Abstract: A method can include classifying, using a compressed and specialized convolutional neural network (CNN), an object of a video frame into classes, clustering the object based on a distance of a feature vector of the object to a feature vector of a centroid object of the cluster, storing top-k classes, a centroid identification, and a cluster identification, in response to receiving a query for objects of class X from a specific video stream, retrieving image data for each centroid of each cluster that includes the class X as one of the top-k classes, classifying, using a ground truth CNN (GT-CNN), the retrieved image data for each centroid, and for each centroid determined to be classified as a member of the class X providing image data for each object in each cluster associated with the centroid.

Abstract: A multi-hop relay network comprises a set of data nodes arranged in a relay topology, wherein the set of data nodes communicate with one another via a data plane comprising a set of high frequency data links, such as mmWave links. An edge node communicates with one or more of the set of data nodes via the data plane and communicates with the set of data nodes over a control plane comprising a set of low frequency wireless links, such as a Wi-Fi network. The edge node determines a path utilization for the set of high frequency wireless links. When one of the high frequency wireless links is over-utilized, the edge node communicates, to a data node in the set of data nodes via the control plane, a command to change the relay topology. The edge node also determines whether the relay topology is operating at a target accuracy. When it is not, the edge node adjusts a data analytics parameter for a node to improve the overall accuracy of the network.

Abstract: This document relates to performing video analytics on a cloud device that preserves privacy. One example uses data-oblivious algorithms to process input video data, where the data-oblivious algorithms can assist with preventing attackers from exploiting side-channels induced by data-dependent access patterns.

Abstract: A multi-hop relay network comprises a set of data nodes arranged in a relay topology, wherein the set of data nodes communicate with one another via a data plane comprising a set of high frequency data links, such as mmWave links. An edge node communicates with one or more of the set of data nodes via the data plane and communicates with the set of data nodes over a control plane comprising a set of low frequency wireless links, such as a Wi-Fi network. The edge node determines a path utilization for the set of high frequency wireless links. When one of the high frequency wireless links is over-utilized, the edge node communicates, to a data node in the set of data nodes via the control plane, a command to change the relay topology. The edge node also determines whether the relay topology is operating at a target accuracy. When it is not, the edge node adjusts a data analytics parameter for a node to improve the overall accuracy of the network.

Abstract: This document relates to performing video analytics on a cloud device that preserves privacy. One example uses data-oblivious algorithms to process input video data, where the data-oblivious algorithms can assist with preventing attackers from exploiting side-channels induced by data-dependent access patterns.

Abstract: This document relates to performing live video stream analytics on edge devices. One example determines resources available to the system, and a video analytics configuration is selected that distributes work between edge devices and cloud devices in a cascading manner, where edge device processing is prioritized over cloud processing in order to conserve resources. This example can dynamically modify the allocation of processing depending on changing conditions, such as network availability.

Abstract: A method can include classifying, using a compressed and specialized convolutional neural network (CNN), an object of a video frame into classes, clustering the object based on a distance of a feature vector of the object to a feature vector of a centroid object of the cluster, storing top-k classes, a centroid identification, and a cluster identification, in response to receiving a query for objects of class X from a specific video stream, retrieving image data for each centroid of each cluster that includes the class X as one of the top-k classes, classifying, using a ground truth CNN (GT-CNN), the retrieved image data for each centroid, and for each centroid determined to be classified as a member of the class X providing image data for each object in each cluster associated with the centroid.

Abstract: A method can include classifying, using a compressed and specialized convolutional neural network (CNN), an object of a video frame into classes, clustering the object based on a distance of a feature vector of the object to a feature vector of a centroid object of the cluster, storing top-k classes, a centroid identification, and a cluster identification, in response to receiving a query for objects of class X from a specific video stream, retrieving image data for each centroid of each cluster that includes the class X as one of the top-k classes, classifying, using a ground truth CNN (GT-CNN), the retrieved image data for each centroid, and for each centroid determined to be classified as a member of the class X providing image data for each object in each cluster associated with the centroid.

Abstract: Latency in responding to queries directed to geographically distributed data can be reduced by allocating individual steps, of a multi-step compute operation requested by the query, among the geographically distributed computing devices so as to reduce the duration of shuffling of intermediate data among such devices, and, additionally, by pre-moving, prior to the receipt of the query, portions of the distributed data that are input to a first step of the multistep compute operation, to, again, reduce the duration of the exchange of intermediate data. The pre-moving of input data occurring, and the adaptive allocation of intermediate steps, are prioritized for high-value data sets. Additionally, a threshold increase in a quantity of data exchanged across network communications can be established to avoid incurring network communication usage without an attendant gain in latency reduction.

Abstract: Latency in responding to queries directed to geographically distributed data can be reduced by allocating individual steps, of a multi-step compute operation requested by the query, among the geographically distributed computing devices so as to reduce the duration of shuffling of intermediate data among such devices, and, additionally, by pre-moving, prior to the receipt of the query, portions of the distributed data that are input to a first step of the multistep compute operation, to, again, reduce the duration of the exchange of intermediate data. The pre-moving of input data occurring, and the adaptive allocation of intermediate steps, are prioritized for high-value data sets. Additionally, a threshold increase in a quantity of data exchanged across network communications can be established to avoid incurring network communication usage without an attendant gain in latency reduction.

