Abstract: Extensible change control management for development and deployment of applications is described. A change control manager may register plugin components, including requester, approver and performer components, to process changes to a resource of a resource type, where registering is performed responsive to request(s) received via programmatic interface(s). Upon receiving a change request from the requester component, the change control manager may send the change request to the approver component to approve the change request and, responsive to notification of approval from the approver component, the change control manager may then send the change request to the performer component to implement the requested change.

Abstract: Systems and methods are provided to perform multiply-accumulate operations of reduced precision numbers in a systolic array. Each row of the systolic array can receive reduced inputs from a respective reducer. The reduced input can include a reduced input data element and/or a reduced weight. The systolic array may lack support for inputs with a first bit-length and the reducers may reduce the bit-length of a given input from the first bit-length to a second shorter bit-length and provide the reduced input to the array. In order to reduce the bit-length, the reducer may reduce the number of trailing bits of the input. Further, the systolic array can receive a reduced and rounded input. The systolic array can propagate the reduced input through the processing elements in the systolic array. Each processing element may include a multiplier and/or an adder to perform arithmetical operations based on the reduced input.

Abstract: A test application programming interface (API) request may be issued to a test service implementation. A test API response, which is responsive to the test API request, may be received from the test service implementation. First contents of the test API response may be compared to second contents of a control API response. The comparing may be performed based at least in part on a set of one or more expected differences associated with the test API response and the control API response. It may be determined, based on the comparing, whether an unexpected difference exists between the test API response and the control API response. An indication may be provided of whether the unexpected difference exists between the test API response and the control API response, and the unexpected difference may be automatically corrected, such as via one or more configuration changes.

Abstract: A transcoding process may be performed in parallel, which may include different nodes processing different segments of a media file at different times. Different time slices of media files may have different properties, which may lead to errors when media files are encoded for later playback. Extracting metadata from frames of media files and evaluating the frames for discrepancies enables creation of a reconciliation manifest to correct discrepancies associated with the media file. As the transcoding process is performed, the reconciliation manifest may be reference to generate an output media file.

Abstract: Metadata indicating that a virtual traffic hub enabling connectivity between a plurality of isolated networks has been established is stored. A determination is made that a first entry of a first isolated network attached to the hub is to be represented in a second routing table of a second isolated network attached to the hub, e.g., to enable network packets originating at resources of the second isolated network to be transmitted via the hub to the first isolated network. A new entry corresponding to the first entry is included in the second routing table.

Abstract: Technologies directed to a slot antenna as a calibration antenna for a phased array antenna are described. A communication system includes an antenna array with a first antenna module with a plurality of antenna elements and a conductive wall structure isolating each of the plurality of antenna elements. The conductive wall structure includes a portion separating a first antenna element and a second antenna element of the plurality of antenna elements. The portion includes a slot antenna.

Abstract: This disclosure describes, in part, context-driven device arbitration techniques to select a speech interface device from multiple speech interface devices to provide a response to a command included in a speech utterance of a user. In some examples, the context-driven arbitration techniques may include executing multiple pipeline instances to analyze audio signals and device metadata received from each of the multiple speech interface devices which detected the speech utterance. A remote speech processing service may execute the multiple pipeline instances and analyze the audio signals and/or metadata, at various stages of the pipeline instances, to determine which speech interface device is to respond to the speech utterance.

Abstract: An indication of semantic relationships among classes is obtained. A neural network whose loss function is based at least partly on the semantic relationships is trained. The trained neural network is used to identify one or more classes to which an input observation belongs.

Abstract: A selection to use a customer-generated matchmaking algorithm is received, by a multi-tenant gaming service, from a first customer. Indications of a plurality of video game player matchmaking requests and corresponding player characteristic data are provided, by the multi-tenant gaming service, for processing by the customer-generated matchmaking algorithm. An indication of a first player match determined by the customer-generated matchmaking algorithm, based on the video game player matchmaking requests and the player characteristic data, is received, by the multi-tenant gaming service. The first player match may be between at least a first player and a second player. Execution of a first video game session to which the first player and the second player are connected is hosted, by the multi-tenant gaming service.

