Abstract: Systems and methods for flexible parameter sharing for multi-task learning are provided. A training method can include obtaining a test input, selecting a particular task from one or more tasks, and training a multi-task machine-learned model for the particular task by performing a forward pass using the test input and one or more connection probability matrices to generate a sample distribution of test outputs, training the components of the machine-learned model based at least in part on the sample distribution, and performing a backwards pass to train a connection probability matrix of the multi-task machine-learned model using a straight-through Gumbel-softmax approximation.

Abstract: Systems and methods of the present disclosure are directed to a method for generating a machine-learned multitask model configured to perform tasks. The method can include obtaining a machine-learned multitask search model comprising candidate nodes. The method can include obtaining tasks and machine-learned task controller models associated with the tasks. As an example, for a task, the method can include using the task controller model to route a subset of the candidate nodes in a machine-learned task submodel for the corresponding task. The method can include inputting task input data to the task submodel to obtain a task output. The method can include generating, using the task output, a feedback value based on an objective function. The method can include adjusting parameters of the task controller model based on the feedback value.

Abstract: A computing system can be configured for low-photon-count visual object classification. The computing system can include a photon detection system including one or more cells. Each of the one or more cells can include one or more photon detectors. Each of the one or more photon detectors can be configured to output photon signatures in response to a photon being incident on the one or more photon detectors. The computing system can include one or more processors and one or more memory devices storing computer-readable data. The data can include a low-photon-count classification model and one or more instructions that, when implemented, cause the one or more processors to perform operations for low-photon-count visual object recognition. The operations can include obtaining a photon signature from a photon detection system. The operations can include providing the photon signature to a low-photon-count classification model.

Abstract: A method of determining a final architecture for a task neural network for performing a target machine learning task is described. The target machine learning task is associated with a target training dataset.

Abstract: Systems and methods for flexible parameter sharing for multi-task learning are provided. A training method can include obtaining a test input, selecting a particular task from one or more tasks, and training a multi-task machine-learned model for the particular task by performing a forward pass using the test input and one or more connection probability matrices to generate a sample distribution of test outputs, training the components of the machine-learned model based at least in part on the sample distribution, and performing a backwards pass to train a connection probability matrix of the multi-task machine-learned model using a straight-through Gumbel-softmax approximation.

Abstract: A demographics analysis trains classifier models for predicting demographic attribute values of videos and users not already having known demographics. In one embodiment, the demographics analysis system trains classifier models for predicting demographics of videos using video features such as demographics of video uploaders, textual metadata, and/or audiovisual content of videos. In one embodiment, the demographics analysis system trains classifier models for predicting demographics of users (e.g., anonymous users) using user features based on prior video viewing periods of users. For example, viewing-period based user features can include individual viewing period statistics such as total videos viewed. Further, the viewing-period based features can include distributions of values over the viewing period, such as distributions in demographic attribute values of video uploaders, and/or distributions of viewings over hours of the day, days of the week, and the like.

Abstract: Provided content is determined to contain an asset represented by reference content by comparing digital fingerprints of the provided content and the reference content. The fingerprints of the reference content and the provided content are generated using a convolutional neural network (CNN). The CNN is trained using a plurality of frame triplets including an anchor frame representing the reference content, a positive frame which is a transformation of the anchor frame, and a negative frame representing content that is not the reference content. The provided content is determined to contain the asset represented by the reference content based on a similarity measure between the generated fingerprints. If the provided content is determined to contain the asset represented by the reference content, a policy associated with the asset is enforced on the provided content.

Abstract: A system for automated annotation of images and videos points a mobile device towards an object of interest, such as a building or landscape scenery, for the device to display an image of the scene with an annotation for the object. An annotation can include names, historical information, and links to databases of images, videos, and audio files. Different techniques can be used for determining positional placement of annotations, and, by using multiple techniques, positioning can be made more precise and reliable. The level of detail of annotation information can be adjusted according to the precision of the techniques used. A trade-off can be taken into account between precision of annotation and communication cost, delay and/or power consumption. An annotation database can be updated in a self-organizing way. Public information as available on the web can be converted to annotation data.

Abstract: Techniques are shown for predicting the number of times a media selection will be consumed by one or more users at a target time. Examples of user behavior during the consumption of a media selection are chosen as input features. A partitioner separates a set of media selections into a training subset and an evaluation subset. The input features are transformed into feature vectors, and a learned function is derived to define a relationship between the feature vector for the training subset and the number of times a media selection from the training subset is consumed. The learned function is then applied to a feature vector for the evaluation subset to test its accuracy.

Abstract: Described herein are techniques related to prediction of video revenue for non-monetized videos. This Abstract is submitted with the understanding that it will not be used to interpret or limit the scope and meaning of the claims. A video-revenue prediction tool predicts revenue for non-monetized videos using historical revenue data of monetized videos.

Abstract: A method of predicting interest levels associated with publication and content item combinations is described. Additionally, a server computing device for predicting interest levels associated with publication and content item combinations is described. Further, a computer-readable storage device having processor-executable instructions embodied thereon is described. The processor-executable instructions are for predicting interest levels associated with publication and content item combinations.

