A method and apparatus for continuous meta-learning based on a dirichlet process

Abstract

The application discloses a kind of based on Dirichlet process's continuous meta-learning method and device, comprising: at least 1 independent meta-knowledge distribution is constructed as initial meta-knowledge distribution;Buffer current time sequential data, based on initial meta-knowledge distribution, the number of new meta-knowledge distribution is dynamically added to the data of current time using Dirichlet process, and new meta-knowledge distribution is randomly initialized;Through Bayesian continuous learning method, the posterior of previous time meta-knowledge distribution is regarded as the prior of current time meta-knowledge distribution, and according to Bayes rule, the iteration update of meta-knowledge distribution is carried out in the data stream buffered, wherein, meta-knowledge distribution includes initial meta-knowledge distribution and the number of new meta-knowledge distribution increased at each time;The corresponding relationship between updated meta-knowledge distribution and target task is constructed, and the meta-knowledge initialization is carried out for the corresponding model of target task according to the corresponding relationship, then a small amount of labeled data is used to learn the model, and the model capable of realizing prediction is obtained.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meta-learning, and particularly relates to a continuous meta-learning method and device based on a Dirichlet process. BACKGROUND

[0002] Meta-learning is a method for few-shot learning. This kind of method is mainly aimed at the problem that deep learning cannot be effectively applied when data is scarce. The core idea is to learn how to learn, that is, to learn a method for capturing data patterns from data (i.e., meta-knowledge). On subsequent few-shot tasks, the learned method for capturing data patterns can quickly capture the data patterns of the few-shot tasks and thus be applied to the target tasks. Meta-learning methods are widely used in various fields due to their strong generalization ability and rapid adaptation ability in few-shot problems, including recommendation systems, user portraits, and small molecule property prediction. For example, patent document CN114187472A discloses a breast cancer molecular subtype prediction method based on model-driven meta-learning, and patent document CN111199458A discloses a recommendation system based on meta-learning and reinforcement learning.

[0003] Traditional meta-learning methods learn a globally shared initialization value from all data and quickly adapt to the target task. However, this kind of traditional meta-learning method has two obvious shortcomings: first, it assumes that all data is available, but in actual applications, data arrives sequentially; and due to limited storage space, not all data is available; second, a single initialization value is shared in all target tasks, but due to the large differences between tasks in the real world, a single initialization value cannot effectively solve the problem.

[0004] Online meta-learning is a meta-learning method for solving data arriving sequentially, which aims to improve the application of traditional meta-learning methods in data streams. However, this kind of existing method has two obvious defects: first, it assumes that all incoming data is available, which is often not satisfied in large-scale data applications, i.e., the size of the storage space determines that only the currently incoming data is available; second, it needs to train on all incoming data, which brings a large computational overhead and is not conducive to practical applications. SUMMARY

[0005] In view of the above, the purpose of the present application is to provide a continuous meta-learning method and device based on a Dirichlet process, which realizes continuous updating of meta-knowledge of data on a low data storage space, and reduces the computational overhead of meta-learning while ensuring the comprehensiveness and correctness of meta-knowledge.

[0006] To achieve the above-mentioned purpose of the application, the embodiments provide the following technical solutions:

[0007] A Dirichlet process-based continuous meta-learning method, comprising the following steps:

[0008] At least one independent meta-knowledge distribution is constructed as an initial meta-knowledge distribution;

[0009] The data sequentially arriving at the current time is cached, and the number of new meta-knowledge distributions is dynamically added to the data at the current time based on the initial meta-knowledge distribution using a Dirichlet process, and the new meta-knowledge distribution is randomly initialized;

[0010] The posterior of the meta-knowledge distribution at the previous time is taken as the prior of the meta-knowledge distribution at the current time by a Bayesian continuous learning method, and the meta-knowledge distribution is iteratively updated in the cached data stream according to the Bayesian rule, wherein the meta-knowledge distribution includes the initial meta-knowledge distribution and the new meta-knowledge distribution added at each time;

[0011] The correspondence between the updated meta-knowledge distribution and the target task is constructed, the meta-knowledge of the corresponding model for the target task is initialized according to the correspondence, and the model is learned using a small amount of labeled data based on the initialized meta-knowledge to obtain a model capable of achieving prediction.

