A method for designing episodic memory networks based on synaptic remodeling model

Through the episodic memory network based on the synaptic remodeling model, combining attention and soft attention mechanism, the shortcomings of traditional neural network models in lifelong learning and persistent memory are solved, and dynamic updates and long-term storage of episodic memory are achieved, reducing the Disastrous forgetting improves the adaptability and identification accuracy of the model.

CN118536545BActive Publication Date: 2024-11-15NORTHEAST DIANLI UNIVERSITY
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410756894.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-13
Publication Date
2024-11-15
Estimated Expiration
2044-06-13

AI Technical Summary

Technical Problem

Traditional neural network models have shortcomings in realizing lifelong learning and lasting memory, especially the catastrophic forgetting problems caused by static weight update mechanisms. New knowledge learning will lead to the loss of existing knowledge.

Method used

Design an episodic memory network based on the synaptic remodeling model, through the episodic feature encoder and external memory units, update the internal memory unit parameters using the synaptic remodeling parameter importance matrix, combine the attention mechanism and soft attention mechanism, and extract the and store visual features and semantic features, form a context feature matrix and key-value pairs, and realize dynamic updates and long-term storage of context memory.

Benefits of technology

The lifelong learning ability of the model is realized, ensuring rapid response to the current situation and coherent accumulation and application of long-term knowledge, reducing catastrophic forgetting, and improving the accuracy and adaptability of the model to identify and classify in complex situations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118536545B_ABST
    Figure CN118536545B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for designing a situational memory network based on a synaptic remodeling model, and relates to the field of artificial intelligence and neural network technology. The present invention comprises the following steps: designing a situational feature encoder as an internal memory unit and using a synaptic remodeling parameter importance matrix to update the internal memory unit parameters; extracting visual features and semantic features of image data from samples to be classified; competing for the extracted features, selecting features with large weights for sparse encoding, and forming a situational feature matrix. The present application enables the model to have lifelong learning capabilities through the memory design of internal memory units and external memory units. The internal memory unit is responsible for short-term and dynamic information processing to ensure the model's rapid response and understanding of the current situational memory; the external memory unit provides long-term storage and recall functions, so that the model can maintain coherent knowledge accumulation and application capabilities in different tasks and scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and neural network, and in particular, relates to a method for designing a situational memory network based on a synaptic remodeling model. Background Art

[0002] Episodic memory is a highly complex cognitive function that involves the memorization and retrieval of specific events and scenes. Although there are many techniques for memory and learning in traditional neural network models, these conventional techniques have some significant shortcomings, especially in achieving lifelong learning and persistent memory. First, traditional models usually rely on static weight update mechanisms, which means that the model needs to retrain the entire network when faced with new information, which is not only time-consuming and labor-intensive, but also prone to catastrophic forgetting, that is, the learning of new knowledge will lead to the loss of existing knowledge.

[0003] In view of this, the present invention is proposed. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a method for designing a contextual memory network based on a synaptic remodeling model, thereby solving the problems raised in the above-mentioned background technology.

[0005] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0006] A method for designing an episodic memory network based on a synaptic remodeling model comprises the following steps:

[0007] Design the context feature encoder as an internal memory unit and use the synaptic remodeling parameter importance matrix to update the internal memory unit parameters;

[0008] Extract visual features and semantic features of image data from samples to be classified; compete the extracted features, select features with large weights for sparse coding, and form a scenario feature matrix

[0009]

[0010] Focus on the subset of images that can mark the scene features and extract the scene feature matrix from the image Retrieve information in

[0011] Attention mechanism uses gate function For the candidate class sample f t Processing to obtain scenario features The scene attention weight is calculated through multiple features, and the context vector is calculated based on the obtained scene attention weight;

[0012] Through the context vector ct and the episodic memory vector m at the previous moment t-1 To update episodic memory t , and finally memorize m j Transfer to external memory unit;

[0013] The design is based on the key-value storage mechanism as an external memory unit, extracts the relationship between visual features and semantic features, forms key-value pairs, and uses the soft attention mechanism for key addressing and hidden state update;

[0014] Combine the classification results with the context information and update the storage value of the external memory unit.

