A robot memory reinforcement method based on a memory palace algorithm

Abstract

The application discloses a robot memory reinforcement method based on a memory palace algorithm, and relates to the technical field of machine learning.The method comprises the following steps: acquiring a memory scene uploaded by a user and outputting a memory subject set; performing entity recognition on the memory scene and obtaining a plurality of entities, analyzing the correlation degree of the entities with the memory subjects to filter out a set of labeled entities; analyzing historical memory data of each memory subject to obtain a memory difficulty coefficient, and combining the correlation degree to calculate a memory resource coefficient; selecting a corresponding model according to the memory resource coefficient, generating visual labeling elements that fuse the memory subjects and the labeled entities; labeling the labeling elements at the entity coordinates, generating a reinforced memory palace scene and showing the user the scene.The application fuses the memory palace and machine learning to construct a personalized memory environment for the user, effectively solves the problems of the traditional memory training method, such as being dull, low in efficiency and lacking in correlation, and improves the memory efficiency and long-term memory retention rate of the user.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and specifically to a method for enhancing robot memory based on the memory palace algorithm. Background Technology

[0002] The memory palace technique, a long-established memory enhancement method, effectively leverages humans' exceptional spatial and visual memory capabilities to improve memory efficiency by associating abstract information with specific locations within familiar spatial scenes. In recent years, with the development of artificial intelligence, applying such efficient memory mechanisms to assisted training systems has become a research hotspot. However, existing memory aids are mostly simple labels or static images, lacking deep spatial association and dynamic generation capabilities.

[0003] Currently, existing memory aids have the following limitations: First, they use fixed templates and rules, making it difficult to adapt to users' personalized needs; second, they rely on users to build memory associations themselves, lacking automated intelligent analysis and generation capabilities; and third, they do not dynamically adjust strategies based on historical memory data, making it impossible to achieve adaptive reinforcement training for different memory difficulties, ultimately resulting in limited effectiveness and a poor user experience. Summary of the Invention

[0004] This invention addresses the technical problems of existing robot memory-assisted training, such as insufficient intelligence, low memory efficiency, poor correlation, weak adaptability, and inability to dynamically adjust training strategies according to users' personalized needs and memory content. It provides a robot memory enhancement method based on the memory palace algorithm.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This invention provides a robot memory enhancement method based on the memory palace algorithm, comprising: Obtain the memory scenes uploaded by the user and output the memory subject set, wherein the memory subject set includes the number of memory subjects; Entity recognition is performed on the memory scene to obtain multiple entities, and the correlation between each entity and multiple memory subjects is analyzed. Entities with the number of memories are selected and used as the labeled entity set. Analyze the memory difficulty coefficient of each memory subject to obtain a set of memory difficulty coefficients. Combine the correlation between multiple labeled entities and multiple memory subjects to configure memory resources. Based on the set of memory subjects and the set of labeled entities, use machine learning to generate multiple labeled elements. By labeling multiple elements at the corresponding entity coordinates of the labeled entity set, a generated memory scene based on the memory palace is obtained and displayed to the user.

[0006] The beneficial effects of this invention are: Compared to existing technologies, this invention firstly introduces the memory palace algorithm combined with real-world scenarios provided by the user to construct a highly personalized memory training environment, effectively solving the problem of weak adaptability in traditional methods. Secondly, it employs machine learning methods to dynamically analyze the semantic correlation between the memory subject and scene entities, and adaptively calculates the memory difficulty coefficient based on the user's historical memory data, achieving intelligent configuration of training strategies. Thirdly, through a deep learning-based generative model, it automatically creates reinforcement annotation elements with visual impact and emotional relevance, significantly improving the attractiveness and memorability of the memory content. Finally, by precisely embedding the generated reinforcement elements into the entity coordinates of the user's scene, it forms intuitive spatial memory cues, providing users with an efficient training path that conforms to the laws of memory, ultimately significantly improving the efficiency, persistence, and personalized experience of memory training. Attached Figure Description

[0007] Figure 1 A flowchart illustrating a robot memory enhancement method based on the memory palace algorithm provided by this invention; Figure 2 This is a logical diagram illustrating a robot memory enhancement method based on the memory palace algorithm provided by the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0010] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0011] Example 1, as Figure 1 , Figure 2 As shown, this embodiment of the invention provides a robot memory enhancement method based on the memory palace algorithm, including: S10: Obtain the memory scenes uploaded by the user and output the memory subject set, wherein the memory subject set includes the number of memory subjects; Specifically, it retrieves user-uploaded memory scenarios and outputs a memory subject set, including: Retrieve user-uploaded memory scenes, where the memory scene is a spatial scene; Output the set of memory subjects to be memorized, wherein the set of memory subjects includes the number of memory subjects to be memorized.

