Learning device, classification device, learning method, classification method, and computer program
The learning device generates synthetic data to selectively forget unwanted classes by combining input data with pseudo data, addressing privacy concerns and data leakage in lifelong learning.
Patent Information
- Application Number
- JP2023508211
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-03-23
- Publication Date
- 2024-11-20
- Estimated Expiration
- 2041-03-23
AI Technical Summary
Existing lifelong learning methods face challenges in selectively forgetting unnecessary knowledge from past tasks while preserving other knowledge, leading to risks of data leakage and privacy violations, especially in practical applications like facial recognition systems.
A learning device and method that generates synthetic data by combining learning input data with class-specific pseudo data, allowing selective forgetting of unwanted classes by ignoring pseudo data associated with those classes during new task learning, using a machine learning model with specific loss functions to update parameters.
Enables selective forgetting of unnecessary past knowledge while maintaining relevant information, enhancing privacy protection and preventing data leaks in lifelong learning scenarios.
Smart Images

Figure 0007587185000010 
Figure 0007587185000011 
Figure 0007587185000012
Abstract
Description
[Technical field]
[0001] The present invention relates to a learning device, a classification device, a learning method, a classification method, and a computer program. [Background technology]
[0002] In deep learning, a type of machine learning, when a network is updated for new learning, its performance on the previous learning content (hereafter referred to as "task") drops dramatically. To mitigate this effect, there is active research on lifelong learning, which aims to update the network to adapt to new tasks without forgetting past learning results. Specific examples of new tasks include new classes and new sets of instances.
[0003] Examples of research on lifelong learning include the contents shown in Non-Patent Documents 1 to 3. For example, Non-Patent Document 1 discloses a method called memory replay-based, which stores a part of past learning samples and replays them when a new task is performed. Non-Patent Document 2 discloses a method for freezing model parameters in deep learning. Non-Patent Document 3 also discloses a method for implicitly utilizing knowledge of previous tasks by introducing an additional regularization term. [Prior art documents] [Non-patent literature]
[0004] [Non-Patent Document 1] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. “iCaRL: Incremental Classifier and Representation Learning”, In Proc.CVPR, 2017. [Non-Patent Document 2] Arun Mallya and Svetlana Lazebnik, “PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning”, In Proc. CVPR, 2018. [Non-Patent Document 3] Zhizhong Li and Derek Hoiem. “Learning without Forgetting”, TPAMI, 40(12):2935-2947, 2017. Summary of the Invention [Problem to be solved by the invention]
[0005] Many of the existing lifelong learning methods mentioned above are designed to learn expressive models for new tasks while retaining all knowledge from previous tasks. On the other hand, artificial intelligence is currently facing new types of problems, and as artificial intelligence becomes more practical and connected to everyday life, various ethical issues such as privacy protection and data leakage prevention have become important themes. This issue is unavoidable even in technology for lifelong learning.
[0006] Maintaining complete knowledge of all previous tasks may lead to the risk of data leakage and privacy violation. Furthermore, it is not always necessary to have complete knowledge of previous tasks. For example, if we consider a face recognition system installed at the entrance of an office, it is not necessarily necessary to remember the faces of people who are transferred to other departments. Therefore, it is desirable to perform learning that can selectively forget unnecessary knowledge from past tasks while preserving other knowledge.
[0007] In view of the above circumstances, an object of the present invention is to provide a technique that enables learning in which unnecessary learning content from past learning contents can be selectively forgotten while the other learning content can be preserved. [Means for solving the problem]
[0008] One aspect of the present invention is a learning device comprising: a data input unit that inputs first data to be learned, second data for identifying the first data, and second past data related to learning content to be saved, the second past data being used during past learning as data for identifying the first data; a synthetic data generation unit that generates synthetic data by synthesizing the first data and the second data; and a parameter update unit that updates parameters of the machine learning model based on characteristics of the synthetic data obtained by inputting the synthetic data and the second past data into a machine learning model and the second past data.
