can perform particularly well when applied to problems that have a natural hierarchical structure. For example, motor vehicle classification can benefit from first learning to recognize wheels, windshields, and other features. These features may be combined in different ways in higher layers to recognize cars, trucks, and airplanes.

[0026] Neural networks may be designed with various connectivity patterns. In feedforward networks, information is passed from lower layers to higher layers, with each neuron in a given layer transmitting to neurons in higher layers. A hierarchical representation may be established in successive layers of a feedforward network, as described above. Neural networks may also have recurrent or feedback (also called top-down) connections. In recurrent connections, the output from a neuron in a given layer may be transmitted to another neuron in the same layer. Recurrent architectures may be useful in recognizing patterns across two or more of the input data chunks delivered to the neural network in sequence. Connections from neurons in a given layer to neurons in lower layers are called feedback (or top-down) connections. Networks with many feedback connections may be useful when recognizing high-level concepts can help distinguish certain low-level features of the input.

[0027] The connections between layers of a neural network may be fully connected or locally connected. FIG. 2A shows an example of a fully connected neural network 202. In a fully connected neural network 202, a neuron in a first layer may transmit its output to every neuron in a second layer, resulting in each neuron in the second layer receiving input from every neuron in the first layer. FIG. 2B shows an example of a locally connected neural network 204. In a locally connected neural network 204, a neuron in a first layer may be connected to a limited number of neurons in the second layer. More generally, the locally connected layers of a locally connected neural network 204 may be configured such that each neuron in a layer has the same or similar connectivity pattern, but with connection strengths that may have different values ​​(e.g., 210, 212, 214, and 216). Because higher layer neurons in a given region may receive inputs that are tuned through training to the properties of a limited subset of the total inputs to the network, the connectivity patterns of local connections may give rise to spatially distinct receptive fields within the higher layers.

[0028] One example of a locally connected neural network is a convolutional neural network. Figure 2C shows an example of a convolutional neural network 206. The convolutional neural network 206 may be configured such that the connection strengths associated with the inputs to each neuron in the second layer are shared (e.g., 208). Convolutional neural networks may be suitable for problems where the spatial location of the inputs is meaningful.

[0029] One type of convolutional neural network is the deep convolutional network (DCN). Figure 2D shows a detailed example of a DCN 200 designed to recognize visual features from images 226 input from an image capture device 230, such as an in-car camera. The DCN 200 in this example may be trained to identify traffic signs and numbers given on traffic signs. Of course, the DCN 200 may also be trained for other tasks, such as identifying lane markings or traffic signals.

[0030] The DCN 200 may be trained using supervised learning. During training, the DCN 200 may be presented with an image, such as a speed limit sign image 226, and then a forward pass may be computed to generate the output 222. The DCN 200 may include a feature extraction section and a classification section. Upon receiving the image 226, the convolutional layer 232 may apply a convolutional kernel (not shown) to the image 226 to generate the first set of feature maps 218. As an example, the convolutional kernel for the convolutional layer 232 may be a 5×5 kernel that generates a 28×28 feature map. In this example, four different feature maps are generated in the first set of feature maps 218, so four different convolutional kernels were applied to the image 226 in the convolutional layer 232. A convolutional kernel may also be referred to as a filter or a convolutional filter.

[0031] The first set of feature maps 218 may be subsampled by a max pooling layer (not shown) to generate a second set of feature maps 220. The max pooling layer reduces the size of the first set of feature maps 218. That is, the size of the second set of feature maps 220, such as 14x14, is smaller than the size of the first set of feature maps 218, such as 28x28. The reduced size provides similar information to subsequent layers while reducing memory consumption. The second set of feature maps 220 may be further convolved through one or more subsequent convolutional layers (not shown) to generate one or more subsequent sets of feature maps (not shown).

[0032] 2D , the second set of feature maps 220 is convolved to generate a first feature vector 224. Furthermore, the first feature vector 224 is further convolved to generate a second feature vector 228. Each feature in the second feature vector 228 may include a number corresponding to a possible feature of the image 226, such as "sign," "60," and "100." A softmax function (not shown) may convert the numbers in the second feature vector 228 into probabilities. Thus, the output 222 of the DCN 200 is the probability that the image 226 contains one or more features.

[0033] In this example, the probabilities in output 222 for "sign" and "60" are higher than the probabilities for other outputs 222, such as "30," "40," "50," "70," "80," "90," and "100." Before training, the output 222 produced by DCN 200 may be inaccurate. Thus, an error may be calculated between output 222 and a target output. The target output is the ground truth of image 226 (e.g., "sign" and "60"). The weights of DCN 200 may then be adjusted so that output 222 of DCN 200 more closely aligns with the target output.