[0012] First, the system acquires user-uploaded memory scenes. Memory scenes refer to spatial scene data containing spatial structure information, typically in the form of two-dimensional floor plans or three-dimensional models. This data includes elements such as spatial layout, object locations, and environmental features, providing the spatial foundation for the subsequent construction of the memory palace. Memory scenes originate from spatial scenes actively provided by users, such as photos of their actual rooms. Uploading these to the robot provides the environmental basis and data source for building a personalized memory palace.

[0013] Secondly, the system outputs the memory subject set to be memorized. The memory subject set refers to the collection of information units that need to be reinforced, containing a specific number of independent memory items. Each independent memory item is a specific object or concept to be memorized, such as a single word from a vocabulary list (banana, milk, sneakers), or a set of abstract information such as key points of an event or a sequence of numbers. The generation of the memory subject set is based on the user's input of memory needs or key information automatically extracted by the robot, providing clear target content for subsequent association analysis and memory reinforcement.

[0014] In summary, by acquiring memory scenario data provided by users, an environmental framework is provided for the subsequent construction of a memory palace based on spatial location. At the same time, the output memory subject set clarifies the specific content goals that need to be strengthened, providing clear operational objects for subsequent association analysis, difficulty assessment, and resource allocation.

[0015] S20: Perform entity recognition on the memory scene to obtain multiple entities, analyze the correlation with multiple memory subjects respectively, and filter to obtain the number of entities in the memory as the labeled entity set; Specifically, entity recognition is performed on the memory scene to obtain multiple entities, and the correlation between each entity and multiple memory subjects is analyzed. Entities with a certain number of memories are then selected as the labeled entity set, including: Construct an entity recognizer, input the memory scene into the entity recognizer, and recognize and output multiple entities; Multiple entities and each memory subject within the memory subject set are combined and input into the association analysis table. Multiple association degrees are output. The association analysis table is constructed by combining the sample entities and sample memory subjects with the labeled sample association degrees. The association degree is greater than or equal to 0 and less than or equal to 1. Select the entity with the highest relevance to each memory subject as the labeled entity to obtain the labeled entity set.

[0016] First, the steps for constructing the entity recognizer include: Collect a set of sample memory scenes, and annotate each pixel in each sample memory scene to obtain multiple sets of sample entities; Based on semantic segmentation, we construct the basic architecture for entity recognizers; The entity recognizer is trained under supervision using the sample memory scene set and multiple sample entity sets. After the training and testing converge, the recognition is completed.

[0017] An entity recognizer is a computer vision model based on deep learning algorithms. It takes memorized scene data as input and, through image recognition and semantic segmentation techniques, automatically identifies and outputs all entity objects contained within the scene. The entity recognizer establishes a mapping relationship between scene images and entity annotations, enabling precise pixel-level recognition and classification extraction of various objects within the memorized scene.

[0018] First, training data preparation is performed. Sample memory scenes are collected to form a sample memory scene set, and each pixel in each sample memory scene is finely annotated as an entity. A sample memory scene refers to two-dimensional or three-dimensional environmental data containing various entity objects with a clear spatial layout, specifically represented by indoor scene images, architectural floor plans, or 3D point cloud models. Entity annotation refers to the operation of assigning specific category labels to each pixel in the image. Category labels include, but are not limited to, specific entity types such as furniture, appliances, decorations, and building structures. The annotation process must clearly define the entity category to which each pixel belongs, thereby obtaining a sample entity set that corresponds one-to-one with the sample scene. The final data with precise annotations provides a basis for supervised training, ensuring the generalization ability and recognition accuracy of the entity recognizer.