[0009] One aspect of the present invention is a classification device that classifies input data using a trained model obtained by the above-mentioned learning device.
[0010] One aspect of the present invention is a learning method that inputs first data to be learned, second data for identifying the first data, and second past data related to learning content to be saved, the second past data being used during past learning as data for identifying the first data, generating synthetic data by synthesizing the first data and the second data, and updating parameters of the machine learning model based on characteristics of the synthetic data obtained by inputting the synthetic data and the second past data into a machine learning model and characteristics of the second past data.
[0011] One aspect of the present invention is a classification method for classifying input data using a trained model obtained by the above-mentioned learning device.
[0012] One aspect of the present invention is On the computer, First data to be learned and second data for identifying the first data. Generate synthetic data by synthesizing the The composite data; and Second past data used in past learning as data for identifying the first data, the second past data being related to the learning content to be saved A process of updating parameters of the machine learning model based on features of the second past data and the synthetic data obtained by inputting the second past data into a machine learning model. To executeIt is a computer program.
[0013] One aspect of the present invention is a computer program for causing a computer to function as the above-described identification device. Effect of the Invention
[0014] According to the present invention, it is possible to carry out learning in which unnecessary learning contents from among past learning contents are selectively forgotten while the other learning contents are preserved. [Brief description of the drawings]
[0015] [Figure 1] FIG. 2 is a block diagram showing a functional configuration of a learning device according to an embodiment. [Diagram 2] FIG. 1 is a diagram for explaining the object of the present invention. [Diagram 3] FIG. 2 is a diagram illustrating a process according to an embodiment. [Figure 4] 4 is a flowchart showing the flow of a learning process of the learning device in the embodiment. [Diagram 5] FIG. 1 is a diagram for explaining the superiority of the present invention. [Figure 6] FIG. 2 is a diagram illustrating an example of a hardware configuration of an identification device according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0016] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. 1 is a block diagram showing the functional configuration of a learning device 1 according to an embodiment. The learning device 1 according to an embodiment is a device that performs so-called lifelong learning, enabling continuous learning. The learning device 1 updates a predetermined machine learning model by machine learning until a predetermined termination condition is satisfied.
[0017] The predetermined machine learning model at the time when the predetermined termination condition is satisfied is the trained model. Therefore, the learning device 1 acquires the trained model by updating the predetermined machine learning model by machine learning until the predetermined termination condition is satisfied.
[0018] For ease of explanation, hereinafter, performing machine learning is also referred to as learning. Updating a machine learning model (hereinafter referred to as a "machine learning model") by machine learning means suitably adjusting the parameter values in the machine learning model. In the following explanation, learning to be A means that the parameter values in the machine learning model are adjusted to satisfy A. A represents a condition. In addition, hereinafter, "for learning" means that it is used to update the machine learning model. Note that a machine learning model is a set including one or more processes whose execution conditions and order are predetermined.
[0019] The machine learning model is represented by, for example, a neural network. A neural network is a circuit such as an electronic circuit, an electric circuit, an optical circuit, or an integrated circuit that represents a machine learning model. The parameters of the neural network are suitably adjusted based on the loss, and the parameters of the network are the parameters of the machine learning model that is represented. The parameters of the network are the parameters of the circuits that constitute the network.
[0020] The machine learning model updated by the learning device 1 is a machine learning model that identifies input data. The trained model obtained by the learning device 1 updating parameters is, for example, a model that performs data classification. Learning input data, learning reference data, and learning pseudo data are input to the machine learning model.
[0021] The learning input data is data to be learned. For example, the learning input data may be two-dimensional or more data such as an image, one-dimensional time series data such as voice, sound, or vibration, or symbolic data such as natural language. In this way, the learning input data may be any data that is generally applicable to machine learning. When one-dimensional time series data such as voice, sound, or vibration is used as the learning input data, it is sufficient to convert it into two dimensions using a ceptrum or the like.