[0034] To adjust the weights, the learning algorithm may calculate a gradient vector for the weights. The gradient may indicate the amount by which the error would increase or decrease if the weights were adjusted. In the top layer, the gradient may correspond directly to the value of the weights connecting activated neurons in the penultimate layer to neurons in the output layer. In lower layers, the gradient may depend on the value of the weights and the calculated error gradient of the upper layer. The weights may then be adjusted to reduce the error. This method of adjusting weights is sometimes called "backpropagation" because it involves a "backward pass" through the neural network.

[0035] In practice, the error gradient of the weights may be calculated over a small number of examples so that the calculated gradient approximates the true error gradient. This approximation method is sometimes called stochastic gradient descent. Stochastic gradient descent may be repeated until the achievable error rate of the overall system stops decreasing or until the error rate reaches a target level. After training, the DCN may be presented with new images, and a forward pass through the network may produce an output 222, which may be considered the DCN's inference or prediction.

[0036] A deep belief network (DBN) is a probabilistic model containing multiple layers of hidden nodes. DBNs may be used to extract hierarchical representations of a training dataset. DBNs may be obtained by stacking layers of restricted Boltzmann machines (RBMs). RBMs are a type of artificial neural network that can learn probability distributions over a set of inputs. Because RBMs can learn probability distributions in the absence of information about the class into which each input should be categorized, RBMs are often used in unsupervised learning. Using a hybrid unsupervised-supervised paradigm, the lower RBM of a DBN may be trained in an unsupervised manner and act as a feature extractor, while the upper RBM may be trained in a supervised manner (on the joint distribution of inputs from previous layers and the target class) and act as a classifier.

[0037] A deep convolutional network (DCN) is a network of convolutional networks composed of additional pooling and normalization layers. DCNs have achieved state-of-the-art performance for many tasks. DCNs can be trained using supervised learning, where both the input and output targets are known for a large number of examples and are used to modify the network weights using gradient descent.

[0038] The DCN may be a feedforward network. Additionally, as described above, connections from neurons in a first layer of the DCN to groups of neurons in the next higher layer are shared across neurons in the first layer. The feedforward and shared connections of the DCN may be exploited for high-speed processing. The computational burden of the DCN may be much less than that of a similarly sized neural network that includes, for example, recurrent or feedback connections.

[0039] The processing in each layer of a convolutional network may be considered a spatially invariant template or basis projection. If the input is first decomposed into multiple channels, such as the red, green, and blue channels of a color image, a convolutional network trained on that input may be considered three-dimensional, with two spatial dimensions along the image axes and a third dimension capturing color information. The output of the convolutional connections may be considered to form a feature map in a subsequent layer, with each element of the feature map (e.g., 220) receiving input from a range of neurons in the previous layer (e.g., feature map 218) and from each of multiple channels. Values ​​in the feature map may be further processed using nonlinearities such as rectification, max(0, x), etc. Values ​​from neighboring neurons may be further pooled (corresponding to downsampling) to achieve further local invariance and dimensionality reduction. Normalization, corresponding to whitening, may also be applied through lateral inhibition between neurons in the feature map.

[0040] The performance of deep learning architectures can improve as more labeled data points become available or as computational power increases. Modern deep neural networks are routinely trained using computational resources thousands of times greater than those available to the average researcher just 15 years ago. New architectures and training paradigms can further improve deep learning performance. Rectified linear units may reduce the training problem known as vanishing gradients. New training techniques may reduce overfitting, thus allowing larger models to achieve better generalization. Encapsulation techniques may extract data within a given receptive field, further improving overall performance.

[0041] 3 is a block diagram illustrating a deep convolutional network 350. The deep convolutional network 350 may include multiple different types of layers based on connectivity and weight sharing. As shown in FIG. 3, the deep convolutional network 350 includes convolutional blocks 354A and 354B. Each of the convolutional blocks 354A and 354B may be composed of a convolutional layer (CONV) 356, a normalization layer (LNorm) 358, and a max pooling layer (MAX POOL) 360.

[0042] The convolutional layer 356 may include one or more convolutional filters that may be applied to input data to generate feature maps. While only two of the convolutional blocks 354A, 354B are shown, the present disclosure is not so limited; instead, any number of convolutional blocks 354A, 354B may be included in the deep convolutional network 350 according to design preference. The normalization layer 358 may normalize the outputs of the convolutional filters. For example, the normalization layer 358 may perform whitening or lateral inhibition. The max-pooling layer 360 may perform downsampling aggregation across space for local invariance and dimensionality reduction.

