Abstract: A system for optimizing and performing a model training session performs at least two profiling sessions. Memory allocation calls are intercepted to generate first and second memory usage profile logs generated during two different sessions, respectively, at different batch sizes. The system classifies memory objects identified in the first and second memory usage profile logs as a same-size object or a proportional-size object, processes the classified memory objects to simulate memory usage for various candidate batch sizes, determines an optimal batch size which maximizes memory usage without exceeding an available memory capacity, and generates a memory schedule containing offset assignments for each classified memory object. Then the system performs the training session at the optimal batch size to train the untrained model, using a custom allocator to perform memory allocation in accordance with the memory schedule.

Abstract: Disclosed herein are systems, methods, and devices for using adaptive learning to identify objects. An object-identifying device performs a first object identification based on one or more features of a first modality of an object retrieved from an image frame including the object and a first database including first modality identification features. A second object identification is performed based on one or more features of a second modality of the object retrieved from the image frame and a second database including second modality identification features. The second database is updated by adaptively learning a new second modality identification feature according to a first identification result of the first object identification. The second object identification is trained with the updated second database and determines a final identification result by integrating a first identification result of the first object identification and a second identification result of the second object identification.

Abstract: Systems, methods, apparatuses, and computer program products to provide stochastic trajectory prediction using social graph networks. An operation may comprise determining a first feature vector describing destination features of a first person depicted in an image, generating a directed graph for the image based on all people depicted in the image, determining, for the first person, a second feature vector based on the directed graph and the destination features, sampling a value of a latent variable from a learned prior distribution, the latent variable to correspond to a first time interval, and generating, based on the sampled value and the feature vectors by a hierarchical long short-term memory (LSTM) executing on a processor, an output vector comprising a direction of movement and a speed of the direction of movement of the first person at a second time interval, subsequent to the first time interval.

Abstract: Performing online learning for a model to detect unseen actions in an action recognition system is disclosed. The method includes extracting semantic features in a semantic domain from semantic action labels, transforming the semantic features from the semantic domain into mixed features in a mixed domain, and storing the mixed features in a feature database. The method further includes extracting visual features in a visual domain from a video stream and determining if the visual features indicate an unseen action in the video stream. If no unseen action is determined, applying an offline classification model to the visual features to identify seen actions, assigning identifiers to the identified seen actions, transforming the visual features from the visual domain into mixed features in the mixed domain, and storing the mixed features and seen action identifiers in the feature database.

Abstract: Disclosed herein are systems, methods, and devices for using adaptive learning to identify objects. An object-identifying device performs a first object identification based on one or more features of a first modality of an object retrieved from an image frame including the object and a first database including first modality identification features. A second object identification is performed based on one or more features of a second modality of the object retrieved from the image frame and a second database including second modality identification features. The second database is updated by adaptively learning a new second modality identification feature according to a first identification result of the first object identification. The second object identification is trained with the updated second database and determines a final identification result by integrating a first identification result of the first object identification and a second identification result of the second object identification.

Abstract: Systems, methods, apparatuses, and computer program products to provide stochastic trajectory prediction using social graph networks. An operation may comprise determining a first feature vector describing destination features of a first person depicted in an image, generating a directed graph for the image based on all people depicted in the image, determining, for the first person, a second feature vector based on the directed graph and the destination features, sampling a value of a latent variable from a learned prior distribution, the latent variable to correspond to a first time interval, and generating, based on the sampled value and the feature vectors by a hierarchical long short-term memory (LSTM) executing on a processor, an output vector comprising a direction of movement and a speed of the direction of movement of the first person at a second time interval, subsequent to the first time interval.

Abstract: A system and a method for rendering an image including an image gateway arranged to receive content data representative of one or more objects and style data representative of one or more desired rendering styles; and an image manipulation engine arranged to process the content data and the style data to render a target image, wherein the target image includes a representation of the one or more objects and includes the one or more desired rendering styles.

Abstract: Examples methods, apparatus, and articles of manufacture corresponding to a spectral nonlocal block have been disclosed. An example apparatus includes a first convolution filter to perform a first convolution using input features and first weighted kernels to generate first weighted input features, the input features corresponding to data of a neural network; an affinity matrix generator to: perform a second convolution using the input features and second weighted kernels to generate second weighted input features; perform a third convolution using the input features and third weighted kernels to generate third weighted input features; and generate an affinity matrix based on the second and third weighted input features; a second convolution filter to perform a fourth convolution using the first weighted input features and fourth weighted kernels to generate fourth weighted input features; and a accumulator to transmit output features corresponding to a spectral nonlocal operator.