[0022] The learning reference data is so-called correct answer data in machine learning. The learning reference data may be information equivalent to a class label such as an attribute, a class name, or a person ID that has been quantified, or may be data similar to the learning input data, i.e., image or voice data. The learning reference data may be any objective variable that is generally applied to machine learning. More specifically, when a label is used as the objective variable, the objective is identification or authentication, and when data similar to the learning input data is used as the variable, the objective is regression.
[0023] Hereinafter, data including at least one pair of learning input data and one pair of learning reference data will be referred to as learning data. That is, the learning data is data including at least a pair of learning input data and learning reference data, and is an example of so-called learning data. The learning data is one aspect of the first data.
[0024] The training pseudo data is data different from the training data. For example, the training pseudo data may be data of a random noise pattern, data of a pattern in which a part of the training input data is degraded, or data of a pattern represented by the average value of the training input data. The pattern represented by the average value of the training input data may be a pattern in which each pixel is represented by the average value of each pixel in a plurality of training input data, or a pattern in which the average value of each pixel of one training input data is assigned to each pixel. Note that the training pseudo data may be a small number of samples from data of past tasks that are guaranteed to be safe from the viewpoint of privacy, or may be a synthetic image generated by a generative model obtained by adversarial learning or the like. The training pseudo data is data for identifying a class to which the training data belongs.
[0025] In this specification, an example is described in which image data is mainly handled as data input to the learning device 1, but the present invention is not limited thereto. In this specification, a case in which deep learning is used is described.
[0026] In the following description, for the sake of clarity, the set of input data will be expressed as in equation (1).
[0027]
number
[0028] For example, each data set D k is the training input data x k and the reference data for learning y k Using the above, we express it as in the following formula (2). i k is the input data for learning. y i k is the reference data for learning.
[0029]
number
[0030] As mentioned above, the purpose of the present invention is to selectively forget unnecessary knowledge (classes in this case) from past tasks while retaining other knowledge (classes in this case). In the following explanation, we will call the classes we want to retain (hereinafter referred to as the "retained set") C P k In this case, the class to be selectively forgotten (hereafter called the "deletion set") is C P k The complement of , i.e., -C P k (- is placed above C).
[0031] The network in the learning device 1 is represented by f, and the parameters are formally represented by θ. That is, the network is θ This is expressed as:
[0032] The objective of the present invention can be expressed as follows using the above notation: P If the input learning data x belongs to the correct class y, the input learning data x is mapped to the correct class y. On the other hand, if the correct class y belongs to the preserved class C P If the input data x does not belong to , then f is used to map the input data x to a class other than the correct class y. θ Learn.
[0033] In achieving the above object, the learning device 1 may set the following constraints.
[0034] (constraints) When learning the kth task, the past training input data {D1,…,D k-1 Do not use}.
[0035] Formally, these objectives can be expressed as shown in Fig. 2. Fig. 2 is a diagram for explaining the objectives of the present invention. In the following explanation, a case where the above-mentioned constraints are applied will be described, but the present invention is also applicable in cases where the above-mentioned constraints are not applied.
[0036] (Specific Configuration of Learning Device 1) Learning device 1 includes an input unit 11, a control unit 12, a communication unit 13, a memory unit 14, and an output unit 15.
[0037] The input unit 11 is composed of a data input unit 111, a pseudo-data input unit 112, and a pseudo-data re-input unit 113.
[0038] The data input unit 111 inputs the data set D shown in the above formula (2) in the k-th task. That is, the data input unit 111 inputs the set of learning input data x k and learning reference data y i k in the k-th task. i k
[0039] The pseudo-data input unit 112 inputs learning pseudo-data. Hereinafter, the set of learning pseudo-data is represented as {ζ k,c}. Note that ζ k,c is the learning pseudo-data of the c-th class in the k-th task. The learning pseudo-data input to the pseudo-data input unit 112 is a form of the second data.
