Unsupervised online learning method and system for optimizing local plasticity rules with meta-learning

Abstract

This invention provides an unsupervised online learning method and system for optimizing local plasticity rules through meta-learning. The method includes the following steps: S1: constructing a dataset based on a quantitative number of example samples; S2: pre-training a feature encoding network based on the base dataset; S3: learning parameterized local plasticity rules based on the meta-learning dataset; and S4: updating the classification network using the feature encoding network and the local plasticity rules. This invention offers the advantages of unsupervised online learning, eliminating the need for manual annotation of newly added unseen samples during online learning, reducing annotation costs, and improving the real-time performance of the online learning process. Simultaneously, it possesses the advantage of rapid online learning with streaming samples, overcoming the shortcomings of traditional batch learning. When new data samples arrive, the model can be updated quickly and effectively. Furthermore, because the model parameters are updated based on the local plasticity mechanism, learning can be achieved in a forward-looking manner, avoiding the need for fine-tuning the network using the backpropagation algorithm.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically, to an unsupervised online learning method and system for meta-learning to optimize local plasticity rules. Background Technology

[0002] Machine learning plays a crucial role in modern data analytics and artificial intelligence applications. Traditional machine learning paradigms typically operate on a batch learning model, where the model follows some learning algorithms to train on a pre-given set of training data, then deploys for inference with minimal updates. Behind modern machine learning achievements lies backpropagation of errors, a key training algorithm used to compute parameter updates. However, backpropagation requires global information about the network when computing backward gradients, resulting in significant routing overhead—a major reason why it's considered not a brain-like learning mechanism. Therefore, due to the high cost of training, modern model training is often performed offline. In the era of big data, real-time data is rapidly growing and evolving, requiring AI-based systems operating in rapidly changing scenarios to adapt in real time. Making machine learning scalable and practical, especially in unsupervised learning from streaming samples, has become a major challenge in the fields of machine learning and artificial intelligence.

[0003] Online learning aims to learn models sequentially from data. It overcomes the shortcomings of traditional batch learning, allowing learners to quickly and effectively update models as new data samples arrive. The brain possesses remarkable online learning capabilities, rapidly adapting to changing scenarios and data, and learning new knowledge and skills. This brain-like online learning ability, combined with neuromorphic hardware, enables efficient on-chip learning on edge devices, showing great promise for applications such as adapting to device drift, environmental changes, and user customization. Therefore, this invention focuses on a key scenario of online learning technology: learning new tasks with a limited number of example samples. Adaptation or learning typically requires fine-tuning the entire or a portion of the network. Due to limitations in power, system architecture memory walls, etc., the adaptability of current edge devices is difficult to adapt to online learning tasks using highly complex backpropagation algorithms. Therefore, an unsupervised online learning method based on meta-learning optimization and local plasticity rules is particularly important in the current environment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide an unsupervised online learning method and system for meta-learning to optimize local plasticity rules.

[0005] An unsupervised online learning method for optimizing local plasticity rules through meta-learning, provided by the present invention, includes:

[0006] Step S1: Construct a dataset based on a quantitative set of example samples;

[0007] The dataset includes a base dataset, a meta-learning dataset, and a test dataset;

[0008] Step S2: Pre-train the feature encoding network based on the basic dataset;

[0009] Step S3: Learn parameterized local plasticity rules based on the meta-learning dataset;

[0010] Step S4: Update the classification network using the feature encoding network and the local plasticity rule.

[0011] Preferably, step S2 includes:

[0012] The data is preprocessed by cropping, randomly flipping horizontally, and normalizing the images to ensure consistent input size and increase data diversity.

[0013] The model is trained, and the cross-entropy loss function is used to evaluate the model's prediction error, measuring the difference between the predicted distribution and the true distribution.

[0014] The model parameters are updated based on the loss value using the backpropagation algorithm, and the loss function is minimized using the gradient descent method to optimize the model parameters.

[0015] Preferably, step S3 includes an internal adaptive loop updating the parameters of the classification network using parameterized local plasticity rules, and an external meta-optimization loop modifying the plasticity meta-parameters.