[0019] Next, the network model is constructed. For example, since the relationship between entity categories and pixel distribution in memory scenes often presents a complex spatial hierarchy, and convolutional neural networks have outstanding advantages in extracting local features and capturing spatial semantic information, a convolutional neural network model is chosen to construct this entity recognizer.

[0020] Specifically, this entity recognizer employs an encoder-decoder architecture. The encoder consists of multiple stacked convolutional and pooling layers, using 3×3 convolutional kernels to extract spatial features at different scales and enhancing non-linear expressiveness through the ReLU activation function. The decoder gradually restores the feature map resolution through deconvolutional layers and upsampling operations, ultimately outputting pixel-level class predictions of the same size as the input image. A skip connection mechanism is introduced into the network to fuse the high-resolution features of the encoder with the semantic features of the decoder, improving the recognition accuracy of small-sized entities.

[0021] During training, key hyperparameters included a learning rate of 0.001, a training epoch count of 200, and a batch size of 32. The learning rate ensured the stability of gradient descent, the number of training epochs ensured sufficient learning of feature extraction capabilities, and the batch size balanced memory usage with training efficiency. Specifically, a supervised learning approach was adopted, constructing training data based on a set of memorized scene images and a set of sample entities. Scene images served as input features, and pixel-level entity annotations served as multi-class labels. All data was randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0022] Next, scene images from the training set are input into the network, with the corresponding entity annotation maps used as supervision signals. The network parameters are iteratively updated using the Adam optimizer via backpropagation. The cross-entropy loss function is used to quantify the difference between the predicted results and the actual annotations, and the training process is monitored using a validation set. Training is terminated when the average intersection-union ratio (IUU) of the validation set no longer improves after several consecutive rounds and reaches a preset accuracy threshold, such as 90%, resulting in a finally converged entity recognizer. This entity recognizer can accurately capture the spatial distribution characteristics of entity objects in the scene, providing reliable entity recognition results for subsequent association analysis.

[0023] Furthermore, the memory scene is input into a trained entity recognizer, which parses the scene using semantic segmentation technology and outputs multiple entities contained in the scene, such as television, dining table, chair, and sofa. Then, each identified entity is paired with each memory subject in the memory subject set, forming multiple entity-memory subject pairs. These pairs are then input into a pre-generated association analysis table for query matching. The association analysis table is essentially an association degree mapping database, storing a large number of manually labeled sample entity and sample memory subject combinations and their corresponding association degree scores. The association degree is used to quantify the semantic relevance strength between sample entities and sample memory subjects, and its value ranges from [0, 1]. The closer the value is to 1, the stronger the association; the closer the value is to 0, the weaker the association. For example, the association degree of wardrobe-clothes is 0.9, and the association degree of wardrobe-shoes is 0.6.

[0024] Finally, for each memory subject in the memory subject set, all entities are traversed, and the entity with the highest correlation score is selected as the labeled entity corresponding to that memory subject. All selected labeled entities together constitute the labeled entity set, which establishes the optimal mapping relationship between memory content and spatial location, providing core data support for generating visual labeled elements in the memory palace. For example, the correlation score between wardrobe and clothes is 0.9, and the correlation score between wardrobe and shoes is 0.6, so clothes are selected as the labeled entity for the memory subject wardrobe.

[0025] S30: Analyze the memory difficulty coefficient of each memory subject to obtain a set of memory difficulty coefficients. Combine the correlation between multiple labeled entities and multiple memory subjects to configure memory resources. Based on the set of memory subjects and the set of labeled entities, use machine learning to generate multiple labeled elements. Analyze the memory difficulty coefficient of each memory subject to obtain a set of memory difficulty coefficients. Combine the correlation between multiple labeled entities and multiple memory subjects to configure memory resources. Based on the set of memory subjects and the set of labeled entities, machine learning is used to generate multiple labeled elements, including: Obtain the memory history data of each memory subject within a historical time period to obtain a set of memory results for multiple memory subjects, where each memory subject's memory result includes yes or no; Calculate the percentage of no memory results in the memory result set for each memory subject to obtain the set of memory difficulty coefficients; By combining the correlation between multiple memory subjects and multiple labeled entities, as well as the set of memory difficulty coefficients, multiple memory resource coefficients are calculated. Based on multiple memory resource coefficients, select multiple corresponding annotation element generation models within the annotation element generation model group, combine the memory subject set and the annotation entity set respectively, input them into the corresponding annotation element generation model, and output multiple annotation elements.

