International Business Machines Corporation

Google Inc.

Waymo LLC

Microsoft Technology Licensing, LLC

Apple Inc.

Fujifilm Corporation

Amazon Technologies, Inc.

Nicira, Inc.

Globalfoundries Inc.

Kyndryl, Inc.

Globalfoundries U.S. 2 LLC

Yang Song

Ramani Ranjan Routray

Rakesh Jain

Ramya Raghavendra

Kang-Won Lee

Ming Zhao

Sumant Padbidri

Gabriel Alatorre

Vasileios Pappas

Bong Jun Ko

Charles Joseph Rosenberg

James E Olson

Dinesh C Verma

Hartwig Adam

Seraphin Bernard Calo

Bernhard Julius Klingenberg

Congcong Li

Henry Allan Rowley

Yushi Jing

Ullas Gargi

Nagapramod S Mandagere

Ann Corrao

Chung-hao Tan

Aameek Singh

Eric Kevin Butler

Jiang Wang

Dakshi Agrawal

Dragomir Dimitrov Anguelov

Divyesh Jadav

Sandeep Gopisetty

Heiko Hary Ludwig

Raghav Gupta

Thomas D Griffin

Ting He

Subhojit Roy

Qiang Zhang

Jay Yagnik

Sergey Ioffe

Min Li

Robert L Rounthwaite

Anca Sailer

Hidayatullah Habeebullah Shaikh

Colin Scott Dawson

Mu Qiao

David Randall Blea

Kavita Chavda

Meng Wang

Li-wei He

Bo Wu

Raheleh B Dilmaghani

Junhua Mao

Thomas Leung

John Zhang

Olivier Soares

Pankaj S Bavishi

Ho Yin Starsky Wong

Graham L Fyffe

Gilles M Cadet

Shaobo Guan

Tom Hagan

Robin H Lewis

Jeff N Marinstein

Mohamed Mohamed

Bo Chen

Xinying Song

Yin Zhou

Dengyong Zhou

Soner Sevinc

Xinwei Shi

Ruizhongtai Qi

Jorge D Acuña

Donnie H Kim

Liwei He

Andrew Yan-Tak Ng

Bingbing Ni

Bryan S Langston

Andre Liang Cornman

Alexander Gorban

Sanqiang Zhao

Ahmed Hassan Awadallah

Michelle Xue Zhou

Thomas K Clark

Troy Chinen

Shimei Pan

Wei Cai

Shixia Liu

Jan Henrik Jonsson

Gabor Madl

Smita J Raut

Mark V Chitti

Gabriel Wolosin

Wei Hong Qian

Steven Pantridge

Zhongli Ding

Scott Kirk Imig

Yuan Li

Shadie Hijazi

Jonathan Stringer

Kuangyou Yao

Davor Golac

Systems and methods for performing inference for word or wordpiece tokenization are disclosed using a left-to-right longest-match-first greedy process. In some examples, the vocabulary may be organized into a trie structure in which each node includes a precomputed token or token_ID and a fail link, so that the tokenizer can parse the trie in a single pass to generate a list of only those tokens or token_IDs that correspond to the longest matching vocabulary entries in the sample string, without the need for backtracking. In some examples, the vocabulary may be organized into a trie in which each node has a fail link, and any node that would share token(s) or token_ID(s) of a preceding node is instead given a prev_match link that points back to a chain of nodes with those token(s) or token_ID(s).

Inference methods for word or wordpiece tokenization

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for estimating a 3-D pose of an object of interest from image and point cloud data. In one aspect, a method includes obtaining an image of an environment; obtaining a point cloud of a three-dimensional region of the environment; generating a fused representation of the image and the point cloud; and processing the fused representation using a pose estimation neural network and in accordance with current values of a plurality of pose estimation network parameters to generate a pose estimation network output that specifies, for each of multiple keypoints, a respective estimated position in the three-dimensional region of the environment.

Multi-modal 3-D pose estimation

Provided is a knowledge distillation technique for training a student language model that, relative to a larger teacher language model, has a significantly smaller vocabulary, lower embedding dimensions, and/or hidden state dimensions. Specifically, aspects of the present disclosure are directed to a dual-training mechanism that trains the teacher and student language models simultaneously to obtain optimal word embeddings for the student vocabulary. In some implementations, this approach can be combined with learning shared projection matrices that transfer layer-wise knowledge from the teacher language model to the student language model. Example experimental results have also demonstrated higher compression efficiency and accuracy when compared with other state-of-the-art compression techniques, including the ability to compress the BERTmodel by more than 60×, with only a minor drop in downstream task metrics, resulting in a language model with a footprint of under 7 MB.

Extreme language model compression with optimal sub-words and shared projections

Various implementations disclosed herein include devices, systems, and methods that present a view of a device user's face portion, that would otherwise be blocked by an electronic device positioned in front of the face, on an outward-facing display of the user's device. The view of the user's face portion may be configured to enable observers to see the user's eyes and facial expressions as if they were seeing through a clear device at the user's actual eyes and facial expressions. Various techniques are used to provide views of the user's face that are realistic, that show the user's current facial appearance, and/or that present the face portion with 3D spatial accuracy, e.g., each eye appearing to be in its actual 3D position. Some implementations combine live data with previously-obtained data, e.g., combining live data with enrollment data.