[0043] For example, the parallel filter bank of the deep convolutional network may be loaded onto the CPU 102 or GPU 104 of the SOC 100 to achieve high performance and low power consumption. In alternative embodiments, the parallel filter bank may be loaded onto the DSP 106 or ISP 116 of the SOC 100. Additionally, the deep convolutional network 350 may access other processing blocks that may be present on the SOC 100, such as the sensor processor 114 and navigation module 120, which are dedicated to sensors and navigation, respectively.

[0044] The deep convolutional network 350 may also include one or more fully connected layers 362 (FC1 and FC2). The deep convolutional network 350 may further include a logistic regression (LR) layer 364. Between each layer 356, 358, 360, 362, 364 of the deep convolutional network 350 are weights (not shown) that are updated. The output of each of the layers (e.g., 356, 358, 360, 362, 364) may serve as input for subsequent ones of the layers (e.g., 356, 358, 360, 362, 364) in the deep convolutional network 350 to learn hierarchical feature representations from the input data 352 (e.g., image, audio, video, sensor data, and / or other input data) provided at the first one of the convolutional blocks 354A. The output of the deep convolutional network 350 is a classification score 366 for the input data 352. The classification score 366 may be a set of probabilities, where each probability is the probability that the input data contains a feature from the set of features.

[0045] Aspects of the present disclosure are directed to a prototype-based personalization pruning (PPP) model that learns a metric space using a prototype representation. PPP considers model complexity as well as personalization. PPP selects a subset graph using a prototype module that uses prototypes of personal identities derived from limited personalization data. For this reason, the model structure of PPP may be dynamically customized to enhance personalization and reduce model complexity. According to aspects of the present disclosure, a prototype of personal identities may be generated for each personality, which may be defined as each identity or group of personal data. The prototype of personal identities represents the characteristics of the personal data. A dynamic path network may be trained for personalization pruning to generate a personalized model, which is a pruned version of the generalized model or full model without fine-tuning using personal data.

[0046] During the training phase, the gating module may be trained to select appropriate channels for pruning. Personalization is also considered at the same time. Personal identity prototypes may be defined using the output of the gating module. The output of the gating module is a binary embedding pattern for each convolution module that indicates whether to prune each channel. To consider performance after pruning using personal identity prototypes during the training phase, the output of the gating module may be regularized to be similar to the output of the personal identity prototypes. As a result, after training, the global model may be more easily pruned by pruning using personal identity prototypes derived from given limited personal data, for example, using a graph generator and a network pruner. Furthermore, after training, PPP may operate without a global model (e.g., a full model) because channel selection is based on personal identity prototypes that represent personal data rather than the incoming input used in conventional dynamic path networks.

[0047] The architecture of the PPP may be, for example, a convolutional neural network (e.g., 350 shown in FIG. 3). The convolutional neural network may be configured with a gate module that controls each channel of the convolutional module. A gate g may be provided for each convolutional layer,

number

number

number

number

number

[0048] The PPP model may be trained based on a global model. The training dataset D={(x1, t1),...,(x N , t N )}, t i ∈{1,…,K}, where K is the number of classes.

number

[0049]

number

[0050] where h and w correspond to the height and width of the activation, respectively, x is the input activation, and y is the output of the convolutional layer.

number

[0051] The output feature maps and embedding vectors may be calculated according to the individual identity for each mini-batch. To train the gating module for channel selection, the pruned channels are passed through the broadcast operator

number

[0052] The prototype may represent personal identification information. In each mini-batch, the number of specific p-th identification data is n p Then, for all zp The average of the embedding vectors in

number

number

number

number

number

number

[0053] The personalization regularization loss may be calculated as follows:

number

[0054] loss L prototype Similar gate patterns, such as prototypes of personal identification information,

number

number

[0055] The target loss may act as a soft constraint on the network utilization. The number of all active channels (filters) in each convolutional layer may be counted, and the utilization may be regularized by a specific target. Thus, the target loss may be given by:

number

[0056] Therefore, the target loss L target may provide a means to adjust or control the network complexity. task Using the , the overall training loss may be expressed as: L total =L task +αL prototype +βL target (6) where α and β are hyperparameters for balancing the loss.

[0057] Unlike conventional techniques that use data-dependent paths that always use the entire network (e.g., by selecting a forward graph based on the input), PPP does not depend on the input after training. Rather, according to aspects of the present disclosure, the forward graph may depend on prototypes of given personal data. Thus, unlike conventional personalization methods, such as using personal data to fine-tune a global model, PPP generates a personal model on the fly, thereby reducing computational complexity.

