Microsoft Technology Licensing, LLC

Madanlal S Musuvathi

Todd Douglas Mytkowicz

Saeed Maleki

Yufei Ding

Shaz Qadeer

Sebastian Burckhardt

Alastair Wolman

Stefan Saroiu

Eduardo Alberto Cuervo Laffaye

Sumit Gulwani

Kristin E Lauter

Hao Chen

Shuvendu K Lahiri

Alice Zheng

Kim Laine

Satish Narayanasamy

Veselin Raychev

Laxmi Narsimha Rao Kakulamarri

Lin Tan

Christopher William Dern

Olli Ilari Saarikivi

Roy Patrick Tan

Nishant A Mehta

Jacob Samuels Burnim

Pravesh Kumar Kothari

Santosh Ganapati Nagarakatte

Daniel Marino

Roshan Dathathri

Olli Llari Saarikivi

Systems, methods, and computer-readable media are disclosed for parallel stochastic gradient descent using linear and non-linear activation functions. One method includes: receiving a set of input examples; receiving a global model; and learning a new global model based on the global model and the set of input examples by iteratively performing the following steps: computing a plurality of local models having a plurality of model parameters based on the global model and at least a portion of the set of input examples; computing, for each local model, a corresponding model combiner based on the global model and at least a portion of the set of input examples; and combining the plurality of local models into the new global model based on the current global model and the plurality of corresponding model combiners.

Systems, methods, and computer-readable media for parallel stochastic gradient descent with linear and non-linear activation functions

Embodiments of the present invention provide systems, methods, and computer storage media directed to optimizing the generation, evaluation, and selection of tensor circuit specifications for a tensor circuit to perform homomorphic encryption operations on encrypted data. A computing device having an improved compiler and runtime configuration can obtain a tensor circuit and associated schema. The computing device can map the obtained tensor circuit to an equivalent tensor circuit, adapted to perform fully homomorphic encryption (FHE) operations, and instantiated based on the obtained associated scheme. The computing device can then monitor a flow of data through the equivalent FHE-adapted tensor circuit utilizing various tensor circuit specifications determined therefor. A cost of each tensor circuit specification can be determined by the computing device based on the monitored flow of data, so as to identify an optimal set of optimal tensor circuit specifications that can be employed by the obtained tensor circuit, to efficiently perform homomorphic encryption operations on encrypted data.

Homomorphic evaluation of tensor programs

Described herein is a system that transmits and combines local models, that individually comprise a set of local parameters computed via stochastic gradient descent (SGD), into a global model that comprises a set of global model parameters. The local models are computed in parallel at different geographic locations along with symbolic representations. The symbolic representations can be used to combine the local models. The global model can determine a likelihood, given a new data instance of a feature set, that a user performs a computer interaction with the content element. For instance, the system can use the model to provide search results in response to a search query submitted by a user. Or, the system can use the model to make a recommendation or suggestion to a user in response to a request for content (e.g., display a targeted advertisement, suggest a news story, etc.).

Determining a likelihood of a user interaction with a content element

Described herein is a system that transmits and combines local models, that individually comprise a set of local parameters computed via stochastic gradient descent (SGD), into a global model that comprises a set of global model parameters. The local models are computed in parallel at different geographic locations along with symbolic representations. Network transmission of the local models and the symbolic representations, rather than transmission of the large training data subsets processed to compute the local models and symbolic representations, conserves resources and decreases latency. The global model can then be used as a model to determine a likelihood of a course of action being successful for an organization. For example, the course of action can be a purchase of a security or a business operation strategy. In another example, the course of action can be a type of medical treatment for a patient.

Determining a course of action based on aggregated data

Systems, methods, and computer media for machine learning through a symbolic, parallelized stochastic gradient descent (SGD) analysis are provided. An initial data portion analyzer can be configured to perform, using a first processor, SGD analysis on an initial portion of a training dataset. Values for output model weights for the initial portion are initialized to concrete values. Local model builders can be configured to perform, using an additional processor for each local model builder, symbolic SGD analysis on an additional portion of the training dataset. The symbolic SGD analysis uses a symbolic representation as an initial state for output model weights for the corresponding portions of the training dataset. The symbolic representation allows the SGD analysis and symbolic SGD analysis to be performed in parallel. A global model builder can be configured to combine outputs of the local model builders and the initial data portion analyzer into a global model.

Machine learning through parallelized stochastic gradient descent

Described herein is a system transmits and combines local models, that individually include a set of local parameters computed via stochastic gradient descent (SGD), into a global model that includes a set of global model parameters. The local models are computed in parallel at different geographic locations (e.g., different instances of computing infrastructure) along with symbolic representations. Network transmission of the local models and the symbolic representations, rather than transmission of the large training data subsets processed to compute the local models and symbolic representations, conserves resources and decreases latency. The global model can then be used as a model to determine a likelihood that at least a portion of current and/or recently received data traffic is illegitimate data traffic that is associated with a cyber attack. In some instances, the system can implement a remedial action to mitigate the effects of the cyber attack on computing infrastructure.