[0040] The pseudo-data re-input unit 113 re-inputs the learning pseudo-data input by the pseudo-data input unit 112. More specifically, the pseudo-data re-input unit 113 may input the learning pseudo-data of {ζ k´,c} (where k´ < k) in the k-th task. That is, the pseudo-data re-input unit 113 inputs the learning pseudo-data used during past learning as data for identifying learning data. The learning pseudo-data input to the pseudo-data re-input unit 113 is a form of the second past data.
[0041] The control unit 12 controls the entire learning device 1. The control unit 12 is configured using a processor such as a CPU (Central Processing Unit) and a memory. The control unit 12 executes a program to realize the functions of an extended data generation unit 121, a feature extraction unit 122, a classification unit 123, a loss acquisition unit 124, a parameter update unit 125, and a termination determination unit 126.
[0042] The extended data generator 121 generates the learning input data x i k and the pseudo data ζ inputted by the pseudo data input unit 112. k,c Specifically, the extended data generating unit 121 generates extended data (synthesized data) by synthesizing the learning input data x i k and the pseudo data for training, ζ k,c The extended data may be generated by blending the above-mentioned data in a ratio determined by a predetermined probability distribution function. The extended data generator 121 is one aspect of the combined data generator.
[0043] The extended data generator 121 will be described using a specific example. As described above, the object of the present invention is to generate a saved set C P k While remembering the classes in P k (- is placed above C) to forget the classes included in the class. To achieve this goal, in the present invention, the information of each class is stored as training pseudo data ζ k,c The system associates the class with a simple piece of data called a pseudo-data set, and uses that pseudo-data set to control whether the class is remembered or forgotten.
[0044] More specifically, a special type of data augmentation is used to perform this control. An overview of this process is shown in FIG. 3. FIG. 3 is a diagram showing a schematic diagram of the processing in the embodiment. When receiving a new task (new task in FIG. 3), the augmented data generator 121 generates class-specific pseudo data for learning ζk,c For each class, we generate an image with random pixel values and embed it into all samples of the corresponding class.
[0045] For example, during the training of the kth task, the original training input data x i k For the training pseudo data ζ k,c By synthesizing the extended data, i k (~ is placed above x) can be generated.
[0046]
number
[0047] In formula (3), λ is a random value between 0 and 1. The extended data generator 121 extracts the original set {(x i k ,y i k ) N i=1} plus extended data ~x i k (~ is above x) i k ,y i k ) N i=1} is stored in a memory (not shown). In a later process, the parameter update unit 125 uses these two sets to update the parameters of the feature extraction unit 122 and the classification unit 123. After the update is completed, the learning pseudo data ζ k,c You just need to leave that behind.
[0048] Such extended data~x i k (~ is placed above x) to train samples of the same class, which are the corresponding training pseudo data ζ k,c Therefore, when learning a new task, the training pseudo data ζ k,c Depending on whether or not you use the pseudo data for training,k,c As a result, we can control whether to locally maintain the feature distribution around the training input data x i k Without using the pseudo data for training, k,c It is possible to have a character learn or forget any class using
[0049] The pseudo data for learning, ζ k,c To generate the training data, we use random color patterns, such as those shown in Figure 3. i k We can assign random colors to each grid of an image of the same size as the training data ζ k,c The advantages of using this method are as follows:
[0050] First, random patterns are easy to generate. Second, the patterns are independent for each class and task. Third, like existing memory-based approaches, we use training input data x i k Unlike the use of parts of a pattern, the pattern itself does not directly represent information in the raw data, which makes it suitable for privacy protection and data leakage prevention.
[0051] The feature extraction unit 122 extracts features by inputting the extended data generated by the extended data generation unit 121 and the learning pseudo data re-input by the pseudo data re-input unit 113. As a method for extracting features, a general deep learning network such as ResNet disclosed in Reference 1 may be used.
[0052] (Reference 1: K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition”, In Proc. CVPR, 2016.)