[0015] Optionally, the steps for calculating the context attention weight are:

[0016] By paying attention to the door and each scenario feature The attention mechanism is realized by the association of With the classification result c and the context memory vector m of the previous moment t-1 The interaction between them is used to calculate and obtain the context attention weight Its expression is:

[0017]

[0018] In the formula, w (1) ,w (2) represents the weight matrix of the neural network, b (1) , b (2) represents the bias vector of the neural network, M i represents the number of candidates on feature i, represents the intermediate representation vector of feature i at time t, Represents the intermediate representation of feature i after being processed by the activation function tanh.

[0019] Optional, the weight of the and processed situational features The weighted sum of t , whose expression is: After getting the context vector c t and the episodic memory vector m at the previous moment t-1 Then through the formula m t =GRU(c t ,m t-1 ) will finally remember m j Transfer to external memory unit.

[0020] Optionally, the external memory unit consists of key-value pairs <k1,v1>,...,<kM ,v M > Compose memory slots to extract the relationship between visual features and semantics. The key and value of the memory slot are obtained by weighted summation of multiple key and value feature vectors. The specific calculation formula is as follows:

[0021]

[0022] in, As the weight, we sum the feature vectors of the key and value to get the key Φ of the memory slot. t (K) is used for key addressing at the next moment, and the value of the memory slot is Φ t (V) is passed as input to the encoder to produce the next category.

[0023] Optionally, the steps to obtain the key are:

[0024] The feature vector f of each sample in the category data set extracted by the scene feature encoder t , extract the feature vector f t Organize them in order, and the organized feature vectors is recorded as a key, indicating the visual context relationship between samples, i.e., key k i is the organized feature vector The expression is

[0025] Optionally, the steps to get the value are:

[0026] Each frame image I i The pre-trained model CLIP is used to extract visual and semantic information to form a visual semantic embedding representation. Representation; then, these embeddings and the corresponding keys k i Combine to form a key-value memory slot <k i ,v i >; Among them, key k i is the feature vector extracted from the scene feature encoder, and the value vi is the visual semantic embedding obtained by the CLIP model.

[0027] Optionally, the steps for key addressing are:

[0028] Use the soft attention mechanism to get the alignment probability qi of each memory, and the key addressing value Φ read at the previous moment t-1 (K) reads past values ​​as input and outputs new hidden states The formula is as follows:

[0029]

[0030] Among them l kis the tanh function. At time t, the classification result C is the decoder hidden state and key-addressable hidden state The weighted sum of is obtained as follows:

[0031]

[0032] According to the obtained classification result C and the key vector k at time i i Through the tanh function, we get the eigenvector k at the i-th moment i Relevance score:

[0033]

[0034] Finally, the Softmax function is used to get the new alignment probability The expression is:

[0035]

[0036] in, For Perform exponential operations, It means that the correlation scores of all N feature vectors are summed after exponential operation. This summation operation is used for normalization to ensure the alignment probability. The sum of is 1.

[0037] Optionally, during the training process, the episodic memory network is trained by combining supervised learning and unsupervised learning. In the initial stage, supervised learning is used to adjust model parameters through labeled data; then the unsupervised learning stage is entered to adaptively adjust and optimize the model through massive unlabeled data.

[0038] Optionally, the scene feature encoder uses a multi-scale feature extraction technique when extracting features to ensure comprehensive capture of features at different scales.

[0039] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0040] This application uses the memory design of internal memory units and external memory units to enable the model to have lifelong learning capabilities. The internal memory unit is responsible for short-term and dynamic information processing, ensuring the model's rapid response and understanding of current situational memory; the external memory unit provides long-term storage and recall functions, allowing the model to maintain coherent knowledge accumulation and application capabilities in different tasks and scenarios.

[0041] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0043] In the figure:

[0044] Figure 1 Design a block diagram for an episodic memory network;

[0045] Figure 2 Design a schematic for the external memory unit.

[0046] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0047] The present invention will now be described in further detail with reference to the accompanying drawings.

[0048] See also Figure 1-2 As shown, in this embodiment, a method for designing an episodic memory network based on a synaptic remodeling model is provided, comprising the following steps:

[0049] Design the context feature encoder as an internal memory unit and use the synaptic remodeling parameter importance matrix to update the internal memory unit parameters;

[0050] It should be noted that by extracting and encoding visual and semantic features through the context feature encoder, the model can more accurately capture and describe complex context information. In addition, the synaptic remodeling parameter importance matrix is ​​used to dynamically adjust the parameters of the internal memory unit, so that the network can effectively identify and retain important features, thereby alleviating the overfitting problem and having stronger adaptability when facing new tasks.

