Efficient data classification method and apparatus based on dictionary contrastive learning via adaptive label embedding

Abstract

To provide an efficient data classification method and apparatus based on dictionary contrastive learning via adaptive label embedding.SOLUTION: A data classification method that is performed by a data classification apparatus includes steps of: extracting features from input data through a learning network model; and outputting prediction results based on the features. The learning network model includes a model that compares local features derived through an individual layer other than a final layer of the learning network model with label embedding vectors corresponding to a classification label.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The embodiments disclosed herein relate to a method and apparatus for classifying input data using a learning network model based on dictionary contrast learning, and more particularly to a method and apparatus for extracting features from each layer of a learning network model to derive local features, and training a learning network model using label embeddings corresponding to each classification label and a contrastive loss function.

[0002] This research was conducted as a result of the "Artificial Intelligence Graduate School Support (Seoul National University)" project (IITP-2021-0-01343) under the Ministry of Science and ICT and the Institute for Information and Communications Technology Planning (IITP)'s ICT Broadcasting Innovation Talent Development Project. [Background technology]

[0003] Basic learning methods in deep learning include the backpropagation (BP) method, the local learning (LL) method, and the forward learning (FL) method.

[0004] First, backpropagation is a method of updating the network weights by performing a forward pass through all layers of the model to derive a final error signal from the last layer, and then performing a backward pass of this signal toward the input layer to adjust the weights. Backpropagation requires symmetry in the weights used during the forward and backward passes. Backpropagation does not begin the backward pass until the forward pass is completely completed, and vice versa. This limits computational efficiency and makes parallel processing difficult. Furthermore, weight gradient calculation requires storing the local activations of each layer, which is inefficient in terms of memory usage.

[0005] The second method, local learning, utilizes a module-wise auxiliary network in the learning network model to mitigate the limitations of backpropagation. The auxiliary network converts the local features extracted from each module into a suitable model for local loss function calculation, and also reduces unnecessary information. However, applying an auxiliary network significantly increases the number of model parameters, resulting in greater memory consumption than the forward learning method.

[0006] The third method, forward learning, uses gradient descent to train the parameters of each layer based on the local error signal of each layer without backpropagation. Because forward learning does not use an auxiliary network, the main challenge is the process of converting local features to be suitable for loss function calculation. Due to the absence of an auxiliary network, performance is lower than that of backpropagation and local learning methods. While this method has the potential to significantly improve computational efficiency, it is necessary to ensure effective conversion of local features and accuracy of training.

[0007] Therefore, there is a need for a model learning method that overcomes the limitations of the backpropagation method and has the advantages of forward learning and local learning.

[0008] For reference, Patent Document 1 is an invention related to a synthetic noise image generation method and device, Patent Document 2 is an invention related to an artificial neural network model training method and system, and Patent Document 3 is an invention related to an artificial neural network training method and an electronic device supporting the same. Patent Documents 1 to 3 only disclose general content for training an artificial neural network, but do not provide a network model training technology that combines the advantages of forward learning and local learning. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] Korean Patent Publication No. 10-2023-0151863 [Patent Document 2] Korean Patent Registration No. 10-2505946 [Patent Document 3] Korean Patent Publication No. 10-2022-0049759 [Non-patent literature]

[0010] [Non-Patent Document 1] Paper Priyank Pathak, et al. "Local Learning on Transformers via Feature Reconstruction". 29 Dec 2022. [Non-patent document 2] Paper Yulin Wang, et al. "Revisiting Locally Supervised Learning: an Alternative to End-to-end Training". 26 Jan 2021. [Non-patent document 3] Paper "Training Neural Networks with Local Error Signals". 7 May 2019. Summary of the Invention [Problem to be solved by the invention]

[0011] The embodiment disclosed in this specification aims to significantly reduce memory consumption while achieving training performance equal to or better than that of the backpropagation method by training a network model based on dictionary contrastive learning using adaptive label embedding.