[0012] In one embodiment, each meta-knowledge distribution represents a method of capturing data patterns from few-shot data, including the initialization value of the model, the gradient calculation of the model, the learning rate of the model, and the architecture selection of the model.

[0013] In one embodiment, the number of new meta-knowledge distributions is dynamically added to the data at the current time based on the initial meta-knowledge distribution using a Dirichlet process, comprising:

[0014] The number of new meta-knowledge distributions for the current task is generated by a Dirichlet process according to the relationship between past tasks and all current meta-knowledge distributions, so that the number of new meta-knowledge distributions created at the current time is determined by the prior value;

[0015] If the number of meta-knowledge distributions generated for the current task by the Dirichlet process is zero, it is considered that no new meta-knowledge distribution is created for the data at the current time.

[0016] In one embodiment, the Dirichlet process contains a Chinese restaurant process or an Indian buffet process to generate the prior of the number of new meta-knowledge distributions for the current task.

[0017] In one embodiment, the meta-knowledge distribution is iteratively updated in the cached data stream according to the Bayesian rule, comprising:

[0018] Based on the cached data, an approximate distribution of the meta-knowledge distribution is calculated by using a structured variational inference algorithm and an implicit reparameterization technique, and the sampling process of the approximate distribution is parameterized by using the reparameterization technique, so that an end-to-end gradient descent method is used to update the approximate distribution of the meta-knowledge distribution to approximate the real distribution of the meta-knowledge.

[0019] In one embodiment, the approximate distribution of the meta-learning based on the Dirichlet process includes a Gaussian distribution, a Beta distribution and a Bernoulli distribution.

[0020] In one embodiment, the reparameterization technique includes an explicit reparameterization technique and an implicit reparameterization technique, the explicit reparameterization technique is used when the approximate distribution is a Gaussian distribution, and the implicit reparameterization technique is used when the approximate distribution is a Beta distribution or a Bernoulli distribution.

[0021] In one embodiment, the construction of the corresponding relationship between the updated meta-knowledge distribution and the target task includes:

[0022] The target task corresponding model is initialized by each meta-knowledge distribution, and the prediction probability of a small amount of labeled data in the model is calculated, the prediction probability is used as the associated probability of each meta-knowledge distribution and the current target task, and at least one meta-knowledge distribution with a large associated probability is selected to construct a corresponding relationship with the target task.

[0023] To achieve the above-mentioned purposes, the embodiments provide a device for continuous meta-learning based on Dirichlet process, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above-mentioned method for continuous meta-learning based on Dirichlet process when executing the computer program.

[0024] Compared with the prior art, the present application has at least the following beneficial effects:

[0025] A small part of the data is cached on a low data storage space, the number of meta-knowledge distributions is dynamically increased for the data at the current time by using the Dirichlet process, and the meta-knowledge distribution is updated by using the Bayesian continuous learning method, so that the meta-knowledge distribution is gradually updated and accurate and comprehensive meta-knowledge distribution is obtained in the data flow process, and it is not necessary to cache all the data to learn the meta-knowledge distribution.

[0026] According to the corresponding relationship between the meta-knowledge distribution and the target task constructed by screening, the model corresponding to the target task is initialized by using the meta-knowledge distribution, and learning of the model on a small amount of labeled data is performed based on this, which greatly shortens the learning time, reduces the computational cost, and also provides learning performance and ensures the robustness of the model. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to make the technical solutions of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0028] Figure 1 is a flowchart of the method for continuous meta-learning based on Dirichlet process provided by the embodiment;

[0029] Figure 2 is a probabilistic graphical model of Dirichlet process provided by the embodiment;

[0030] Figure 3 is a flowchart of the method for continuous meta-learning based on Dirichlet process provided by the embodiment, wherein D represents a task data set, D s and D Q represent two subsets of the task data set, and theta represents meta-knowledge, and phi represents task-specific parameters. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the protection scope of the present application.