[0016] Preferably, the test dataset includes a support set and an unlabeled query set; the weights of the classification network are initialized using the support set of the test dataset, and the weights of the classification network are updated unsupervised using the unlabeled query set of the test dataset to improve the recognition ability of the pre-trained model, and the performance of the method is evaluated on the unlabeled query set.

[0017] Preferably, when new data samples arrive, the model parameters are updated based on the local plasticity parameter update rule, and learning is performed in a forward-looking manner.

[0018] An unsupervised online learning system for meta-learning optimization of local plasticity rules, provided by the present invention, includes:

[0019] Module M1: Constructs a dataset based on a quantitative number of example samples;

[0020] The dataset includes a base dataset, a meta-learning dataset, and a test dataset;

[0021] Module M2: A feature encoding network pre-trained on the base dataset;

[0022] Module M3: Learns parameterized local plasticity rules on a meta-learning dataset;

[0023] Module M4: Updates the classification network using a feature encoding network and local plasticity rules.

[0024] Preferably, the module M2 includes:

[0025] The data is preprocessed by cropping, randomly flipping horizontally, and normalizing the images to ensure consistent input size and increase data diversity.

[0026] The model is trained, and the cross-entropy loss function is used to evaluate the model's prediction error, measuring the difference between the predicted distribution and the true distribution.

[0027] The model parameters are updated based on the loss value using the backpropagation algorithm, and the loss function is minimized using the gradient descent method to optimize the model parameters.

[0028] Preferably, the module M3 includes an internal adaptive loop that updates the parameters of the classification network using parameterized local plasticity rules, and an external meta-optimization loop that modifies the plasticity meta-parameters.

[0029] Preferably, the test dataset includes a support set and an unlabeled query set; the weights of the classification network are initialized using the support set of the test dataset, and the weights of the classification network are updated unsupervised using the unlabeled query set of the test dataset to improve the recognition ability of the pre-trained model, and the performance of the method is evaluated on the unlabeled query set.

[0030] Preferably, when new data samples arrive, the model parameters are updated based on the local plasticity parameter update rule, and learning is performed in a forward-looking manner.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This invention proposes a novel local fine-tuning scheme based on a pre-trained model. The pre-training + on-chip local learning framework is suitable for deep learning tasks in resource-constrained environments. First, the model is pre-trained using a large-scale dataset. Then, local learning is used to fine-tune the model parameters to adapt to specific tasks or domains. Pre-training allows the model to learn general features and patterns on large-scale datasets, thereby improving the model's generalization ability. Simultaneously, on-chip local learning can process data and perform local inference on edge devices, enabling the model to better adapt to specific tasks or environments, further improving model performance. Experiments show that compared to existing few-shot learning schemes based on transmission learning, the local plasticity rule based on meta-learning optimization proposed in this invention achieves approximately 10% improvement in recognition accuracy in several unbalanced streaming few-shot learning tasks.

[0033] 2. This invention offers the advantage of unsupervised online learning. During online learning, manual annotation of newly added unseen samples is unnecessary, reducing annotation costs and improving the real-time performance of the online learning process. Experiments show that, compared to existing few-shot learning schemes based on transfer learning, this invention achieves convergence in approximately 15 iterations.

[0034] 3. This invention has the advantage of rapid online learning from streaming samples, overcoming the shortcomings of traditional batch learning. When new data samples arrive, the model can be updated quickly and effectively. Furthermore, since the model parameters are updated based on the local plasticity mechanism, learning can be achieved in a forward-looking manner, avoiding the need to use the backpropagation algorithm to fine-tune the network, making the online fine-tuning time less than 1 minute.

[0035] Other beneficial effects of the present invention will be explained in detail through the introduction of specific technical features and technical solutions in specific embodiments. Those skilled in the art should be able to understand the beneficial technical effects brought about by these technical features and technical solutions through the introduction of these technical features and technical solutions. Attached Figure Description

[0036] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0037] Figure 1 This is an overall block diagram of an unsupervised online learning method based on local plasticity rules optimized by meta-learning in this invention.

[0038] Figure 2 This is a flowchart of the algorithm for discovering locally plasticity rules with biological interpretability based on the meta-learning method of this invention.

[0039] Figure 3 This is a flowchart of an unsupervised online learning method based on meta-learning optimization of local plasticity rules in this invention.

