Abstract: A method for training a video large multimodal model (VLMM) by iterative self-retrospective judgment (i-SRT) is provided. The method includes steps of: (i) feeding a (k?1)-th self-retrospective context, video data, and query data into a (k?1)-th trained VLMM, instructing the (k?1)-th trained VLMM to generate a k_th self-retrospective context as to the video data and generate a k_1-st and a k_2-nd responses as to the query data, and thus generating a k_th preference data set, (ii) feeding the k_th preference data set into the (k?1)-th trained VLMM, instructing the (k?1)-th trained VLMM to determine a k_th preference response and a k_th non-preference response, and generating a k_th preference feedback data, and (iii) generating a k_th loss as to the k_th preference feedback data by using a DPO (Direct Preference Optimization) and updating parameters of the (k?1)-th trained VLMM by using the k_th loss, thereby generating a k_th trained VLMM.

Abstract: There is provided a conversation method and system for operating a conversation model according to the presence or absence of related knowledge. A conversation method according to an embodiment may receive an input of a user utterance, may determine a conversation model based on an inputted user utterance content, may create a conversation content by using the determined conversation model, and may convert the created conversation content into a voice, and may output the voice. Accordingly, different conversation models may be operated according to the presence or absence of related knowledge, so that an empathic conversation may be generated instead of generating an inappropriate response in a knowledge-based conversation, and user's satisfaction on a conversation may be enhanced.

Abstract: A method for subdivided representation reinforcement of an image/text representation vector through an attribute value of an object in an image-language alignment model is provided. The method for training an image-language alignment model, according to an embodiment of the present invention, generates, in an input image, object-specific representation vectors of the image, generates, in an input text, object-specific representation vectors of the text, and uses the generated object-specific representation vectors so as to train an image-language align model through a contrast loss function. Therefore, object-specific attribute representation is reinforced such that each attribute is represented to be subordinate to the objects, and thus accurate image searches can be performed for more complex natural language queries by means of the image-language alignment model, and accurate natural language searches can be performed for images having various objects.

Abstract: Provided are a method and a system for generating a human animation from SMPL series parameters. An animation generation method according to an embodiment includes: extracting, by a motion extraction module, motion information from an image content; converting, by a motion conversion module, the extracted motion information into parameters required by an animation generation module; and generating, by the animation generation module, a motion image of a human object from the converted motion information. Accordingly, an animation accurately showing motions of an object appearing in an image content by a more realistic 3D model may be generated and digital human contents of high reality may be created and used.

Abstract: There is provided a multi-modal knowledge-based conversation data generation and additional information labeling system. A conversation data generation method according to an embodiment includes: receiving input of a user utterance; searching pieces of text knowledge related to the user utterance; searching images related to the user utterance; receiving input of an answer to the user utterance referring to the searched text knowledge and images; and collecting the user utterances, the pieces of text knowledge, the images, and the answer to the user utterances as conversation data. Accordingly, open domain knowledge-based conversation data may be established by using both images and texts, and an image may be used as utterance and information acquired from the image may be used as utterance, so that AI conversation data similar to actual conversations can be established.

Abstract: Provided is an accelerator provided in an electronic device and configured to perform simultaneous localization and mapping (SLAM), the accelerator including a factor graph database, a memory, and a back-end processor, wherein the back-end processor is configured to receive a first piece of data corresponding to map points and camera positions from the factor graph database, convert the received first piece of data into a matrix for the map points and a matrix for the camera positions, store, in the memory, results obtained by performing an optimization calculation on the matrix for the map points and a matrix for at least one camera position, among the camera positions, corresponding to the map points, and obtain a second piece of data optimized with respect to the first piece of data based on the results stored in the memory.

Abstract: There are provided a method and a system for automatically generating a back-channel in an interactive agent system. According to an embodiment of the disclosure, an automatic back-channel generation method includes: predicting a back-channel by analyzing an utterance of a user inputted in a back-channel prediction model; and generating the predicted back-channel, and the back-channel prediction model is an AI model that is trained to predict a back-channel to express from the utterance of the user. Accordingly, a back-channel is automatically generated by utilizing a back-channel prediction module which is based on a language model, so that a natural dialogue interaction with a user may be implemented in an interactive agent system, and quality of a dialogue service provided to a user may be enhanced.

Abstract: There is provided a conversation method and system for operating a conversation model according to the presence or absence of related knowledge. A conversation method according to an embodiment may receive an input of a user utterance, may determine a conversation model based on an inputted user utterance content, may create a conversation content by using the determined conversation model, and may convert the created conversation content into a voice, and may output the voice. Accordingly, different conversation models may be operated according to the presence or absence of related knowledge, so that an empathic conversation may be generated instead of generating an inappropriate response in a knowledge-based conversation, and user's satisfaction on a conversation may be enhanced.

Abstract: There is a method for monitoring emotion and behavior occurring during a conversation with a user who is in need of protection. According to an embodiment, a user monitoring method receives an input of a user utterance and generates a conversation content by using a conversation model based on a user utterance content, converts the generated conversation content into a voice and outputs the voice, and recognizes emotion of the user by using an emotion analysis model based on the user utterance content and monitors the emotion. Accordingly, the method may monitor behavior/emotion of a user through a natural conversation while allowing a specialist/guardian to monitor an abnormal sign of a user who is in need of protection.

