Microsoft Technology Licensing, LLC

Amazon Technologies, Inc.

Harvard College

Wei Ting Chen

Hanumantha Rao Kodavalla

Panagiotis Antonopoulos

Girish Mittur Venkataramanappa

Peter Byrne

Raghavendra Thallam Kodandaramaih

Adrian-Leonard Radu

Federico Capasso

Shane Kumar Mainali

Maneesh Sah

Mohammadreza Khorasaninejad

Zichen Sun

Jun Tse Chen

Georgi M Chalakov

Thomas Leo Marquardt

Shaoyu Zhang

Esfandiar Manii

Saurabh Pant

Dana Yulian Kaban

Saher B Ahwal

Da Zhou

Amit Pratap Singh

Junhua Gu

Giuseppe Di Fabbrizio

Tie-Yan Liu

Craig Steven Freedman

Imre Attila Kiss

Anjishnu Kumar

Quan Zhang

Tyler Ogren Vrooman

Brad Nathaniel Arnold

Alexander Yutong Zhu

Varun Kunjbihari Tibrewal

Jingchao Zhang

Zhujun Shi

Jaewon Oh

Sajesh Kumar Saran

Charles Roques-Carmes

Jack G Song

Todd Nelson Schoepflin

Qiwei Ye

Prashanth Purnananda

Vyshakh Sanjeev

Benjamin Edward Lambert

Qi Meng

Chaitanya Sreenivas Ravella

Ishan Mishra

Robert Charles Devlin

A computing device is provided, including non-volatile memory storing a database including a table having a plurality of rows. The computing device may further include a processor configured to receive a request to perform a first transaction on a row. The processor may assign a first transaction identifier (TID) of the first transaction to the row. The processor may impose a first exclusive TID lock on the first TID of the row and may impose a first exclusive data lock associated with the first transaction on the row. The processor may perform the first transaction on the row. Performing the first transaction may include modifying the table as stored in the memory. In response to completing the first transaction, the processor may release the first exclusive data lock on the row and release the first exclusive TID lock on the first TID of the row.

Transaction identifier locking with data row locks

In implementations of the present disclosure, a solution for optimization of a learning network in an equivalent class space is provided. In this solution, base paths running through layers of a learning network are determined. Each node utilizes an activation function with a scaling invariant property to process an input from a node of a previous layer, each base path comprises a single node in each layer, and processing in the base paths is linearly independent from each other. A combined value of parameters associated with nodes in each base path is updated. A parameter associated with a node is used to adjust an input obtained from a node of a previous layer. Values of parameters associated with nodes in the base paths are updated based on updated combined values of parameters. Through this solution, optimization efficiency can be improved and more accurate optimized values of parameters are achieved.

Optimization of neural network in equivalent class space

A relational database system that implements persistent version storage may include an in-row module that is executable by one or more processors to implement an in-row scheme by storing in-row previous version information within a payload of a row within a database page. The relational database system may additionally include an off-row module that is executable by the one or more processors to implement an off-row scheme by storing off-row previous version information in an off-row page that is separate from the database page. The relational database system may additionally include a storage policy that defines when previous version information is stored in accordance with the in-row scheme and when the previous version information is stored in accordance with the off-row scheme. The relational database system may additionally include a cleanup module that cleans up older versions when they are deemed unnecessary by the system.

Persistent version storage for relational database management system

Database recovery can be performed in substantially constant time. From a database transaction log, the state of one or more transactions is identified including those that were active at a time of a crash but not committed. Transactions can be reapplied sequentially starting from the beginning of a last successful checkpoint to the end of the transaction log. Uncommitted user transactions are not undone synchronously, but simply marked as aborted. Subsequently, versions of affected data elements can be reverted back to a prior saved version asynchronously.