[0040] Figure 4 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0042] Reference Figure 1 As shown, an unsupervised online learning method for meta-learning to optimize local plasticity rules includes:

[0043] Step 1: Construct a dataset for a new task learning problem based on a small number of example samples;

[0044] In a new task learning setting with a limited number of example samples, the dataset is divided into a base dataset, a meta-learning dataset, and a test dataset, with no overlap in the categories of these three datasets. This invention uses the base dataset to pre-train the model to obtain the feature extraction network, the meta-learning dataset to discover local plasticity parameter update rules, the support set of the test dataset to initialize the weights of the classification network, and the unlabeled query set of the test dataset to unsupervisedly update the classification network weights to improve the recognition ability of the pre-trained model. The performance of the method is then evaluated on the unlabeled query set.

[0045] Step 2: Pre-train the feature encoding network on the base dataset;

[0046] First, in the data preprocessing stage, images are cropped, randomly horizontally flipped, and normalized to ensure consistent input size, increase data diversity, and meet the input requirements of the pre-trained model. These measures help improve the training efficiency and stability of the model. Second, during model training, the cross-entropy loss function is used to evaluate the prediction error, and gradient descent is used to minimize the loss, ensuring that the model learns effectively in classification tasks. These steps provide a foundation for the pre-training of the feature encoding network, enabling it to learn general features and patterns on the base dataset, thereby improving the model's generalization ability.

[0047] Step 3: Learn parameterized local plasticity rules on the meta-learning dataset;

[0048] The parameter update rule based on local plasticity is a parameter update method that utilizes pre- and post-synaptic correlations. For unsupervised online learning tasks, this invention introduces a competition suppression mechanism into the prediction distribution of the classification network. By passing the prediction distribution through a parameterized piecewise linear function, the encouraged output distribution of the classification network is obtained. The parameters of the local plasticity rule are learned using a meta-learning method. Meta-learning is a machine learning paradigm that aims to learn elements in the learning process. This framework consists of a two-layer learning scheme: an inner adaptive loop updates the parameters of the classification network using parameterized local plasticity rules, and an outer meta-optimization loop modifies the plasticity meta-parameters, i.e., the parameters of the local plasticity rule.

[0049] Step 4: On the test dataset, i.e., on new classes that have not been seen before during meta-learning and pre-training, update the classification network parameters in an unsupervised streaming manner to enhance the recognition ability.

[0050] Step 2 yields a pre-trained feature encoding network on labeled base class data. Then, on the test dataset—new classes not encountered during training—with only one or a few labeled samples (support set) for each new class, the weights of the classification network are initialized. Subsequently, the classification network learns sequentially on unlabeled samples of the new classes (query set), where parameter updates rely on the local plasticity rules optimized by meta-learning in step 3, and the algorithm's performance is evaluated on the query set.

[0051] This invention offers the advantages of unsupervised online learning, eliminating the need for manual annotation of newly added unseen samples during the online learning process, thus reducing annotation costs and improving the real-time performance of the online learning process. It also boasts the advantage of rapid online learning with streaming samples, overcoming the shortcomings of traditional batch learning. When new data samples arrive, the model can be updated quickly and effectively. Furthermore, because the model parameters are updated based on the local plasticity mechanism, learning can be achieved in a forward-looking manner, avoiding the need to fine-tune the network using the backpropagation algorithm.

[0052] The above are basic embodiments of the present invention. The technical solution of the present invention will be further described below through a preferred embodiment.

[0053] Example 1

[0054] Reference Figure 1 As shown, an unsupervised online learning method for meta-learning to optimize local plasticity rules includes:

[0055] Step 1: Construct a dataset for a new task learning problem based on a small number of example samples.

[0056] Suppose we have a labeled training set:

[0057]

[0058] Where, x i Let y be the original feature of sample i. i It is its corresponding one-hot encoded label.

[0059] In the field of few-shot learning, this labeled set is often referred to as the base dataset. The few-shot scenario assumes there is a test dataset:

[0060]

[0061] It uses a completely new set of classes, making Create few-shot learning tasks from random sampling, each task having some labeled examples.

[0062] Specifically, each K-way N S The -shot task involves drawing N from K distinct categories.S The labeled samples are also randomly selected. Let... The set representing these labeled samples is called the support set, and its size is .