[0026] After obtaining the labeled entity set, it is also necessary to analyze the memory difficulty coefficient of each memory subject. Specifically, by acquiring the historical memory data of each memory subject, the percentage of memory failures is calculated, thus forming a memory difficulty coefficient set. This memory difficulty coefficient set, together with the aforementioned correlation coefficient, constitutes the decision-making basis for memory resource allocation.

[0027] First, memory history data for each subject within a historical time period is acquired, forming a memory result set for each subject. Each memory result includes "yes" or "no," representing success or failure in a binary format. Memory history data refers to the sequence of user feedback or system-automated evaluation results recorded by the robot in past memory tasks; the memory result refers to the success or failure judgment of each memory attempt for a specific subject, recorded as "yes" for success and "no" for failure. Then, based on the historical data, the proportion of "no" (failure) memories for each subject out of the total number of attempts is calculated, serving as the memory difficulty coefficient for that subject. The memory difficulty coefficient = number of failures / total number of attempts, used to quantitatively characterize the difficulty of long-term stable memorization for that subject. The difficulty coefficients of all subjects are then compiled into a memory difficulty coefficient set.

[0028] Furthermore, by integrating the correlation between multiple memory subjects and multiple labeled entities, as well as the set of memory difficulty coefficients, multiple memory resource coefficients are calculated. Memory resource coefficient = [(1 - correlation) + memory difficulty coefficient] / 2. Memory subject refers to the specific information unit that needs to be reinforced, such as vocabulary, concepts, or key points of events; the correlation of a labeled entity refers to the quantitative value of the semantic relevance between the entity and the memory subject, with a value range of [0, 1]; the memory difficulty coefficient refers to the quantitative index of the difficulty of memorizing the memory subject, calculated based on the historical memory failure rate, with a value range of [0, 1]. The memory resource coefficient is obtained by averaging the correlation complement (1 - correlation) with the memory difficulty coefficient. This memory resource coefficient is used to comprehensively characterize the intensity of memory reinforcement resources that should be allocated to the memory subject-labeled entity combination. The higher the value, the more computational resources and generation complexity are required to ensure the memory effect.

[0029] Finally, based on the numerical range of the memory resource coefficient, a matching annotation element generation model is selected from the pre-trained annotation element generation model group. Different annotation element generation models have different complexities and expressiveness; the higher the resource coefficient, the more complex the annotation element generation model is selected. After pairing and combining the memory subject with the annotation entity, the input is fed into the selected annotation element generation model, which can output annotation elements with different feature strengths. This achieves high-resource reinforcement of difficult-to-remember content and low-resource processing of easy-to-remember content, thus optimizing the user's overall memory efficiency.

[0030] Specifically, based on multiple memory resource coefficients, multiple corresponding annotation element generation models are selected from the annotation element generation model group. The memory subject set and the annotation entity set are combined accordingly and input into the corresponding annotation element generation model. The output yields multiple annotation elements, including: Based on deep learning, multiple annotation element generation models are trained to obtain a group of annotation element generation models, in which the amount of training data for each annotation element generation model is arranged from largest to smallest. Based on the number of models generated by multiple labeled elements, multiple resource coefficient intervals are constructed and mapped and associated with multiple labeled element generation models. Among them, the labeled element generation model with more training data corresponds to a larger resource coefficient interval. Based on the resource coefficient range that each memory resource coefficient falls into, select the corresponding annotation element to generate the model, combine the corresponding memory subject and annotation entity and input them, and output multiple annotation elements.

[0031] First, based on deep learning, multiple annotation element generation models are trained to obtain a group of annotation element generation models, including: A network architecture for generating multiple labeled elements based on deep learning was constructed. Collect the sample memory subject set and the sample labeled entity set, and randomly combine them to obtain the sample labeled memory combination set. Obtain the sample labeled elements of each sample labeled memory combination to obtain the sample labeled element set. The sample annotation memory set and the sample annotation element set are split multiple times to obtain multiple sets of annotation element training data, in which the data volume of the multiple sets of annotation element training data is arranged from largest to smallest; The multiple sets of labeled element training data are used to supervise the training of multiple labeled element generation models until convergence, thus obtaining a group of labeled element generation models.