[0032] The embodiment provides a method for continuous meta-learning based on Dirichlet process, which performs meta-learning under the condition that sequential data is available and sequential data is available through Dirichlet process, wherein sequential available is understood as data arriving in sequence and only data arriving at the current time being available. As shown in Figure 1 and Figure 2 include the following steps:

[0033] Step 1, constructing at least one independent meta-knowledge distribution as an initial meta-knowledge distribution.

[0034] In the embodiment, each meta-knowledge distribution represents a method for capturing data patterns from few-shot data, including initialization value of the model, gradient calculation of the model, learning rate of the model, architecture selection of the model, etc. For different tasks, the composition of the meta-knowledge distribution is different, and each task needs to have customized meta-knowledge distribution, so as to facilitate the subsequent learning of few-shot data. The meta-knowledge distribution captured in different data streams is not single.

[0035] In the embodiment, at least one independent meta-knowledge distribution is constructed as an initial meta-knowledge distribution through random initialization, and the initial meta-knowledge distribution is used as the basis for constructing and updating the meta-knowledge distribution in the subsequent data stream.

[0036] Step 2, cache the data sequentially arrived at the current time, and dynamically increase the number of meta-knowledge distributions for the data at the current time based on the initial meta-knowledge distribution using the Dirichlet process, and randomly initialize the new meta-knowledge distribution.

[0037] In the embodiment, a low data cache space is provided, which can cache a small amount of sequentially arrived data, and can at least cache the cache data arrived at one time. The initial meta-knowledge distribution is taken as the meta-knowledge distribution of the data at time 0, and based thereon, the Dirichlet process is used to dynamically increase the number of meta-knowledge distributions for the data at the current time, which specifically includes: generating the prior of the number of new meta-knowledge distributions for the current task through the Dirichlet process according to the relationship between the past tasks and all the meta-knowledge distributions at the current time, so as to determine the number of new meta-knowledge distributions created at the current time through the prior value; if the number of meta-knowledge distributions generated for the current task through the Dirichlet process is zero, it is considered that no new meta-knowledge distribution is created for the data at the current time.

[0038] In the embodiment, the number of new meta-knowledge distributions is dynamically increased by generating the number of new meta-knowledge distributions for the current task, so as to capture incremental knowledge. The Dirichlet process includes Chinese Restaurant Process (CRP) and Indian Buffet Process (IBP). The embodiment can use the CRP or the IBP to generate the prior of the number of new meta-knowledge distributions for the current task. When the CRP is used, one meta-knowledge distribution is added each time and the prior of one-to-one matching relationship between the task and the meta-knowledge is provided, and when the IBP is used, multiple meta-knowledge distributions are added each time and the prior of many-to-many matching relationship between the task and the meta-knowledge is provided.

[0039] Based on the Dirichlet process, the prior of the number of meta-knowledge distributions can be provided when a new task is contacted, so as to ensure that the number of meta-knowledge distributions can be dynamically increased in the setting of the task stream to capture incremental information.

[0040] Step 3, through the Bayesian continuous learning method, the posterior of the meta-knowledge distribution at the previous time is taken as the prior of the meta-knowledge distribution at the current time, and the iterative update of the meta-knowledge distribution is performed in the cached data stream according to the Bayesian rule.

[0041] In the embodiment, the meta-knowledge distribution is updated according to the data cached at each time, that is, the iterative update of the meta-knowledge distribution in the cached data stream is realized. The meta-knowledge distribution updated by the data stream includes the initial meta-knowledge distribution and the meta-knowledge distribution increased at each time.