[0012] Other objects and advantages of the present invention will become apparent from the following description and examples, and it will be readily apparent that the objects and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims. [Means for solving the problem]

[0013] As a technical means for achieving the above-mentioned technical object, a data classification method executed by a data classification device includes a step of extracting features from input data via a learning network model and outputting a prediction result based on the features, and the learning network model may include a model that compares local features derived via an individual layer that is not the final layer of the learning network model with a label embedding vector corresponding to a classification label.

[0014] According to another embodiment, a data classification device includes a memory for storing a training network model having multiple layers; and a control unit for extracting features from input data through the training network model and outputting a prediction result based on the features, wherein the training network model may include a model for comparing local features derived through individual layers other than the final layer of the training network model with label embedding vectors corresponding to classification labels.

[0015] According to yet another embodiment, the recording medium is a computer-readable recording medium having a program recorded thereon for executing the data classification method.

[0016] According to yet another embodiment, the computer program is a computer program stored on a recording medium for being executed by a data classifying device and for performing the data classifying method. [Effects of the Invention]

[0017] According to any one of the above-mentioned means for solving the problem, a data classification method and apparatus can be provided that can improve classification performance to the same level as or better than the backpropagation method while minimizing the number of model parameters and memory consumption by training a network model based on dictionary comparison learning while directly comparing local features derived from individual layers with adaptive label embedding vectors.

[0018] The effects obtained by the disclosed embodiments are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art to which the disclosed embodiments pertain from the following description. [Brief explanation of the drawings]

[0019] The accompanying drawings below illustrate preferred embodiments disclosed in this specification and serve to further understand the technical ideas disclosed in this specification, along with specific content for implementing the invention, and therefore the content disclosed in this specification should not be interpreted as being limited solely to the matters depicted in the drawings.

[0020] [Figure 1] FIG. 10 is a diagram illustrating a data flow of the backpropagation method. [Figure 2] FIG. 1 illustrates an example data flow for a local learning method. [Figure 3] FIG. 1 illustrates an example data flow of a forward learning method. [Figure 4]FIG. 2 is a block diagram illustrating a functional configuration of a data classification device according to an embodiment. [Figure 5] FIG. 2 is a diagram illustrating a data flow of a learning network model processed by a data classifier according to an embodiment. [Figure 6] 1 is a flowchart illustrating basic and learning operations of a data classification method according to one embodiment. [Figure 7] 1 is a flowchart illustrating basic and learning operations of a data classification method according to one embodiment. [Figure 8] FIG. 10 is a diagram illustrating simulated learning performance according to an embodiment. [Figure 9] FIG. 10 is a diagram illustrating simulated learning performance according to an embodiment. [Figure 10] FIG. 10 is a diagram illustrating simulated learning performance according to an embodiment. [Figure 11] FIG. 10 is a diagram illustrating simulated learning performance according to an embodiment. [Figure 12] FIG. 10 is a diagram illustrating simulated learning performance according to an embodiment. [Figure 13] FIG. 10 is a diagram illustrating simulated learning performance according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0021] Various embodiments will be described in detail below with reference to the accompanying drawings. The embodiments described below may be implemented in various modified forms. In order to more clearly describe the features of the embodiments, detailed descriptions of matters that are well known to those skilled in the art to which the following embodiments pertain will be omitted. In addition, parts of the drawings that are not relevant to the description of the embodiments will be omitted, and similar parts will be designated by similar reference numerals throughout the specification.

[0022] Throughout the specification, when a certain component is said to be "connected" to another component, this includes not only "directly connected" but also "connected via another component in between." Furthermore, when a certain component is said to "include" another component, this does not exclude the other component, but means that the other component may also be included, unless otherwise specified.

[0023] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings.

[0024] FIG. 1 is a diagram illustrating a data flow in the backpropagation method, FIG. 2 is a diagram illustrating a data flow in the local learning method, and FIG. 3 is a diagram illustrating a data flow in the forward learning method.

[0025] JPEG2025168206000002.jpg20150

[0026] A "network model" is a model that can detect features from input data and classify the input data based on the features. Various types of deep learning network models can be applied as needed. Among the various deep learning network models, convolutional neural network models are mainly used when processing image and video data. Convolutional neural network models are also called CNNs (Convolutional Neural Networks) and can be used as image feature extraction models, image identification models, image classification models, etc.

