Artificial intelligence training method and device based on selective memory reinforcement, and medium

Abstract

The present disclosure provides an artificial intelligence training method and device based on selective memory reinforcement and a medium, relating to the technical field of artificial intelligence. The method comprises: acquiring multi-modal interaction data of an interaction object; encoding the multi-modal interaction data to generate corresponding interaction feature vectors, and calculating an emotional energy value and a valence of a current interaction event; calculating an importance score of the current interaction event based on the interaction feature vectors, the emotional energy value and the valence, and performing selective memory storage according to the importance score to generate structured memory data; generating multi-modal feedback instructions based on the interaction feature vectors, the structured memory data and a predefined personality vector; executing the multi-modal feedback instructions, and updating the personality vector based on response data of the interaction object to the execution result of the feedback instructions. The present disclosure can realize efficient memory management according to the importance score, and the real-time updated personality vector enables the AI to form an interaction style adapted to the user.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to an artificial intelligence training method, apparatus and medium based on selective memory reinforcement. Background Technology

[0002] Currently, companion-style AI systems built on large language models still have significant shortcomings in data processing, such as human-computer interaction data processing and memory information storage. These systems typically store all interaction data in a memory bank for retrieval, leading to a bloated memory bank, reduced retrieval efficiency, and an inability to simulate the human mechanism for reinforcing important memories. Furthermore, their interaction data management and retrieval logic is simplistic and rigid. Simultaneously, the personality of such AI systems is often defined by static prompts, lacking the ability to evolve with long-term user interaction, resulting in a lack of personalized growth and emotional depth in the interaction. Summary of the Invention

[0003] To overcome the problems existing in related technologies, this disclosure provides an artificial intelligence training method, device and medium based on selective memory reinforcement.

[0004] According to a first aspect of the present disclosure, an artificial intelligence training method based on selective memory reinforcement is provided, comprising: Obtain multimodal interaction data of interactive objects; The multimodal interaction data is encoded to generate corresponding interaction feature vectors, and the emotional energy value and valence of the current interaction event are calculated. Based on the interaction feature vector, emotional energy value, and valence, the importance score of the current interaction event is calculated, and selective memory storage is performed on the feature data corresponding to the current interaction event according to the matching result of the importance score and the preset threshold, generating structured memory data. By integrating the interaction feature vector, structured memory data, and predefined personality vector, a behavioral decision vector is generated, wherein the personality vector is used to characterize the personality traits of companion-type artificial intelligence; Based on the aforementioned behavior decision vector, multimodal feedback instructions are generated; The multimodal feedback instruction is executed, and the personality vector is updated based on the response data of the interaction object to the execution result of the feedback instruction.

[0005] In some embodiments, calculating the importance score of the current interaction event based on the interaction feature vector, emotional energy value, and valence includes: The importance score of the current interaction event is calculated using the following formula: e=w1×Emotion+w2×TopicRelevance Where e is the importance score of the current interaction event, w1 and w2 are preset weight coefficients; Emotion is the emotion score of the current interaction event, which is related to the emotional energy value and valence of the current interaction event; TopicRelevance is the relevance score between the current interaction event and the topic of interest of the interaction object.

[0006] In some embodiments, the preset threshold includes: a high importance threshold and a low importance threshold; The step of selectively storing feature data corresponding to the current interaction event based on the matching result of the importance score and the preset threshold, and generating structured memory data, includes: If the importance score of the current interaction event is greater than the high importance threshold, the feature data corresponding to the current interaction event will be stored in the long-term memory pool. If the importance score of the current interaction event is greater than or equal to the low importance threshold and less than the high importance threshold, the feature data corresponding to the current interaction event is stored in the intermediate memory pool, and after the storage time reaches the preset time, a natural forgetting operation is performed. If the importance score of the current interaction event is less than the low importance threshold, the feature data corresponding to the current interaction event is stored in the short-term memory pool, and a natural forgetting operation is performed along with the dialog sliding window. The structured memory data is generated based on the feature data stored in the long-term memory pool, the medium-term memory pool, and the short-term memory pool.

[0007] In some embodiments, after storing the feature data corresponding to the current interaction event into a long-term memory pool, the method further includes: Clustering operations are performed on the feature data stored in the long-term memory pool to integrate discrete interaction event feature data into user cognitive feature data, thus completing the memory distillation operation.

[0008] In some embodiments, updating the personality vector based on the response data of the interactive object to the result of the feedback instruction includes: The personality vector is updated using the following formula: P(t+1)=P(t)+α×R×(U(t)-P(t)) Where P(t) is the personality vector before the update, and P(t+1) is the personality vector after the update; α is the preset fitness coefficient; R is the feedback value generated based on the response data; U(t) is the interaction feature vector corresponding to the current interaction event; and t is the time.

[0009] In some embodiments, the multimodal feedback instructions include: voice feedback instructions and motion feedback instructions; The generation of multimodal feedback instructions based on the behavioral decision vector includes: The behavior decision vector is input into the large language model to generate text response content corresponding to the behavior decision vector; the speech output parameters of the speech synthesis model are adjusted according to the behavior decision vector, and the text response content is converted into a speech feedback command corresponding to the current interaction event through the adjusted speech synthesis model. Based on the behavior decision vector, a matching servo motor instruction sequence is retrieved from a predefined embodied action library to generate an action feedback instruction.

[0010] In some embodiments, the multimodal interaction data includes: voice data, image data, and environmental data; The process of encoding the multimodal interaction data to generate corresponding interaction feature vectors includes: Extract semantic information and acoustic features from the speech data to generate a speech feature vector; Extract facial features and body movement features of the interactive objects from the image data to generate an image feature vector; The environmental data is normalized to generate an environmental feature vector; The speech feature vector, image feature vector, and environmental feature vector are fused to generate the interaction feature vector.

[0011] In some embodiments, the method further includes: Acquire physiological sensor data of the interactive object, including heart rate variability data; Based on the physiological sensor data, adjust the calculation parameters of the importance score and / or the generation logic of the behavioral decision vector to adapt to the physiological state of the interactive object.