[0063]

[0064] In addition, each task has a user The query set represented by the query set is composed of

[0065]

[0066] It consists of unlabeled (unseen) examples from K classes. To support meta-learning to discover local plasticity parameter update rules, an additional set of entirely new classes is used.

[0067]

[0068] A meta-learning dataset was constructed.

[0069] This invention uses a base dataset to pre-train the model to obtain a feature extraction network, uses a meta-learning dataset to discover local plasticity parameter update rules, uses a support set to initialize the weights of the classification network, uses an unlabeled query set to unsupervisedly update the classification network weights to improve the recognition ability of the pre-trained model, and evaluates the performance of the method on the unlabeled query set.

[0070] Step 2: Pre-train the feature extraction network.

[0071] The base dataset χ of the standard dataset base The feature encoder is trained on a network (such as mini-ImageNet or tiered-ImageNet). The specific steps are as follows:

[0072] 1. Data Preprocessing

[0073] Cropping: Each image is cropped and adjusted to the uniform size required for model input. This step ensures that all input images have a consistent size, facilitating model processing.

[0074] Random horizontal flipping: During the training phase, images are randomly flipped horizontally to enhance data diversity and help the model learn stronger feature representations.

[0075] Normalization: Normalizes the pixel values ​​of the image to meet the input requirements of the pre-trained model. This step can accelerate model training and improve stability.

[0076] 2. Model Training

[0077] Loss function: The cross-entropy loss function is used to evaluate the prediction error of the model. It is widely used in classification tasks and effectively measures the difference between the predicted distribution and the true distribution.

[0078] Backpropagation: The model parameters are updated based on the loss value using the backpropagation algorithm, and the loss function is minimized using gradient descent, thereby optimizing the model parameters.

[0079] The above steps provide a foundation for the pre-training of the feature encoding network, enabling it to learn general features on a standard dataset. This lays a solid foundation for subsequent specialized training and optimization, ensuring that the model can efficiently adapt to different tasks and domains.

[0080] Step 3: Meta-learning methods are used to discover local plasticity rules with biological interpretability.

[0081] Neural networks are typically trained to classify discrete outputs by adding a softmax output layer. This is a linear mapping that projects the network's d-dimensional hidden outputs onto m outputs, where m is the number of distinct classes. The softmax operator is then applied to generate the probability distributions for each class. The parameters in this softmax layer are typically optimized using gradient descent along with the network parameters.

[0082] The weights of the softmax layer This can be viewed as a set of m vectors θ[i]; i = 1, ..., m, each vector corresponding to a given class. When trained using supervised loss (such as cross-entropy), each step of gradient descent pulls the parameter θ[y] (corresponding to the class label y) towards a direction with a larger inner product with the network output h, and pushes the inner product of other parameters θ[j], j ≠ y, with h to be smaller. One drawback of neural network classifiers trained with backpropagation is that they require many input samples to predict a given class with reasonable accuracy.

[0083] Inspired by associative memory, a simple optimization process is that the network directly activates h. t The weights θ[y] are directly accumulated to the softmax layer. t In this context, accumulating or smoothing network activations into weights actually corresponds to the well-known Hebbian learning update rule.

[0084]

[0085] In the special case of the output classification layer, where Corresponding to θ,h t ,y t .

[0086]

[0087] in The label y represents sample j. t,j The one-hot tag, because y t,j =i, only θ[i] was updated.

[0088] For unsupervised online learning tasks, the idea of ​​pseudo-labels is used to obtain the true category information y. t Network predictions can be used. Replacement is performed. Due to network predictions... It is a probability distribution processed by softmax, and there are various ways to handle it. It is then combined with the activation h... t To achieve binding, one approach is to adopt a "winner-takes-all" method, encoding the predicted values ​​using one-hot encoding. Only the category with the highest predicted probability corresponds to It will only be updated then;

[0089] Second, an interlayer lateral inhibition competition mechanism is introduced. When a neuron is tuned to a certain input pattern, the intralayer lateral competitive inhibition will prevent other neurons from selectively adopting the same pattern.