[0058] In a testing phase where little personal data is given, a prototype of personal identity information may be calculated using Equation 2. Once the prototype of personal identity information is calculated, a pruned personalization model may be generated without additional training by removing or pruning convolutional layer filters having a given binary pattern for each specific personalization using a graph generator and a network pruner. In some aspects, pruning may be repeated to further strengthen the personalization model.

[0059] FIG. 4 is a block diagram illustrating an example pruning of a graph of a full model or global model 400 to generate a graph of a personalized model 402, according to an embodiment of the present disclosure. Referring to FIG. 4, the graph of the full model 400 may include multiple nodes (e.g., 406a) with interconnections (e.g., 408a) between the nodes (e.g., 406a). The graph of the full model 400 may represent a convolutional neural network (e.g., 350 shown in FIG. 3). Each column of nodes (e.g., 406a) of the full model 400 may correspond, for example, to a convolutional layer (e.g., 356 shown in FIG. 3) of the convolutional neural network. Each interconnection 408a may represent a channel to the node. A node (e.g., 406z) may be considered removed if all channels to or from the node (e.g., 406z) have been pruned or removed.

[0060] According to aspects of the present disclosure, the full model 400 may be pruned using personal data to remove or delete interconnections 408a or channels. As described, the convolutional neural network may be configured with a gate module that controls each channel of the convolution module. A gate g may be provided for each convolutional layer,

number

[0061] Next, each node that has no connection to another node (indicated by dashed line 406z) may also be removed. By removing interconnections 408a and nodes 406a, a personalized model 402 may be generated. In doing so, network size and computational complexity may be reduced.

[0062] FIG. 5 is a block diagram 500 illustrating an example implementation of a personalization model according to aspects of the present disclosure. Referring to FIG. 5, a smartphone 502 may, for example, register an enrollment for voice-assisted activation. A prototype generator 504 may create a user prototype based on the enrollment sample. A graph generator 506 may determine a network graph (structure) based on the user prototype of the personal identification information. The graph structure may be provided to a network pruner 508, which may then generate a personalization model by pruning the network based on the determined network graph. In some aspects, the pruned model may also be fine-tuned using user samples. Thus, the user's voice may be processed by a small personalization network.

[0063] 6A and 6B are block diagrams illustrating exemplary prototype-based personalization pruning (PPP) models 600 and 650, respectively, according to an embodiment of the present disclosure. Referring to FIG. 6A, the PPP model 600 includes a prototype generator 602, a graph generator 604, and a network pruner 606. The prototype generator 602 may generate prototypes of personal identification information based on personal data samples, such as those via a sensor 608 (e.g., a camera or microphone). The prototype generator 602 may provide the personal prototypes to the graph generator 604. The graph generator 604 may then determine a network graph structure based on the personal prototypes. The graph structure may then be provided to the network pruner 606, for example, as meta-information. Using the full model 610, the network pruner 606 performs a binary decision to remove channels of the full model 610 according to the determined graph to generate a smaller personalized model (shown as a pruned model 612). In operation, input data may be fed to pruned model 612 via sensor 608 to generate a prediction (e.g., a classification of an image or an indication that a keyword has been detected).

[0064] 6B, PPP model 650 may be further configured for knowledge transfer from full model 652. PPP model 650 may have a configuration and functionality similar to that of PPP model 600. During the teaching phase, data from sensors 656 (e.g., cameras or microphones) may be provided to full model 652. Full model 652 may generate predictions, which may then be provided to pruned model 654 to further train the personalized model.

[0065] 6C is a block diagram illustrating an example residual block 670 of a neural network according to an embodiment of the present disclosure. Referring to FIG. 6C, the example residual block 670 of a neural network (e.g., 350 shown in FIG. 3) includes a gating module 672a, 672b and a prototype generator 674a, 674b in each convolutional layer 676a, 676b of the neural network. While two convolutional layers are shown, this is merely an example for ease of explanation. In this configuration, the gating modules 672a, 672b control each channel of the convolutional block, rather than the entire residual block 670.

[0066] The prototype generator 674 generates continuous prototypes using the output of the gating module 672 for one or more personal identity data, for example, based on Equation 2. The prototype generator 674 may also convert the continuous prototypes into discrete prototypes, for example, by the step function of Equation 3.

[0067] 7 is a flowchart illustrating an example method 700 for generating a personalized network model according to an aspect of the present disclosure. As shown in FIG. 7, a personal data sample is received at block 702. In one example, the personal data sample may be generated via a sensor such as a camera or microphone.