[0053] The identification unit 123 identifies input data using the features extracted by the feature extraction unit 122. As a method for identifying from features, for example, the final layer of a network in general deep learning, such as ResNet disclosed in Non-Patent Document 4, may be used. Alternatively, as a method for identifying from features, for example, a method such as a support vector machine may be used. Note that, although the identification unit 123 in the embodiment has been described as estimating a certain class label, the present invention is not limited to this. For example, the identification unit 123 may perform regression processing instead of performing identification processing. Furthermore, a case where the structure of the network in the identification unit 123 is a multi-head structure will be described, but the present invention is not limited to this.
[0054] The loss calculation unit 124 calculates losses using the learning data input by the data input unit 111, the learning pseudo data input by the pseudo data re-input unit 113, and the results of identification by the identification unit 123. For example, the loss calculation unit 124 may include regularization for restricting the range of parameters to be updated by the parameter update unit 125.
[0055] More specifically, the operation of the loss calculation unit 124 will be described. The loss calculation unit 124 calculates the loss based on the formula (4). The total loss function L for learning shown in the formula (4) is the classification loss L C , Pseudo data loss L M , selective forgetting loss L F , regularization term L R It may consist of four terms:
[0056]
number
[0057] Classification loss L C and pseudo data loss L M The term is for learning a new task. Selective forgetting loss L F and the regularization term L RThe items below are for maintaining the previous tasks. Each of these items will be explained in detail below.
[0058] Classification loss L C is introduced with the aim of accurately classifying new tasks. Specifically, the classification loss L C can be given, for example, as equation (5).
[0059]
number
[0060] In equation (5), N k is the number of training samples for the kth task, and l(x,y) is the training input data x i k and the learning reference data y i k (e.g., class label y). For example, the loss function can be softmax cross entropy (CE) or additive margin softmax (AMS) loss.
[0061] Simulated Data Loss L M is the training input data x i k Classification loss L C In addition to the loss L, another loss using augmented data with the training pseudo data may be used to link each training pseudo data to the corresponding class. M is expressed as the following equation (6).
[0062]
number
[0063] The loss function is the classification loss L C A similar one may be used.
[0064] Selective forgetting loss LF is the preserved set C P k Then, only the classes belonging to the deletion set -C P k (- is placed above C) is introduced for the purpose of forgetting classes. Specifically, this is achieved by learning only with pseudo data for training that correspond to classes in the retention set, and discarding other pseudo data for training (e.g., pseudo data for training that correspond to classes in the deletion set). More specifically, x i p The pseudo data for training to generate i p Then, the loss can be given by, for example, the following equation (7).
[0065]
number
[0066] In equation (7), N p is the number of training samples in the pth task, γ SF is the weight for adjusting the strength of this loss. This loss function is i k Do not use any of the above. And, the deletion set-C P k By not using the training pseudo data for classes in (- is placed above C), these classes are catastrophically forgotten. This causes the training input data x i k It is possible to realize selective forgetting of previous tasks without using any
[0067] Regularization term L R can be achieved by using general regularization (e.g., Non-Patent Document 3, References 2 and 3) that is introduced to prevent catastrophic forgetting.
[0068] (Reference 2: James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al, “Overcoming catastrophic forgetting in neural networks”, PNAS, 114(13):3521-3526, 2017.)
[0069] (Reference 3: Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. “Memory Aware Synapses: Learning what (not) to forget”, ECCV, pages 139-154, 2018.)
[0070] In this embodiment, the regularization term is the preserved set C p k Since it is necessary to prevent catastrophic forgetting only for these preserved sets, the method may be modified to provide regularization only for these preserved sets. More specifically, in the case of Non-Patent Document 3, for example, the following formula (8) may be used as the regularization term.
[0071]
number
[0072] In equation (8), γ is the weight of the term, and i is the index of the class label. Here, the saved set C P k i ∈ C P k Change the sum as follows: y in Eq. (8) ´ o (i) and^y´ o (i) (^ above y) is a correction between the recorded probability and the current probability. In the case of Non-Patent Document 5, the following equation (9) can be used as the regularization term.