[0090]

[0091] The above rules extend to the well-known Oja rule, the first of which is presynaptic neural activity h. t With postsynaptic neuron activity The product of the two terms, the second term being to ensure the weights of the hidden neurons converge to a vector of 1s. Lateral competition inhibition relies on a function g, where it is desired that the synaptic weights of strongly driven hidden units are pushed toward the modes that drive them, while the synaptic weights of weakly driven hidden units are pushed away from these modes, thus making the synaptic weights different for each hidden unit and specific to the features of the data. This activation function can be designed manually, for example, at a threshold h. * The above is positive, within the threshold h. * The following are negative, thus leading to biomimetic excitation-inhibition learning respectively.

[0092] To find a more robust function g, this invention utilizes a meta-learning method to discover locally plasticity rules with biological interpretability. Meta-learning is a machine learning paradigm that aims to learn elements of the learning process. The framework consists of a two-layer learning scheme: an inner adaptive loop uses parameterized locally plasticity rules. Learning model f θ The external meta-optimization loop modifies the plasticity meta-parameter Θ. The meta-training dataset contains a set of tasks. Each task consists of training data (X) of varying lengths. train ) and Q query data (X query ,Yquery The former is used to train model f. θ The latter is used to optimize the meta-parameter Θ. A parameterized model of the local plasticity rule is constructed using piecewise linear functions:

[0093]

[0094] Then Θ={m0,m1,m2,…,m k-1 ,b0,b1,b2,…,b k-1} is a plasticity element parameter, such as Figure 2 As shown, the meta-learning framework proposed in this invention is described in detail.

[0095] In each meta-iteration, the initial model f θ Training is performed on an unsupervised streaming online training dataset, with each adaptive iteration using a single sample x. train To update θ, in order to make the learned local plasticity rules more adaptive, X train The distribution of each category is unbalanced, and the number of samples for different tasks is random.

[0096] Step 4: Update the classification network using local plasticity rules based on meta-learning optimization

[0097] As described in step one, for a given new task learning based on a small number of example samples, there is a support set. and query set Let X represent Let Y = {1,…,K} be a random variable related to the original features, and let it be a random variable related to the label. Let φ represent the feature encoder function of the deep neural network in step two, where φ represents the trainable parameters. The feature set representing the hidden space.

[0098] This invention first utilizes a pre-trained model to collect feature representations of the support set and query set in the latent space:

[0099]

[0100] Secondly, the classification network is initialized using the support set, and a classification network W := [w1,…,w K ]for:

[0101]

[0102] Finally, the weights of the classification network are updated on the query set using the local plasticity rule optimized by the ternary learning in step 3, such as... Figure 3 As shown.

[0103] The present invention also provides an unsupervised online learning system for optimizing local plasticity rules through meta-learning. The unsupervised online learning system for optimizing local plasticity rules through meta-learning can be implemented by executing the process steps of the unsupervised online learning method for optimizing local plasticity rules through meta-learning. That is, those skilled in the art can understand the unsupervised online learning method for optimizing local plasticity rules through meta-learning as a preferred embodiment of the unsupervised online learning system for optimizing local plasticity rules through meta-learning.

[0104] Specifically, an unsupervised online learning system for meta-learning optimization of local plasticity rules includes:

[0105] Module M1: Constructs a dataset based on a quantitative number of example samples;

[0106] The dataset includes a base dataset, a meta-learning dataset, and a test dataset;

[0107] Module M2: A feature encoding network pre-trained on the base dataset;

[0108] Module M3: Learns parameterized local plasticity rules on a meta-learning dataset;

[0109] Module M4: Updates the classification network using a feature encoding network and local plasticity rules.

[0110] The module M2 includes:

[0111] The data is preprocessed by cropping, randomly flipping horizontally, and normalizing the images to ensure consistent input size and increase data diversity.

[0112] The model is trained, and the cross-entropy loss function is used to evaluate the model's prediction error, measuring the difference between the predicted distribution and the true distribution.

[0113] The model parameters are updated based on the loss value using the backpropagation algorithm, and the loss function is minimized using the gradient descent method to optimize the model parameters.

[0114] The module M3 includes an internal adaptive loop that updates the parameters of the classification network using parameterized local plasticity rules, and an external meta-optimization loop that modifies the plasticity meta-parameters.