Constant time database recovery

Methods, systems, and computer programs are presented for providing file system functions on a cloud storage system based on blob storage. A cloud storage system comprises storage clusters, and each storage cluster comprises a plurality of storage nodes (each including a partition layer that stores data in blobs organized within containers, and a stream layer for streaming blob data), a hierarchical namespace service for implementing file system functions on the data stored in the blobs and, a front end layer for processing user requests to access and manage data stored on the blobs. The file system functions include flat namespace functions and hierarchical namespace functions. Further, the front end layer accesses the partition layer to access the blobs based on addresses of blobs and accesses the hierarchical namespace service to identify the blobs.

Cloud storage distributed file system

Methods, systems, and programs provide for executing distributed transactions in a cloud storage system with a hierarchical namespace. One method includes receiving a request with operations to be executed atomically. Further, nodes are identified for executing the operations, each node having a respective clock and having at least part of a transactions table for controlling updates to entities. Each clock is one of a loosely-synchronized, a strictly-synchronized clock, a logical, or a physical clock. Additionally, the nodes process the operations, which includes setting a commit timestamp (CS) to a value of the clock in the node if the node is a first node in the processing. One node coordinates the transactions, and may be one of the nodes executing transactions. If the clock in the node is less than a current value of the CS, the node waits for the clock to reach the current value of the CS and the CS is updated. The transactions table is updated based on the value of the CS, the atomic execution is committed utilizing the final value of the CS, and a status is returned.

Distributed transactions in cloud storage with hierarchical namespace

A meta-lens having a phase profile includes a substrate and a plurality of nanostructures disposed on the substrate. The nanostructures together define the phase profile of the meta-lens. The phase profile achieves an off-axis focus. Each nanostructure is designed according to at least one design parameter of the nanostructure that imparts a phase shift of light passing through the nanostructure.

Super-dispersive off-axis meta-lenses for high resolution compact spectroscopy

An optical device includes a substrate, a reflective layer disposed over the substrate, and a metalens disposed over the reflective layer. The metalens includes a plurality of nanopillars, the plurality of nanopillars together specifying a phase profile such that the metalens has a focal length that is substantially constant over a wavelength range of an incident light of about 490 nm to about 550 nm.

Achromatic metalens and metalens with reverse chromatic dispersion

This disclosure describes techniques for synchronizing independent data streams. In some instances, a computing device couples to multiple independent sensors, such as cameras, and applies accurate timestamp information to the individual frames of sensor data from the independent sensors. After aligning these data streams by applying these accurate timestamps, the computing device may, in some instances, encode and transmit these timestamped data streams to one or more entities for further processing. In one example, a first camera (e.g., a depth camera configured to generate a depth map) may capture images of an environment, as may a second camera (e.g., an Red-Green-Blue (RGB) camera configured to generate color images). The resulting images may be temporally aligned with one another via the timestamping, and the resulting aligned images from both the depth sensor and the RGB camera may be used to create a three-dimensional (3D) model of the environment.

Synchronizing data streams

Features are disclosed for generating models, such as conditional random field ('CRF') models, that consume less storage space and/or transmission bandwidth than conventional models. In some embodiments, the generated CRF models are composed of fewer or alternate components in comparison with conventional CRF models. For example, a system generating such CRF models may forgo the use of large dictionaries or other cross-reference lists that map information extracted from input (e.g., 'features') to model parameters; reduce in weight (or exclude altogether) certain model parameters that may not have a significant effect on model accuracy; and/or reduce the numerical precision of model parameters.

Conditional random field model compression

Features are disclosed for processing user utterances and applying user-supplied corrections to future user utterances. If a user utterance is determined to relate to a speech processing error that occurred when processing a previous utterance, information about the error or a correction thereto may be stored. Such information may be referred to as correction information. Illustratively, the correction information may be stored in a skip list. Subsequent utterances may be processed based at least partly on the correction information. For example, speech processing results generated from processing subsequent utterances that include a term associated with the error may be removed or re-scored in order to reduce or prevent the chance that an error will be repeated.

Speech processing using skip lists

Growing community of inventors

Cambridge, MA, United States of America

Wei Ting Chen