[0012] According to a second aspect of the present disclosure, an artificial intelligence training device based on selective memory reinforcement is provided, comprising: The acquisition module is used to acquire multimodal interaction data of interactive objects; The encoding module is used to encode the multimodal interaction data, generate corresponding interaction feature vectors, and calculate the emotional energy value and valence of the current interaction event. The memory module is used to calculate the importance score of the current interaction event based on the interaction feature vector, emotional energy value and valence, and to perform selective memory storage on the feature data corresponding to the current interaction event according to the matching result of the importance score and the preset threshold, so as to generate structured memory data. The decision-making module is used to integrate the interaction feature vector, structured memory data and predefined personality vector to generate a behavior decision vector, wherein the personality vector is used to represent the personality characteristics of the companion-type artificial intelligence; The generation module is used to generate multimodal feedback instructions based on the behavior decision vector; The execution module is used to execute the multimodal feedback instructions and update the personality vector based on the response data of the interaction object to the execution result of the feedback instructions.

[0013] According to a third aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the method of the first aspect of the present disclosure.

[0014] The technical solutions provided in this disclosure may have the following beneficial effects: The solution disclosed herein acquires multimodal interaction data of the interacting object; encodes the multimodal interaction data to generate corresponding interaction feature vectors, and calculates the emotional energy value and valence of the current interaction event; based on the interaction feature vectors, emotional energy value, and valence, calculates the importance score of the current interaction event, and performs selective memory storage on the feature data corresponding to the current interaction event according to the matching result of the importance score and a preset threshold, generating structured memory data; integrates the interaction feature vectors, structured memory data, and predefined personality vectors to generate a behavioral decision vector, which is used to represent the personality characteristics of the companion-style artificial intelligence; generates multimodal feedback instructions based on the behavioral decision vectors; executes the multimodal feedback instructions, and updates the personality vector based on the response data of the interacting object to the execution result of the feedback instructions. This disclosure can filter and store the feature data corresponding to the interaction event according to the importance score, achieving efficient and human-like memory management. Furthermore, based on the personality vector updated in real time with user feedback, the AI ​​can form and continuously evolve into an interaction style adapted to the user, enhancing the emotional stickiness of long-term companionship and providing users with a natural and authentic embodied intelligent interaction experience. Attached Figure Description

[0015] Figure 1 The diagram illustrates a flowchart of an artificial intelligence training method based on selective memory reinforcement, as described in an embodiment of this disclosure.

[0016] Figure 2 A schematic flowchart of a memory distillation method according to an embodiment of the present disclosure is shown.

[0017] Figure 3 The diagram shows a flowchart of a personality vector update method according to an embodiment of this disclosure. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0019] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0020] The following will describe the exemplary implementation method in detail with reference to the accompanying drawings and embodiments.

[0021] First, this disclosure provides an artificial intelligence training method based on selective memory reinforcement. This method can be executed by any electronic device with computing power, and can be applied to various scenarios such as embodied intelligent companion robots, virtual companion digital humans, and home intelligent interactive terminals.

[0022] Specifically, the electronic device used to execute this method can be a physical robot or a virtual digital human. A physical robot can interact with the user by executing corresponding action commands, in conjunction with a movement mechanism, display screen, speaker, and other output components. A virtual digital human can display its own image on a display screen and present interactive actions in the form of animation, while also interacting with the user in conjunction with speaker and other output components; this disclosure does not limit the specific implementation of these methods.

[0023] Specifically, Figure 1 This diagram illustrates a flowchart of an artificial intelligence training method based on selective memory reinforcement, as shown in an embodiment of this disclosure. Figure 1 As shown, the artificial intelligence training method based on selective memory reinforcement provided in this embodiment includes the following steps S101 to S106.

[0024] S101, Obtain multimodal interaction data of the interactive object.

[0025] In some embodiments, multimodal interaction data can be multi-dimensional data acquired through a supporting acquisition device, capable of fully representing the state of the interactive object and information about the interactive scene, and is also the foundation for artificial intelligence to perceive the external environment and interactive objects. For example, multimodal interaction data may include voice data, image data, and environmental data.

[0026] For example, the acquisition of multimodal data can be carried out in real time and synchronously. The interactive system will mark all different modal data acquired at the same time with a unified timestamp to ensure that the data of different dimensions such as voice, image, environment, etc. are fully aligned in the time dimension, and avoid deviations in subsequent feature fusion and emotion calculation due to time asynchrony.

[0027] For example, after acquiring the raw data of multimodal interaction data, the interaction system can perform preprocessing operations on the raw data. For instance, for audio data, operations such as noise reduction, echo cancellation, and filtering out invalid silent segments can be performed; for image data, operations such as face detection and alignment, and human key point localization can be performed; and for environmental data, operations such as outlier filtering can be performed. Through these preprocessing operations, clean and effective raw data can be provided for subsequent encoding processes.

[0028] S102, encode the multimodal interaction data to generate the corresponding interaction feature vector, and calculate the emotional energy value and valence of the current interaction event.

[0029] It should be noted that encoding can be understood as the process of converting raw data into standardized feature vectors by employing corresponding feature extraction algorithms tailored to the characteristics of different modal data. The final generated interaction feature vector is a high-dimensional vector formed by fusing multiple single-modal feature vectors, capable of fully representing all the core information of the interaction object in the current interaction event, including semantic content, voice emotion, facial expressions, body movements, and scene context.

[0030] While performing encoding processing, the system simultaneously calculates the emotional energy value and valence of the current interaction event. These two parameters are the core dimensions of the two-dimensional emotion model, capable of quantifying human emotional states. Valence measures the polarity and pleasure of emotion, with a value range of [-1, +1], where -1 represents extremely negative emotion, 0 represents neutral emotion, and +1 represents extremely positive emotion. Emotional energy, also known as arousal, measures the physiological activation intensity and arousal level of emotion, with a value range of [0, 1], where 0 represents complete calm with no arousal, and 1 represents extremely strong emotional arousal.

[0031] In actual execution, the system performs independent feature extraction for each modality of the multimodal data, generating corresponding single-modal feature vectors. All single-modal features are then mapped to the same feature space for fusion, generating the final interaction feature vector. Simultaneously, based on the acoustic, facial, and semantic features extracted during encoding, the system uses a pre-trained emotion regression model to output the emotional energy and valence of the current interaction event, thus quantifying the emotional state.

[0032] S103: Based on the interaction feature vector, emotional energy value and valence, calculate the importance score of the current interaction event, and perform selective memory storage on the feature data corresponding to the current interaction event according to the matching result of the importance score and the preset threshold, to generate structured memory data.

