Microsoft Technology Licensing, LLC

Lin Xiao

Jie Jerry Liu

Feng Zhao

Suman Kumar Nath

Leonidas Rigas

Wenbo He

Gong Chen

Li Deng

Jianfeng Gao

Xiaodong He

Jeremy Eric Elson

Xinying Song

Jianshu Chen

Yelong Shen

Ji He

A processing unit can successively operate layers of a multilayer computational graph (MCG) according to a forward computational order to determine a topic value associated with a document based at least in part on content values associated with the document. The processing unit can successively determine, according to a reverse computational order, layer-specific deviation values associated with the layers based at least in part on the topic value, the content values, and a characteristic value associated with the document. The processing unit can determine a model adjustment value based at least in part on the layer-specific deviation values. The processing unit can modify at least one parameter associated with the MCG based at least in part on the model adjustment value. The MCG can be operated to provide a result characteristic value associated with test content values of a test document.

Training and operating multi-layer computational models

A technology is presented by which a learned mechanism is developed by solving a minimization problem by using regularized dual averaging methods to provide regularized stochastic learning and online optimization. An objective function sums a loss function of the learning task and a regularization term. The regularized dual averaging methods exploit the regularization structure in an online learning environment, in a manner that obtains desired regularization effects, e.g., sparsity under L-regularization.

Regularized dual averaging method for stochastic and online learning

The claimed subject matter provides a system and/or a method facilitates energy-aware connection distribution among a plurality of servers in a cluster. A set of busy servers in the cluster can be provided that each handle a high number of connections. In addition, a set of tail servers in the cluster can be managed that each maintain a low number of connections. A load skewing component gives priority to at least a subset of the set of busy servers when dispatching new connection requests from a plurality of users. In addition, the load skewing component controls the number of tail servers to maintain a sufficient number for energy-aware operation.

Load skewing for power-aware server provisioning

The claimed subject matter provides a system and/or a method that facilitates managing a number of active servers in a cluster. A forecast component can predict at least one of login rate or number of connections in the cluster at a future time. A dynamic load analysis component can evaluate dynamic behaviors in login rate and number of connections in the cluster as a result of load dispatching. Moreover, a provisioning component can determine a number of servers in the cluster needed based at least in part on the prediction and dynamic behavior analysis. In addition, the provisioning component can include an additional margin in the number of servers needed in accordance with multiplicative factors.

Framework for joint analysis and design of server provisioning and load dispatching for connection-intensive server

Growing community of inventors

Redmond, WA, United States of America

Lin Xiao