Efficient data classification method and apparatus based on dictionary-reference learning using 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 in this specification relate to a method and apparatus for classifying input data using a learning network model based on dictionary contrastive learning. More specifically, features are extracted from each layer of the learning network model to derive local features, and the learning network model can be learned using label embeddings corresponding to each classification label and a contrastive loss function.

[0002] This research was conducted as a research result of the "Support for Artificial Intelligence Graduate School (Seoul National University)" project (IITP-2021-0-01343) of the Information and Communication Broadcasting Human Resources Development Project of the Ministry of Science and ICT and the Institute for Information & communications Technology Planning & Evaluation (IITP).

Background Art

[0003] As basic learning methods of deep learning, there are a backpropagation (BP) method, a local learning (LL) method, and a forward learning (FL) method.

[0004] Firstly, the backpropagation method updates the network weights by performing a forward pass across all layers of the model to derive the final error signal from the last layer, and then performing a backward pass of this signal towards the input layers to adjust the weights. Backpropagation requires weight symmetry in the weights used during the forward and backward passes. The backpropagation method does not start the backward pass until the forward pass is completely finished, and vice versa. This limits computational efficiency and makes parallel processing difficult. Furthermore, the gradient calculation of the weights must store the local activation of each layer, making it inefficient in terms of memory usage.

[0005] Secondly, the local learning method utilizes a module-wise auxiliary network in the learning network model to mitigate the limitations of backpropagation. The auxiliary network transforms the local features extracted from each module to be suitable for calculating the local loss function, and also serves to reduce unnecessary information. However, applying the auxiliary network significantly increases the number of parameters in the model, leading to increased memory consumption compared to the forward learning method.

[0006] Thirdly, the forward learning method learns the parameters of each layer using a sloped descent method based on the local error signals of each layer, without backpropagation. Since the forward learning method does not use an auxiliary network, the process of transforming local features to be suitable for loss function calculation becomes the main challenge, and due to the absence of an auxiliary network, its performance is lower compared to the backpropagation method and the local learning method. It has the potential to greatly improve computational efficiency, but it is necessary to ensure the effective transformation of local features and the accuracy of learning.

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

[0008] For reference, Patent Document 1 relates to a method and apparatus for generating synthesized noise images, Patent Document 2 relates to an artificial neural network model learning method and system, and Patent Document 3 relates to an artificial neural network learning method and an electronic device supporting it. Patent Documents 1 to 3 only disclose general information on learning artificial neural networks and do not provide a network model learning technique that combines the advantages of forward learning and local learning. [Prior art documents] [Patent Documents]

[0009] [Patent Document 1] Korean Published Patent Publication No. 10-2023-0151863 [Patent Document 2] Korean Registered Patent Publication No. 10-2505946 [Patent Document 3] Korean Published Patent 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. [Overview of the Initiative] [Problems that the invention aims to solve]

[0011] The embodiments disclosed herein aim to achieve learning performance equal to or better than that of backpropagation while significantly reducing memory consumption by training a network model based on dictionary-reference learning using adaptive label embedding.

[0012] Other objectives and advantages of the present invention can be understood from the following description and will be more clearly evident from one embodiment. Furthermore, it will be readily apparent that the objectives and advantages of the present invention can be realized by the means and combinations thereof described in the claims. [Means for solving the problem]

[0013] As a technical means to achieve the technical challenges described above, the data classification method performed by the data classification device includes the steps of extracting features from input data via a learning network model and outputting prediction results based on the features, wherein the learning network model may include a model that compares local features derived via individual layers other than the final layer of the learning network model with label embedding vectors corresponding to classification labels.

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

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

[0016] According to still another embodiment, the computer program is a computer program that is executed by a data classification device and is stored on a recording medium for executing the data classification method.

Advantages of the Invention

[0017] According to any one of the above-described problem-solving means, by learning a network model based on dictionary comparison learning while directly comparing the local features derived from individual layers with the adaptive label embedding vectors, it is possible to present a data classification method and apparatus capable of improving the classification performance to be equal to or higher than that of the backpropagation method while minimizing the number of model parameters and memory consumption.

[0018] The effects obtained in the disclosed embodiments are not limited to the effects mentioned above, and other effects not mentioned will be clearly understandable to those skilled in the art in the technical field to which the disclosed embodiments belong from the following description.