[0027] A "feature" is an output extracted through a model layer, contains information that accurately represents an object, and is mainly used in the form of a vector. A "local feature" can refer to a feature extracted by an individual layer (e.g., an intermediate layer) that is not the final layer. When a model layer derives a local feature, it can also apply a receptive field. The deeper the layer from which a feature is extracted, the larger the receptive field of each vector contained in the feature becomes, allowing the vector to contain information from a wider range.

[0028] "Embedding" transforms data using a latent space so that a model can understand the relationships between the data, and "embedding vector" is information expressed as a vector through embedding. For example, it can be understood as a form of dimensionality reduction or data compression. "Latent space" is a distribution space of features that well represent an object, and is also called embedding space.

[0029] The network model is formed into a network structure in which multiple layers are connected, and each layer includes nodes as building blocks. The model may have parameters to be learned, and the parameters may include weights and biases.

[0030] "Weight" is a parameter that adjusts the degree to which the input influences the output at the nodes of a layer, and "bias" is a parameter that adjusts how easily the nodes of a layer activate (output as 1).

[0031] An "activation function" is a function that converts a linear input into a nonlinear output, taking into account weights and biases. Layers that output linear values ​​without applying an activation function are also possible.

[0032] "Supervised learning" is a method of training a model using input data that has labels indicating the correct answer for the data.

[0033] "Label" means each class assigned to data, "Class" means the group to which data belongs in a dataset, "Correct label" means the actual label treated as the correct answer, and "Predicted label" means the label inferred by the model.

[0034] An "error signal" refers to the difference between a model's predicted value and its actual value. The error signal is primarily calculated by a loss function and can be propagated through the connections between layers. Because node operations depend on the output of the previous node, a backpropagation method can be used to resolve the complexity of gradient operations, which are the rate of change of the error signal.

[0035] Referring to FIG. 1, the backpropagation method performs a forward pass through all layers of a model to derive a final error signal from the last layer to update the network weights. Then, this signal is passed backward toward the input layer to adjust the weights. Backpropagation requires symmetry in the weights used during the forward and backward passes. This means that the same weights are used in the forward and backward passes. However, such symmetry in weights is considered biologically unrealistic. In fact, biological neural networks, such as the human brain, do not use the same paths and weights for forward and backward signal transmission. Therefore, the symmetry in weights applied to the backpropagation method makes it difficult to accurately mimic the learning mechanism of an actual brain.

[0036] The backpropagation method has two problems: forward locking, which allows the backward pass to start when the forward pass is completely completed, and backward locking, which is the opposite case. This limits computational efficiency and makes parallel processing difficult.

[0037] Referring to FIG. 2, the local learning method utilizes an auxiliary network in a learning network model to alleviate the limitations of backpropagation. The learning network model is composed of multiple modules or layers. Here, a module refers to a unit composed of one or more layers. The auxiliary network also converts local features extracted from each module so that they are suitable for calculating a local loss function and reduces unnecessary information. In the local learning method, learning is performed by backpropagating local error signals on a module-by-module basis based on the local loss function calculated through the auxiliary network. By performing a backward pass only on a module-by-module basis, memory efficiency can be improved compared to backpropagation learning.

[0038] Referring to Figure 3, the forward learning method uses gradient descent to learn the parameters of each layer based on the local error signal of each layer without backpropagation. Because the forward learning method does not use an auxiliary network, the main challenge is the process of converting local features to be suitable for loss function calculation. The absence of an auxiliary network results in lower performance compared to backpropagation and local learning methods. While this method has the potential to significantly improve computational efficiency, it is necessary to ensure effective conversion of local features and accuracy of learning.

[0039] This embodiment aims to train a model that combines the advantages of forward learning and local learning while overcoming the limitations of the backpropagation method. By directly comparing local features derived from individual layers with adaptive label embedding vectors and training a network model based on dictionary comparison learning, classification performance can be improved to the same level as or even better than that of the backpropagation method while minimizing the number of model parameters and memory consumption.