Rendering of enrolled user's face for external display

Various implementations disclosed herein include devices, systems, and methods that present a view of a device user's face portion, that would otherwise be blocked by an electronic device positioned in front of the face, on an outward-facing display of the user's device. The view of the user's face portion may be configured to enable observers to see the user's eyes and facial expressions as if they were seeing through a clear device at the user's actual eyes and facial expressions. Various techniques are used to provide views of the user's face that are realistic, that show the user's current facial appearance, and/or that present the face portion with 3D spatial accuracy, e.g., each eye appearing to be in its actual 3D position.

Rendering of a guest user's face for external display

Systems and methods are described for merging customer profiles, such as may be implemented by a computer-implemented contact center service. In some aspects, a subset of profiles may be determined that satisfy merging criteria, where individual profiles include a plurality of data fields. At least one value in a first data field that conflicts between at least two profiles may be identified. Next a merged value may be selected for the first data field based on data deduplication criteria, where the data deduplication criteria includes at least one indicator of accuracy of values of the plurality of data fields. As a result of a determination that at least the subset of profiles of the group of profiles meet the merging criteria, at least the subset of profiles may be combined into a combined profile using the merged value.

Merging duplicate customer data

Aspects of the subject matter disclosed herein include methods, systems, and other techniques for training, in a first phase, an object classifier neural network with a first set of training data, the first set of training data including a first plurality of training examples, each training example in the first set of training data being labeled with a coarse-object classification; and training, in a second phase after completion of the first phase, the object classifier neural network with a second set of training data, the second set of training data including a second plurality of training examples, each training example in the second set of training data being labeled with a fine-object classification.

Neural networks for coarse- and fine-object classifications

Provided is a knowledge distillation technique for training a student language model that, relative to a larger teacher language model, has a significantly smaller vocabulary, lower embedding dimensions, and/or hidden state dimensions. Specifically, aspects of the present disclosure are directed to a dual-training mechanism that trains the teacher and student language models simultaneously to obtain optimal word embeddings for the student vocabulary. In some implementations, this approach can be combined with learning shared projection matrices that transfer layer-wise knowledge from the teacher language model to the student language model. Example experimental results have also demonstrated higher compression efficiency and accuracy when compared with other state-of-the-art compression techniques, including the ability to compress the BERTmodel by more than 60&#xd7;, with only a minor drop in downstream task metrics, resulting in a language model with a footprint of under 7 MB.

Systems and methods for performing inference for word or wordpiece tokenization are disclosed using a left-to-right longest-match-first greedy process. In some examples, the vocabulary may be organized into a trie structure in which each node includes a precomputed token or token ID and a fail link, so that the tokenizer can parse the trie in a single pass to generate a list of only those tokens or token IDs that correspond to the longest matching vocabulary entries in the sample string, without the need for backtracking. In some examples, the vocabulary may be organized into a trie in which each node has a fail link, and any node that would share token(s) or token_ID(s) of a preceding node is instead given a prev_match link that points back to a chain of nodes with those token(s) or token_ID(s).

Methods, computer program products, and systems are presented. The methods include, for instance: analyzing a dataset associated with a service provided by the data protection service provider in order to determine a policy for when and how to replicate the respective components of the dataset corresponding to the service from a source site to a target site, such that the target site may perform the service with a minimum cost.

Optimized disaster-recovery-as-a-service system

Systems and methods are disclosed to implement a Boolean query evaluation system that allows clients to specify a latency constraint for evaluating Boolean queries. In embodiments, the system evaluates queries using data from multiple external data sources. If some data sources fail to return data in a timely fashion to satisfy the specified latency constraint, the system will evaluate the query using the data that it was able to obtain and generate a partial answer to the query&#x2014;a partial true or a partial false. In embodiments, the query response may include a reason for the partial answer, and indicate the list of data sources that failed to provide timely results. In embodiments, the system may publish a table indicating different latency constraints specified for a category of queries and the types of answers returned for those queries. Clients may use the table select latency constraints for future queries.

Generating partial boolean query results based on client-specified latency constraints

A method of enforcing security rules for a packet on a host is provided. The method at a security service dispatcher, determines a dispatching action on a packet for each of a group of security services. Each security service is for enforcing a set of security rules on each packet. The method for each security service, sends the packet to the security service when the dispatch rule for the security service indicates that the set of security rules of the security service has to be enforced on the packet. The method for each security service, bypasses the enforcement of the security rules of the security service when the dispatch rule for the security service indicates that the set of security rules of the security service has to be bypassed for the packet.

Third-party service chaining using packet encapsulation in a flow-based forwarding element

Growing community of inventors

San Jose, CA, United States of America

Yang Song