[0073]
number
[0074] In equation (9), γ is the weight of the regularization term, Γ p,q is the q-th parameter on the diagonal ^θ p,q (^ is placed on θ) is the element of the Fisher matrix of the pth previous task. Note that the saved set C P k The Fisher matrix may be modified so that the regularization term is evaluated only for the class corresponding to the regularization term. A combination of these multiple regularizations may be used as the regularization term by linear sum or the like.
[0075] The parameter update unit 125 updates the machine learning model so that the loss acquired by the loss acquisition unit 124 is reduced. The update of the machine learning model based on the loss is specifically a process of updating the parameter values of a neural network representing the machine learning model according to a predetermined rule based on the loss. More specifically, the update of the parameter values of the neural network is, for example, an update of the parameter values corresponding to each of the feature extraction unit 122 and the identification unit 123. When updating the parameters, the parameter update unit 125 may update the parameters using an existing method such as a stochastic gradient method.
[0076] The end determination unit 126 determines whether a predetermined end condition is satisfied. The predetermined end condition may be set in advance. For example, the predetermined end condition may be that the loss is reduced by a threshold or more, or that the parameters are updated a predetermined number of times.
[0077] The communication unit 13 includes a communication interface for connecting the learning device 1 to an external device. The communication unit 13 communicates with the external device via wired or wireless communication. The external device is, for example, a device that transmits learning data.
[0078] The storage unit 14 is configured using a non-transitory computer-readable storage medium device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 14 stores various information related to the learning device 1. The storage unit 14 stores information input via, for example, the input unit 11 or the communication unit 13. The storage unit 14 stores, for example, a machine learning model. The storage unit 14 stores, for example, various information generated by the execution of the machine learning model.
[0079] It should be noted that the learning data and the learning pseudo data do not necessarily have to be input only to the input unit 11, and do not necessarily have to be input only to the communication unit 13. The learning data and the learning pseudo data may be input from either the input unit 11 or the communication unit 13. For example, the learning data may be input to the input unit 11, and the learning pseudo data may be input to the communication unit 13. The learning data and the learning pseudo data may be stored in advance in the memory unit 14.
[0080] The output unit 15 outputs various information. The output unit 15 includes a display device such as a CRT (Cathode Ray Tube) display, a liquid crystal display, or an organic EL (Electro-Luminescence) display. The output unit 15 may be configured as an interface that connects these display devices to the learning device 1. The output unit 15 outputs information input to the input unit 11, for example. The output unit 15 may display learning data input to the input unit 11 or the communication unit 13, for example. The output unit 15 may display the execution result of a machine learning model, for example.
[0081] FIG. 4 is a flowchart showing the flow of the learning process of the learning device 1 in the embodiment. The data input unit 111 inputs learning input data for a new task (step S101). The data input unit 111 outputs the input learning input data to the control unit 12. The pseudo data input unit 112 inputs learning pseudo data corresponding to the new task (step S102). The pseudo data input unit 112 outputs the input learning pseudo data to the control unit 12. The pseudo data re-input unit 113 inputs learning pseudo data corresponding to a past task (step S103). The pseudo data re-input unit 113 outputs the input learning pseudo data to the control unit 12.
[0082] The extended data generating unit 121 acquires the learning input data and the learning pseudo data output from the pseudo data input unit 112. The extended data generating unit 121 generates extended data by synthesizing the acquired learning input data and the learning pseudo data (step S104). The extended data generating unit 121 outputs the generated extended data to the feature extracting unit 122.
[0083] The feature extraction unit 122 extracts features using the extended data output from the extended data generation unit 121 and the learning pseudo data corresponding to the past task output from the pseudo data re-input unit 113 (step S105). The identification unit 123 performs identification from the features extracted by the feature extraction unit 122 (step S106). The loss acquisition unit 124 acquires a loss based on the identification result obtained by the identification unit 123, the learning input data output from the data input unit 111, and the learning pseudo data corresponding to the past task output from the pseudo data re-input unit 113 (step S107). For example, the loss acquisition unit 124 acquires the loss by the above formula (4).