[0042] In the embodiment, the iterative updating of the meta-knowledge distribution in the cached data stream according to the Bayesian rule comprises: based on the cached data, an approximate distribution of the meta-knowledge distribution is calculated by using a structured variational inference algorithm and an implicit reparameterization technique. The structure of the approximate distribution is as follows:

[0043]

[0044] where q(·) refers to the approximate distribution of the task-specific parameter distribution, the meta-knowledge distribution, the relationship between the task and the meta-knowledge, and the underlying Beta distribution. At the same time, the sampling process of the approximate distribution is parameterized by using the reparameterization technique, so that the approximate distribution of the meta-knowledge distribution is updated by using an end-to-end gradient descent method to approximate the real meta-knowledge distribution which is difficult to calculate in reality:

[0045] q(v t,k )=Beta(α t,k ,β t,k ),

[0046] q(z t,k |π t,k )=Bern(π t,k ), where π t,k =v t,k ,

[0047]

[0048]

[0049] The approximate distribution described above includes Gaussian distribution, Beta distribution, Bernoulli distribution, etc. The reparameterization technique is used to sample the meta-knowledge distribution, parameterize the sampled distribution, so that the training of the model can follow the end-to-end gradient descent training. The reparameterization technique includes explicit reparameterization technique and implicit reparameterization technique. When the approximate distribution is Gaussian distribution, the explicit reparameterization technique is used. When the approximate distribution is Beta distribution or Bernoulli distribution, the implicit reparameterization technique is used.

[0050] By updating the meta-knowledge distribution by the Bayesian continuous learning method, the meta-knowledge distribution of the newly arrived data can be incrementally learned without forgetting the meta-knowledge distribution captured from the past data.

[0051] Figure 3 is a probabilistic graphical model of the Dirichlet process provided by the embodiment, which describes how to capture the meta-knowledge distribution under the condition that the data is sequentially available. For each time t, the following process is used:

[0052] First, the number K t,new of new meta-knowledge distributions to be added is determined by the Chinese restaurant process.

[0053]

[0054] Alternatively, the number K of newly added meta-knowledge distributions can be determined through the Indian buffet process. t,new :

[0055]

[0056] Here, α is a hyperparameter used to control the growth rate of the meta-knowledge distribution, and Possion(·) represents the Poisson distribution.

[0057] Each new meta-knowledge distribution is initialized using a beta distribution. The task-related Bernoulli distribution of each meta-knowledge distribution is determined using a stick-breaking method, i.e., with respect to z. t,k The distribution of z t,k Indicate whether the corresponding meta-knowledge distribution is relevant to the task:

[0058] v k ~Beta(α,1)for k=1,…,∞

[0059]

[0060] z t,k ~Bern(π) k ), for k=1,…,∞

[0061] Where Beta(·) and Bern(·) refer to the Beta distribution and Bernoulli distribution, respectively, k represents the index of the meta-knowledge distribution, and v i Let v represent the variable from the i-th Beta distribution. k Let the variable representing the k-th Beta distribution be...

[0062] Initialize the meta-knowledge distribution using a normal distribution. via z t,k The knowledge distribution of elements θ t,k Together, they determine the multimodal meta-knowledge distribution provided to the task, and based on this multimodal meta-knowledge distribution model, the optimal parameter φ is obtained under limited sample data. t The optimal parameter φ t This will be used for subsequent prediction tasks: φ t |θ t ,z t ~p(φ t |θ t ,z t ), p(φ t |θ t ,z t This can be understood as the data being at the optimal parameter φ. tthe optimal parameter φ t is determined by z t,k and the meta-knowledge distribution θ t,k is initialized by the multi-modal meta-knowledge distribution model and is determined by parameter optimization.

[0063] Based on the above Dirichlet process probability graph model, the structured variational inference algorithm is used to update the model parameters using the data arriving at each time, Figure 3 where x i represents the data, y i represents the prediction of the model to the data x i , N t and M t represent the number of samples used for training and testing in the training task, respectively.

[0064] Step 4, build the correspondence relationship between the updated meta-knowledge distribution and the target task, initialize the meta-knowledge of the corresponding model of the target task according to the correspondence relationship, and on the basis of the initialized meta-knowledge, use a small amount of labeled data to learn the model to obtain a model that can realize prediction.