[0068] At block 704, a personal identity prototype is generated based on the personal data samples. The personal identity prototype is trained to reflect the user's personal characteristics. For example, as described with reference to FIG. 6A, the prototype generator 602 may generate a personal identity prototype based on the personal data samples, such as via a sensor 608 (e.g., a camera or microphone). A personal identity prototype may be generated for each personality, which may be defined as each identity or group of personal data. That is, a prototype may represent a personal identity. In each mini-batch, the number of specific identity data is n p Then, for all z p The average of the embedding vectors in

number

number

number

[0069] A network graph is generated based on the personal identity prototypes at block 706. For example, as described with reference to Figure 6A, the graph generator 604 may receive the prototypes from the prototype generator 602. The graph generator 604 may determine a network graph structure based on the personal prototypes.

[0070] At block 708, channels of the global network are pruned based on the network graph to generate a personalized model. For example, as described with reference to FIG. 6A, the graph structure generated via graph generator 604 may be provided to network pruner 606 as, for example, meta-information. Using full model 610, network pruner 606 performs a binary decision to remove channels of full model 610 according to the determined graph to generate a smaller personalized model (shown as pruned model 612).

[0071] Implementation examples are provided in the numbered clauses below. 1. receiving one or more personal data samples; generating a prototype of personal identifying information based on the personal data sample, the prototype of personal identifying information being trained to reflect personal characteristics of the user; generating a network graph based on the prototype personal identification information; pruning one or more channels of the global network based on the network graph to generate a personalized model; 10. A computer-implemented method comprising: 2. The computer-implemented method of clause 1, wherein the network graph is generated based on the task and target size of the personalization model. 3. The computer-implemented method of clause 1 or 2, wherein the personally identifiable information prototype is learned by minimizing the distance between features of the personal data sample and the personally identifiable information prototype corresponding to the user. 4. receiving input via a personalization model; generating inferences via a personalization network based on the input; 4. The computer-implemented method of any of clauses 1-3, further comprising: 5. receiving an output from a global network; fine-tuning the personalization network based on the output; 5. The computer-implemented method of any of clauses 1 to 4, further comprising: 6. The computer-implemented method of any of clauses 1 to 5, wherein the prototype of personally identifiable information is the mean vector of personal data samples within the same batch. 7. The computer-implemented method of any of clauses 1-6, wherein one or more channels are pruned based on a binary decision according to a prototype of personally identifiable information. 8. An apparatus comprising: Memory and and at least one processor coupled to the memory, wherein the at least one processor: receiving one or more personal data samples; generating a prototype of personal identifying information based on the personal data sample, the prototype of personal identifying information being trained to reflect personal characteristics of the user; generating a network graph based on the prototype of personal identification information; pruning one or more channels of the global network based on the network graph to generate a personalized model; An apparatus configured to: 9. The apparatus of clause 8, wherein the at least one processor is further configured to generate a network graph based on the task and target size of the personalization model. 10. The apparatus of clause 8 or 9, wherein the at least one processor is further configured to learn the personally identifiable information prototype by minimizing a distance between features of the personal data sample and the personally identifiable information prototype corresponding to the user. 11. At least one processor: receiving input via a personalization model; generating inferences via a personalization network based on the input; 11. The apparatus of any of clauses 8 to 10, further configured to: 12. At least one processor: receiving an output from a global network; Fine-tuning the personalization network based on the output 12. The apparatus of any of clauses 8 to 11, further configured to: 13. The apparatus of any of clauses 8 to 12, wherein the prototype of personal identification information is the average vector of personal data samples within the same batch. 14. The apparatus of any of clauses 8-13, wherein one or more channels are pruned based on a binary decision according to a prototype of personal identification information. 15. means for receiving one or more personal data samples; means for generating a personal identifying information prototype based on the personal data sample, the personal identifying information prototype being trained to reflect the personal characteristics of the user; means for generating a network graph based on prototypes of personal identification information; means for pruning one or more channels of the global network based on the network graph to generate a personalized model; An apparatus comprising: 16. The apparatus of clause 15, further comprising means for generating a network graph based on the task and target size of the personalization model. 17. The apparatus of clause 15 or 16, further comprising means for learning a prototype of personally identifiable information by minimizing the distance between characteristics of the personal data sample and a prototype of personally identifiable information corresponding to the user. 18. means for receiving input via a personalization model; means for generating inferences via a personalization network based on the input; 18. The device of any one of clauses 15 to 17, further comprising: 19. means for receiving an output from a global network; A means for fine-tuning the personalization network based on the output; and 19. The apparatus of any one of clauses 15 to 18, further comprising: 20. The device of any of clauses 15 to 19, wherein the prototype of personal identification information is the average vector of personal data samples within the same batch. 21. The device of any of clauses 15-20, wherein one or more channels are pruned based on a binary decision according to a prototype of personal identification information. 22. A non-transitory computer-readable medium encoding program code, the program code being executed by a processor; program code for receiving one or more personal data samples; program code for generating a prototype of personal identifying information based on the personal data sample, the prototype of personal identifying information being trained to reflect personal characteristics of the user; program code for generating a network graph based on prototypes of personal identification information; program code for pruning one or more channels of a global network based on the network graph to generate a personalized model; 1. A non-transitory computer-readable medium comprising: 23. The non-transitory computer-readable medium of clause 22, further comprising program code for generating a network graph based on the task and target size of the personalization model. 24. The non-transitory computer-readable medium of clause 22 or 23, further comprising program code for learning a prototype of personally identifiable information by minimizing the distance between features of the personal data sample and a prototype of personally identifiable information corresponding to the user. twenty five. program code for receiving input via a personalization model; program code for generating inferences via a personalization network based on input; The non-transitory computer-readable medium of any of clauses 22 to 24, further comprising: 26. program code for receiving output from the global network; Program code for fine-tuning the personalization network based on the output, and The non-transitory computer-readable medium of any of clauses 22 to 25, further comprising: 27. The non-transitory computer-readable medium of any of clauses 22 to 26, wherein the prototype of personally identifiable information is the average vector of personal data samples within the same batch. 28. The non-transitory computer-readable medium of any of clauses 22-27, wherein one or more channels are pruned based on a binary decision according to a prototype of personally identifiable information.

