Microsoft Technology Licensing, LLC

Google Inc.

Hewlett-packard Development Company, L.p.

Michael Isard

Andrew David Birrell

Yuan Yu

Paul Ronald Barham

Qifa Ke

Ulfar Erlingsson

Frank David McSherry

Gaurav Sareen

Mihai Budiu

Rebecca Isaacs

Dennis Craig Fetterly

Derek G Murray

Jeffrey Adgate Dean

Jian Qiao Sun

Sanjay Ghemawat

Arthur Charles Tomlin

Yonghui Wu

Ashraf A Michail

Jeffrey Neil Margolis

Martin Abadi

Ruoming Pang

Chandramohan Amyangot Thekkath

Nathan Ackerman

Martin Thomas Shetter

Georg Klein

Andrew Martin Pearson

Zhong Wu

Ryan Sepassi

Kevin Woo

Sudip Roy

Calvin Chan

Oliver M C Williams

Aakanksha Chowdhery

Laurent El Shafey

Oliver Michael Christian Williams

Douglas Kevin Service

Hyeontaek Lim

Parker Edward Schuh

Tuan Wong

Brennan Saeta

Daniel William Hurt

Henricus Johannes Maria Meijer

Shahram Izadi

Michael Burrows

Timothy Lawrence Harris

Mark David Lillibridge

Mark Steven Manasse

Roger Shane Barga

Alexandru Octavian Balan

Jason S Flaks

Alexander Sasha Stojanovic

Charles Patrick Thacker

Simon John Baker

David Kim

David G Molyneaux

Vijayan Prabhakaran

Kunal Talwar

Carl M Carter-Schwendler

Oliver M Williams

Alan Bryant Heirich

Girish Kumar

Jonathan James Currey

Jia H Deng

Pradeep Kumar Gunda

Christopher John Rossbach

Yinzhe Yu

Ehud Wieder

Savas Parastatidis

Jean-Philippe Martin

Steve Hodges

Otmar Hiliges

Derek Gordon Murray

David Changsoo Lee

Johnson Hsieh

Mark A Shand

Ashraf Ayman Michail

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for distributing machine learning workloads, e.g., computations for training a neural network or computing an inference using a neural network, across multiple hardware accelerators. One of the systems comprises a plurality of accelerator islands, each hardware accelerator island comprising a respective plurality of hardware devices that include a plurality of hardware accelerators and a corresponding host for each of the plurality of hardware accelerators; and a respective scheduler for each of the accelerator islands that is configured to schedule workloads across the plurality of accelerators and corresponding hosts in the accelerator island, wherein the system is configured to: receive data representing a machine learning workload; and assign a respective portion of the machine learning workload to each of the plurality of accelerator islands for scheduling by the respective scheduler for the accelerator island.

Asynchronous distributed data flow for machine learning workloads

General-purpose distributed data-parallel computing using a high-level language is disclosed. Data parallel portions of a sequential program that is written by a developer in a high-level language are automatically translated into a distributed execution plan. The distributed execution plan is then executed on large compute clusters. Thus, the developer is allowed to write the program using familiar programming constructs in the high level language. Moreover, developers without experience with distributed compute systems are able to take advantage of such systems.

General purpose distributed data parallel computing using a high level language

A computer-readable storage medium stores computer-executable instructions that, when executed by a processor, perform operations including scheduling first and second threads to operate independently on first and second partitions of data. The operations include beginning a first operation on the first and second partitions by the first and second threads, respectively. The operations include tracking progress of the first operation by the first and second threads using a replicated data structure. The operations include, for a record on which the first operation will be performed, adding an entry to the replicated data structure with a timestamp indicating an epoch and iteration. The operations include determining a number of yet-to-be-processed records for a selected entry of the replicated data structure. The selected entry has the most recent timestamp for the first thread. The operations include terminating the first thread when the number of yet-to-be-processed records for the selected entry is zero.

Reachability-based coordination for cyclic dataflow

Various embodiments provide techniques for working with large-scale collections of data pertaining to real world systems, such as a social network, a roadmap/GPS system, etc. The techniques perform incremental, iterative, and interactive parallel computation using a coordination clock protocol, which applies to scheduling computations and managing resources such as memory and network resources, etc., in cyclic graphs including those resulting from a differential dataflow model that performs computations on differences in the collections of data.

Methods for generating and displaying images associated with one or more virtual objects within an augmented reality environment at a frame rate that is greater than a rendering frame rate are described. The rendering frame rate may correspond with the minimum time to render images associated with a pose of a head-mounted display device (HMD). In some embodiments, the HMD may determine a predicted pose associated with a future position and orientation of the HMD, generate a pre-rendered image based on the predicted pose, determine an updated pose associated with the HMD subsequent to generating the pre-rendered image, generate an updated image based on the updated pose and the pre-rendered image, and display the updated image on the HMD. The updated image may be generated via a homographic transformation and/or a pixel offset adjustment of the pre-rendered image.

Late stage reprojection

An automatic mutual exclusion computer programming system is disclosed which allows a programmer to produce concurrent programming code that is synchronized by default without the need to write any synchronization code. The programmer creates asynchronous methods which are not permitted make changes to shared memory that they cannot reverse, and can execute concurrently with other asynchronous methods. Changes to shared memory are committed if no other thread has accessed shared memory while the asynchronous method executed. Changes are reversed and the asynchronous method is re-executed if another thread has made changes to shared memory. The resulting program executes in a serialized order. A blocking system method is disclosed which causes the asynchronous method to re-execute until the blocking method's predicate results in an appropriate value. A yield system call is disclosed which divides asynchronous methods into atomic fragments. When a yield method call is made, shared memory changes are committed if possible or reversed and the atomic fragment is re-executed.

Automatic mutual exclusion

Image search results are obtained by providing weights to visual features to emphasize features corresponding to objects of interest while simultaneously deemphasizing irrelevant or inconsistent features that lead to poor search results. In order to minimize the impact of visual features that are unreliable or irrelevant with respect to the objects of interest in the image, context-dependent weights are provided to detect visual features such that those visual features pertaining to the objects of interest are more heavily weighted than those visual features that pertain to irrelevant or unreliable portions of the image. Visual features may be weighted for images in a searchable database. Training data may be obtained and used in weighting visual features in a query image and, alternatively, in searchable database images.

Image retrieval using discriminative visual features

The techniques discussed herein efficiently perform data-parallel computations on collections of data by implementing a differential dataflow model that performs computations on differences in the collections of data. The techniques discussed herein describe defined operators for use in a data-parallel program that performs the computations on the determined differences between the collections of data by creating a lattice and indexing the differences in the collection of data according to the lattice.

Differential dataflow

A system and method for providing an augmented reality environment in which the environmental mapping process is decoupled from the localization processes performed by one or more mobile devices is described. In some embodiments, an augmented reality system includes a mapping system with independent sensing devices for mapping a particular real-world environment and one or more mobile devices. Each of the one or more mobile devices utilizes a separate asynchronous computing pipeline for localizing the mobile device and rendering virtual objects from a point of view of the mobile device. This distributed approach provides an efficient way for supporting mapping and localization processes for a large number of mobile devices, which are typically constrained by form factor and battery life limitations.

Distributed asynchronous localization and mapping for augmented reality

Growing community of inventors

San Francisco, CA, United States of America

Michael Isard