Implementing network security measures in response to a detected cyber attack

Described herein is a system that transmits and combines local models, that individually comprise a set of local parameters computed via stochastic gradient descent (SGD), into a global model that comprises a set of global model parameters. The local models are computed in parallel at different geographic locations along with symbolic representations. Network transmission of the local models and the symbolic representations, rather than transmission of the large training data subsets processed to compute the local models and symbolic representations, conserves resources and decreases latency. The global model can then be used as a model to determine a likelihood of a monitored resource or a user of the monitored resource experiencing a problem with respect to performance or completion of one or more operations. The system can also implement an action to assist in resolving or avoiding the problem.

Determining a likelihood of a resource experiencing a problem based on telemetry data

Concepts and technologies are described herein providing technologies for mining patterns in temporal data streams. Data is broken into data sub-portions. Dependencies in computation between one or more of the data sub-portions are broken using symbolic data types. Symbolic summaries of computations of sub-portions are performed in parallel and are reduced to generate an output.

Technologies for mining temporal patterns in big data

Various technologies described herein pertain to performing collaborative rendering. A GPU of a mobile device can generate a mobile-rendered video stream based on a first instance of an application executed on the mobile device. A GPU of a server can generate one or more server-rendered video streams based on instance(s) of the application executed on the server. Based on the one or more server-rendered video streams, the server can generate a compressed server-manipulated video stream. The mobile device can further combine the mobile-rendered video stream and the compressed server-manipulated video stream to form a collaborative video stream, and a display screen of the mobile device can be caused to display the collaborative video stream. The mobile-rendered video stream can have a first level of a quality attribute and the collaborative video stream can have a second level of the quality attribute greater than the first level of the quality attribute.

Server GPU assistance for mobile GPU applications

Various technologies described herein pertain to parallel local sequence alignment that aligns a query sequence with a database sequence. The database sequence is segmented into a plurality of stripes. A first processing unit can compute Smith-Waterman values for a first stripe of the database sequence across the query sequence based on a cost function that models biological similarity between sequences. Moreover, a second processing unit can compute Smith-Waterman values for a second stripe of the database sequence across the query sequence based on the cost function. Further, a subset of the Smith-Waterman values for the second stripe of the database sequence across the query sequence can be re-computed based on the cost function (e.g., by the first processing unit or the second processing unit). The subset of the Smith-Waterman values to be re-computed can be determined based on a query sequence length and the cost function.

Parallel local sequence alignment

The techniques and/or systems described herein implement parallel processing of a dynamic programming problem across stages and/or clusters by breaking dependencies between stages and/or clusters. For instance, the techniques and/or systems may identify dependencies between sub-problems of the dynamic programming problem and group the sub-problems into stages. The techniques and/or systems may also group the stages into clusters (e.g., at least two clusters to be parallel processed). Then, the techniques and/or systems generate one or more solutions to use instead of actual solutions so that the dynamic programming problem can be parallel processed across stages and/or clusters.

Parallel dynamic programming through rank convergence

The described implementations relate to analysis of computing programs. One implementation provides a technique that can include accessing values of input variables that are processed by test code and runtime values that are produced by the test code while processing the input variables. The technique can also include modeling relationships between the runtime values and the values of the input variables. The relationships can reflect discontinuous functions of the input variables.

Modeling software behavior using learned predicates

Techniques for providing a fair stateless model checker are disclosed. In some aspects, a schedule is generated to allocate resources for threads of a multi-thread program in lieu of having an operating system allocate resources for the threads. The generated schedule is both fair and exhaustive. In an embodiment, a priority graph may be implemented to reschedule a thread when a different thread is determined not to be making progress. A model checker may then implement one of the generated schedules in the program in order to determine if a bug or a livelock occurs during the particular execution of the program. An output by the model checker may facilitate identifying bugs and/or livelocks, or authenticate a program as operating correctly.

Fair stateless model checking

A concurrency assertions system disclosed herein provides for atomic evaluation of an assertion expression by locking an assertion lock upon initiating an assertion and thereby protecting the assertion evaluation from concurrent modifications to the variables in the assertion expressions. When a violation of an assertion is detected, the concurrency assertions system ensures that the exception statistics at the time of the assertion violation represents a program state where the assertion is violated, thus improving analysis of assertion violations. Furthermore, the concurrency assertions system continuously evaluates an expression for an assertion for a time period while other threads in the program are being executed.

Growing community of inventors

Redmond, WA, United States of America

Madanlal S Musuvathi