[0033] It should be noted that the importance score is a core indicator used to quantify the memory value of the current interactive event for the interactive object. Its value range can be [0,1]. The higher the score, the higher the emotional value and long-term memory significance of the current interactive event for the interactive object.

[0034] Selective memory storage involves performing differentiated storage and management operations on interaction data based on the ranking results of event importance scores, mimicking the human memory structure that includes short-term and long-term memory. The resulting structured memory data is a complete memory set formed after hierarchical storage and integration. It includes both discrete single-event memories of the interaction objects and deep cognitive characteristics of the long-term behaviors and preferences of the interaction objects, serving as one of the foundations for subsequent decision-making.

[0035] In actual execution, the system first calculates the importance score of the current event using a preset importance scoring formula, combined with interaction feature vectors, emotional energy values, and valence. The score result is then matched against preset grading thresholds. Based on the matching results, the feature data corresponding to the current event is stored in corresponding memory pools, with data in different memory pools exhibiting different forgetting patterns. Finally, all contents from different memory pools are integrated to generate updated structured memory data, completing the memory processing for a single interaction event.

[0036] The feature data corresponding to the interaction event can be understood as a set of features that can characterize the intention, emotional state, and context of the interaction object, including the interaction feature vector generated in S102. It may also include valence, emotional energy value, importance score, and raw multimodal interaction data calculated during the interaction process; this disclosure does not limit this.

[0037] The embodiments disclosed herein can simulate the memory screening mechanism of the human hippocampus, quantify the memory value of current interactive events, and perform hierarchical selective memory storage based on the evaluation results, thereby achieving long-term retention of important events and natural forgetting of unimportant events.

[0038] S104 integrates interactive feature vectors, structured memory data, and predefined personality vectors to generate behavioral decision vectors.

[0039] Among them, personality vectors are used to characterize the personality traits of companion-type artificial intelligence.

[0040] It should be noted that the personality vector is the carrier for realizing a dynamically evolving personality in this disclosure. For example, the personality vector can be a multi-dimensional floating-point vector, where each dimension corresponds to a core personality trait of the artificial intelligence, and the value range of each dimension is [0,1]. By combining the values ​​of different dimensions, the personality traits of the artificial intelligence can be accurately represented. During the initialization of the interactive system, initial values ​​can be configured for the personality vector according to the application scenario.

[0041] In some embodiments, the final generated behavioral decision vector is a unified decision-making carrier formed by integrating multi-dimensional information, fully encompassing decision information from the three core dimensions of content decision, voice decision, and action decision. Based on the same behavioral decision vector, feedback instructions in the three dimensions of text, voice, and action can be generated simultaneously, ensuring a high degree of coordination and consistency among all outputs.

[0042] In actual execution, the system first performs feature extraction and dimension alignment on the current interaction feature vector, structured memory data, and personality vector, mapping them to the same feature space. Then, the aligned features are concatenated and fused to generate a fused feature vector. Finally, the fused feature vector is input into a pre-trained decision network, which outputs the final behavioral decision vector, completing the generation of the interaction decision.

[0043] The embodiments disclosed herein can deeply integrate interactive feature vectors, structured memory data, and personality vectors to generate a unified behavioral decision vector, providing a unique decision basis for the generation of subsequent multimodal feedback instructions and ensuring the consistency of expression of subsequent different modal feedback instructions.

[0044] S105 generates multimodal feedback instructions based on behavioral decision vectors.

[0045] In some embodiments, multimodal feedback instructions include at least two types: voice feedback instructions and motion feedback instructions, corresponding to the two core interaction forms of voice interaction and embodied motion interaction, respectively. The generation process of the two types of instructions is based on the same behavioral decision vector and executed in parallel to ensure complete synchronization and coordination between the two.

[0046] For example, the interactive system can construct a corresponding large language model for prompt word input based on the content decision information in the behavior decision vector, generating text response content that matches the current interaction scenario, the needs of the interaction object, and the personality characteristics of the artificial intelligence. Simultaneously, based on the speech decision information in the behavior decision vector, the system adjusts the output parameters of the speech synthesis model, converting the validated text content into speech feedback instructions matching the current interaction state. Based on the action decision information in the behavior decision vector, the system retrieves matching servo motor instruction sequences from a predefined embodied action library, and after smooth concatenation and parameter adjustment, generates corresponding action feedback instructions. While generating these two types of instructions, the system simultaneously matches the execution timing, ensuring that key action nodes completely correspond to key nodes in the speech content, achieving deep collaboration between language and behavior.

[0047] S106, execute the multimodal feedback instruction and update the personality vector based on the response data of the interactive object to the execution result of the feedback instruction.

[0048] In some embodiments, during the execution of multimodal feedback instructions, the system can output the generated voice feedback instructions to the speaker for playback, and simultaneously output the motion feedback instructions to the servo motor drive module to control the device to perform corresponding limb movements. During the execution process, the voice playback progress and the motion execution progress are synchronized in real time to ensure complete coordination between the two.

[0049] In some embodiments, after the feedback instruction is executed, the system can restart the multimodal data acquisition process to continuously acquire the response data of the interactive object to the feedback, extract the corresponding response feature vector, calculate the change in the emotional state of the interactive object after the response, and identify the semantic feedback content. Subsequently, the personality vector can be iteratively updated based on this data to achieve smooth updates of the personality vector and realize the continuous adaptation of the AI ​​personality to the preferences of the interactive object.

[0050] The raw data of the response data can be consistent with the data type of the multimodal interaction collected in S101. It is used to characterize the attitude, emotion and satisfaction of the interaction object in the execution of this multimodal feedback instruction. It is the core basis for evaluating the effect of this interaction and optimizing personality traits.

[0051] This embodiment of the disclosure achieves adaptive growth of the AI ​​personality by updating the personality vector, enabling the AI ​​personality to adapt to the user's personalized needs without being constrained by pre-configured prompts.

[0052] The above combination Figure 1 The overall inventive concept of this disclosure has been described. For ease of understanding, the technical details involved will be further explained below.

[0053] In some embodiments, in S103 above, the importance score of the current interaction event can be calculated using the following formula: e=w1×Emotion+w2×TopicRelevance Where e is the importance score of the current interaction event, w1 and w2 are preset weight coefficients; Emotion is the emotion score of the current interaction event, which is related to the emotional energy value and valence of the current interaction event; TopicRelevance is the relevance score between the current interaction event and the topic of interest of the interaction object.