[0040] The algorithm for forward learning based on dictionary contrastive learning according to this embodiment can be called DCL (Dictionary Contrastive Learning).

[0041] FIG. 4 is a block diagram illustrating the functional configuration of a data classification device according to an embodiment.

[0042] Referring to FIG. 4, a data classifying apparatus 100 according to an embodiment may include an input / output unit 110, a memory 120, a control unit 130, and a communication unit 140.

[0043] The input / output unit 110 may include an input unit for receiving an input from a user and an output unit for displaying information such as a result of execution of a task or a status of the data classifying device 100. That is, the input / output unit 110 is configured to receive input data and output a result of arithmetic processing performed on the input data. The data classifying device 100 according to the embodiment may receive a request for learning or inference through the input / output unit 110.

[0044] The input / output unit 110 can provide a user interface for inputting data to be classified or inputting a training network model, and can provide a user interface for outputting features or labels inferred by the training network model.

[0045] The memory 120 is configured to store files and programs and may be configured with various types of memory. In particular, the memory 120 may store data and programs that enable the control unit 130 (described later) to perform operations for model learning and data classification according to algorithms presented below.

[0046] The memory 120 can store a training network model having multiple layers, input data (e.g., images or videos) to be input to the training network, and features or prediction results output from the training network model.

[0047] The control unit 130 includes at least one processor such as a CPU, a GPU, etc., and can control the overall operation of the data classifying apparatus 100. That is, the control unit 130 can control other components included in the data classifying apparatus 100 to perform operations for model learning and data classification. The control unit 130 can execute a program stored in the memory 120 to perform operations for model learning and data classification according to the algorithms presented below.

[0048] The communication unit 140 may perform wired or wireless communication with other devices or networks. For example, if a specific device that collects or processes input data is implemented as a separate device, the communication unit 140 may receive input data through communication and provide a result of inference based on the input data to another device or a user terminal.

[0049] For this purpose, the communication unit 140 may include a communication module supporting at least one of various wired and wireless communication methods, and the communication module may be implemented in the form of a chipset. The mobile communication or wireless communication supported by the communication unit 140 may be, for example, an N-generation mobile communication protocol, WiFi (Wireless Fidelity), Wi-Fi Direct, Bluetooth, UWB (Ultra-Wide Band), or NFC (Near Field Communication).

[0050] The control unit 130 can extract features from input data via a learning network model and output a prediction result based on the features.

[0051] The control unit 130 can extract features from each layer of the training network model. The control unit 130 can derive local features through individual layers that are not the final layer of the training network model and compare label embedding vectors corresponding to classification labels with the local features.

[0052] The control unit 130 can remove the dependency on the operation graph used for gradient calculation so that the operation values ​​processed by at least one layer of the learning network model are not tracked, and can be configured so that the error signal of the local feature derived from at least one layer is not propagated in the direction of the previous layer.

[0053] The control unit 130 may directly or indirectly connect the label embedding dictionary, which maps the label embedding vectors, to at least one layer of the network model, and directly compare the label embedding vectors and local features using the label embedding dictionary.

[0054] The control unit 130 can adaptively and dynamically update the label embedding vectors of the label embedding dictionary based on the error signals of the local features.

[0055] The control unit 130 may configure a path such that at least one layer of the learning network model receives an error signal of a local feature from a loss function set based on dictionary contrastive learning.

[0056] The control unit 130 can update the parameters of the learning network model so as to maximize the similarity between the label embedding vectors and local features corresponding to the local features in the label embedding dictionary through the dictionary contrast loss function, while minimizing the similarity between the label embedding vectors and local features that do not correspond to the local features in the label embedding dictionary.

[0057] The control unit 130 calculates a final error signal for the final layer of the training network model through a final loss function, and separates the backpropagation path between the previous layer and the final layer to prevent the final error signal from being propagated to the intermediate layer of the training network model.