[0072] In one aspect, the receiving means, the means for generating a prototype of personally identifiable information, the means for generating a network graph, and / or the pruning means may be the CPU 102, a program memory associated with the CPU 102, the dedicated memory block 118, and / or the NPU 108 configured to perform the recited functions. In another configuration, the above-mentioned means may be any module or any device configured to perform the recited functions by the above-mentioned means.

[0073] The various operations of the methods described above may be performed by any suitable means capable of performing the corresponding functions. The means may include various hardware and / or software components and / or modules, including, but not limited to, circuits, application specific integrated circuits (ASICs), or processors. Generally, where there are operations illustrated in figures, those operations may have corresponding counterpart means-plus-function components that are similarly numbered.

[0074] The term "determining" as used herein encompasses a wide variety of actions. For example, "determining" may include calculating, computing, processing, deriving, investigating, looking up (e.g., looking up in a table, database, or another data structure), ascertaining, etc. Additionally, "determining" may include receiving (e.g., receiving information), accessing (e.g., accessing data in a memory), etc. Further, "determining" may include resolving, selecting, choosing, establishing, etc.

[0075] As used herein, a phrase referring to "at least one of" a list of items refers to any combination of those items, including single members. As an example, "at least one of a, b, or c" is intended to encompass a, b, c, ab, ac, bc, and abc.

[0076] The various example logic blocks, modules, and circuits described in connection with this disclosure may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device (PLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the described functions. A general-purpose processor may be a microprocessor, but alternatively, the processor may be any commercially available processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0077] The steps of a method or algorithm described in connection with the present disclosure may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in any form of storage medium known in the art. Some examples of storage media that may be used include random access memory (RAM), read-only memory (ROM), flash memory, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, CD-ROMs, etc. A software module may comprise a single instruction, or many instructions, and may be distributed across several different code segments, among different programs, and across multiple storage media. A storage medium may be coupled to a processor such that the processor can read information from, and write information to, the storage medium. Alternatively, the storage medium may be integral to the processor.

[0078] The methods disclosed herein include one or more steps or actions for achieving the described method. Method steps and / or actions may be interchanged with one another without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of specific steps and / or actions may be modified without departing from the scope of the claims.

[0079] The described functions may be implemented in hardware, software, firmware, or any combination thereof. If implemented in hardware, an exemplary hardware configuration may comprise a processing system within a device. The processing system may be implemented using a bus architecture. The bus may include any number of interconnected buses and bridges, depending on the particular application and overall design constraints of the processing system. The bus may link various circuits together, including processors, machine-readable media, and bus interfaces. The bus interface may be used to, among other things, connect a network adapter to the processing system via the bus. The network adapter may be used to implement signal processing functions. In some aspects, a user interface (e.g., keypad, display, mouse, joystick, etc.) may also be connected to the bus. The bus may also link various other circuits, such as timing sources, peripherals, voltage regulators, power management circuits, etc., which are well known in the art and therefore will not be described further.