[0054] Specifically, the value of e can range from [0,1]. The sum of w1 and w2 is fixed at 1, and the default value can be adjusted according to the application scenario. For example, in an elderly companionship scenario centered on emotional support, emotions have a higher impact on memory value, so w1=0.7 and w2=0.3 can be set; in a children's companionship scenario centered on educational guidance, the weight of topic relevance can be appropriately increased, so w1=0.5 and w2=0.5 can be set. Interactive objects can also manually adjust the values ​​of the two weight coefficients according to their own memory needs.

[0055] Emotion represents the emotional score of the current interaction event, ranging from [0,1]. It can be calculated using the formula Emotion = |Valence| × Energy. Here, Valence is the emotional valence calculated in S202, and Energy is the emotional energy value calculated in S202. In other words, the higher the intensity of the emotion evoked by the event, whether positive or negative, the higher its memory value for the interacting object. For example, strong sadness triggered by mentioning the death of a loved one or strong joy triggered by sharing family happy events will both receive high emotional scores, while weak emotional fluctuations triggered by casually mentioned daily trivia will have an emotional score close to 0.

[0056] TopicRelevance scores the relevance between the current interactive event and the topics of interest to the interactive object, with a value ranging from [0,1]. In actual calculations, the interactive system can periodically model the feature data in the long-term memory pool using topic models, extracting the core topics of interest to the interactive object and generating a corresponding topic feature vector for each topic. When calculating the relevance of the current event, the cosine similarity between the semantic feature vector of the current interactive event and the feature vectors of each topic is calculated. The maximum similarity is taken and mapped to the interval [0,1] through a linear transformation, resulting in the final TopicRelevance value. The higher this value, the stronger the relevance between the current event and the core topics of interest to the interactive object, and the higher the memory value.

[0057] In some embodiments, for the generation of structured memory data, selective memory storage of feature data corresponding to the current interaction event can be achieved by setting two preset thresholds, including a high importance threshold and a low importance threshold, so as to generate structured memory data.

[0058] Specifically, if the importance score of the current interaction event is greater than the high importance threshold, the feature data corresponding to the current interaction event is stored in the long-term memory pool. If the importance score of the current interaction event is greater than or equal to the low importance threshold and less than the high importance threshold, the feature data corresponding to the current interaction event is stored in the intermediate memory pool, and after the storage time reaches the preset time, the natural forgetting operation is performed. If the importance score of the current interaction event is less than the low importance threshold, the feature data corresponding to the current interaction event is stored in the short-term memory pool, and a natural forgetting operation is performed as the dialog sliding window moves. Structured memory data is generated based on the feature data stored in the long-term memory pool, the medium-term memory pool, and the short-term memory pool.

[0059] For example, the values ​​of the high importance threshold and the low importance threshold can be in the range of [0,1], with the high importance threshold being greater than the low importance threshold. For instance, the high importance threshold can be set to 0.6, and the low importance threshold can be set to 0.2. The interactive object can manually adjust the values ​​of the two thresholds according to its own memory needs. If you want the artificial intelligence to remember more content, you can appropriately lower the high importance threshold; if you want the artificial intelligence to retain only the most essential memory content, you can appropriately increase the high importance threshold.

[0060] When performing selective memory storage, the system executes a three-level storage strategy based on the matching results of the importance score and two thresholds. If the importance score of the current interaction event is greater than or equal to the high importance threshold, the event is determined to be of core importance to the interaction object. The feature data, original text content, timestamp, sentiment parameters, and importance score corresponding to the current interaction event are all stored in the long-term memory pool. The data stored in the long-term memory pool will be permanently retained and will not be naturally forgotten, serving as the core memory data of the interaction object.

[0061] If the importance score of the current interaction event is less than the low importance threshold, the event is determined to be a trivial event with no long-term memory value, and the feature data corresponding to the current interaction event is stored in the short-term memory pool. The short-term memory pool is managed using a fixed-length dialog sliding window, with a default window size of 20 dialogue rounds. When new dialogue content is stored in the short-term memory pool, low-importance content exceeding the window length is automatically cleared, thus simulating natural human forgetting.

[0062] If the importance score of a current interaction event is greater than or equal to the low importance threshold, the event is determined to be a medium importance event with some memory value but not requiring permanent storage, and the corresponding feature data is stored in the medium-term memory pool. For example, the medium-term memory pool can be set to a default storage period of 7 days. Content that is not mentioned again by the interacting object after the storage period will be automatically deleted. If it is mentioned again by the interacting object within the storage period, the storage period of the content will be refreshed, and its importance score will be recalculated. If the recalculated score exceeds the high importance threshold, it will be transferred to the long-term memory pool for permanent storage.

[0063] Ultimately, the system will integrate all valid data from the long-term memory pool, medium-term memory pool, and short-term memory pool to generate complete structured memory data, providing support for subsequent interactive decision-making.

[0064] In some embodiments, after storing the feature data corresponding to the current interaction event into the long-term memory pool, clustering operations can be performed on the feature data stored in the long-term memory pool to integrate the discrete interaction event feature data into user cognitive feature data, thereby completing the memory distillation operation.

[0065] For example, the memory distillation operation can be performed during system idle periods, such as the charging period of the robot device, the rest period of the interactive object, or the idle period of more than 10 consecutive minutes without interaction, thereby avoiding the use of the device's computing resources by performing the memory distillation operation during the interaction process, which would affect the real-time response of the interaction.

[0066] For example, during memory distillation, the interactive system can standardize the feature data in the long-term memory pool, extracting the interaction feature vector, topic tag, sentiment parameter, and timestamp for each event to form a standardized memory dataset. Subsequently, a density clustering algorithm (e.g., DBSCAN) is used to cluster the interaction feature vectors in the memory dataset. Density clustering algorithms automatically determine the number of clusters without requiring pre-setting the number of clusters, adapting to the dynamic growth of long-term memory data and grouping events with similar themes and related content into the same category.

[0067] For each event category obtained from clustering, the interaction system can use a large language model to summarize, generalize, and abstract the content of all events within the category, generating corresponding user cognitive feature data. For example, for an interaction where a user asks about dietary restrictions for diabetes more than 10 times, the system can distill the information to generate a deep understanding that "the user has diabetes and needs to strictly control their intake of sugar and high-carbohydrate foods." Finally, the generated user cognitive feature data is updated into the structured memory data, completing the entire memory distillation process.