[0058] FIG. 5 is a diagram illustrating a data flow of a learning network model processed by a data classifier according to an embodiment.

[0059] Contrastive learning is a powerful tool for representation learning and can be used in local learning and forward learning. The local contrast loss function (Lcontrast) for the local output h∈RC×H×W arrangement of the forward pass layer used in local learning can be defined as shown in Equation 1.

[0060]

number

[0061] where τ is the temperature hyperparameter that controls the probability distribution, y∈{1, ..., Z} is the ground truth label, and fφ is the auxiliary network. ai and aj are positive features. The objective of the local contrastive loss function is to maximize the similarity between positive features and minimize the similarity between negative features.

[0062] The loss function Lfeat denotes the local contrast loss function (Lcontrast), where fφ(h)=h.

[0063] The auxiliary network applied to local contrastive learning has a significant impact on performance, and it is necessary to verify the function of the auxiliary network in order to improve the performance of the forward learning method that uses contrastive learning even without the auxiliary network.

[0064] The performance difference between Lcontrast with an auxiliary network and Lfeat without an auxiliary network is the result of the presence of mutual information I(h,r) (see 910 in Figure 9), where r is a nuisance and x is a task-irrelevant variable. Given a task-related variable y, mutual information I indicates the amount of information gained about one random variable by observing other variables, so I(r,y) = 0.

[0065] Lfeat maximizes the similarity between local features (h + Thp) rather than the similarity between local feature h and label y. Therefore, maximizing the similarity between local features increases I(r + , rp), which can lead the model to mistakenly consider task-irrelevant information as meaningful features.

[0066] In this respect, the auxiliary network functions to filter out the disturbing factor r and reduce its influence in local learning (LL). However, in forward learning (FL), where the auxiliary network cannot be used, r has a more adverse effect.

[0067] This embodiment can directly maximize the similarity between the local feature h and the embedding vector corresponding to the target label to solve the problem of distractor r in forward learning (FL).

[0068] In this embodiment, a command (e.g., detach() function) that detaches all inputs before starting forward propagation can be used to prevent a layer from propagating an error signal backward. That is, the inputs of the layer or the operation values ​​of the layer are separated from the operation graph to prevent the error signal from propagating backward. A path can be formed such that the final layer receives an error signal from a final loss function (e.g., cross entropy) for final linear classification, and other layers other than the final layer receive an error signal from a dictionary-based loss function. The dictionary-based loss function optimizes the similarity between local features per layer and label embedding vectors.

[0069] This embodiment can map labels to embedding vectors.

[0070] JPEG2025168206000004.jpg44170

[0071] In this embodiment, the label embedding vector can be initialized in two main ways.

[0072] According to the first initialization method, Z embedding vectors can be initialized to be orthogonal to each other. The label embedding dictionary D⊥ can include orthogonal vectors.

[0073] According to the second initialization method, each element of the embedding vector can be sampled and initialized from a standard normal distribution. The label embedding dictionary DN can include a standard normal probability vector.

[0074] After initialization, the embedding vectors can be scaled by matching the embedding vector norms.

[0075] This embodiment can adaptively update the label embedding.

[0076] The label embeddings in the label embedding dictionary are dynamic concepts that are updated at each iteration. The basic method (DCL) updates the label embeddings through the forward pass of each hidden layer. Using the DCL method, the label embeddings can be updated using layer-specific gradients averaged across all hidden layers. However, the averaging process, which simultaneously integrates the error signals of all layers, can have a negative impact on the weight update.

[0077] JPEG2025168206000005.jpg26150

[0078] This embodiment can optimize the similarity of the label embedding vector and the local features.

[0079] JPEG2025168206000006.jpg23150

[0080] JPEG2025168206000007.jpg18150

[0081] JPEG2025168206000008.jpg39150

[0082] To prevent backpropagation between layers, we can use the Stop Gradient Operator sg[·], which prevents the gradient from passing through a certain part of the function, and set hl=fl(sg[hl-1]).

[0083] This embodiment can perform training using a dictionary-based loss function.