[0080] The processor may be responsible for managing the bus and general processing, including executing software stored on a machine-readable medium. The processor may be implemented using one or more general-purpose and / or special-purpose processors. Examples include microprocessors, microcontrollers, DSP processors, and other circuitry capable of executing software. Software shall be broadly construed to mean instructions, data, or any combination thereof, whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise. The machine-readable medium may include, by way of example, random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, magnetic disk, optical disk, hard drive, or any other suitable storage medium, or any combination thereof. The machine-readable medium may be embodied in a computer program product. The computer program product may comprise packaging materials.

[0081] In a hardware implementation, the machine-readable medium may be part of a processing system separate from the processor. However, as one skilled in the art will readily appreciate, the machine-readable medium, or any portion thereof, may be external to the processing system. By way of example, the machine-readable medium may include a transmission line, a carrier wave modulated by data, and / or a computer product separate from the device, all of which may be accessed by the processor through a bus interface. Alternatively or additionally, the machine-readable medium, or any portion thereof, may be integrated into the processor, such as a cache and / or general-purpose register file. While the various components discussed may be described as having particular locations, such as local components, these components may also be configured in various ways, such as some components configured as part of a distributed computing system.

[0082] The processing system may be configured as a general-purpose processing system having one or more microprocessors providing processor functionality and external memory providing at least a portion of the machine-readable medium, all linked together with other support circuitry through an external bus architecture. Alternatively, the processing system may include one or more neuromorphic processors for implementing the neuron models and neural system models described herein. As another alternative, the processing system may be implemented using an application-specific integrated circuit (ASIC) having the processor, bus interface, user interface, support circuitry, and at least a portion of the machine-readable medium integrated into a single chip, or using one or more field programmable gate arrays (FPGAs), programmable logic devices (PLDs), controllers, state machines, gate logic, discrete hardware components, or any other suitable circuitry or combination of circuitry capable of performing the various functions described throughout this disclosure. Those skilled in the art will recognize how to best implement the described functionality for the processing system depending on the particular application and the overall design constraints imposed on the entire system.

[0083] The machine-readable medium may comprise several software modules. The software modules include instructions that, when executed by a processor, cause the processing system to perform various functions. The software modules may include a transmitting module and a receiving module. Each software module may reside in a single storage device or may be distributed across multiple storage devices. As an example, a software module may be loaded into RAM from a hard drive when a trigger event occurs. During execution of a software module, the processor may load some of the instructions into a cache to increase access speed. One or more cache lines may then be loaded into a general-purpose register file for execution by the processor. When referring hereinafter to functionality of a software module, it will be understood that such functionality is implemented by the processor when executing instructions from that software module. Furthermore, it should be appreciated that aspects of the present disclosure provide improved functionality of a processor, computer, machine, or other system implementing such aspects.

[0084] If implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media, including any medium that facilitates transfer of a computer program from one place to another. Storage media may be any available medium that can be accessed by a computer. By way of example, and not limitation, such computer-readable media may comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Additionally, any connection is properly termed a computer-readable medium. For example, if software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared (IR), radio, and microwave, the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. As used herein, disk and disc include compact discs (CDs), laser discs, optical discs, digital versatile discs (DVDs), floppy disks, and Blu-ray® discs, where disks typically reproduce data magnetically and discs reproduce data optically using lasers. Thus, in some aspects, computer-readable medium may comprise non-transitory computer-readable medium (e.g., tangible media). Additionally, in other aspects, computer-readable medium may comprise transitory computer-readable medium (e.g., a signal). Combinations of the above should also be included within the scope of computer-readable medium.

[0085] Accordingly, some aspects may comprise a computer program product for performing the presented operations. For example, such a computer program product may comprise a computer-readable medium having stored thereon (and / or encoded thereon) instructions executable by one or more processors to perform the described operations. In some aspects, the computer program product may include packaging materials.

[0086] Furthermore, it should be appreciated that modules and / or other suitable means for performing the described methods and techniques may be downloaded and / or otherwise obtained by a user terminal and / or base station, where applicable. For example, such devices may be coupled to a server to facilitate the transfer of means for performing the described methods. Alternatively, the various described methods may be provided via a storage means such that the user terminal and / or base station can obtain the various methods upon coupling or providing the storage means (e.g., RAM, ROM, physical storage medium such as a compact disc (CD) or floppy disk, etc.) to the device. Furthermore, any other suitable technique for providing the described methods and techniques to a device may be utilized.