[0068] Specifically, Figure 2This diagram illustrates a flow chart of a memory distillation method according to an embodiment of the present disclosure, as shown below. Figure 2 As shown, the method includes the following steps S201 to S205.

[0069] S201, when the interactive device is in a preset idle period, trigger the memory distillation operation.

[0070] For example, the criteria for determining idle time periods may include at least one of the following: the interactive device does not receive a valid human-computer interaction operation within a preset time period, the interactive device is connected to the charging circuit and there is no real-time interactive task running, or a pre-configured rest period.

[0071] S202, preprocess the interaction event feature data in the long-term memory pool to generate a memory dataset containing interaction feature vectors, topic tags, and emotion parameters.

[0072] For example, the long-term memory pool stores feature data of interactive events whose importance scores exceed a high importance threshold. Its preprocessing operations include dimensional alignment of discrete interactive feature vectors, temporal sorting of event timestamps, and filtering out outliers and duplicate data. Topic tags can be pre-generated using an LDA topic model and bound to corresponding interactive events. Emotional parameters include the valence and emotional energy values ​​corresponding to the interactive events. Preprocessing ensures the standardization and consistency of the memory dataset, providing relatively standardized input data for subsequent clustering operations.

[0073] S203 uses a density clustering algorithm to cluster the interaction feature vectors in the memory dataset, grouping topic-related interaction events into the same category.

[0074] For example, the DBSCAN density clustering algorithm can be used. This algorithm does not require pre-setting the number of cluster categories and can automatically adapt to the dynamic growth characteristics of long-term memory data. By pre-setting the neighborhood radius and minimum sample number threshold, interactive events with high feature similarity and strong topic relevance are clustered into the same category. At the same time, it automatically identifies and removes outlier isolated event data to avoid invalid data interfering with the clustering results, ensuring that events within the same category have a unified user cognitive orientation.

[0075] S204: Extract features from all interactive event content for each cluster category to generate deep user cognitive feature data for the corresponding category.

[0076] For example, for each cluster category, the original text content, emotional features, and interactive behavior features of all events within the category can be extracted. The quantified large language model is then used to perform induction, summarization, and abstraction to eliminate the accidental details of individual events and extract deep cognitive information such as users' stable preferences, habits, taboos, and core areas of focus.

[0077] S205 integrates user cognitive feature data with effective data from each memory pool to update structured memory data.

[0078] For example, the generated deep user cognitive feature data can be integrated with core event memories in the long-term memory pool, moderately important data in the medium-term memory pool, and dialogue context data in the short-term memory pool to form updated structured memory data. Simultaneously, a higher retrieval weight is assigned to the user cognitive feature data than to single event memories. In subsequent interaction decision generation, the deep user cognitive features are prioritized for matching, ensuring that the AI's response aligns with the user's long-term stable needs, rather than merely matching superficial information from a single event.

[0079] By performing memory distillation, discrete single-interaction event feature data in the long-term memory pool can be integrated into abstract, deep user cognitive feature data, enabling artificial intelligence to achieve a deep understanding of the interactive object, rather than simple memory retelling.

[0080] In some embodiments, in S106 above, the personality vector can be updated using the following formula: P(t+1)=P(t)+α×R×(U(t)-P(t)) Where P(t) is the personality vector before the update, and P(t+1) is the personality vector after the update; α is the preset fitness coefficient; R is the feedback value generated based on the response data; U(t) is the interaction feature vector corresponding to the current interaction event; and t is the time.

[0081] Specifically, P(t) is the personality vector before the update, which is the baseline personality vector of the artificial intelligence before this interaction, and P(t+1) is the personality vector after the update, which is the personality vector iteratively optimized after this interaction.

[0082] α is a preset fitness coefficient, with a value ranging from (0,1), used to control the magnitude of each personality vector update. A larger value results in faster personality updates and a more sensitive response to feedback from interacting objects; a smaller value results in slower personality updates and higher long-term personality stability. For example, α can be set to 0.05 to ensure smooth and continuous personality changes without sudden personality jumps; in the first 100 rounds of system initialization, α can be set to 0.1 to accelerate the initial adaptation of the personality to the preferences of interacting objects; after long-term interaction, α can be reduced to 0.02 to further improve personality stability.

[0083] R is the satisfaction feedback value generated based on the response data of the interacting object. Its value ranges from [-1, +1] and is the core guideline for personality updating. Here, +1 represents that the interacting object is extremely satisfied with the feedback, -1 represents that the interacting object is extremely dissatisfied with the feedback, and 0 represents neutral with no clear tendency.

[0084] For example, the value of R is obtained by weighted summation of three core dimensions: emotion change index, semantic feedback index, and body behavior index. The emotion change index compares the change in emotional valence of the interacting object before and after feedback; a positive value indicates improvement in emotion, while a negative value indicates deterioration. The semantic feedback index identifies the semantic level of satisfaction based on the interacting object's verbal response content using an emotion classification model. The body behavior index identifies the acceptance of feedback based on changes in the interacting object's body movements. After weighted summation of the three dimensions, a linear transformation is applied to map them to the interval [-1, +1] to obtain the final R value.

[0085] Specifically, Figure 3 This diagram illustrates a flowchart of a personality vector update method according to an embodiment of the present disclosure. Figure 3 As shown, the method includes the following steps S301 to S303.

[0086] S301: Collect the multimodal response data of the interactive object in response to this feedback, and extract the response feature vector after completing the encoding process.

[0087] For example, after the multimodal feedback instruction is executed, the multimodal interaction data of the interactive object can be acquired again and used as the response data of the interactive object to the result of the feedback instruction. Subsequently, a process completely consistent with the aforementioned multimodal interaction data encoding process can be performed on the response data to extract speech features, image features, and environmental features respectively and fuse them to generate a response feature vector representing the user's feedback attitude, providing basic data for subsequent satisfaction assessment.

[0088] S302, based on the emotional changes, semantic feedback and body behavior of the interactive object shown by the response feature vector, a weighted satisfaction feedback value R is generated.

[0089] For example, the emotion change index can be calculated by comparing the change in emotional valence before and after user feedback; a positive value indicates an improvement in emotion, while a negative value indicates a decline. The semantic feedback index is obtained by identifying the emotional tendency of the user's response text using an emotion classification model. The body behavior index is obtained by identifying acceptability through changes in the user's body movements and posture. By weighted summing of the three types of indices shown in the response feature vector and linearly transforming them to the interval [-1, +1], the final satisfaction feedback value R can be obtained.