[0051] Extract visual features and semantic features of image data from samples to be classified; compete the extracted features, select features with large weights for sparse coding, and form a scenario feature matrix

[0052]

[0053] It should be noted that: through screening and sparse coding, the scenario feature matrix can not only effectively represent the core features of the sample and improve the recognition accuracy of the model, but also reduce the computational burden and improve processing efficiency. This feature selection and encoding strategy ensures that the model can focus on the most representative features when dealing with complex scenarios, thereby improving the accuracy and robustness of classification tasks, helping to achieve more efficient and accurate scenario memory and classification capabilities, and making the model more adaptable and generalizable in various application scenarios.

[0054] Focus on the subset of images that can mark the scene features and extract the scene feature matrix from the image Retrieve information in

[0055] It should be noted that by focusing on the key subset of images, the model can ignore a large amount of irrelevant or redundant information, thereby reducing the consumption of computing resources. At the same time, this strategy of focusing on iconic scene features allows the model to more accurately capture and utilize key information in the image, improving the results of retrieval and classification. The use of the scene feature matrix further enhances the model's memory and recognition capabilities, enabling it to maintain efficient performance in large-scale data processing tasks.

[0056] Attention mechanism uses gate function For the candidate class sample f t Processing to obtain scenario features The scene attention weight is calculated through multiple features, and the context vector is calculated based on the obtained scene attention weight;

[0057] It should be noted that the context vector is obtained by calculating the context attention weights through multiple features, and the advantage of this method is that it can significantly improve the accuracy and efficiency of the model in complex classification tasks. The use of gate functions ensures that the model can flexibly select and weight important candidate samples, thereby identifying the most representative context features in a diverse and complex feature space. Through the comprehensive calculation of multiple features, the context attention weight can more accurately reflect the importance of each feature, thereby improving the model's attention allocation and information focusing capabilities. Finally, the context vector obtained by calculating the context attention weights not only contains rich semantic information, but also can effectively integrate the information of multiple features, improving the model's memory and reasoning capabilities.

[0058] Through the context vector c t and the episodic memory vector m at the previous moment t-1 To update episodic memory t , and finally memorize m j Transfer to external memory unit;

[0059] It should be noted that the advantage of updating the episodic memory through the context vector and the episodic memory vector of the previous moment and transferring the final memory to the external memory unit is that it can significantly enhance the model's memory and reasoning capabilities and achieve more accurate and efficient scene recognition and classification. The introduction of the context vector enables the model to integrate the current input information and the previous memory state to form a richer scene representation, thereby having a stronger ability to cope with continuous and complex tasks. By fusing the episodic memory of the previous moment with the current context vector, the model can continuously update and optimize the internal memory state, enabling it to gradually accumulate and utilize historical information and improve long-term memory capabilities.

[0060] The design is based on the key-value storage mechanism as an external memory unit, extracts the relationship between visual features and semantic features, forms key-value pairs, and uses the soft attention mechanism for key addressing and hidden state update;

[0061] It should be noted that: through the key-value storage mechanism, the relationship between visual features and semantic features is structured into key-value pairs, so that the model can efficiently store and manage a large amount of memory information. This structured storage not only improves the efficiency of data organization and access, but also ensures that high-precision association and retrieval capabilities can be maintained when processing complex multimodal information. The use of soft attention mechanism for key addressing enables the model to flexibly select and access related memory items, and determine the most relevant keys by calculating the alignment probability, thereby improving the accuracy and speed of retrieval. In addition, the introduction of hidden state updates enables the model to dynamically adjust and optimize its internal state in each retrieval and storage operation, ensuring the continuous updating and optimization of memory.

[0062] Combine the classification results with the context information and update the storage value of the external memory unit.

[0063] The steps for calculating the scene attention weight in this embodiment are:

[0064] By paying attention to the door and each scenario feature The attention mechanism is realized by the association of With the classification result c and the context memory vector m of the previous moment t-1 The interaction between them is used to calculate and obtain the context attention weight Its expression is:

[0065]

[0066] Explanation: Generate a comprehensive intermediate representation vector by combining context features, classification results, and memory from the previous moment

[0067]

[0068] Explanation: The intermediate representation is transformed nonlinearly through a two-layer neural network and the activation function tanh to further extract and process feature information.