Abstract: A system for creating audiences for a shared content publisher, includes a data store comprising a computer readable medium storing a program of instructions for audiences for a shared content publisher; a processor that executes the program of instructions; a data processor to monitor access to an Internet web site by a first set of users for a reference time period; a window extraction module, based on a reference split period, to divide the monitored access into a vector X and a vector Y, wherein vector X is defined by accesses by the first set of users before the reference split period, and vector Y is defined by accesses by the first user after the reference split period; and a data analysis module to create a model based on the vector X and the vector Y, to evaluate the model based on a second set of users accessing content similar to vector X, to create a final model based on the evaluation, and to score a group of users associated with the shared content publisher based on the final model, the data pr

Abstract: A method of predicting interest levels associated with publication and content item combinations is described. Additionally, a server computing device for predicting interest levels associated with publication and content item combinations is described. Further, a computer-readable storage device having processor-executable instructions embodied thereon is described. The processor-executable instructions are for predicting interest levels associated with publication and content item combinations.

Abstract: A classifier training system trains classifiers for inferring the geographic locations of videos. A number of classifiers are provided, where each classifier corresponds to a particular location and is trained from a training set of videos that have been labeled as representing the location. In one embodiment, the training set is further restricted to those videos in which a landmark matching the location label is detected. The classifier training system extracts, from each of these videos, features that characterize the video, such as audiovisual features, text features, address features, landmark features, and category features. Based on these features, the classifier training system trains a location classifier for the corresponding location. Each of the location classifiers can be applied to videos without associated location labels to predict whether, or how strongly, the video represents the corresponding location.

Abstract: A demographics analysis trains classifier models for predicting demographic attribute values of videos and users not already having known demographics. In one embodiment, the demographics analysis system trains classifier models for predicting demographics of videos using video features such as demographics of video uploaders, textual metadata, and/or audiovisual content of videos. In one embodiment, the demographics analysis system trains classifier models for predicting demographics of users (e.g., anonymous users) using user features based on prior video viewing periods of users. For example, viewing-period based user features can include individual viewing period statistics such as total videos viewed. Further, the viewing-period based features can include distributions of values over the viewing period, such as distributions in demographic attribute values of video uploaders, and/or distributions of viewings over hours of the day, days of the week, and the like.

Abstract: Photographic images can be used to enhance three-dimensional (3D) virtual models of a physical location. In an embodiment, a method of generating a 3D scene geometry includes obtaining a first plurality of images and corresponding distance measurements for a first vehicle trajectory; obtaining a second plurality of images and corresponding distance measurements for a second vehicle trajectory, the second vehicle trajectory intersecting the first vehicle trajectory; registering a relative vehicle position and orientation for one or more segments of each of a first vehicle trajectory and a second vehicle trajectory; generating a three-dimensional geometry for each vehicle trajectory; mapping the three-dimensional geometries for each vehicle trajectory onto a common reference system based on the registering; and merging the three-dimensional geometries from both trajectories to generate a complete scene geometry.

Abstract: A system and method generates super-resolution images and videos using motion-compensated low-resolution images and videos. An image is selected as a primary image from a plurality of low-resolution images and the rest of the low-resolution images are registered as secondary images with respect to the primary image. Each registered secondary image is transformed to a motion compensated image. A mask value for a pixel in each motion compensated image is estimated. The super-resolution image of the primary image is generated by combining the mask values and the motion compensated secondary images. Similarly, a low-resolution video is segmented into a plurality of video objects, each of which is represented by an alpha layer. A super-resolution frame of the segmented video object is generated. The super-resolution frames of each segmented video object are combined using the alpha layers to create a super-resolution frame of the resulting video.

Abstract: A demographics analysis trains classifier models for predicting demographic attribute values of videos and users not already having known demographics. In one embodiment, the demographics analysis system trains classifier models for predicting demographics of videos using video features such as demographics of video uploaders, textual metadata, and/or audiovisual content of videos. In one embodiment, the demographics analysis system trains classifier models for predicting demographics of users (e.g., anonymous users) using user features based on prior video viewing periods of users. For example, viewing-period based user features can include individual viewing period statistics such as total videos viewed. Further, the viewing-period based features can include distributions of values over the viewing period, such as distributions in demographic attribute values of video uploaders, and/or distributions of viewings over hours of the day, days of the week, and the like.

Abstract: A classifier training system learns classifiers for categories by combining data from a category-instance repository comprising relationships between categories and more specific instances of those categories with a set of video classifiers for different concepts. The category-instance repository is derived from the domain of textual documents, such as web pages, and the concept classifiers are derived from the domain of video. Taken together, the category-instance repository and the concept classifiers provide sufficient data for obtaining accurate classifiers for categories that encompass other lower-level concepts, where the categories and their classifiers may not be obtainable solely from the video domain.

Abstract: In previously known imaging devices as in still and motion cameras, for example, image sensor signal response typically is linear as a function of intensity of incident light. Desirably, however, akin to the response of the human eye, response is sought to be nonlinear and, more particularly, essentially logarithmic. Preferred nonlinearity is realized in image sensor devices of the invention upon severely limiting the number of pixel states, combined with clustering of pixels into what may be termed as super-pixels.