[0090] S303, based on the formula P(t+1)=P(t)+α×R×(U(t)-P(t)), determine the updated personality vector P(t+1).

[0091] For example, the core logic of this formula is as follows: when the satisfaction feedback value R is positive, it means that the user approves of the personality performance this time, and the baseline personality vector moves closer to the personality performance vector of this time, strengthening the personality characteristics preferred by the user; when R is negative, it means that the user does not approve of the personality performance this time, and the baseline personality vector is adjusted in the opposite direction of the personality performance vector of this time, weakening the personality characteristics rejected by the user.

[0092] After obtaining the updated personality vector P(t+1) through calculation, the original baseline personality vector P(t) can be replaced to complete this iterative update. The updated personality vector will serve as the baseline personality basis for the generation of behavioral decision vectors and multimodal feedback instructions in all subsequent interactions, enabling the AI ​​personality to continuously adapt to user preferences.

[0093] By updating the personality vector using the above formula, when the interaction object provides positive feedback, the personality vector will move closer to the currently used personality expression vector, strengthening the personality traits that the interaction object likes. Conversely, when the interaction object provides negative feedback, the personality vector will adjust in the opposite direction of the currently used personality expression vector, weakening the personality traits that the interaction object dislikes. Through these small adjustments with each interaction, the AI ​​personality can continuously adapt to the preferences of the interaction object.

[0094] In some embodiments, multimodal feedback instructions may include voice feedback instructions and action feedback instructions, which are generated in parallel based on the same behavior decision vector to ensure a high degree of coordination between voice and action.

[0095] For example, for voice feedback commands, a behavior decision vector can be input into a large language model to generate a text response corresponding to the behavior decision vector. Then, based on the behavior decision vector, the voice output parameters of the voice synthesis model are adjusted, and the adjusted voice synthesis model converts the text response into a voice feedback command corresponding to the current interaction event.

[0096] Specifically, the interaction system can extract corresponding content decision features based on behavioral decision vectors, and combine this with the text content of the current interaction, structured memory data, and the feature dimensions of the personality vector to construct standardized prompt words. The prompt word content includes four core parts: personality setting, memory context, current interaction content, and response requirements. The constructed prompt words are input into a large language model to generate text response content matching the current interaction scenario. Subsequently, the interaction system can extract corresponding speech decision features based on the behavioral decision vectors and adjust the speech output parameters of the speech synthesis model, specifically including four core parameters: speech rate, tone, volume, and emotional tone. Speech rate corresponds to the speech rate preference dimension in the personality vector, tone corresponds to the tone fluctuation dimension in the personality vector, volume is adaptively adjusted according to the noise level of the current environment, and emotional tone matches the emotional state of the current interaction. The validated text response content and the adjusted speech output parameters are input into a pre-trained speech synthesis model to generate natural, fluent speech audio data that matches the emotional tone of the current interaction, i.e., the speech feedback command.

[0097] For example, for motion feedback instructions, a matching servo motor instruction sequence can be retrieved from a predefined embodied motion library based on the behavior decision vector to generate motion feedback instructions.

[0098] When generating motion feedback commands, the interactive system predefines an embodied motion library, which stores hundreds of standardized motion units. Each motion unit corresponds to a motion feature vector and a servo motor command sequence, covering all categories of motions, including head movements, upper limb movements, body postures, and facial expression animations. The interactive system extracts the corresponding motion decision features based on the behavior decision vector. By calculating the cosine similarity between the motion decision features and the feature vectors of each motion unit in the motion library, it selects one or more motion units with the highest similarity, and configures corresponding parameters such as motion amplitude and motion frequency to generate the final motion feedback command.

[0099] For example, while generating the two types of instructions, the interactive system also synchronizes the execution timing, and synchronizes the playback progress of the voice feedback instructions to the action execution module in real time, ensuring that the key action nodes correspond completely with the key nodes of the voice content, thereby solving the problem of the separation between language and behavior.

[0100] In some embodiments, the multimodal interaction data includes: voice data, image data, and environmental data. When performing S102 above, the interaction feature vector corresponding to the multimodal interaction data can be generated in the following manner: Extract semantic information and acoustic features from speech data to generate speech feature vectors; Extract facial and limb movement features of interactive objects from image data to generate image feature vectors; The environmental data is normalized to generate environmental feature vectors; The speech feature vector, image feature vector and environmental feature vector are fused to generate an interaction feature vector.

[0101] For example, for speech data, a speech recognition model can be used to convert the speech data into text content, and then a pre-trained BERT model can be used to convert the text content into semantic feature vectors. Simultaneously, the speech data is converted into a Mel-spectrum image, which is then input into a convolutional neural network to extract acoustic feature vectors. Concatenating the semantic feature vectors with the acoustic feature vectors yields the speech feature vector.

[0102] For image data, standardized face images that have undergone face detection and alignment are input into a pre-trained ResNet-18 residual network. After removing the final classification layer, the output of the average pooling layer is taken to obtain the facial feature vector. Simultaneously, standardized multi-frame human skeletal keypoint data are input into a pre-trained ST-GCN spatiotemporal graph convolutional network to extract limb motion feature vectors. The facial feature vector and the limb motion feature vector are then concatenated to obtain the image feature vector.

[0103] For environmental data, the system employs corresponding normalization and encoding processes based on the data type. For continuous numerical data such as light intensity, ambient temperature and humidity, and ambient noise, the Min-Max normalization method is used to map all values ​​to the 0-1 range. For time features such as hours and days of the week, a sine-cosine periodic encoding method is used to fully preserve the periodicity of time. For discrete classification features such as room location, a one-hot encoding method is used to fully preserve the classification features of location. All processed values ​​are then concatenated to obtain the environmental feature vector.

[0104] Finally, the system uses a fully connected layer to map speech feature vectors, image feature vectors, and environmental feature vectors into the same feature space, and then concatenates and fuses the mapped feature vectors to generate the final interactive feature vector.

[0105] In some embodiments, physiological sensory data of the interactive object can also be acquired; based on the physiological sensory data, the calculation parameters of the importance score and / or the generation logic of the behavioral decision vector are adjusted to adapt to the physiological state of the interactive object.