[0065] When the updated meta-knowledge distribution is applied to the target task, the association probability between the meta-knowledge distribution and the target task is constructed according to the performance of each updated meta-knowledge distribution on a small amount of labeled data of the current target task, and the specific process includes: initializing the target task corresponding model with each meta-knowledge distribution, and calculating the prediction probability p(y|x, φ t ) of a small amount of labeled data in the model, which is used as the association probability between each meta-knowledge distribution and the current target task, and at least one meta-knowledge distribution with a large association probability is selected to build a corresponding relationship z t between the target task.

[0066] It should be noted that the correspondence relationship between the meta-knowledge distribution and the target task can be a many-to-many relationship, a one-to-many relationship, or a many-to-one relationship. Among them, the many-to-many relationship is understood as that a similar task can be associated with multiple meta-knowledge distributions, and the meta-knowledge distribution can be associated with different classes of tasks at the same time. The one-to-many relationship is understood as that one meta-knowledge distribution corresponds to multiple different target tasks. The many-to-one relationship is understood as that multiple meta-knowledge distributions correspond to one target task.

[0067] The many-to-many relationship is constructed on the premise of avoiding forgetting caused by meta-knowledge distribution conflict, and encourages the sharing of meta-knowledge distribution between tasks. That is, the meta-knowledge distribution commonly associated between two tasks promotes the sharing of meta-knowledge distribution, and different associated meta-knowledge distributions retain the difference between tasks.

[0068] In the embodiments, after the correspondence relationship is obtained, the meta-knowledge is initialized for the corresponding model of the target task according to the correspondence relationship, that is, a method for providing a data pattern capturing method for a small amount of samples of the target task, including at least one of determining an initialization value of the model, a learning rate of the model, gradient calculation of the model, and architecture of the model. Specifically, by sampling the corresponding relationship z t to obtain a plurality of distributions associated with each other to form a multi-peak distribution, which provides an initialization value for capturing a data pattern for the target task.

[0069] On the basis of initializing the meta-knowledge, the model is learned by using a small amount of labeled data, that is, a gradient descent of a limited number of steps is performed on a small amount of labeled data of the target task, to obtain a model with better performance and capable of realizing prediction.

[0070] Based on the same inventive concept, the embodiments also provide a Dirichlet process-based continuous meta-learning device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the Dirichlet process-based continuous meta-learning method when executing the computer program.

[0071] In practical applications, the computer memory can be a proximal volatile memory such as RAM, and can also be a non-volatile memory such as ROM, FLASH, floppy disk, mechanical hard disk, etc., and can also be a remote storage cloud. The computer processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), that is, the steps of a Dirichlet process-based continuous meta-learning can be implemented by these processors.

[0072] In small molecule property prediction, whether a molecule has a certain property and the relationship between the molecule and the new property are not all known at once, but are constantly increasing with the arrival of experimentally labeled samples. The Dirichlet process-based continuous meta-learning method can meet the updating of meta-knowledge under the setting that the known property data of the molecule is constantly arriving. The updated meta-knowledge can rely on a small amount of labeled samples (whether it has the property) to perform property prediction on the new property.

[0073] The method and device for continuous meta-learning based on Dirichlet process provided by the above embodiments cache a small part of data in sequence in a low data storage space, dynamically increase the number of meta-knowledge distributions for the data at the current time by using Dirichlet process, and update the meta-knowledge distribution by using the Bayesian continuous learning method. In this way, the meta-knowledge distribution is gradually updated and accurate and comprehensive meta-knowledge distribution is obtained in the data flow process, and it is not necessary to cache all the data to learn the meta-knowledge distribution. According to the correspondence between the meta-knowledge distribution constructed by screening and the target task, the meta-knowledge distribution is used to initialize the model corresponding to the target task, and learning of the model on a small amount of labeled data is performed on the basis of the initialization, which greatly shortens the learning time, reduces the computing cost, and also provides the learning performance and ensures the robustness of the model.

[0074] The above detailed description of the specific embodiments has described the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application and is not used to limit the present application. Any modification, supplement and equivalent replacement within the principle range of the present application should be included in the protection range of the present application.