[0069]

[0070] In the formula, w (1) ,w (2) represents the weight matrix of the neural network, b (1) , b (2) represents the bias vector of the neural network, M i represents the number of candidates on feature i, represents the intermediate representation vector of feature i at time t, Represents the intermediate representation of feature i after being processed by the activation function tanh.

[0071] It should be noted that the above formula can dynamically adjust and optimize the model's attention to different scenario features, so that important features gain more weight in the memory and classification process. This mechanism ensures that the model can effectively identify and utilize key information, thereby improving classification accuracy and memory efficiency and stability.

[0072] The weights obtained in this embodiment and processed situational features The weighted sum of t , whose expression is: Context vector c t All important situational features are integrated to represent the comprehensive situational information at the current moment. t and the episodic memory vector m at the previous moment t-1 Then through the formula m t =GRU(c t ,m t-1 ) will finally remember m j Transfer to external memory unit.

[0073] It should be noted that the introduction of GRU enables the model to effectively capture the dependencies in the time series, ensuring the continuity and dynamic update of memory. In addition, the use of GRU makes the context memory vector m t It can be updated dynamically to capture dependencies in time series and ensure the consistency and accuracy of memory in long time series.

[0074] More specifically, the context vector c is generated by weighted summation t ,The model can effectively integrate and utilize multiple situational features and improve the ,ability to understand and handle complex situations.

[0075] More specifically, transferring the final memory to the external memory unit significantly expands the model's memory capacity, enabling it to handle larger data sizes and more complex tasks.

[0076] The external memory unit of this embodiment consists of key-value pairs <k1,v1>,...,<k M ,v M >compose memory slots, each <k i ,v i > Both contain embedded representations of visual features and semantic features, extracting the relationship between visual features and semantics. The key and value of the memory slot are obtained by weighted summation of multiple key and value feature vectors. The specific calculation formula is as follows:

[0077]

[0078] in, As the weight, we sum the feature vectors of the key and value to get the key Φ of the memory slot. t (K) is used for key addressing at the next moment to improve the retrieval efficiency and accuracy of memory. The value of the memory slot Φ t (V) is passed as input to the encoder to generate the next category, thereby dynamically updating and optimizing the memory storage.

[0079] It should be noted that: through weighted summation, the model can effectively integrate multiple feature information, generate representative keys and values, and improve the accuracy and richness of memory representation. The dynamic addressing of keys enables the model to quickly find relevant information when it needs to retrieve memory, reducing retrieval time and improving efficiency. The use of external memory units significantly expands the memory capacity of the model, and can flexibly adapt to different types of tasks and data, improving the adaptability and generalization ability of the model. By dynamically updating the value of the memory slot, the model can continuously optimize its memory and classification capabilities and improve overall performance.

[0080] The steps of obtaining the key in this embodiment are:

[0081] The feature vector f of each sample in the category data set extracted by the scene feature encoder t , extract the feature vector f t Organize them in order, and the organized feature vectors is recorded as a key, indicating the visual context relationship between samples, i.e., key k i is the organized feature vector The expression is

[0082] It should be noted that the feature vector is extracted and organized by the scenario feature encoder so that the features of each sample can be arranged in a certain order to form a structured feature representation. This structured feature expression method helps to better capture the relationship between samples and improve the quality of feature representation. i Using , simplifies the process of feature processing and key-value mapping. This direct mapping method reduces intermediate steps, improves processing efficiency, and reduces possible error accumulation. Using , simplifies the process of feature processing and key-value mapping. This direct mapping method reduces intermediate steps, improves processing efficiency, and reduces possible error accumulation.

[0083] The steps of obtaining the value in this embodiment are:

[0084] Each frame image I i The pre-trained model CLIP is used to extract visual and semantic information to form a visual semantic embedding representation. Representation; then, these embeddings and the corresponding keys k i Combine to form a key-value memory slot <k i ,v i >; Among them, key k i is the feature vector extracted from the scene feature encoder, and the value vi is the visual semantic embedding obtained by the CLIP model.