[0032] The annotation element generation model is a generative artificial intelligence model based on a deep learning architecture. Its core function is to transform abstract memory subjects and spatial entities into concrete and visual memory reinforcement elements. This annotation element generation model learns from a large amount of sample data how to fuse textual or conceptual memory subjects (such as "banana") with image-type annotation entities (such as "table") to generate annotation elements with visual salience, emotional appeal, or absurdity (e.g., "a giant banana dancing on a table"). The specific training steps of the annotation element generation model are as follows: First, a network architecture for multiple annotation element generation models is constructed based on deep learning technology. Each annotation element generation model can adopt the same initial network structure configuration, or different architecture designs with varying complexity can be adopted according to expected performance requirements, such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or Diffusion Models.

[0033] For example, Generative Adversarial Networks (GANs) are chosen as the basic architecture for the annotation element generation model because of their significant advantages in generating high-resolution, highly realistic images, and because their discriminator-generator adversarial training mechanism can effectively guarantee the visual quality and diversity of the generated annotation elements. In this network architecture, the generator is responsible for fusing the features of the memorized main text with the features of the annotated entity image to generate the target annotation elements, while the discriminator distinguishes the differences between the generated elements and the real annotation elements. Through continuous adversarial optimization, the generation quality is continuously improved.

[0034] Specifically, this generative adversarial network employs a conditional generative architecture. The generator part uses a U-Net structure, whose input is a concatenated vector of labeled entity image features and remembered subject text features. It gradually generates target labeled elements through multiple layers of transposed convolutions and upsampling operations. The discriminator part uses a PatchGAN structure to perform local perception-based realism discrimination on the generated images, improving the realism of the generated details. An attention mechanism is introduced into the network, enabling the generator to dynamically focus on key association regions between the remembered subject and the labeled entity, enhancing the semantic consistency of the generated elements.

[0035] In the training dataset construction phase, firstly, a set of sample memory entities and a set of sample labeled entities are collected as basic data. These two types of sample data are then randomly paired and combined to form a sample labeled memory combination set. For each combination, corresponding sample labeled elements are manually labeled or generated, thereby obtaining a complete set of sample labeled elements. This set of sample labeled elements provides a reference for real values ​​in supervised training.

[0036] Secondly, the sample annotation memory set and sample annotation element set are partitioned multiple times to generate multiple annotation element training datasets with decreasing data volume. To build a differentiated model group, the annotation element training datasets are divided according to the proportion of data volume from most to least. The training set with the largest proportion of data volume contains all or most of the sample data, while the training set with the smallest proportion of data volume contains only a small number of core samples. Each subset is used to independently train an annotation element generation model.

[0037] The training dataset uses labeled elements with different proportions, such as 20%, 40%, and 60%, to independently train multiple labeled element generation models. During training, key hyperparameters include a learning rate of 0.0002 (the generator and discriminator use the same learning rate), 500 training epochs, and a batch size of 16. The learning rate setting follows common GAN training configurations, the number of training epochs ensures sufficient adversarial optimization between the generator and discriminator, and the batch size balances training stability and memory efficiency.

[0038] Specifically, a supervised learning approach is adopted, constructing training data based on a set of sample-labeled memory combinations and a set of sample-labeled elements. The memory subject-labeled entity pairs serve as conditional inputs, while manually labeled elements are the true output targets. All data is randomly divided into training, validation, and test sets in a ratio of 7:2:1. The memory subject-labeled entity pairs from the training set are input into the generator, with the corresponding labeled elements serving as supervision signals. The network parameters are optimized by minimizing the generative adversarial loss function. The Adam optimizer is used to alternately update the generator and discriminator parameters, and a learning rate decay strategy is employed to refine model performance in the later stages of training until the training loss converges and the validation set performance stabilizes. Training is then terminated, yielding the finally converged labeled element generation model. This labeled element generation model accurately captures the semantic relationships between the memory subject and the labeled entities, generating visual elements with significant memory enhancement effects.