Claims

1. A method of continual meta-learning based on Dirichlet processes, characterized in that, The method comprises the following steps: Constructing at least one independent meta-knowledge distribution as an initial meta-knowledge distribution; Caching data sequentially arrived at the current moment, dynamically adding the number of new meta-knowledge distributions for the data at the current moment based on the initial meta-knowledge distribution by using a Dirichlet process, and randomly initializing the new meta-knowledge distribution, wherein the Dirichlet process contains a Chinese restaurant process or an Indian buffet process to generate a prior of the number of new meta-knowledge distributions for the current task, when the Chinese restaurant process is used, one meta-knowledge distribution is added each time and a prior of a one-to-one matching relationship between the task and the meta-knowledge is provided, and when the Indian buffet process is used, multiple meta-knowledge distributions are added each time and a prior of a many-to-many matching relationship between the task and the meta-knowledge is provided; By using a Bayesian continuous learning method, the posterior of the meta-knowledge distribution at the previous moment is taken as the prior of the meta-knowledge distribution at the current moment, and the meta-knowledge distribution is iteratively updated in the cached data stream according to the Bayesian rule, wherein the meta-knowledge distribution comprises the initial meta-knowledge distribution and the new meta-knowledge distribution added at each moment; Constructing a corresponding relationship between the updated meta-knowledge distribution and the target task, comprising: initializing a target task corresponding model with each meta-knowledge distribution, and calculating the prediction probability of a small amount of labeled data in the model, the prediction probability is taken as the association probability of each meta-knowledge distribution and the current target task, and at least one meta-knowledge distribution with a large association probability is selected to construct a corresponding relationship with the target task, wherein the corresponding relationship between the meta-knowledge distribution and the target task is a many-to-many relationship, a one-to-many relationship or a many-to-one relationship; According to the corresponding relationship, the meta-knowledge of the target task corresponding model is initialized, and based on the initialized meta-knowledge, the model is learned by using a small amount of labeled data to obtain a model capable of realizing prediction.

2. The Dirichlet process based continual meta-learning method of claim 1, wherein, Each meta-knowledge distribution represents a method for capturing data patterns from a small amount of sample data, including the initialization value of the model, the gradient calculation of the model, the learning rate of the model and the architecture selection of the model.

3. The Dirichlet process based continual meta-learning method of claim 1, wherein, The number of new meta-knowledge distributions added for the data at the current moment based on the initial meta-knowledge distribution by using the Dirichlet process comprises: According to the relationship between the past tasks and all the current meta-knowledge distributions, the prior of the number of new meta-knowledge distributions for the current task is generated by using the Dirichlet process, so that the number of new meta-knowledge distributions created at the current moment is determined by the prior value; If the number of meta-knowledge distributions generated for the current task by using the Dirichlet process is zero, it is considered that no new meta-knowledge distribution is created for the data at the current moment.

4. The Dirichlet process based continual meta-learning method of claim 1, wherein, The iterative updating of the meta-knowledge distribution in the cached data stream according to the Bayesian rule comprises: Based on the cached data, the structured variational inference algorithm is used to calculate the approximate distribution of the meta-knowledge distribution, and the sampling process of the approximate distribution is parameterized by using the reparameterization technique, so that the end-to-end gradient descent method is used to update the approximate distribution of the meta-knowledge distribution to approximate the real distribution of the meta-knowledge.

5. The Dirichlet process based continual meta-learning method of claim 4, wherein, The approximate distribution of the meta-learning based on the Dirichlet process comprises a Gaussian distribution, a Beta distribution and a Bernoulli distribution.

6. The Dirichlet process based continual meta-learning method of claim 5, wherein, The reparameterization technique includes an explicit reparameterization technique and an implicit reparameterization technique, the explicit reparameterization technique is used when the approximate distribution is a Gaussian distribution, and the implicit reparameterization technique is used when the approximate distribution is a Beta distribution or a Bernoulli distribution.

7. A Dirichlet process based continual meta-learning apparatus, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the steps of the Dirichlet process based continual meta-learning method according to any one of claims 1-6 when executing the computer program.

Images