Hulu, LLC

Nokia Technologies Oy

Lenovo (beijing) Limited

Panasonic Corporation

Nokia Corporation

Tsinghua University

Beijing Lenovo Software Ltd.

Xiaohui Xie

Kongqiao Wang

Yunsheng Jiang

Yingfei Liu

Wei Wang

Zhen Liu

Bin Wang

Zhigang Li

Xia Wang

Huanglingzi Liu

Chunyang Wei

Yuezhong Tang

Dingming Wu

Xianghui Mao

Pengyu Zhao

Chunxu Xu

Honglei Zhang

Yanming Zou

Shulei Ma

Brian T Morrison

Pengfei Gao

Yaqi Wang

Fanding Li

Kaiwen Deng

Bo Liu

Peng Wang

Liangliang Li

Min Zhang

Qiang Wei

Weijie Liu

Christopher Russell Kehler

Yiqun Liu

Gang Wu

Cailiang Liu

Xiaorong Gao

Jiarui Yang

Manjun Xiao

Rui Lu

Xuesong Chen

Yin Zheng

Qingying Chen

Zhiwei Zeng

Xiao Zeng

Zhicheng Zhao

Anni Cai

Youlong Cheng

Jing Ruan

Diman Shen

Chao Yi

Hongxiang Chen

Ziyi Ye

Weihang Su

Ran Cao

Shuqi Zhu

Shaoping Ma

Yike Sun

Hanqi Wang

In some embodiments, a method receives a first sequence of inputs for processing via a sub-model of a plurality of sub-model. The plurality of sub-models are part of a main model. An input in the sequence of inputs is masked with a masked value to generate a second sequence of inputs. The method processes the second sequence of inputs using the sub-model to generate a sequence of features that correspond to the second sequence of inputs and processes the sequence of features to generate a first output. The first output is processed to generate a second output of the main model. The sub-model is trained based on a feature in the sequence of features that corresponds to the masked input and the second output.

Masked model training of a prediction network

In some embodiments, a method receives a line item factor for a line item at a first model. The method receives a competition factor at a second model. The first model generates a first value for a representation parameter based on the line item factor. The representation parameter is used to generate a representation for estimating selection of the instance of supplemental content for the line item for delivery. The second model generates a modification factor based on the competition factor. The modification factor is configured to modify the first value for the representation parameter based on competition from other line items to select the instance of supplemental content for the line item for delivery. The method modifies the first value for the representation parameter using the modification factor to generate a second value for the representation parameter where the second value is used to generate the representation.

Representation parameter adjustment using separate models

The present disclosure relates to a method and an apparatus, an electronic device and a storage medium for web search. The method includes: acquiring a steady-state visual evoked potential in EEG information, where the steady-state visual evoked potential is generated when a user gazes at a key on a query inputting keyboard in a visual spelling page; sending the steady-state visual evoked potential to a server, such that the server, based on the steady-state visual evoked potential, determines a character string inputted by the user, and a landing page corresponding to the character string; and in response to receiving the landing page sent by the server, displaying the landing page.

Method, apparatus, electronic device, and storage medium for web search

In some embodiments, a method determines a reward metric based on feedback for an instance of content. A delivery status for a delivery constraint of the instance of content is applied to the reward metric to generate a constrained reward metric. The method uses the constrained reward metric to train a model. The model is used to select from a plurality of instances of content. One of the plurality of instances of content is selected for delivery using the model.

Reinforcement learning for delivery of content

In some embodiments, a method receives a sequence of subsidiary features that are associated with a sequence of main features. A subsidiary feature provides subsidiary information for a main feature. A sequence of first weights for the sequence of subsidiary features is generates where a first weight in the sequence of first weights is generated based on a respective subsidiary feature. The method processes the sequence of first weights to generate a sequence of second weights. The processing uses relationships in the sequence of first weights to generate values of the second weights. The method uses the sequence of second weights to process the sequence of main features to generate an output for the sequence of main features.

Reweighting network for subsidiary features in a prediction network

In some embodiments, a method receives a history of videos that were viewed on a video delivery system as a first sentence in a sequential order and a target video as a second sentence as input to a prediction network. The prediction network analyzes representations for the history of videos and a representation of the target video. The prediction network generates a session representation based on bidirectionally analyzing a sequence of the first representations and the second representation. The method uses the session representation to determine whether to recommend the target video.