[0106] Specifically, the interactive system can acquire physiological sensing data in real time from the wearable physiological sensing device worn by the interactive object through the wireless communication module. This data includes at least heart rate variability data, and may also include other physiological indicators such as heart rate, blood pressure, skin conductance level, and respiratory rate.

[0107] For example, the interactive system can identify the physical and emotional states of the interacting object in real time based on the acquired physiological sensor data. When it detects a decrease in the user's heart rate variability and an increase in skin conductance, it determines that the user is in a state of high stress and tension. At this time, the calculation parameters of the importance score can be adjusted to increase the weight coefficient of the emotion score and strengthen the memory of the user's emotion-related events. At the same time, the generation logic of the behavioral decision vector can be adjusted to make the AI's response more soothing and calming, reduce the speech rate and the frequency of proactive questioning, and adapt to the user's physiological and emotional state.

[0108] When the system detects a user's heart rate continuously increasing, breathing rate accelerating, accompanied by painful facial expressions and groans, it will trigger emergency response logic, prioritize inquiring about the user's physical condition, and, if necessary, push emergency contact information and contact details of nearby medical institutions to meet the emergency needs of elderly care scenarios.

[0109] This disclosure constructs a complete technical closed loop of "input of multimodal interactive data - importance scoring and hierarchical memory storage - memory distillation - fusion of memory and personality vector decision-making - output of multimodal feedback instructions - personality vector update based on feedback instruction response data", thereby realizing anthropomorphic memory management and personalized companionship experience adapted to users.

[0110] Specifically, this disclosure achieves three levels of selective memory storage—long-term, medium-term, and short-term—through importance scoring. Combined with memory distillation during idle periods, it can form higher-level user profiles based on long-term memory through clustering and induction. This significantly surpasses the efficiency and accuracy of full-data recall in related technologies, while also significantly reducing device storage load and computing power consumption, thus meeting the needs of long-term companionship scenarios. Furthermore, this disclosure generates a unified behavioral decision vector by integrating real-time interaction features, structured memory, and personality vectors. Based on the same decision, it synchronously generates multimodal feedback commands, ensuring behavioral consistency between voice and action feedback commands. Simultaneously, iteratively updating the personality vector based on real-time user feedback enables smooth adaptive growth of the AI ​​personality. Lightweight lifelong learning can be achieved on the device without large-scale model fine-tuning, effectively improving the naturalness of interaction, emotional engagement, and personalized adaptation capabilities in long-term companionship scenarios.

[0111] Based on the same inventive concept, this disclosure also provides an artificial intelligence training device based on selective memory reinforcement, as shown in the following embodiment. Since the principle of this interactive device embodiment in solving the problem is the same as that described above... Figure 2 The method embodiments shown are similar, therefore the implementation of this interactive device embodiment can be found in the above description. Figure 2 The implementation of the method embodiments shown will not be repeated here.

[0112] Specifically, the artificial intelligence training device based on selective memory reinforcement provided in this disclosure includes: an acquisition module, an encoding module, a memory module, a decision-making module, a generation module, and an execution module.

[0113] The acquisition module is used to acquire multimodal interaction data of the interactive object; The encoding module is used to encode multimodal interaction data, generate corresponding interaction feature vectors, and calculate the emotional energy value and valence of the current interaction event. The memory module is used to calculate the importance score of the current interaction event based on the interaction feature vector, emotional energy value and valence, and perform selective memory storage on the feature data corresponding to the current interaction event according to the matching result of the importance score and the preset threshold, so as to generate structured memory data. The decision-making module is used to integrate interactive feature vectors, structured memory data and predefined personality vectors to generate behavioral decision vectors. The personality vectors are used to represent the personality characteristics of companion-type artificial intelligence. The generation module is used to generate multimodal feedback instructions based on behavioral decision vectors; The execution module is used to execute multimodal feedback instructions and update the personality vector based on the response data of the interaction object to the execution results of the feedback instructions.

[0114] In some embodiments, the memory module is used to calculate the importance score of the current interaction event using the following formula: e=w1×Emotion+w2×TopicRelevance Where e is the importance score of the current interaction event, w1 and w2 are preset weight coefficients; Emotion is the emotion score of the current interaction event, which is related to the emotional energy value and valence of the current interaction event; TopicRelevance is the relevance score between the current interaction event and the topic of interest of the interaction object.

[0115] In some embodiments, the preset thresholds include a high importance threshold and a low importance threshold. The memory module is configured to: store the feature data corresponding to the current interaction event in a long-term memory pool when the importance score of the current interaction event is greater than or equal to the high importance threshold; store the feature data corresponding to the current interaction event in a medium-term memory pool when the importance score of the current interaction event is greater than or equal to the low importance threshold but less than the high importance threshold, and perform a natural forgetting operation after the storage period reaches a preset duration; store the feature data corresponding to the current interaction event in a short-term memory pool when the importance score of the current interaction event is less than the low importance threshold, and perform a natural forgetting operation along with the dialog sliding window; and generate structured memory data based on the feature data stored in the long-term memory pool, medium-term memory pool, and short-term memory pool.

[0116] In some embodiments, the memory module is used to perform clustering operations on the feature data stored in the long-term memory pool in order to integrate discrete interaction event feature data into user cognitive feature data, thereby completing the memory distillation operation.

[0117] In some embodiments, the execution module is configured to update the personality vector using the following formula: P(t+1)=P(t)+α×R×(U(t)-P(t)) Where P(t) is the personality vector before the update, and P(t+1) is the personality vector after the update; α is the preset fitness coefficient; R is the feedback value generated based on the response data; U(t) is the interaction feature vector corresponding to the current interaction event; and t is the time.

[0118] In some embodiments, the multimodal feedback instructions include: voice feedback instructions and action feedback instructions. The generation module is used to input the behavior decision vector into a large language model to generate text response content corresponding to the behavior decision vector; adjust the voice output parameters of the voice synthesis model according to the behavior decision vector, and convert the text response content into a voice feedback instruction corresponding to the current interaction event through the adjusted voice synthesis model; and, based on the behavior decision vector, retrieve a matching servo motor instruction sequence from a predefined embodied action library to generate action feedback instructions.