Brief Description of the Drawings

[0019] Hereinafter, the accompanying drawings illustrate preferred embodiments disclosed in this specification and serve to better understand the technical idea disclosed in this specification together with the specific content for implementing the invention. Therefore, the content disclosed in this specification should not be construed as being limited only to the matters described in the drawings.

[0020] [Figure 1] It is a diagram illustrating the data flow of the backpropagation method. [Figure 2] It is a diagram illustrating the data flow of the local learning method. [Figure 3] It is a diagram illustrating the data flow of the forward learning method. [Figure 4]This is a block diagram illustrating the functional configuration of a data classification device according to one embodiment. [Figure 5] This figure illustrates the data flow of a learning network model processed by a data classification device according to one embodiment. [Figure 6] This flowchart illustrates the basic operation and learning process of a data classification method according to one embodiment. [Figure 7] This flowchart illustrates the basic operation and learning process of a data classification method according to one embodiment. [Figure 8] This figure illustrates the learning performance simulated by the example. [Figure 9] This figure illustrates the learning performance simulated by the example. [Figure 10] This figure illustrates the learning performance simulated by the example. [Figure 11] This figure illustrates the learning performance simulated by the example. [Figure 12] This figure illustrates the learning performance simulated by the example. [Figure 13] This figure illustrates the learning performance simulated by the example. [Modes for carrying out the invention]

[0021] The following describes various embodiments in detail based on the attached drawings. The embodiments described below can be modified and implemented in various forms. In order to more clearly describe the characteristics of the embodiments, detailed explanations of matters that are commonly known to those who have ordinary skill in the art to which the embodiments belong are omitted. Furthermore, parts of the drawings that are not relevant to the description of the embodiments are omitted, and similar parts are denoted by similar reference numerals throughout the specification.

[0022] Throughout the specification, when we say that one configuration is "linked" to another, this includes not only cases where they are "directly linked," but also cases where they are "linked with another configuration in between." Furthermore, when we say that one configuration "includes" another, unless otherwise stated, this does not mean that it excludes other configurations, but rather that it may include even more configurations.

[0023] The following describes an embodiment in detail based on the attached drawings.

[0024] Figure 1 illustrates the data flow of the backpropagation method, Figure 2 illustrates the data flow of the local learning method, and Figure 3 illustrates the data flow of the forward learning method.

[0025] JPEG0007853392000001.jpg20150

[0026] A "network model" is a model that can detect features from input data and classify the input data based on those 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 primarily used when processing image and video data. Convolutional neural network models, also known as CNNs (Convolutional Neural Networks), can be used as image feature extraction models, image recognition models, and image classification models.

[0027] A "feature" is an output extracted by passing through the layers of a model, containing information that accurately represents the subject, and is primarily used in vector form. A "local feature" can refer to a feature extracted by an individual layer other than the final layer (e.g., an intermediate layer). When a model layer derives a local feature, it may also apply concepts such as a receptive field. The deeper the layer from which the 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 of regions.

[0028] "Embedding" is the process of transforming data using a latent space so that a model can understand the relationships between the data, and "embedding vectors" are the information represented as vectors 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 the subject, and is also called the embedding space.

[0029] A network model is formed as a network structure in which multiple layers are connected, and each layer contains nodes, which are constituent units. A model can have parameters to be learned, and these parameters can include weights and biases.

[0030] "Weight" is a parameter that adjusts the degree to which the input influences the output in a layer node, while "bias" is a parameter that adjusts how easily a layer node is activated (outputs as 1).

[0031] An "activation function" is a function that transforms a linear input, taking into account weights and biases, into a nonlinear output. 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 answers for each data point.

[0033] "Label" refers to the class assigned to each data point, while "Class" refers to the group to which the data belongs within the dataset. "Ground truth label" refers to the actual label treated as the correct answer, while "predicted label" refers to the label inferred by the model.

