Microsoft Technology Licensing, LLC

Baidu USA LLC

Forster Energy LLC

Other

Qualcomm Incorporated

Ji Li

Amit Srivastava

Adit Krishnan

James F Coward

Furu Wei

Laurence J Newell

Peter H Chang

Ting K Yee

Derek Martin Johnson

Xingxing Zhang

Ming Zhou

Manjiang Zhang

Xu Zhou

Youjun Liu

Han Hu

Konstantin Seleskerov

Gencheng Wu

Qi Dai

Junwei Bao

Norman L Swenson

Wesley M Shao

Gregory M Cutler

Davy Huang

Huan Yang

David Alan Pechner

Jianlong Fu

Weiyu Wendy Lu

Michael W Rowan

Priyanka Vikram Sinha

Slava Yazhgur

Fatima Zohra Daha

Yixuan Wei

Xiaozhi Yu

Tin Vo

Robert K McElheny

Muin Barkatali Momin

Mingxi Cheng

Xiao Lin Sun

Matthew Charles Bashaw

Sheng Zhao

Robert Alexander Sim

Dachuan Zhang

Dolly Sobhani

Tiffany Zhang

Min Guo

Gregory Allen Mitchell

Dhruv Joshi

Shengjin Zhou

Bei Liu

Oh Kwan

Ramya Tridandapani

Youling Zou

Fu-Kang Liao

Tapan Bohra

Gregory J Mitchell

Yue Cao

Bohan Li

Sumithra Bhakthavatsalam

John Christian Leone

Shin-Sheng Tarng

Huey-Ru Tsai

Wendy Lu

Sindhu Vigasini Jambunathan

Xinfeng Chen

Varun Tandon

Antonio Andre Monteiro Manoel

Gregory Alexander DePaul

Chin-Sheng Tarng

Matthew C Bradshaw

Muqi Li

Emily Lauren Tohir

SivaPriya Kalyanaraman

Ran Zhang

Andrew Yao

Aman Malik

Yanfei Cai

Qiang Dai

A system for development of an Artificial Intelligence (AI) model while protecting sensitive user information includes: a confidential computing environment in which original prompts to the AI model written by users are collected; a trained synthetic prompt generator to generate synthetic prompts based on the original prompts, wherein the synthetic prompt generator generates anonymized synthetic prompts without sensitive user information identifiable from the original prompts; and a developer computing environment in which the synthetic prompts are submitted to the AI model under development to generate a dataset that includes the synthetic prompts and corresponding AI model output for analysis to determine updates for the AI model while protecting the sensitive user information of actual users.

Protecting sensitive user information in developing artificial intelligence models

A data processing system implements obtaining textual content in a first language from a first client device and segmenting the textual content into a plurality of first tokens. The system also implements translating the first tokens from the first language to a second language using a bilingual dictionary, extracting features information from the second tokens to create a features vector, providing the feature vector to a first natural language processing model trained to analyze textual input in the second language and to output contextual information indicating one or more topics or subject matter of the first textual content, and providing the contextual information to a first machine learning model configured to analyze the contextual information and to identify one or more content items predicted to be relevant to the contextual information. The system further implements providing the information identifying the one or more content items to the first client device.

Multilingual support for natural language processing applications

A system for retrieving multimodal assets using domain-specific knowledge includes receiving a search query for searching for multimodal assets; encoding the search query into a first query representation via a first trained query representation machine-learning (ML) model and a second query representation via a second trained query representation ML model; comparing the first query representation to a plurality of multimodal representations to calculate a first similarity score, each of the plurality of multimodal representations being a representation of one of the plurality of candidate multimodal assets; comparing the second query representation to a plurality of domain-specific representations to calculate a second similarity score, the domain-specific representations being representations of domain-specific data associated with one or more of the plurality of the multimodal representations; calculating a third similarity score based on keyword matching between the domain-specific data and the one or more search terms in the search query; aggregating the first, second and third similarity scores to calculate a total similarity score for each of the plurality of candidate multimodal assets; ranking the plurality of candidate multimodal assets based on the total similarity scores to identify search results for the search query; and providing the identified candidate multimodal assets for display as the search results.