[0087] It is to be understood that the claims are not limited to the precise configuration and components illustrated above. Various modifications, changes and variations may be made in the arrangement, operation and details of the methods and apparatus described above without departing from the scope of the claims. [Explanation of symbols]

[0088] 100 System on a Chip (SOC), SOC 102 Central Processing Unit (CPU), CPU, General Purpose Processor 104 Graphics Processing Unit (GPU), GPU 106 Digital Signal Processor (DSP), DSP 108 Neural Processing Unit (NPU), NPU 110 Connectivity Blocks 112 Multimedia Processor 114 Sensor Processor 116 Image Signal Processor (ISP), ISP 118 memory blocks, dedicated memory blocks 120 Navigation Module 200 DCN 202 Fully Connected Neural Network 204 Locally Connected Neural Networks 206 Convolutional Neural Networks 208, 210, 212, 214, 216 bond strength 218 First set of feature maps, feature maps 220 Second set of feature maps, feature maps 222 output 224 First feature vector 226 images 228 Second feature vector 230 Image Capture Device 232 Convolutional Layer 350 Deep Convolutional Network, Convolutional Neural Network, Neural Network 352 input data 354A, 354B convolution blocks 356 Convolutional Layer (CONV), Convolutional Layer, Layer 358 normalization layer (LNorm), normalization layer, layer 360 Max Pooling Layer (MAX POOL), Max Pooling Layer, Layer 362 fully connected layer, layer 364 Logistic Regression (LR) Layer, Layer 366 Classification Score 400 full model, global model 402 Personalization Model 406a, 406z nodes 408a Interconnection 500 Block Diagram 502 smartphones 504 Prototype Generator 506 Graph Generator 508 Network Pruner 600 Prototype-based Personalized Pruning (PPP) Model, PPP Model 602 Prototype Generator 604 Graph Generator 606 Network Pruner 608 Sensor 610 full model 612 pruned models 650 Prototype-based personalized pruning (PPP) model, PPP model 652 full model 654 pruned models 656 Sensor 670 Residual Blocks 672, 672a, 672b Gate Modules 674, 674a, 674b Prototype Generator 676a, 676b convolutional layers 700 methods

Claims

1. 1. A computer-implemented method comprising: receiving one or more personal data samples; generating a prototype of personal identifying information based on the personal data sample, wherein the prototype of personal identifying information is trained to reflect personal characteristics of a user; generating a network graph based on the prototype of the personal identification information; pruning one or more channels of a global network based on the network graph to generate a personalized model; receiving an output from the global network; fine-tuning the personalization model based on the output; A method comprising:

2. The method of claim 1 , wherein the network graph is generated based on a task and a target size of the personalization model.

3. The method of claim 1 , wherein the prototype of the personally identifying information is learned by minimizing the distance between features of the personal data sample and the prototype of the personally identifying information corresponding to the user.

4. receiving input via the personalization model; generating an inference via the personalization model based on the input; The method of claim 1 further comprising:

5. The method of claim 1 , wherein the prototype of the personal identifying information is a mean vector of personal data samples within the same batch.

6. The method of claim 5 , wherein the one or more channels are pruned based on a binary decision according to the prototype of the personal identifying information.

7. 1. An apparatus comprising: Memory and at least one processor coupled to the memory, receiving one or more personal data samples; generating a prototype of personal identifying information based on the personal data sample, wherein the prototype of personal identifying information is trained to reflect personal characteristics of the user; generating a network graph based on the prototype of the personal identification information; pruning one or more channels of a global network based on the network graph to generate a personalized model; receiving an output from the global network; fine-tuning the personalization model based on the output; and The apparatus is further configured to:

8. 8. The apparatus of claim 7, wherein the at least one processor is further configured to generate the network graph based on a task and a target size of the personalization model.

9. 8. The apparatus of claim 7, wherein the at least one processor is further configured to learn the prototype of the personal identifying information by minimizing a distance between features of the personal data sample and the prototype of the personal identifying information corresponding to the user.

10. the at least one processor: receiving input via the personalization model; generating an inference via the personalization model based on the input; and The apparatus of claim 7, further configured to:

11. The apparatus of claim 7 , wherein the prototype of the personal identity information is a mean vector of personal data samples within the same batch.

12. The apparatus of claim 7 , wherein the one or more channels are pruned based on a binary decision according to the prototype of the personal identification information.

13. 10. A non-transitory computer readable storage medium encoding program code, the program code being executed by a processor and comprising the method of any one of claims 1 to 6.

Images