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

[0035] Referring to Figure 1, the backpropagation method updates the network's weights by performing a forward pass across all layers of the model to derive the final error signal from the last layer, and then performing a backward pass of this signal back towards the input layers to adjust the weights. Backpropagation requires weight symmetry in the weights used during the forward and backward passes. This means that the same weights are used in both the forward and backward passes. However, such weight symmetry seems to be a biologically unrealistic element. In reality, biological neural networks like the human brain do not use the same paths and weights for forward and backward signal transmission. Therefore, the weight symmetry applied to the backpropagation method makes it difficult to accurately mimic the learning mechanisms of the actual brain.

[0036] The backpropagation scheme involves forward locking, where the backward path can only begin when the forward path has completely finished, and backward locking, the opposite of forward locking. This limits computational efficiency and makes parallel processing difficult.

[0037] Referring to Figure 2, the local learning method utilizes an auxiliary network in the learning network model to mitigate the limitations of backpropagation. The learning network model consists of numerous modules or layers. Here, a module refers to a unit consisting of one or more layers. The auxiliary network transforms the local features extracted from each module to be suitable for calculating the local loss function and also serves to reduce unnecessary information. In the local learning method, learning is performed by backpropagating the local error signal on a module-by-module basis based on the local loss function calculated via the auxiliary network, and by executing the 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 learns the parameters of each layer using a sloped descent method based on the local error signals of each layer, without backpropagation. Since the forward learning method does not use an auxiliary network, the process of transforming local features to be suitable for loss function calculation becomes the main challenge, and its performance is lower compared to the backpropagation method and the local learning method due to the absence of an auxiliary network. Although it has the potential to greatly improve computational efficiency, it is necessary to ensure the effective transformation of local features and the accuracy of learning.

[0039] This embodiment aims to achieve model learning 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 the network model based on dictionary-reference learning, it is possible to improve classification performance to the same level as or better than the backpropagation method while minimizing the number of model parameters and memory consumption.

[0040] The algorithm for forward learning based on dictionary contrast learning in this example can be called DCL (Dictionary Contrastive Learning).

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

[0042] Referring to Figure 4, a data classification device 100 according to one 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 input from a user and an output unit for displaying information such as the result of an operation or the status of the data classification device 100. In other words, the input / output unit 110 is configured to receive data input and output the result of processing it. The data classification device 100 according to this embodiment can receive requests for learning or inference, etc., via the input / output unit 110.

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

[0045] Memory 120 is configured to store files and programs and can be composed of various types of memory. In particular, memory 120 can store data and programs that enable the control unit 130, described later, to perform calculations for model learning and data classification using the algorithms presented below.

[0046] Memory 120 can store a learning network model having multiple layers. Memory 120 can store input data (e.g., images or videos) that is input to the learning network. Memory 120 can also store features or prediction results output from the learning network model.

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

[0048] The communication unit 140 can perform wired wireless communication with other devices or networks. For example, if a specific device that collects or processes input data is embodied as a separate device, the communication unit 140 can receive input data by communication and provide the results inferred based on the input data to other devices or the user's terminal.

[0049] For this purpose, the communication unit 140 may include a communication module that supports at least one of a variety of wired and wireless communication methods, and the communication module may be embodied in the form of a chipset. The mobile or wireless communication supported by the communication unit 140 may be, for example, N-generation mobile communication protocols, 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 the input data via a learning network model and output prediction results based on those features.

[0051] The control unit 130 can extract features from each layer of the learning network model. The control unit 130 can derive local features through individual layers other than the final layer of the learning 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 computation graph used for gradient calculation so that computation values ​​processed by at least one layer of the learning network model are not tracked, and can configure the system so that error signals of local features derived from at least one layer are not propagated back to previous layers.

[0053] The control unit 130 directly or indirectly connects a label embedding dictionary, which maps label embedding vectors, to at least one layer of the network model, and can directly compare 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 error signals of local features.

[0055] The control unit 130 can form a path so that at least one layer of the learning network model receives error signals for local features from a loss function set based on dictionary contrast learning.

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

[0057] The control unit 130 calculates the final error signal for the final layer of the learning network model via the final loss function, and separates the backpropagation paths between the final layer and the previous layers, thereby preventing the final error signal from propagating to the intermediate layers of the learning network model.

[0058] Figure 5 illustrates the data flow of a learning network model processed by a data classification device according to one embodiment.

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

[0060]

number

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

[0062] The loss function Lfeat represents a loss function where the local contrast loss function (Lcontrast) is fφ(h)=h.