Abstract: A method can include classifying, using a compressed and specialized convolutional neural network (CNN), an object of a video frame into classes, clustering the object based on a distance of a feature vector of the object to a feature vector of a centroid object of the cluster, storing top-k classes, a centroid identification, and a cluster identification, in response to receiving a query for objects of class X from a specific video stream, retrieving image data for each centroid of each cluster that includes the class X as one of the top-k classes, classifying, using a ground truth CNN (GT-CNN), the retrieved image data for each centroid, and for each centroid determined to be classified as a member of the class X providing image data for each object in each cluster associated with the centroid.

Abstract: Latency in responding to queries directed to geographically distributed data can be reduced by allocating individual steps, of a multi-step compute operation requested by the query, among the geographically distributed computing devices so as to reduce the duration of shuffling of intermediate data among such devices, and, additionally, by pre-moving, prior to the receipt of the query, portions of the distributed data that are input to a first step of the multistep compute operation, to, again, reduce the duration of the exchange of intermediate data. The pre-moving of input data occurring, and the adaptive allocation of intermediate steps, are prioritized for high-value data sets. Additionally, a threshold increase in a quantity of data exchanged across network communications can be established to avoid incurring network communication usage without an attendant gain in latency reduction.

Abstract: A global manager communicates with various local managers to receive and process video queries. The video queries identify components that process live video streams, placement options for where the components of the video query may be executed, and various video query plans. The video query plans include options such as framerate and video quality. As the global manager processes the video queries, the global manager determines an initial set of video query configurations that identify a video query plan and placement option for each component of a given video query. Using the initial set of video query configurations, the global manager then determines an optimal set of video query configurations for the received set of video queries. The global manager communications instructions to the local managers to execute the components of the video queries using the video query plans and placement options from the optimal set of video query configurations.

Abstract: Various technologies described herein pertain to performing video analytics. The approaches set forth herein support live video analytics at scale with approximate and delay-tolerant processing. Video streams can be captured by multiple cameras and continuously streamed to a video analytics computing system; the video streams can be received at the video analytics computing system. Multiple video analytics queries can be executed on the video streams. The multiple video analytics queries can be concurrently executed by the video analytics computing system on the video streams as the video streams are continuously streamed to the video analytics computing system. The multiple video analytics queries can be executed utilizing resources of the video analytics computing system allocated between the multiple video analytics queries. Execution of the multiple video analytics queries can return respective results for the multiple video analytics queries.

Abstract: A global manager communicates with various local managers to receive and process video queries. The video queries identify components that process live video streams, placement options for where the components of the video query may be executed, and various video query plans. The video query plans include options such as framerate and video quality. As the global manager processes the video queries, the global manager determines an initial set of video query configurations that identify a video query plan and placement option for each component of a given video query. Using the initial set of video query configurations, the global manager then determines an optimal set of video query configurations for the received set of video queries. The global manager communications instructions to the local managers to execute the components of the video queries using the video query plans and placement options from the optimal set of video query configurations.

Abstract: Techniques for determining network paths for voice calls include analyzing network path measurements of the network paths to determine historical network performance data for the network paths, and identifying a group of top-k network paths based on those network paths having better network performance data other network paths. A particular network path may be selected using various techniques, such as selecting the network path with the best historical network performance, selecting a network path by applying a multi-armed bandit algorithm to select the path from the group of top-k network paths, or selecting a network path at random. The selected network path may be used to by a source-destination pair of computing devices for a voice call, and a record of that voice call may be used to update network performance information for the particular network path.

Abstract: Various technologies described herein pertain to performing video analytics. The approaches set forth herein support live video analytics at scale with approximate and delay-tolerant processing. Video streams can be captured by multiple cameras and continuously streamed to a video analytics computing system; the video streams can be received at the video analytics computing system. Multiple video analytics queries can be executed on the video streams. The multiple video analytics queries can be concurrently executed by the video analytics computing system on the video streams as the video streams are continuously streamed to the video analytics computing system. The multiple video analytics queries can be executed utilizing resources of the video analytics computing system allocated between the multiple video analytics queries. Execution of the multiple video analytics queries can return respective results for the multiple video analytics queries.

Abstract: Techniques for determining network paths for voice calls include analyzing network path measurements of the network paths to determine historical network performance data for the network paths, and identifying a group of top-k network paths based on those network paths having better network performance data other network paths. A particular network path may be selected using various techniques, such as selecting the network path with the best historical network performance, selecting a network path by applying a multi-armed bandit algorithm to select the path from the group of top-k network paths, or selecting a network path at random. The selected network path may be used to by a source-destination pair of computing devices for a voice call, and a record of that voice call may be used to update network performance information for the particular network path.

Abstract: Latency in responding to queries directed to geographically distributed data can be reduced by allocating individual steps, of a multi-step compute operation requested by the query, among the geographically distributed computing devices so as to reduce the duration of shuffling of intermediate data among such devices, and, additionally, by pre-moving, prior to the receipt of the query, portions of the distributed data that are input to a first step of the multistep compute operation, to, again, reduce the duration of the exchange of intermediate data. The pre-moving of input data occurring, and the adaptive allocation of intermediate steps, are prioritized for high-value data sets. Additionally, a threshold increase in a quantity of data exchanged across network communications can be established to avoid incurring network communication usage without an attendant gain in latency reduction.

Abstract: The described implementations relate to distributed computing. One implementation provides a system that can include an outlier detection component that is configured to identify an outlier task from a plurality of tasks based on runtimes of the plurality of tasks. The system can also include a cause evaluation component that is configured to evaluate a cause of the outlier task. For example, the cause of the outlier task can be an amount of data processed by the outlier task, contention for resources used to execute the outlier task, or a communication link with congested bandwidth that is used by the outlier task to input or output data. The system can also include one or more processing devices configured to execute one or more of the components.