[0039] Similarly, each model iteratively optimizes its parameters on the corresponding training set using supervised learning until the training loss converges and the validation set performance stabilizes. Through this tiered training approach, a group of labeled element generation models with different generation capabilities and complexities is ultimately obtained, providing a technical foundation for subsequent model selection based on resource coefficients.

[0040] Furthermore, based on the number of annotation element generation models in the group, the resource coefficient range is divided into several continuous intervals, such as 0-20% and 20-40%, with each interval mapped to a specific annotation element generation model. It should be noted that annotation element generation models with larger training data volumes have higher representational capabilities and generation complexity, and are therefore mapped to larger resource coefficient intervals to handle more difficult combinations with lower correlation or higher memorization difficulty. Therefore, when the memory resource coefficient is high, the more capable annotation element generation model will be automatically selected for processing.

[0041] In practice, based on the memory resource coefficient value corresponding to each memory subject-label entity pair, its resource coefficient range is determined, and a corresponding label element generation model is selected. The memory subject and label entity combination to be processed are input into the selected label element generation model, which then outputs the corresponding label elements. In this way, a precise match between memory resources and model processing capabilities is achieved, ensuring that high resource coefficient combinations yield more refined label element generation results.

[0042] S40: Label multiple annotation elements at the entity coordinates corresponding to the annotation entity set to obtain a generated memory scene based on the memory palace and display it to the user.

[0043] Multiple labeled elements are marked at the entity coordinates corresponding to the labeled entity set to obtain a generated memory scene based on the memory palace, which is then displayed to the user, including: Obtain the coordinates of multiple labeled entities within the memory scene; Multiple annotation elements are labeled at the entity coordinates of each corresponding annotation entity within the annotation entity set to obtain the generated memory scene, which is then displayed to the user.

[0044] First, the specific coordinates of each entity in the original memory scene are obtained from the labeled entity set. These coordinates are determined by pixel-level positioning information generated during the entity recognition stage, providing a spatial reference for the precise placement of the labeled elements. Second, the previously generated labeled elements are spatially bound to their corresponding labeled entities, accurately labeling each element at the coordinates of the corresponding entity. The labeling process must maintain visual harmony between the elements and the scene, ensuring that the labeled elements are highlighted without disrupting the structural integrity of the original scene.

[0045] The final generated memory palace scene retains the layout features of the original spatial environment while enhancing the visual salience of memory points through labeled elements. The completed memory palace scene is then presented to users through a visual interface, providing intuitive spatial memory cues to strengthen and consolidate the memory content. This transforms abstract memory associations into a concrete spatial visualization, completing the final transformation from data processing to user interaction.

[0046] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this invention firstly constructs a highly personalized memory training environment by integrating the memory palace algorithm with user-defined scenarios, effectively overcoming the shortcomings of traditional methods in terms of adaptability. Secondly, it uses machine learning technology to dynamically analyze the semantic correlation between the memory subject and scene entities, and adaptively calculates the memory difficulty coefficient based on the user's historical memory performance, achieving precise and intelligent configuration of training strategies. Thirdly, relying on a deep learning-based generative model, it automatically constructs reinforcement annotation elements with visual saliency and emotional relevance features, significantly enhancing the attractiveness of the memory content and the memory retention effect. Finally, by accurately mapping the generated high-quality reinforcement elements to the entity coordinates in the user's scene, it forms intuitive spatial memory cues, providing users with an efficient training path that conforms to cognitive laws, thereby comprehensively improving the efficiency, persistence, and personalization of user memory training.