[0063] The auxiliary networks applied to local controlled learning have a significant impact on performance, and it is necessary to examine the functionality of these auxiliary networks in order to improve the performance of forward learning methods that use controlled learning even without them.

[0064] The performance difference between Lcontrast, which uses an auxiliary network, and Lfeat, which does not, is a result of the existence of mutual information I(h,r) (see 910 in Figure 9). Here, r is a nuisance and x represents a variable unrelated to the task. Given a task-related variable y, the mutual information I represents the amount of information obtained for 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 information irrelevant to the task as meaningful features.

[0066] In this respect, auxiliary networks serve to filter out disruptive elements r, thereby reducing their impact in local learning (LL). However, in forward learning (FL), where auxiliary networks cannot be used, r has a more detrimental effect.

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

[0068] This embodiment can configure layers to prevent error signals from propagating backward by using an instruction (e.g., the detach() function) that isolates all inputs before forward propagation begins. That is, error signals are prevented from propagating backward by isolating the layer's inputs or the layer's calculated values ​​from the calculation graph. The final layer can be routed to receive the error signal from the final loss function for final linear classification (e.g., cross-entropy), while other layers that are not the final layer receive the error signal from the lexicographical loss function. The lexicographical loss function optimizes the similarity between layer-specific local characteristics and label embedding vectors.

[0069] This embodiment allows for mapping labels to embedding vectors.

[0070] JPEG0007853392000003.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 contain orthogonal vectors.

[0073] According to the second initialization method, each element of the embedding vector can be initialized by sampling it using a standard normal distribution. The label embedding dictionary DN can contain standard normal probability vectors.

[0074] After initialization, the embedding vector can be scaled by matching it with the embedding vector norm.

[0075] This embodiment allows for adaptive updating of label embedding.

[0076] Label embedding in a label embedding dictionary is a dynamic concept that is updated in each iterative stage. The basic method (DCL) updates label embedding through a forward pass across each intermediate layer. Using the DCL method, label embedding can be updated by layer-specific gradients averaged across all intermediate layers. However, the averaging process, which simultaneously integrates error signals from all layers, can negatively impact weighted value updates.

[0077] JPEG0007853392000004.jpg26150

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

[0079] JPEG0007853392000005.jpg23150

[0080] JPEG0007853392000006.jpg18150

[0081] JPEG0007853392000007.jpg39150

[0082] To prevent backpropagation between layers, a Stop Gradient Operator (Sg[·]) can be used to prevent the gradient from passing through a specific part of the function, such as hl=fl(sg[hl-1]).

[0083] This embodiment allows training to be performed using a lexicographical-controlled loss function.

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

[0085] JPEG0007853392000008.jpg20150

[0086]

number

[0087] JPEG0007853392000010.jpg15150

[0088] JPEG0007853392000011.jpg20150

[0089] JPEG0007853392000012.jpg13170

[0090] The label embedding vector t is adaptive, updating its weights based on Ldict's error signals.

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

[0092] Minimizing Ldict maximizes the similarity between local features h and the corresponding label embedding vector t+, while minimizing similarity with non-corresponding label embedding vectors. Leveraging this attribute of Ldict allows DZ to be used in inference without a final linear classifier fL. Predictions can then be generated by selecting the target label with the highest similarity to the feature vector.

[0093]

number

[0094] Therefore, prediction is possible at all layers. Furthermore, this allows the weighted sum of layer-specific predictions to function as a global prediction. Such an approach can surpass predictions made at the fL level alone.

[0095] Figures 6 and 7 are flowcharts illustrating the basic operation and learning process 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 a stage in which the data is processed chronologically by the data classification device shown in Figures 4 and 5. Even if some details are omitted below, the information described above regarding the data classification device 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] Referring to Figure 6, in step S610, the data classification device can extract features from the input data via the learned network model. In step S620, the data classification device can output prediction results based on the features. Here, the learned network model can be a model trained according to the procedure in Figure 7.

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

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

[0100] At stage S730, the data classification device can derive local features from individual layers of the learning network model and directly compare the label embedding vectors corresponding to the classification labels with the local features.

[0101] At stage S730, the data classification device can ensure that at least one layer of the learning network model receives an error signal for local features from the lexicographical control loss function (Ldict).