[0085] It should be noted that the use of the CLIP model enables the model to utilize the large amount of knowledge accumulated during the pre-training process to improve the quality and diversity of feature extraction. The structured storage of key-value pairs not only improves the efficiency of information organization and access, but also ensures that high-precision association and retrieval capabilities can be maintained when processing complex multimodal information. The model can find relevant memory items more quickly, improving retrieval efficiency. By leveraging the rich features extracted by the pre-trained model CLIP, the model can show stronger adaptability and generalization capabilities on different tasks and data sets. This helps the model maintain high accuracy and stability when dealing with diverse scenarios in practical applications.

[0086] The steps of key addressing in this embodiment are:

[0087] Use the soft attention mechanism to get the alignment probability qi of each memory, and the key addressing value Φ read at the previous moment t-1 (K) reads past values ​​as input and outputs new hidden states The formula is as follows:

[0088]

[0089] Among them l kis the tanh function. At time t, the classification result C is the decoder hidden state and key-addressable hidden state The weighted sum of is obtained as follows:

[0090]

[0091] According to the obtained classification result C and the key vector k at time i i Through the tanh function, we get the eigenvector k at the i-th moment i Relevance score:

[0092]

[0093] Finally, the Softmax function is used to get the new alignment probability The expression is:

[0094]

[0095] in, For Perform exponential operations, It means that the correlation scores of all N feature vectors are summed after exponential operation. This summation operation is used for normalization to ensure the alignment probability. The sum of is 1.

[0096] It should be noted that: through the soft attention mechanism and Softmax function, important feature vectors can be accurately selected and weighted, which improves the accuracy of feature selection and weighting. By combining the hidden state of the current moment and the memory of the previous moment, the model can better integrate and utilize multiple feature information and improve the representation and classification capabilities of scenario features. Through precise attention allocation and weighted summation, the model can more effectively process large-scale data, improve the efficiency of memory and classification, and reduce unnecessary computational burden.

[0097] During the training process of this embodiment, the situational memory network is trained by combining supervised learning and unsupervised learning. In the initial stage, supervised learning is used to adjust the model parameters through labeled data; then the unsupervised learning stage is entered to adaptively adjust and optimize the model through massive unlabeled data.

[0098] It should be noted that through the initial stage of supervised learning, the model can use the labeled data to adjust parameters. These labeled data provide clear guidance, allowing the model to quickly learn basic patterns and features and establish a solid foundation. This stage ensures the initial accuracy and stability of the model. After the foundation is established in the supervised learning stage, it enters the unsupervised learning stage and is trained with massive unlabeled data. Unsupervised learning enables the model to discover potential structures and patterns in the data, further improving the generalization ability of the model. The model can better adapt and perform when faced with new data. The unsupervised learning stage uses a large amount of unlabeled data, and the model can continuously update and improve its parameters through adaptive adjustment and optimization. In this way, the model can continue to learn and improve in a dynamic environment, improving its performance in different scenarios and tasks.

[0099] The scene feature encoder of this embodiment adopts a multi-scale feature extraction technology when extracting features to ensure comprehensive capture of features at different scales.

[0100] The present invention combines features with larger weights in a sparse coding manner by adopting a competition mechanism to form a scenario feature matrix. Specifically, the features that contribute most to model decision-making are screened out through the competition mechanism, and these features are sparsely processed to improve the efficiency and accuracy of the model. The constructed scenario feature matrix can better capture important information in the scenario and provide high-quality input data for the model.

[0101] The present invention uses the scenario feature encoder as the internal memory unit and uses the synaptic remodeling parameter importance matrix to update the internal memory unit parameters. This design allows the model to dynamically adjust the internal memory unit parameters and update them in real time according to scenario changes, thereby improving the model's learning and adaptability. Through the synaptic remodeling parameter importance matrix, the model can effectively identify and retain features that are critical to scenario understanding, further enhancing the model's ability to handle complex scenarios.

[0102] The present invention uses an external memory unit based on a key-value storage mechanism, which consists of key-value pairs to form memory slots, and aligns the extracted visual features and semantic features according to the context information. The key-value storage mechanism provides an efficient storage and retrieval method, where the key corresponds to the key information in the context and the value is the detailed feature data related to it. In this way, the model can not only remember the key information and store it for a long time, but also efficiently retrieve and use this information when needed.