[0047] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0048] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0049] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for enhancing robot memory based on the memory palace algorithm, characterized in that, The method includes: Obtain the memory scenes uploaded by the user and output the memory subject set, wherein the memory subject set includes the number of memory subjects; Entity recognition is performed on the memory scene to obtain multiple entities, and the correlation between each entity and multiple memory subjects is analyzed. Entities with the number of memories are selected and used as the labeled entity set. Analyze the memory difficulty coefficient of each memory subject to obtain a set of memory difficulty coefficients. Combine the correlation between multiple labeled entities and multiple memory subjects to configure memory resources. Based on the set of memory subjects and the set of labeled entities, machine learning is used to generate multiple labeled elements, including: Obtain the memory history data of each memory subject within a historical time period to obtain a set of memory results for multiple memory subjects, where each memory subject's memory result includes yes or no; Calculate the percentage of no memory results in the memory result set for each memory subject to obtain the set of memory difficulty coefficients; By combining the correlation between multiple memory subjects and multiple labeled entities, as well as the set of memory difficulty coefficients, multiple memory resource coefficients are calculated. Based on multiple memory resource coefficients, select multiple corresponding annotation element generation models from the annotation element generation model group, combine the memory subject set and the annotation entity set accordingly, input them into the corresponding annotation element generation model, and output multiple annotation elements, including: Based on deep learning, multiple annotation element generation models are trained to obtain a group of annotation element generation models, in which the amount of training data for each annotation element generation model is arranged from largest to smallest. Based on the number of models generated by multiple labeled elements, multiple resource coefficient intervals are constructed and mapped and associated with multiple labeled element generation models. Among them, the labeled element generation model with more training data corresponds to a larger resource coefficient interval. Based on the resource coefficient range that each memory resource coefficient falls into, select the corresponding annotation elements to generate the model, combine the corresponding memory subject and annotation entity and input them, and output multiple annotation elements; By labeling multiple elements at the corresponding entity coordinates of the labeled entity set, a generated memory scene based on the memory palace is obtained and displayed to the user.

2. The robot memory reinforcement method based on the memory palace algorithm according to claim 1, characterized in that, Obtain the memory scenes uploaded by the user and output the memory subject set, including: Retrieve user-uploaded memory scenes, where the memory scene is a spatial scene; Output the set of memory subjects to be memorized, wherein the set of memory subjects includes the number of memory subjects to be memorized.

3. The robot memory reinforcement method based on the memory palace algorithm according to claim 1, characterized in that, Entity recognition is performed on the memory scene to obtain multiple entities. The correlation between each entity and the multiple memory subjects is analyzed, and entities representing the number of memories are selected as the labeled entity set, including: Construct an entity recognizer, input the memory scene into the entity recognizer, and recognize and output multiple entities; Multiple entities and each memory subject within the memory subject set are combined and input into the association analysis table. Multiple association degrees are output. The association analysis table is constructed by combining the sample entities and sample memory subjects with the labeled sample association degrees. The association degree is greater than or equal to 0 and less than or equal to 1. Select the entity with the highest relevance to each memory subject as the labeled entity to obtain the labeled entity set.

4. The robot memory enhancement method based on the memory palace algorithm according to claim 3, characterized in that, The steps for constructing the entity recognizer include: Collect a set of sample memory scenes, and annotate each pixel in each sample memory scene to obtain multiple sets of sample entities. Based on semantic segmentation, we construct the basic architecture for entity recognizers; The entity recognizer is trained under supervision using the sample memory scene set and multiple sample entity sets. After the training and testing converge, the recognition is completed.

5. The robot memory reinforcement method based on the memory palace algorithm according to claim 1, characterized in that, Based on deep learning, multiple annotation element generation models are trained to obtain a group of annotation element generation models, including: A network architecture for generating multiple labeled elements based on deep learning was constructed. Collect the sample memory subject set and the sample labeled entity set, and randomly combine them to obtain the sample labeled memory combination set. Obtain the sample labeled elements of each sample labeled memory combination to obtain the sample labeled element set. The sample annotation memory set and the sample annotation element set are split multiple times to obtain multiple sets of annotation element training data, in which the data volume of the multiple sets of annotation element training data is arranged from largest to smallest; The multiple sets of labeled element training data are used to supervise the training of multiple labeled element generation models until convergence, thus obtaining a group of labeled element generation models.

6. The robot memory enhancement method based on the memory palace algorithm according to claim 1, characterized in that, Multiple labeled elements are marked at the entity coordinates corresponding to the labeled entity set to obtain a generated memory scene based on the memory palace, which is then displayed to the user, including: Obtain the coordinates of multiple labeled entities within the memory scene; Multiple annotation elements are labeled at the entity coordinates of each corresponding annotation entity within the annotation entity set to obtain the generated memory scene, which is then displayed to the user.

Images