[0102] At stage S730, the data classification system can update the parameters of some layers based on a dictionary-controlled loss function to maximize the similarity between label embedding vectors corresponding to local features in the label embedding dictionary and the local features, while minimizing the similarity between label embedding vectors that do not correspond to local features in the label embedding dictionary and the local features. Some layers are at least one layer, and may be the remaining layers excluding the final layer, or some intermediate layers.

[0103] At stage S740, the data classification device can adaptively and dynamically update the label embedding vectors in the label embedding dictionary based on error signals from local features.

[0104] At step S750, the data classification device can calculate the final error signal for the final layer of the learning network model based on the final loss function (Lfinal) and update the parameters of the final layer to minimize the final error signal. Since the backpropagation paths between the previous layers and the final layer are separated, the final error signal is not propagated to the intermediate layers of the learning network model.

[0105] At stage S760, the data classification device determines whether the iteration termination conditions are met. For example, it can determine whether the number of iterations is met or whether the criteria for minimizing the loss function are met.

[0106] If the iterative termination condition is not met at stage S760, the process of updating the model parameters is repeated. For example, steps S730, S740, and S750 can be repeated. If the iterative termination condition is met at stage S760, the training can be terminated.

[0107] Figures 8 to 13 illustrate the simulated learning performance obtained by the embodiment.

[0108] Examples DCL, DCL-O, and DCL-LD are forward learning algorithms based on dictionary-reference learning, while 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 from mathematical formula 1, LL-contrec is a local learning algorithm using Lcontrast from mathematical formula 1 and the image reconstruction loss function from Non-Patent Literature 2, and LL-predisim, LL-bpf, LL-pred, and LL-sim are local learning algorithms using Non-Patent Literature 3.

[0109] Figure 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 errors. Figure reference numeral 910 represents the result of comparing Lcontrast and Lfeat, and Figure reference numerals 920 and 930 represent the result of comparing Ldict and Lfeat. Although the embodiment is a forward learning algorithm, it can be confirmed that it has performance comparable to local learning even without using an auxiliary network.

[0111] Figure 10 shows task-irrelevant information acquired in the intermediate layers of the model. Figure 1010 is an estimate of the mutual information between local feature h and input image x, figure 1020 is an estimate of the mutual information between local feature h and label y, and figure 1030 is an estimate of the mutual information between local feature h and interfering element r. In particular, it can be seen 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 dictionary DZ, which has adaptive embedding vectors, performs better than dictionary DN, which has fixed-type embedding with probability vectors, and dictionary D⊥, which has orthogonal vectors.

[0113] Figure 12 shows a saliency map corresponding to the dot product of the embedding vector and individual local feature vectors for a single label. It can be confirmed that the results predicted by the example in first place match the ground truth.

[0114] Figure 13 shows the semantic attributes of adaptive embedding. It can be seen that the examples clearly distinguish the semantic relationships of multiple superlabels, including multiple sublabels.

[0115] As used in the above embodiments, the term "~part" refers to software or hardware components such as FPGAs (field programmable gate arrays) or ASICs, and "~part" plays a role. However, "~part" is not limited to software or hardware. "~part" can be configured to reside in an addressable storage medium, or to regenerate one or more processors. Thus, as an example, "~part" 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 programmatic code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables.

[0116] Components and the functions provided within a "~part" can be combined with a smaller number of components and "~parts" or separated from additional components and "~parts".

[0117] Furthermore, the components and "~parts" can also be embodied to regenerate one or more CPUs within the device or security multimedia card.

[0118] On the other hand, the data classification method according to one embodiment described herein can also be embodied in the form of a computer-readable medium that stores computer-executable instructions and data. Here, the instructions and data can be stored in the form of program code, which, when executed by a processor, can generate a predetermined program module and perform a predetermined operation. The computer-readable medium may be any available medium accessible by a computer, and may include volatile and non-volatile media, and isolated and non-isolated media. The computer-readable medium may also be a computer recording medium. The computer recording medium may include any volatile and non-volatile, isolated and non-isolated media embodied by any method or technique for storing computer-readable information such as instructions, data structures, program modules, or other data. For example, the computer recording medium may be a magnetic storage medium such as an HDD and an SSD, an optical recording medium such as a CD, DVD, and a Blu-ray disc, or memory contained in a server accessible via a network.

