Abstract: The technology relates to methods and systems for implicit calibration for gaze tracking. This can include receiving, by a neural network module, display content that is associated with presentation on a display screen (1202). The neural network module may also receive uncalibrated gaze information, in which the uncalibrated gaze information includes an uncalibrated gaze trajectory that is associated with a viewer gaze of the display content on the display screen (1204). A selected function is applied by the neural network module to the uncalibrated gaze information and the display content to generate a user-specific gaze function (1206). The user-specific gaze function has one or more personalized parameters. And the neural network module can then apply the user-specific gaze function to the uncalibrated gaze information to generate calibrated gaze information associated with the display content on the display screen (1208).

Abstract: Technologies are described here for, among other things, improving search query relevance by executing a query on a search engine, retrieving search-page-data generated from executing the query, the search-page-data including document-titles and universal resource locators (URLs), each document-title being a title of a document associated with a URL, determining relevant-entity-words in the query from an entity relevance score for matching search terms in the query, Domain-URLs, and Domain-Titles, determining relevant-intent-words in the query from an intent-word relevance score based on a number of times a search term appears in the query and the URLs relative to other search terms in the query and the URLs, comparing each of the determined relevant-entity-words and each of the determined relevant-intent-words with a plurality of stored past-user queries, retrieving the plurality of stored past-user search queries including the relevant-entity-words and the relevant-intent-words, and transmitting a set of qu

Abstract: Systems and methods for a weakly supervised action localization model are provided. Example models according to example aspects of the present disclosure can localize and/or classify actions in untrimmed videos using machine-learned models, such as convolutional neural networks. The example models can predict temporal intervals of human actions given video-level class labels with no requirement of temporal localization information of actions. The example models can recognize actions and identify a sparse set of keyframes associated with actions through adaptive temporal pooling of video frames, wherein the loss function of the model is composed of a classification error and a sparsity of frame selection. Following action recognition with sparse keyframe attention, temporal proposals for action can be extracted using temporal class activation mappings, and final time intervals can be estimated corresponding to target actions.

Abstract: Systems and methods for a weakly supervised action localization model are provided. Example models according to example aspects of the present disclosure can localize and/or classify actions in untrimmed videos using machine-learned models, such as convolutional neural networks. The example models can predict temporal intervals of human actions given video-level class labels with no requirement of temporal localization information of actions. The example models can recognize actions and identify a sparse set of keyframes associated with actions through adaptive temporal pooling of video frames, wherein the loss function of the model is composed of a classification error and a sparsity of frame selection. Following action recognition with sparse keyframe attention, temporal proposals for action can be extracted using temporal class activation mappings, and final time intervals can be estimated corresponding to target actions.

Abstract: Systems and methods for a weakly supervised action localization model are provided. Example models according to example aspects of the present disclosure can localize and/or classify actions in untrimmed videos using machine-learned models, such as convolutional neural networks. The example models can predict temporal intervals of human actions given video-level class labels with no requirement of temporal localization information of actions. The example models can recognize actions and identify a sparse set of keyframes associated with actions through adaptive temporal pooling of video frames, wherein the loss function of the model is composed of a classification error and a sparsity of frame selection. Following action recognition with sparse keyframe attention, temporal proposals for action can be extracted using temporal class activation mappings, and final time intervals can be estimated corresponding to target actions.

Abstract: A computing system receives an indication that a user has selected a passage shown on a search engine results page (SERP) presented on a display. Upon receiving the indication, the computing system identifies a plurality of suggested queries related to the passage, where the plurality of suggested queries are generated based upon the passage and an entry for an entity in a knowledge graph. Upon identifying the plurality of suggested queries, the computing system presents the plurality of suggested queries in a pop-up graphical element that overlays a portion of the SERP, where the pop-up graphical element is located proximate to the passage shown on the SERP. When a query in the plurality of suggested queries is selected, a second SERP is presented on the display, where the second SERP is based upon the query.

Abstract: Technologies are described here for, among other things, improving search query relevance by executing a query on a search engine, retrieving search-page-data generated from executing the query, the search-page-data including document-titles and universal resource locators (URLs), each document-title being a title of a document associated with a URL, determining relevant-entity-words in the query from an entity relevance score for matching search terms in the query, Domain-URLs, and Domain-Titles, determining relevant-intent-words in the query from an intent-word relevance score based on a number of times a search term appears in the query and the URLs relative to other search terms in the query and the URLs, comparing each of the determined relevant-entity-words and each of the determined relevant-intent-words with a plurality of stored past-user queries, retrieving the plurality of stored past-user search queries including the relevant-entity-words and the relevant-intent-words, and transmitting a set of qu

Abstract: Technologies are described here for, among other things, improving search query relevance by executing a query on a search engine, retrieving search-page-data generated from executing the query, the search-page-data including document-titles and universal resource locators (URLs), each document-title being a title of a document associated with a URL, determining relevant-entity-words in the query from an entity relevance score for matching search terms in the query, Domain-URLs, and Domain-Titles, determining relevant-intent-words in the query from an intent-word relevance score based on a number of times a search term appears in the query and the URLs relative to other search terms in the query and the URLs, comparing each of the determined relevant-entity-words and each of the determined relevant-intent-words with a plurality of stored past-user queries, retrieving the plurality of stored past-user search queries including the relevant-entity-words and the relevant-intent-words, and transmitting a set of qu

Abstract: Technologies are described here for, among other things, improving search query relevance by executing a query on a search engine, retrieving search-page-data generated from executing the query, the search-page-data including document-titles and universal resource locators (URLs), each document-title being a title of a document associated with a URL, determining relevant-entity-words in the query from an entity relevance score for matching search terms in the query, Domain-URLs, and Domain-Titles, determining relevant-intent-words in the query from an intent-word relevance score based on a number of times a search term appears in the query and the URLs relative to other search terms in the query and the URLs, comparing each of the determined relevant-entity-words and each of the determined relevant-intent-words with a plurality of stored past-user queries, retrieving the plurality of stored past-user search queries including the relevant-entity-words and the relevant-intent-words, and transmitting a set of qu

Abstract: Systems and methods for a weakly supervised action localization model are provided. Example models according to example aspects of the present disclosure can localize and/or classify actions in untrimmed videos using machine-learned models, such as convolutional neural networks. The example models can predict temporal intervals of human actions given video-level class labels with no requirement of temporal localization information of actions. The example models can recognize actions and identify a sparse set of keyframes associated with actions through adaptive temporal pooling of video frames, wherein the loss function of the model is composed of a classification error and a sparsity of frame selection. Following action recognition with sparse keyframe attention, temporal proposals for action can be extracted using temporal class activation mappings, and final time intervals can be estimated corresponding to target actions.