[0119] In some embodiments, the multimodal interaction data includes: speech data, image data, and environmental data. An encoding module is used to extract semantic information and acoustic features from the speech data to generate a speech feature vector; extract facial features and body movement features of the interactive object from the image data to generate an image feature vector; normalize the environmental data to generate an environmental feature vector; and fuse the speech feature vector, image feature vector, and environmental feature vector to generate an interaction feature vector.

[0120] In some embodiments, the acquisition module is used to acquire physiological sensor data of the interactive object, including heart rate variability data; and based on the physiological sensor data, adjust the calculation parameters of the importance score and / or the generation logic of the behavioral decision vector to adapt to the physiological state of the interactive object.

[0121] Those skilled in the art will understand that various aspects of this disclosure can be implemented as a system, method, or program product. Therefore, various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system."

[0122] In exemplary embodiments of this disclosure, a computer-readable storage medium is also provided, which may be a readable signal medium or a readable storage medium. A program product capable of implementing the methods described above is stored thereon. In some possible implementations, various aspects of this disclosure may also be implemented as a program product including program code, which, when run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure.

[0123] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. An artificial intelligence training method based on selective memory reinforcement, characterized in that, include: Obtain multimodal interaction data of interactive objects; The multimodal interaction data is encoded to generate corresponding interaction feature vectors, and the emotional energy value and valence of the current interaction event are calculated. Based on the interaction feature vector, emotional energy value and valence, the importance score of the current interaction event is calculated, and according to the matching result of the importance score and the preset threshold, selective memory storage is performed on the feature data corresponding to the current interaction event to generate structured memory data. By integrating the interaction feature vector, structured memory data, and predefined personality vector, a behavioral decision vector is generated, wherein the personality vector is used to characterize the personality traits of companion-type artificial intelligence; Based on the aforementioned behavior decision vector, multimodal feedback instructions are generated; The multimodal feedback instruction is executed, and the personality vector is updated based on the response data of the interaction object to the execution result of the feedback instruction.

2. The method according to claim 1, characterized in that, The step of calculating the importance score of the current interaction event based on the interaction feature vector, emotional energy value, and valence includes: The importance score of the current interaction event is calculated using the following formula: e=w1×Emotion+w2×TopicRelevance Where e is the importance score of the current interaction event, w1 and w2 are preset weight coefficients; Emotion is the emotion score of the current interaction event, which is related to the emotional energy value and valence of the current interaction event; TopicRelevance is the relevance score between the current interaction event and the topic of interest of the interaction object.

3. The method according to claim 1 or 2, characterized in that, The preset thresholds include: a high importance threshold and a low importance threshold; The step of selectively storing feature data corresponding to the current interaction event based on the matching result of the importance score and the preset threshold, and generating structured memory data, includes: If the importance score of the current interaction event is greater than or equal to the high importance threshold, the feature data corresponding to the current interaction event will be stored in the long-term memory pool. If the importance score of the current interaction event is greater than or equal to the low importance threshold and less than the high importance threshold, the feature data corresponding to the current interaction event is stored in the intermediate memory pool, and after the storage time reaches the preset time, a natural forgetting operation is performed. If the importance score of the current interaction event is less than the low importance threshold, the feature data corresponding to the current interaction event is stored in the short-term memory pool, and a natural forgetting operation is performed along with the dialog sliding window. The structured memory data is generated based on the feature data stored in the long-term memory pool, the medium-term memory pool, and the short-term memory pool.

4. The method according to claim 3, characterized in that, After storing the feature data corresponding to the current interaction event into the long-term memory pool, the method further includes: Clustering operations are performed on the feature data stored in the long-term memory pool to integrate discrete interaction event feature data into user cognitive feature data, thus completing the memory distillation operation.

5. The method according to claim 1, characterized in that, The process of updating the personality vector based on the response data of the interactive object to the execution result of the feedback command includes: The personality vector is updated using the following formula: P(t+1)=P(t)+α×R×(U(t)-P(t)) Where P(t) is the personality vector before the update, and P(t+1) is the personality vector after the update; α is the preset fitness coefficient; R is the feedback value generated based on the response data; U(t) is the interaction feature vector corresponding to the current interaction event; and t is the time.

6. The method according to claim 1, characterized in that, The multimodal feedback commands include: voice feedback commands and motion feedback commands; The step of generating multimodal feedback instructions based on the behavioral decision vector includes: The behavior decision vector is input into the large language model to generate text response content corresponding to the behavior decision vector; the speech output parameters of the speech synthesis model are adjusted according to the behavior decision vector, and the text response content is converted into a speech feedback command corresponding to the current interaction event through the adjusted speech synthesis model. Based on the behavior decision vector, a matching servo motor instruction sequence is retrieved from a predefined embodied action library to generate an action feedback instruction.

7. The method according to claim 1, characterized in that, The multimodal interaction data includes: voice data, image data, and environmental data; The process of encoding the multimodal interaction data to generate corresponding interaction feature vectors includes: Extract semantic information and acoustic features from the speech data to generate a speech feature vector; Extract facial features and body movement features of the interactive objects from the image data to generate an image feature vector; The environmental data is normalized to generate an environmental feature vector; The speech feature vector, image feature vector, and environmental feature vector are fused to generate the interaction feature vector.

8. The method according to claim 1, characterized in that, The method further includes: Acquire physiological sensor data of the interactive object, including heart rate variability data; Based on the physiological sensor data, adjust the calculation parameters of the importance score and / or the generation logic of the behavioral decision vector to adapt to the physiological state of the interactive object.

9. An artificial intelligence training device based on selective memory reinforcement, characterized in that, include: The acquisition module is used to acquire multimodal interaction data of interactive objects; The encoding module is used to encode the multimodal interaction data, generate corresponding interaction feature vectors, and calculate the emotional energy value and valence of the current interaction event. The memory module is used to calculate the importance score of the current interaction event based on the interaction feature vector, emotional energy value and valence, and to perform selective memory storage on the feature data corresponding to the current interaction event according to the matching result of the importance score and the preset threshold, so as to generate structured memory data. The decision-making module is used to integrate the interaction feature vector, structured memory data and predefined personality vector to generate a behavior decision vector, wherein the personality vector is used to represent the personality characteristics of the companion-type artificial intelligence; The generation module is used to generate multimodal feedback instructions based on the behavior decision vector; The execution module is used to execute the multimodal feedback instructions and update the personality vector based on the response data of the interaction object to the execution result of the feedback instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 8.

Images