[0119] Furthermore, the data classification method according to one embodiment described herein can also be embodied in a computer program (or computer program product) that includes computer-executable instructions. A computer program includes programmable machine instructions that are processed by a processor and can be embodied in a high-level programming language, object-oriented programming language, assembly language, or machine language. A computer program can also be recorded on a type of computer-readable recording medium (e.g., memory, hard disk, magnetic / optical media, or SSD (Solid-State Drive)).

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

[0121] Here, the processor can process instructions within the computing device. Such instructions may include, for example, instructions stored in memory or storage devices 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, a number of processors and / or a number of buses can be appropriately used together with a number of memories and memory forms. The processor can also be embodied in a chipset consisting of chips containing a number of independent analog and / or digital processors.

[0122] Memory also stores information within a computing device. For example, memory can consist of volatile memory units or a collection thereof. Alternatively, memory can consist of non-volatile memory units or a collection thereof. Furthermore, memory may be other forms of computer-readable media, such as magnetic or optical discs.

[0123] Furthermore, the storage device can provide computing devices with a large-capacity storage space. The storage device may be a computer-readable medium or a configuration including such a medium, and may include, for example, a device in a SAN (Storage Area Network) or other configurations, and may be a floppy disk drive, a hard disk drive, an optical disk drive, or a tape drive, flash memory, or other similar semiconductor memory device or device array. The embodiments described above are illustrative, and those with ordinary skill in the art to which the embodiments belong will understand that they can be easily modified into other specific forms without altering the technical idea or essential features of the embodiments. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not limiting. For example, each component described as a single type can be implemented in a distributed manner, and similarly, components described as distributed can be implemented in a combined manner.

[0124] The scope of protection sought by this specification shall be determined by the claims set forth below rather than by the detailed description above, and all forms of modification or variation derived from the meaning and scope of the claims and their equivalent concepts shall be interpreted as being included within the scope of the invention. [Explanation of Symbols]

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

Claims

1. A data classification method performed by a data classification device, This includes the step of extracting features from input data via a learning network model and outputting prediction results based on those features, The learning network model compares local features derived through individual layers other than the final layer of the learning network model with label embedding vectors corresponding to classification labels, and is a model that directly compares the label embedding vectors and the local features using a label embedding dictionary that is linked to at least one layer of the learning network model and maps the label embedding vectors. A data classification method in which the label embedding vectors of the label embedding dictionary are adaptively and dynamically updated based on error signals of the local features.

2. The data classification method according to claim 1, wherein the learning network model is configured such that the computational values ​​processed by at least one layer of the learning network model are not tracked, and the dependency on the computation graph used for gradient calculation is removed, so that the error signals of the local features derived from the at least one layer are not propagated in the direction of previous layers.

3. The data classification method according to claim 1, wherein the learning network model is a model in which at least one layer of the learning network model receives an error signal of the local feature from a loss function set based on dictionary-contrast learning.

4. The data classification method according to claim 3, wherein the learning network model is a model in which the parameters of the learning network model are updated so as to maximize the similarity between the label embedding vectors corresponding to the local features and the local features in the label embedding dictionary, while minimizing the similarity between the label embedding vectors that do not correspond to the local features in the label embedding dictionary and the local features.

5. The data classification method according to claim 1, wherein the learning network model calculates a final error signal for the final layer of the learning network model, the backpropagation paths between the previous layers of the final layer and the final layer are separated, and the final error signal is not propagated to the intermediate layers of the learning network model.

6. Memory for storing a learning network model with multiple layers, Includes a control unit that extracts features from input data via the learning network model and outputs prediction results based on the features, The learning network model compares local features derived through individual layers other than the final layer of the learning network model with label embedding vectors corresponding to classification labels, and is a model that directly compares the label embedding vectors and the local features using a label embedding dictionary that is linked to at least one layer of the learning network model and maps the label embedding vectors. A data classification device in which the label embedding vectors of the label embedding dictionary are adaptively and dynamically updated based on error signals of the local features.

7. A computer-readable recording medium on which a program for performing the method described in claim 1 is recorded.

8. A computer program executed by a data classification device and stored on a recording medium for performing the method according to claim 1.

Images