[0084] The weights of the final prediction layer fL can be updated using a final loss function, for example, the cross-entropy loss function used in backpropagation training in existing classification tasks.

[0085] JPEG2025168206000009.jpg20150

[0086]

number

[0087] JPEG2025168206000011.jpg15150

[0088] JPEG2025168206000012.jpg20150

[0089] JPEG2025168206000013.jpg13170

[0090] The label embedding vector t is adaptive, updating the weights according to the error signal of Ldict.

[0091] The efficiency of Ldict can vary depending on the number of classes. Compared to using static label embedding vectors, the performance tends to become increasingly significant as the number of label classes Z increases, but Ldict can achieve competitive performance even with fewer classes than existing control loss functions.

[0092] Minimizing Ldict maximizes the similarity between the local feature h and the corresponding label embedding vector t+, while minimizing the similarity with non-corresponding label embedding vectors. By utilizing these properties of Ldict, DZ can be used for inference without the final linear classifier fL. A prediction can be generated by selecting the target label that is most similar to the feature vector.

[0093]

number

[0094] Therefore, prediction is possible at all layers. This also allows the weighted sum of layer-specific predictions to serve as global prediction. This approach goes beyond fL-only prediction.

[0095] 6 and 7 are flowcharts illustrating the basic and learning operations of a data classification method according to one embodiment.

[0096] The data classification method according to the embodiment shown in Figures 6 and 7 includes steps that are processed in time series by the data classifier shown in Figures 4 and 5. Although omitted below, the above description of the data classifier shown in Figures 4 and 5 can also be applied to the data classification method according to the embodiment shown in Figures 6 and 7.

[0097] 6, in step S610, the data classifier may extract features from input data through a training network model. In step S620, the data classifier may output a prediction result based on the features. Here, the training network model may be a model trained according to the procedure of FIG. 7.

[0098] 7, the data classifier may separate backpropagation paths between layers in step S710. In step S710, the data classifier may remove dependency on the operation graph used for gradient calculation so that operation values ​​processed by at least one layer of the learning network model are not tracked, and may set up so that error signals of local features derived from at least one layer are not propagated toward previous layers.

[0099] In step S720, the data classifier may generate a label embedding dictionary that is directly or indirectly connected to at least one layer of the training network model. In step S720, the data classifier may map label embedding vectors to the label embedding dictionary. In step S720, the data classifier may initialize label embedding vectors in the label embedding dictionary.

[0100] In step S730, the data classifier can derive local features from individual layers of the training network model and directly compare the label embedding vectors corresponding to the classification labels with the local features.

[0101] In step S730, the data classifier may allow at least one layer of the learning network model to receive local feature error signals from a dictionary-based loss function (Ldict).

[0102] In operation S730, the data classifier may update parameters of some layers based on the dictionary contrast loss function so as to maximize similarity between the local features and label embedding vectors corresponding to the local features in the label embedding dictionary while minimizing similarity between the local features and label embedding vectors not corresponding to the local features in the label embedding dictionary. The some layers may be at least one layer, and may be the remaining layers excluding the final layer, or some intermediate layers.

[0103] In step S740, the data classifier may adaptively and dynamically update the label embedding vectors of the label embedding dictionary based on the error signals of the local features.

[0104] In step S750, the data classifier calculates a final error signal for the final layer of the training network model based on the final loss function (Lfinal) and updates the parameters of the final layer to minimize the final error signal. Because the backpropagation paths between the previous layer and the final layer are separated, the final error signal is not propagated to intermediate layers of the training network model.

[0105] In step S760, the data classifier determines whether an iteration termination condition is satisfied. For example, the data classifier may determine whether a condition is satisfied, such as whether the number of iterations is satisfied or whether a reference value for minimizing a loss function is satisfied.

[0106] If the iteration termination condition is not met in step S760, the model parameter update step is repeated. For example, steps S730, S740, and S750 may be repeated. If the iteration termination condition is met in step S760, the learning may be terminated.

[0107] 8 to 13 are diagrams illustrating simulated learning performance according to the embodiment.