[0103] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A method for designing an episodic memory network based on a synaptic remodeling model, characterized in that: The following steps are involved: Design the context feature encoder as an internal memory unit and use the synaptic remodeling parameter importance matrix to update the internal memory unit parameters; Extract visual features and semantic features of image data from samples to be classified; compete the extracted features, select features with large weights for sparse coding, and form a scenario feature matrix ; Focus on the subset of images that can mark the scene features and extract the scene feature matrix from the image Retrieve information in Attention mechanism uses gate function For candidate category samples Processing to obtain scenario features , and calculate the scene attention weight through multiple features, and calculate the context vector according to the obtained scene attention weight; Through the context vector and the episodic memory vector of the previous moment To update episodic memory , and will eventually remember Transfer to external memory unit; The design is based on the key-value storage mechanism as an external memory unit, extracts the relationship between visual features and semantic features, forms key-value pairs, and uses the soft attention mechanism for key addressing and hidden state update; Combine the classification results with the context information and update the storage value of the external memory unit; The external memory unit consists of key-value pairs to form memory slots, which extract the relationship between visual features and semantics. It is obtained by weighted summing of multiple key and value feature vectors. The specific calculation formula is as follows: in, As the weight, the feature vectors of the key and value are weighted and summed to get the key of the memory slot. Used for key addressing at the next moment, the value of the memory slot Passed as input to the encoder to produce the next category; The steps to get the key are: The feature vector of each sample in the category data set extracted by the scene feature encoder , the extracted feature vector Organize them in order, and the organized feature vectors is recorded as a key, indicating the visual context relationship between samples, i.e., key is the organized feature vector , the expression is ; The steps to get the value are: Each frame of image The pre-trained model CLIP is used to extract visual and semantic information to form a visual semantic embedding representation. Representation; then, these embeddings and the corresponding keys Combine to form a key-value memory slot ; Among them, the key is the feature vector extracted from the scene feature encoder, and the value It is the visual semantic embedding obtained by the CLIP model; The steps for key addressing are: Use soft attention mechanism to get the alignment probability of each memory , the key addresses the value read at the previous moment Read past values ​​as input and output new hidden states The formula is as follows: in is the tanh function; Moment, classification results is the decoder hidden state and key-addressable hidden state The weighted sum of is obtained as follows: According to the classification results obtained and time The key vector Through the tanh function, we get Moment feature vector Relevance score: Finally, the Softmax function is used to get the new alignment probability , the expression is: in, For Perform exponential operations, Indicates that all The correlation scores of the feature vectors are exponentially calculated and then summed. This summation operation is used for normalization to ensure the alignment probability The sum of is 1.

2. The episodic memory network design method based on synaptic remodeling model according to claim 1, characterized in that: The steps to calculate the context attention weight are: By paying attention to the door and each scenario feature The attention mechanism is realized by the association of And the classification results and the episodic memory vector of the previous moment The interaction between them is used to calculate and obtain the context attention weight , whose expression is: In the formula, , represents the weight matrix of the neural network, , represents the bias vector of the neural network, Indicated in the feature The number of candidates on Indicates at time Moment, Features The intermediate representation vector of Represents the features after being processed by the activation function tanh The intermediate representation of .

3. The episodic memory network design method based on synaptic remodeling model according to claim 2, characterized in that: The weight obtained and processed situational features The weighted sum of , whose expression is: , after getting the context vector and the episodic memory vector of the previous moment Then through the formula The final memory Transfer to external memory unit.

4. The episodic memory network design method based on synaptic remodeling model according to claim 3, characterized in that: During the training process, the episodic memory network is trained through a combination of supervised learning and unsupervised learning. In the initial stage, supervised learning is used to adjust the model parameters through labeled data; then the unsupervised learning stage is entered to adaptively adjust and optimize the model through massive unlabeled data.

5. The episodic memory network design method based on synaptic remodeling model according to claim 4, characterized in that: When extracting features, the scene feature encoder uses multi-scale feature extraction technology to ensure comprehensive capture of features at different scales.

Citation Information

Patent Citations

  • Robot skill acquisition method based on meta-learning under guidance of scene memory

    CN113657573A

  • Intelligent question and answer method and device, device and storage medium

    CN114064864A

  • Scene memory guided robot operation skill efficient learning method

    CN114161419A

  • Devices and methods for using mechanical affective touch therapy to reduce stress, anxiety and depression

    US20220088345A1