[0084] The parameter update unit 125 updates the parameters of the machine learning model based on the loss acquired by the loss acquisition unit 124 (step S108). The termination determination unit 126 determines whether or not the termination condition is satisfied (step S109). If the termination condition is not satisfied (step S109-NO), the learning device 1 executes the processes from step S104 onward.
[0085] On the other hand, if the termination condition is satisfied (step S109-YES), the data output section 30 outputs the classification result of the classification section 123 (step S110).
[0086] Next, the advantages of the technology of the present invention will be explained with reference to FIG. 5. In FIG. 5, the features on the final layer of the shared backbone at the end of each task are visualized using, for example, t-SNE (Stochastic Neighbor Embedding). FIG. 5(A) is a diagram visualizing the features on the final layer when only the classification loss LC is used. FIG. 5(B) is a diagram visualizing the features on the final layer when a general lifelong learning method is used. FIG. 5(C) is a diagram visualizing the features on the final layer when the method of the present invention is used.
[0087] If we only use classification loss, we forget all the past tasks because all the samples from the past tasks are randomly arranged, as shown in Figure 5(A).
[0088] When general lifelong learning is used, as shown in Figure 5(B), samples are aggregated for each class in all past tasks, and the structure is easily separable for each class. This means that lifelong learning is possible, but at the same time, it is not possible to forget a specific class.
[0089] In contrast to the above two, when the method of the present invention is used, as shown in Figure 5(C), samples of classes to be memorized are kept aggregated in the feature space and the classes to be memorized can be quickly dispersed, demonstrating that it is possible to carry out lifelong learning while forgetting only the classes that one wishes to forget.
[0090] 6 is a diagram showing an example of a hardware configuration of the classification device 2 in an embodiment. The classification device 2 performs classification processing using a trained model obtained by the learning device 1. The classification device 2 includes a control unit 21 including a processor 26 such as a CPU and a memory 27 connected by a bus, and executes a program. The classification device 2 functions as a device including the control unit 21, an input unit 22, a communication unit 23, a storage unit 24, and an output unit 25 by executing the program.
[0091] More specifically, the processor 26 reads out the program stored in the storage unit 24, and stores the read out program in the memory 27. The processor 26 executes the program stored in the memory 27, whereby the identification device 2 functions as a device including the control unit 21, the input unit 22, the communication unit 23, the storage unit 24, and the output unit 25.
[0092] The control unit 21 controls the operation of various functional units included in the classification device 2. The control unit 21 classifies input data, for example, by using a trained model obtained by the learning device 1. Specifically, the control unit 21 identifies the data by inputting the data input to the classification device 2 into the trained model. For example, the control unit 21 performs class classification of the data.
[0093] The input unit 22 includes input devices such as a mouse, a keyboard, and a touch panel. The input unit 22 may be configured as an interface that connects these input devices to the device itself. The input unit 22 accepts input of various information for the device itself. The input unit 22 accepts, for example, an input instructing the start of processing. The input unit 22 accepts, for example, an input of data to be processed.
[0094] The communication unit 23 includes a communication interface for connecting the device to an external device. The communication unit 23 communicates with the external device via a wired or wireless connection. The external device with which the communication unit 23 communicates is, for example, an output destination of the processed data. In such a case, the communication unit 23 outputs the processed data to the external device by communicating with the external device.
[0095] The communication unit 23 may communicate with, for example, the learning device 1. In this case, the communication unit 23 acquires a trained model obtained by, for example, the learning device 1.
[0096] The storage unit 24 is configured using a non-transitory computer-readable storage medium device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 24 stores various information related to the identification device 2. The storage unit 24 stores, for example, a trained model acquired via the communication unit 23. The storage unit 24 stores, for example, data to be processed that is input to the input unit 22.