[0115] The test dataset includes a support set and an unlabeled query set. The weights of the classification network are initialized using the support set of the test dataset, and the weights of the classification network are updated unsupervised using the unlabeled query set of the test dataset to improve the recognition ability of the pre-trained model. The performance of the method is evaluated on the unlabeled query set.

[0116] When new data samples arrive, the model parameters are updated based on the local plasticity parameter update rule, and learning is performed in a forward-looking manner.

[0117] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0118] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An unsupervised online learning method for meta-learning to optimize local plasticity rules, characterized in that, include: Step S1: Construct a dataset based on a quantitative set of example samples; The dataset includes a base dataset, a meta-learning dataset, and a test dataset; Step S2: Pre-train the feature encoding network based on the basic dataset; Step S3: Learn parameterized local plasticity rules based on the meta-learning dataset; Step S4: Update the classification network using the feature encoding network and the local plasticity rule.

2. The unsupervised online learning method for optimizing local plasticity rules through meta-learning according to claim 1, characterized in that, Step S2 includes: The data is preprocessed by cropping, randomly flipping horizontally, and normalizing the images to ensure consistent input size and increase data diversity. The model is trained, and the cross-entropy loss function is used to evaluate the model's prediction error, measuring the difference between the predicted distribution and the true distribution. The model parameters are updated based on the loss value using the backpropagation algorithm, and the loss function is minimized using the gradient descent method to optimize the model parameters.

3. The unsupervised online learning method for optimizing local plasticity rules through meta-learning according to claim 2, characterized in that, Step S3 includes an internal adaptive loop updating the parameters of the classification network using parameterized local plasticity rules, and an external meta-optimization loop modifying the plasticity meta-parameters.

4. The unsupervised online learning method for optimizing local plasticity rules through meta-learning according to claim 1, characterized in that, The test dataset includes a support set and an unlabeled query set. The weights of the classification network are initialized using the support set of the test dataset, and the weights of the classification network are updated unsupervised using the unlabeled query set of the test dataset to improve the recognition ability of the pre-trained model. The performance of the method is evaluated on the unlabeled query set.

5. The unsupervised online learning method for optimizing local plasticity rules through meta-learning according to claim 4, characterized in that, When new data samples arrive, the model parameters are updated based on the local plasticity parameter update rule, and learning is performed in a forward-looking manner.

6. An unsupervised online learning system for meta-learning to optimize local plasticity rules, characterized in that, include: Module M1: Constructs a dataset based on a quantitative number of example samples; The dataset includes a base dataset, a meta-learning dataset, and a test dataset; Module M2: A feature encoding network pre-trained on the base dataset; Module M3: Learns parameterized local plasticity rules on a meta-learning dataset; Module M4: Updates the classification network using a feature encoding network and local plasticity rules.

7. The unsupervised online learning system for meta-learning optimization of local plasticity rules according to claim 6, characterized in that, The module M2 includes: The data is preprocessed by cropping, randomly flipping horizontally, and normalizing the images to ensure consistent input size and increase data diversity. The model is trained, and the cross-entropy loss function is used to evaluate the model's prediction error, measuring the difference between the predicted distribution and the true distribution. The model parameters are updated based on the loss value using the backpropagation algorithm, and the loss function is minimized using the gradient descent method to optimize the model parameters.

8. The unsupervised online learning system for meta-learning optimization of local plasticity rules according to claim 7, characterized in that, The module M3 includes an internal adaptive loop that updates the parameters of the classification network using parameterized local plasticity rules, and an external meta-optimization loop that modifies the plasticity meta-parameters.

9. The unsupervised online learning system for meta-learning optimization of local plasticity rules according to claim 6, characterized in that, The test dataset includes a support set and an unlabeled query set. The weights of the classification network are initialized using the support set of the test dataset, and the weights of the classification network are updated unsupervised using the unlabeled query set of the test dataset to improve the recognition ability of the pre-trained model. The performance of the method is evaluated on the unlabeled query set.

10. The unsupervised online learning system for meta-learning optimization of local plasticity rules according to claim 9, characterized in that, When new data samples arrive, the model parameters are updated based on the local plasticity parameter update rule, and learning is performed in a forward-looking manner.

Images