Abstract: There is a method for monitoring emotion and behavior occurring during a conversation with a user who is in need of protection. According to an embodiment, a user monitoring method receives an input of a user utterance and generates a conversation content by using a conversation model based on a user utterance content, converts the generated conversation content into a voice and outputs the voice, and recognizes emotion of the user by using an emotion analysis model based on the user utterance content and monitors the emotion. Accordingly, the method may monitor behavior/emotion of a user through a natural conversation while allowing a specialist/guardian to monitor an abnormal sign of a user who is in need of protection.

Abstract: A preconditioned conjugate gradient (PCG) solver, embedded in an electronic device to perform a simultaneous localization and mapping (SLAM) operation, includes an image database, a factor graph database, and a back-end processor, wherein the back-end processor is configured to receive an image from the image database to perform re-localization, receive, from the factor graph database, data for calculating six degrees of freedom (DoF)-related components, construct a matrix including the six degrees of freedom-related components based on the received data, and load and rearrange the matrix and a vector, to perform calculation on each block of each row of the matrix and the vector, then output first data, and shift second data to a location of the first data.

Abstract: There are provided a method and a system for automatically generating a back-channel in an interactive agent system. According to an embodiment of the disclosure, an automatic back-channel generation method includes: predicting a back-channel by analyzing an utterance of a user inputted in a back-channel prediction model; and generating the predicted back-channel, and the back-channel prediction model is an AI model that is trained to predict a back-channel to express from the utterance of the user. Accordingly, a back-channel is automatically generated by utilizing a back-channel prediction module which is based on a language model, so that a natural dialogue interaction with a user may be implemented in an interactive agent system, and quality of a dialogue service provided to a user may be enhanced.

Abstract: Provided is a processor configured to compute elements affecting an optimization matrix in connection with a first measurement, among elements of a Hessian matrix, instead of generating a whole Hessian matrix for a map point and a camera pose based on all measurements, and accumulate the computed elements over the optimization matrix used to perform optimization operations in relation to states of the map point and the camera pose.

Abstract: This application relates to an adaptive inference system and an operation method therefor. In one aspect, the system includes a user terminal for collecting multi-modal information including at least visual information, voice information and text information. The system may also include an inference support device for receiving the multi-modal information from the user terminal, and inferring the intention of a user on the basis of pre-stored history information related to the user terminal, individualized information and the multi-modal information.

Abstract: Provided is an accelerator provided in an electronic device and configured to perform simultaneous localization and mapping (SLAM), the accelerator including a factor graph database, a memory, and a back-end processor, wherein the back-end processor is configured to receive a first piece of data corresponding to map points and camera positions from the factor graph database, convert the received first piece of data into a matrix for the map points and a matrix for the camera positions, store, in the memory, results obtained by performing an optimization calculation on the matrix for the map points and a matrix for at least one camera position, among the camera positions, corresponding to the map points, and obtain a second piece of data optimized with respect to the first piece of data based on the results stored in the memory.

Abstract: This application relates to an adaptive inference system and an operation method therefor. In one aspect, the system includes a user terminal for collecting multi-modal information including at least visual information, voice information and text information. The system may also include an inference support device for receiving the multi-modal information from the user terminal, and inferring the intention of a user on the basis of pre-stored history information related to the user terminal, individualized information and the multi-modal information.

Abstract: This application relates to an apparatus for adaptive conversation. In one aspect, the apparatus includes a server communication circuit that forms a communication channel with a user terminal and a server processor functionally connected to the server communication circuit. The server processor may receive, from the user terminal, a user utterance and surrounding external information acquired at a point in time at which the user utterance is collected. The server processor may also generate one word input by combining the surrounding external information with input information generated by performing natural language processing on the user utterance. The server processor may further generate a response sentence by applying the word input to a neural network model, and transmit the response sentence to the user terminal.

Abstract: A preconditioned conjugate gradient (PCG) solver, embedded in an electronic device to perform a simultaneous localization and mapping (SLAM) operation, includes an image database, a factor graph database, and a back-end processor, wherein the back-end processor is configured to receive an image from the image database to perform re-localization, receive, from the factor graph database, data for calculating six degrees of freedom (DoF)-related components, construct a matrix including the six degrees of freedom-related components based on the received data, and load and rearrange the matrix and a vector, to perform calculation on each block of each row of the matrix and the vector, then output first data, and shift second data to a location of the first data.

Abstract: A remote monitoring method includes the steps of: (c) receiving a control data for changing an image area captured by the smart phone from a communication channel, which transmits and receives data through an Internet, by the smart phone; (d) controlling a cradle through a short-range network using a pan/tilt control data for controlling one or more of pan and tilt of the cradle, on which the smart phone is mounted, using the received control data, by the smart phone; and (e) encoding an image captured through a provided camera module and outputting the encoded image through the Internet, after controlling the cradle, by the smart phone.

Abstract: The present invention provides a fingerprint sensor module assembly integrated with a cover window for an electronic device. The fingerprint sensor module assembly, according to one embodiment of the present invention, comprises: the cover window; a fingerprint sensor module; and an adhesion part. The cover window is disposed on a front surface of the electronic device. An image is generated from a display module and displayed on the cover window. The fingerprint sensor module is coupled to a seating part which is disposed on a lower surface of the cover window. The fingerprint sensor module comprises a fingerprint sensor having a sensing unit for sensing fingerprints; and a substrate electrically connected to the fingerprint sensor. The adhesion part is disposed between the seating part and the fingerprint sensor module and configured to fix the fingerprint sensor module in the seating part.