[0097] The output unit 25 outputs various information. The output unit 25 includes a display device such as a CRT display, a liquid crystal display, or an organic EL display. The output unit 25 may be configured as an interface that connects these display devices to the output unit 25. The output unit 25 outputs information input to the input unit 22, for example.
[0098] According to the learning device 1 configured as above, it becomes possible to perform learning that can selectively forget unnecessary knowledge from past tasks while preserving other knowledge. The reason for this is that when updating a model (i.e., the feature extraction unit and the identification unit), the extended data generation unit 121 performs data extension that embeds task-specific signals called pseudo input data, for example, all input data of a corresponding class. Then, when learning a new task, the learning device 1 learns by ignoring (not using) the learning pseudo data of a class to be deleted. This is because it is possible to intentionally induce forgetting for a class to be deleted. This is because it is possible to realize selective forgetting for a previous task.
[0099] The learning device 1 may be implemented using a plurality of information processing devices communicably connected via a network. In this case, each functional unit of the learning device 1 may be distributed and implemented in the plurality of information processing devices.
[0100] All or part of the functions of the learning device 1 and the identification device 2 may be realized using hardware (including circuitry) such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may be recorded on a computer-readable recording medium. Examples of computer-readable recording media include portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into a computer system. The program may be transmitted via a telecommunications line.
[0101] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Industrial Applicability]
[0102] The present invention is applicable to lifelong learning. [Explanation of symbols]
[0103] 1...Learning device, 2...Classification device, 11...Input unit, 12...Control unit, 13...Communication unit, 14...Memory unit, 15...Output unit, 21...Control unit, 22...Input unit, 23...Communication unit, 24...Memory unit, 25...Output unit, 26...Processor, 27...Memory, 111...Data input unit, 112...Pseudo data input unit, 113...Pseudo data re-input unit, 121...Augmented data generation unit, 122...Feature extraction unit, 123...Classification unit, 124...Loss acquisition unit, 125...Parameter update unit, 126...Termination determination unit
Claims
1. a data input unit for inputting first data to be learned, second data for identifying the first data, and second past data used in past learning as data for identifying the first data, the second past data being related to learning content to be saved; a composite data generating unit that generates composite data by combining the first data and the second data; a parameter update unit that updates parameters of the machine learning model based on features of the synthetic data obtained by inputting the synthetic data and the second past data into a machine learning model and the second past data; A learning device comprising:
2. The second data is any one of data of a random noise pattern, data of a pattern obtained by partially changing the learning input data, and data of a pattern represented by an average value of the learning input data. The learning device according to claim 1 .
3. an identification unit that identifies the first data by using features of the composite data and the second past data; a loss acquiring unit that acquires a loss by using the first data, the second past data, and a result of the identification by the identification unit; Further comprising: The learning device according to claim 1 or 2.
4. the loss acquisition unit acquires the loss based on regularization that restricts a range of parameters updated by the parameter update unit, a loss based on a classification accuracy in the synthetic data, a loss based on a classification accuracy in the first data, and a loss based on a classification accuracy in the second data. The learning device according to claim 3 .
5. A classification device that classifies input data using a trained model obtained by the learning device according to any one of claims 1 to 4.
6. inputting first data to be learned, second data for identifying the first data, and second past data used in past learning as data for identifying the first data, the second past data relating to learning content to be saved; generating composite data by combining the first data and the second data; updating parameters of the machine learning model based on features of the synthetic data obtained by inputting the synthetic data and the second past data into a machine learning model and the second past data; How to learn.
7. A classification method for classifying input data using a trained model obtained by the learning device according to any one of claims 1 to 4.
8. A computer comprising: generating composite data by combining first data to be learned and second data for identifying the first data; A computer program for executing a process of updating parameters of the machine learning model based on the characteristics of the synthetic data obtained by inputting the synthetic data and second past data related to the learning content to be saved, the second past data being used during past learning as data for identifying the first data, into a machine learning model.
9. A computer program for causing a computer to function as the identification device according to claim 5.
Citation Information
Patent Citations
System and method for improving deep neural network performance
US20200074305A1