Method and system of using domain specific knowledge in retrieving multimodal assets

A system and method and for retrieving one or more one or more multimodal assets includes receiving a search query for searching for one or more multimodal assets from among a plurality of candidate multimodal assets, encoding the search query into one or more query embedding representations via a trained query representation machine-learning (ML) model, comparing, via a matching unit, the one or more query embedding representations to a plurality of multimodal tensor representations, each of the plurality of multimodal tensor representations being a representation of one of the plurality of candidate multimodal assets, and identifying, based on the comparison, at least one of the plurality of the candidate multimodal assets as a search result for the search query, and providing the at least one of the plurality of the candidate multimodal assets for display as the search result.

Method and system of retrieving multimodal assets

A data processing system implements a dynamic knowledge distillation process including dividing training data into a plurality of batches of samples and distilling a student model from a teacher model using an iterative knowledge distillation. The process includes instantiating an instance of the teacher model and the student model in a memory of the data processing system and obtaining a respective batch of training data from the plurality of batches of samples in the memory. The process includes training the teacher and student models using each of the samples in the respective batch of the training data, evaluating the performance of the student model compared with the performance of the teacher model, and providing feedback to student model to adjust the behavior of the student model based on the performance of the student model.

Scalable knowledge distillation techniques for machine learning

A system and method and for retrieving assets from a personalized asset library includes receiving a search query for searching for assets in one or more asset libraries, the one or more asset libraries including a personalized asset library; encoding the search query into embedding representations via a trained query representation machine-learning (ML) model; comparing, via a matching unit, the query embedding representations to a plurality of asset representations, each of the plurality of asset representations being a representation of one of the plurality of candidate assets; identifying, based on the comparison, at least one of the plurality of the candidate assets as a search result for the search query; and providing the identified plurality of candidate assets for display as the search result. The plurality of asset representations for the one or more assets in the personalized content library are generated automatically without human labeling.

Method and system of retrieving assets from personalized asset libraries

A data processing system for generating training data for a multilingual NLP model implements obtaining a corpus including first and second content items. The first content items are English-language textual content, and the second content items are translations of the first content items in one or more non-English target languages. The system further implements selecting a first content item from the first content items, generating a plurality of candidate labels for the first content item by analyzing the first content item with a plurality of first English-language NLP models, selecting a first label from the plurality of candidate labels, generating first training data by associating the first label with the first content item, generating second training data by associating the first label with a second content item of the second content items, and training a pretrained multilingual NLP model with the first training data and the second training data.

Multilingual model training using parallel corpora, crowdsourcing, and accurate monolingual models

A data processing system implements obtaining first textual content in a first language from a first client device; determining that the first language is supported by a first machine learning model; obtaining a guard list of prohibited terms associated with the first language; determining that the textual content does not include one or more prohibited terms associated based on the guard list; providing the first textual content as an input to the first machine learning model responsive to the textual content not including the one or more prohibited terms; analyzing the first textual content with the first machine learning model to obtain a first content recommendation; obtaining a first content recommendation policy that identifies content associated with the first language that may not be provided as a content recommendation; determining that the first content recommendation is not prohibited; and providing the first content recommendation to the first client device.

Multilingual content recommendation pipeline

Systems and methods for providing a machine learning-powered framework to transform overloaded text documents is provided. The system generates a plurality of candidate templates offline. During runtime, the system accesses a text document and analyzes the text document to identify segmentation data. The segmentation data can indicate a plurality of segments derived from the text document. The system then accesses a plurality of candidate templates, whereby each candidate template comprises a plurality of pages having a different background element that shares a common theme. The plurality of candidate templates are ranked based on at least the segmentation data. The network then generates multiple presentation pages for each of a predetermined number of top ranked candidate templates by incorporating each of the plurality of segments into a corresponding page of the plurality of pages for each of the top ranked candidate templates. The multiple presentation pages are presented for each of the top ranked candidate templates as a recommendation.

