Abstract: A method may include receiving, by a virtual assistant of a user device, an input from a user, the virtual assistant being based on software. The method may include obtaining, by the virtual assistant of the user device and via a sensor of the user device, audio information or video information of the user. The method may include determining, by the virtual assistant of the user device, an identity of the user based on the audio information or the video information of the user and a set of facial embeddings and speech embeddings that is correlated with the user, the set of facial embeddings and speech embeddings being generated using a facial embedding model, a speech embedding model, and a sound source localization model. The method may include performing, by the virtual assistant of the user device, an action based on the input and the identity of the user.

Abstract: A method for multimodal content retrieval, may include: receiving a search query corresponding to a request for content; aggregating word features extracted from the search query based on a first set of learned weights; aggregating region features extracted from each of a plurality of images, based on a second set of learned weights, independently of the word features; computing a similarity score between the aggregated words features and the aggregated region features for each of the plurality of images; selecting candidate images from the plurality of images based on the similarity scores between each of the plurality of images and the search query; and selecting at least one final image from the candidate images as a response to the search query, based on attended similarity scores of the candidate images with respect to the search query.

Abstract: Training a classifier using embeddings and building a latent space is disclosed. The embeddings may be based on weights in a trained machine learning model. Also, operation of the classifier to process video segments in real-time using the using the weights and the latent space is disclosed. The embeddings and the latent space allow the classification to be performed at an overall reduced dimensionality. The latent space is designed to efficiently scale with an increasing number of queries to permit fast search through the space. Embodiments permit real-time operation on video with dynamic features. The classifier reduces the bandwidth demand of video camera-equipped devices at a network edge by setting aside, accurately, non-informative video sequences rather than uploading video too many things over the network. Applications include security cameras, robots and augmented reality glasses.

Abstract: A method of real-time video event detection includes: obtaining, based on a natural language query, a query vector; performing multimodal feature extraction on a video stream to obtain a video vector, obtaining a similarity score by comparing the query vector to the video vector; comparing the similarity score to a predetermined threshold; and activating, based on the similarity score being above the predetermined threshold, an action trigger. The multimodal feature extraction is performed using a plurality of overlapping windows that include sequential frames of the video stream.

Abstract: Training a classifier using embeddings and building a latent space is disclosed. The embeddings may be based on weights in a trained machine learning model. Also, operation of the classifier to process video segments in real-time using the using the weights and the latent space is disclosed. The embeddings and the latent space allow the classification to be performed at an overall reduced dimensionality. The latent space is designed to efficiently scale with an increasing number of queries to permit fast search through the space. Embodiments permit real-time operation on video with dynamic features. The classifier reduces the bandwidth demand of video camera-equipped devices at a network edge by setting aside, accurately, non-informative video sequences rather than uploading video too many things over the network. Applications include security cameras, robots and augmented reality glasses.

Abstract: A method may include receiving, by a virtual assistant of a user device, an input from a user, the virtual assistant being based on software. The method may include obtaining, by the virtual assistant of the user device and via a sensor of the user device, audio information or video information of the user. The method may include determining, by the virtual assistant of the user device, an identity of the user based on the audio information or the video information of the user and a set of facial embeddings and speech embeddings that is correlated with the user, the set of facial embeddings and speech embeddings being generated using a facial embedding model, a speech embedding model, and a sound source localization model. The method may include performing, by the virtual assistant of the user device, an action based on the input and the identity of the user.