[0108] In the examples, DCL, DCL-O, and DCL-LD are forward learning algorithms based on dictionary contrast learning, while in the comparative examples, LL-cont, LL-contrec, LL-predisim, LL-bpf, LL-pred, and LL-sim are local learning algorithms. LL-cont is a local learning algorithm using Lcontrast in mathematical formula 1, LL-contrec is a local learning algorithm using Lcontrast in mathematical formula 1 and the image reconstruction loss function in non-patent document 2, and LL-predisim, LL-bpf, LL-pred, and LL-sim are local learning algorithms using non-patent document 3.

[0109] 8 shows memory usage and the number of model parameters, where Δθ indicates the increase in the number of parameters compared to the basic VGG8B model. It can be seen that the embodiment is superior in terms of memory usage and model parameters.

[0110] Figure 9 shows the incorrectly predicted test error. Reference numeral 910 shows the result of comparing Lcontrast and Lfeat, and reference numerals 920 and 930 show the result of comparing Ldict and Lfeat. Although the embodiment uses a forward learning algorithm, it can be confirmed that the performance is comparable to local learning even without using an auxiliary network.

[0111] Figure 10 shows task-irrelevant information acquired in the intermediate layer of the model. Reference numeral 1010 denotes an estimate of the mutual information between the local feature h and the input image x, reference numeral 1020 denotes an estimate of the mutual information between the local feature h and the label y, and reference numeral 1030 denotes an estimate of the mutual information between the local feature h and the distractor r. In particular, it can be confirmed that Ldict effectively reduces task-irrelevant information as the layer index increases.

[0112] Figure 11 shows the performance of different types of label embedding dictionaries. It can be seen that the dictionary DZ with adaptive embedding vectors is improved over the dictionary DN with fixed embedding probability vectors and the dictionary D⊥ with orthogonal vectors.

[0113] Figure 12 shows a saliency map corresponding to the inner product of the embedding vector and the individual local feature vector for one label. It can be seen that the predicted result of the embodiment matches the correct answer.

[0114] 13 shows the semantic properties of adaptive embeddings. It can be seen that the embodiment clearly distinguishes the semantic relationship of multiple super-labels with multiple sub-labels.

[0115] The term "module" used in the above embodiments refers to software or hardware components such as FPGAs (field programmable gate arrays) or ASICs, and the "module" performs a certain function. However, the term "module" is not limited to software or hardware. A "module" may be configured to reside on an addressable storage medium or to execute one or more processors. Thus, by way of example, "module" includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables.

[0116] The functionality provided within components and units may be combined into fewer components and units or separated into additional components and units.

[0117] Furthermore, the components and "units" may be implemented to implement one or more CPUs within a device or a secure multimedia card.

[0118] Meanwhile, the data classification method according to an embodiment described herein may also be embodied in the form of a computer-readable medium storing computer-executable instructions and data. Here, the instructions and data may be stored in the form of program code, which, when executed by a processor, may generate a predetermined program module and perform a predetermined operation. Furthermore, the computer-readable medium may be any available medium accessible by a computer, including both volatile and nonvolatile media, and both separable and non-separable media. The computer-readable medium may also be a computer recording medium. The computer recording medium may include both volatile and non-volatile, separable and non-separable media embodied in any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. For example, the computer recording medium may be a magnetic storage medium such as a hard disk drive (HDD) or solid-state drive (SSD), an optical storage medium such as a CD, DVD, or Blu-ray disc, or a memory included in a server accessible via a network.

[0119] Furthermore, the data classification method according to an embodiment described herein may be embodied as a computer program (or a computer program product) including computer-executable instructions. The computer program includes programmable machine instructions to be processed by a processor, and may be embodied in a high-level programming language, an object-oriented programming language, an assembly language, a machine language, or the like. The computer program may be recorded on any type of computer-readable recording medium (e.g., memory, a hard disk, a magnetic / optical medium, or a solid-state drive (SSD)).