Machine learning-powered framework to transform overloaded text documents

A system and method and for retrieving one or more visual assets includes receiving a search query for the one or more visual assets, the search query including textual data, encoding the textual data into one or more text embedding representations via a trained text representation machine-learning (ML) model, transmitting the one or more text embedding representations to a matching and selection unit, providing visual embedding representations of one or more visual assets to the matching and selection unit, comparing, by the matching and selection unit, the one or more text embedding representations to the visual embedding representations to identify one or more visual asset search results, and providing the one or more visual asset search results for display.

Method and system of content retrieval for visual data

Automatic generation of intelligent content is created using a system of computers including a user device and a cloud-based component that processes the user information. The system performs a process that includes receiving an input document and parsing the input document to generate inputs for a natural language generation model using a text analysis model. The natural language generation model generates one or more candidate presentation scripts based on the inputs. A presentation script is selected from the candidate presentation scripts and displayed. A text-to-speech model may be used to generate a synthesized audio presentation of the presentation script. A final presentation may be generated that includes a visual display of the input document and the corresponding audio presentation in sync with the visual display.

Automated script generation and audio-visual presentations

The present disclosure relates to processing operations that execute image classification training for domain-specific traffic, where training operations are entirely compliant with data privacy regulations and policies. Image classification model training, as described herein, is configured to classify meaningful image categories in domain-specific scenarios where there is unknown data traffic and strict data compliance requirements that result in privacy-limited image data sets. Iterative image classification training satisfies data compliance requirements through a combination of online image classification training and offline image classification training. This results in tuned image recognition classifiers that have improved accuracy and efficiency over general image recognition classifiers when working with domain-specific data traffic. One or more image recognition classifiers are independently trained and tuned to detect an image class for image classification. Training of independent image recognition classifiers is also utilized for training and tuning of deeper learning models for image classification.

Image classification modeling while maintaining data privacy compliance

A scheme for modifying an image is disclosed, which includes receiving a source image having a first image configuration; determining a second image configuration for a target image; providing the received source image to an AI engine trained to identify, based on a set of rules related to visual features, candidate regions from the source image; generating proposal images based on the candidate regions, respectively; determining, based on prior aesthetical evaluation data, an aesthetical value of each regional proposal image; and selecting, based on the determined aesthetical value of each regional proposal image, one of the regional proposal images as the target image; extracting, from the AI engine, the target image; and causing the target image to be displayed via a display of a user device.

AI-based aesthetical image modification

Shortwave infrared (SWIR) hyperspectral imaging (HSI) systems comprise a supercontinuum laser source configured to illuminate objects and a receiver comprising a spectrometer configured to receive light reflected from the objects. In some cases, hyperspectral images can be created by raster scanning of the source/receiver across a scene. The supercontinuum laser source provides active illumination to allow collection of hyperspectral imagery during day (including overcast conditions) and night.

Hyperspectral imaging systems

A method and system for transforming an input image via a plurality of image transformation stylizers includes receiving the input image; providing the input image, information about the plurality of image transformation stylizers and at least one of user data, history data, and contextual data to a trained machine-learning (ML) model for selecting a subset of the plurality of image transformation stylizers; receiving as an output from the ML model the subset of image transformation stylizers; executing the subset of the image transformation stylizers on the input image to generate a plurality of transformed output images; ranking the plurality of transformed output images based on at least one of the input image, the user data, the history data, and the contextual data; and providing the ranked plurality of transformed output images for display.

Image transformation infrastructure

Growing community of inventors

San Jose, CA, United States of America

Ji Li