Content recommendation generation using content based on bi-directional prediction

In some embodiments, a method receives a request for supplemental content to be provided in association with main content. The method selects an instance of supplemental content based on a long-term reward metric and a short-term reward metric. The long-term reward metric is based on feedback from delivery of a plurality of instances of supplemental content and a delivery status for a delivery constraint of one instance of supplemental content. The short-term reward metric is based on feedback from delivery of the one instance of supplemental content. The long-term reward metric is based on feedback from delivery of a plurality of instances of supplemental content and the short-term reward metric is based on feedback from delivery of one instance of supplemental content. The instance of supplemental content is provided to a client device.

Reinforcement learning for guaranteed delivery of supplemental content

In some embodiments, a method selects a sequence of programs watched by a user account. The method calculates a first set of weights based on comparing content of the sequence of programs to content of a target program and calculates a second set of weights based on an order of the sequence of the programs and the first of weights. The first set of weights and the second set of weights are applied to the sequence of programs to generate a prediction of a similarity of the sequence of programs to the target program. Then, the method outputs the prediction of the similarity for use in determining a recommendation for the user account.

Watch sequence modeling for recommendation ranking

In one embodiment, a method receives a set of frames from a video at a first classifier. The first classifier classifies the set of frames with classification scores that indicate a confidence that a frame contains end credit content using the first classifier using a first model that classifies content from the set of frames. A second classifier then refines the classification scores from neighboring frames in the set of frames using a second classifier using a second model that classifies classification scores from the first classifier. A boundary point is selected between a frame in the set of frames considered not including end credit content and a frame in the set of frames including end credit content based on the refined classification scores.

Content type detection in videos using multiple classifiers

In some embodiments, a method maps attributes of metadata for a plurality of content instances to metadata nodes. The metadata nodes are connected to type nodes that define a type of metadata for each metadata node and a respective content node for a respective content instance. The method generates a plurality of sample paths using the content nodes, the metadata nodes, and the type nodes from the mapping of attributes of the metadata. A similarity of content nodes is analyzed using the plurality of sample paths. Then, the method generates a representation of each of the plurality of content instances using the similarity of the content nodes. The representation represents the similarity between content instances in the plurality of content instances.

Representation of content based on content-level features

In some embodiments, a first detector generates a first output based on a first probability that an image was inserted in a video. The first detector is trained with a set of known images to detect the set of known images. A second detector generates a second output based on a second probability that an image was inserted in the video. The second detector is used to detect the set of unknown images without training. The method analyzes the first output from the first detector based on the probability of the image existing in the video and the second output from the second detector based on the probability of the image existing in the video to generate a combined score from the first output and the second output. An indication of whether the image is detected in the video is output based on the combined score.

Image detection using multiple detection processes

In some embodiments, a method detects a first set of frames in a video that include lines of text, the detecting performed at a frame level on each individual frame. A first representation is generated from the first set of frames and a second representation is generated from the first set of frames. The method filters the first representation based on a number of lines of text within a space in the space dimension to select a second set of frames and filters the second representation based on a number of frames within time intervals in the time dimension to select a third set of frames. Frames in both the second set of frames and the third set of frames are analyzed to determine whether the lines of text in both the second set of frames and the third set of frames are burned-in subtitles.

Frame level and video level text detection in video

In some embodiments, a method trains a first prediction network to predict similarity between images in videos. The training uses boundaries detected in the videos to train the prediction network to predict images in a same scene to have similar feature descriptors. The first prediction network generates feature descriptors that describe library images from videos in a video library offered to users of a video delivery service. A search image is received and the prediction network predicts one or more library images for one or more videos that are predicted to be similar to the received image. The one or more library images for the one or more videos are provided as a search result.

Scene level video search

In one embodiment, a system detects objects in an image and generates attention regions that are positioned in the image based on first positions of the objects in the image. Focus points for the objects are generated for the attention regions at one or more second positions. Focus boxes are generated using the second positions of the focus points. Then, the system generates information for a motion effect using content of the image based on a number of the focus boxes and third positions of the focus boxes.

Video motion effect generation based on content analysis

Growing community of inventors

Beijing, China

Xiaohui Xie