[0120] Therefore, the data classification method according to one embodiment described herein can be implemented by executing the computer program as described above on a computing device. The computing device can include at least a processor, a memory, a storage device, a high-speed interface connecting to the memory and a high-speed expansion port, and a low-speed interface connecting to a low-speed bus and the storage device. Each of these components is connected to each other using various buses and can be mounted on a common motherboard or in other suitable manners.

[0121] Here, the processor may process instructions within a computing device. Such instructions may include instructions stored in a memory or storage device for displaying graphic information to provide a GUI (Graphical User Interface) on an external input and output device, such as a display connected to a high-speed interface. In other embodiments, multiple processors and / or multiple buses may be used, along with multiple memories and memory types, as appropriate. Also, the processor may be implemented as a chipset consisting of chips containing multiple independent analog and / or digital processors.

[0122] Also, memory stores information within a computing device. As an example, memory may be comprised of a volatile memory unit or collection thereof. As another example, memory may be comprised of a non-volatile memory unit or collection thereof. Memory may also be in other forms of computer-readable media, such as magnetic or optical disks.

[0123] The storage device can provide a large amount of storage space to a computing device. The storage device may be a computer-readable medium or a configuration that includes such a medium, such as a device in a Storage Area Network (SAN) or other configuration, and may be a floppy disk drive, hard disk drive, optical disk drive, tape drive, flash memory, or other similar semiconductor memory device or device array. The above-described embodiments are merely illustrative, and those skilled in the art will understand that the above-described embodiments may be easily modified into other specific forms without changing the technical ideas or essential features of the above-described embodiments. Therefore, it should be understood that the above-described embodiments are illustrative in all respects and are not limiting. For example, each component described as a single component may be implemented in a distributed form, and similarly, each component described as a distributed component may be implemented in a combined form.

[0124] The scope of protection sought by this specification is determined by the claims set forth below rather than the above detailed description, and all modifications or variations derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included within the scope of the present invention. [Explanation of symbols]

[0125] 100 Data Classification Device 110 Input / output section 120 memory 130 Control Unit 140 Communications Department

Claims

1. A data classification method executed by a data classification device, extracting features from input data through a learning network model and outputting a prediction result based on the features; A data classification method, wherein the learning network model is a model that compares local features derived through individual layers that are not the final layer of the learning network model with label embedding vectors that correspond to classification labels.

2. 2. The data classification method of claim 1, wherein the learning network model is configured to remove dependency on the operation graph used for gradient calculation so that operation values ​​processed by at least one layer of the learning network model are not tracked, and to prevent error signals of the local features derived from the at least one layer from being propagated toward the previous layer.

3. 2. The data classification method of claim 1, wherein the learning network model is a model that is connected to at least one layer of the learning network model and directly compares the label embedding vectors and the local features using a label embedding dictionary to which the label embedding vectors are mapped.

4. The method of claim 3 , wherein the label embedding vectors of the label embedding dictionary are adaptively and dynamically updated based on the local feature error signals.

5. 4. The data classification method of claim 3, wherein at least one layer of the learning network model receives the error signal of the local feature from a loss function set based on dictionary contrastive learning.

6. 6. The data classification method of claim 5, wherein the parameters of the training network model are updated to maximize the similarity between the local features and label embedding vectors corresponding to the local features in the label embedding dictionary, while minimizing the similarity between the local features and label embedding vectors that do not correspond to the local features in the label embedding dictionary.

7. 4. The data classification method of claim 3, wherein the learning network model calculates a final error signal for the final layer of the learning network model, a backpropagation path between the previous layer of the final layer and the final layer is separated, and the final error signal is not propagated to an intermediate layer of the learning network model.

8. a memory for storing a training network model having a plurality of layers; a control unit that extracts features from input data through the learning network model and outputs a prediction result based on the features; A data classification device, wherein the learning network model is a model that compares local features derived through individual layers that are not the final layer of the learning network model with label embedding vectors that correspond to classification labels.

9. A computer-readable recording medium having a program recorded thereon for executing the method of claim 1.

10. A computer program stored on a recording medium for executing the method of claim 1 when executed by a data classifier.

Images