Abstract: Methods, systems, and computer-readable media for energy-optimizing placement of resources in data centers are disclosed. A resource placement manager determines information descriptive of energy usage by one or more data centers. The one or more data centers comprise a plurality of computing resources in a plurality of corresponding locations. The resource placement manager selects, from the plurality of computing resources in the plurality of corresponding locations, a particular computing resource in a particular location for performing one or more computing tasks. The particular computing resource in the particular location is selected based at least in part on reducing energy usage associated with the one or more data centers according to the information descriptive of energy usage. The particular computing resource in the particular location is used to perform the one or more computing tasks.

Abstract: At an artificial intelligence system, during a labeling feedback session, a visualization data set is presented via a programmatic interface. The visualization data set comprises a representation of data items for which labeling feedback is requested for generating a training set of a classifier. At least one of the data items is selected based on an estimated rank with respect to a metric associated with including the data item in a training set. During the session, respective labels for the data items and a filter criterion to be used to select additional data items are obtained. A classifier trained using the labels is stored.

Abstract: At a first layer of an echo canceler, a first compensation for a first set of properties of output of an audio capture device of a first communication environment is applied. The first set of properties includes a property resulting from a difference in clock speeds of an audio capture device and an audio rendering device of the first communication environment. At a second layer of the echo canceler, at which output of the first layer is received, a second compensation for a second set of properties of the output of the first layer is applied. The second set of properties includes an echo. Applying the compensations comprises modifying neural network weights. Output of the second layer is transmitted to a second communication environment.

Abstract: Systems and methods are disclosed to implement, in a multi-writer database system, a database engine node that optimistically sends queued purge operations to a backend storage system without waiting for acknowledgement of previous operations from the storage system. If a conflict is later detected for a particular operation, the database engine node may retry some portion of the purge operations in the queue. In embodiments, the purge operations are performed as part of a garbage collection phase of a multi-version concurrency control (MVCC) protocol, where unneeded versions of objects previously generated for isolated snapshots are deleted. In embodiments, the operations in the queue may be grouped so that a rejection of one operation in the group will cause the entire group to be retried. Because these purge operations are idempotent and directed to objects that are invisible, they may be freely retried in arbitrary order, without maintaining any rollback state.

Abstract: An acceleration engine with multiple accelerators may share a common set of data that is used by each accelerator to perform computations on input data. The set of shared data can be loaded into the acceleration engine from an external memory. Instead of accessing the external memory multiple times to load the set of shared data into each accelerator, the external memory can be accessed once using direct memory access to load the set of shared data into the first accelerator. The set of shared data can then be serially loaded from one accelerator to the next accelerator in the acceleration engine using direct memory access. To achieve data parallelism and reduce computation time, a runtime driver may split the input data into data batches, and each accelerator can perform computations on a different batch of input data with the common set of shared data.

Abstract: In one example, an apparatus comprises: a local on-chip memory; a computation engine configured to generate local data and to store the local data at the local on-chip memory; and a controller. The apparatus is configured to be coupled with a second device via an interconnect, the second device comprising a local memory. The controller is configured to: fetch the local data from the local on-chip memory; fetch remote data generated by another device from a local off-chip memory; generate output data based on combining the local data and the remote data; and store, via the interconnect, the output data at the local memory of the second device.

Abstract: Managed materialized views may be generated from across heterogeneous data sources. A request to create a materialized view may be received and performed by a materialized view management platform, which may obtain and generate the materialized view from different data sources and store the materialized view in a target system. Changes to the data sources may be obtained at the materialized view management platform and updates to the materialized view may be determined. The materialized view in the target system may be updated.

Abstract: A simulation environment (e.g., multi-player game) hosted by a provider network may implement componentized entities to reduce the amount of resource usage for a simulation (e.g., by reducing the amount of input/state data transmitted through the use of dynamically changing input structures). A user may add or remove any number of components to an entity that is simulated at the local client device. When inputs are received for one or more components, values for predictive states are locally determined for each component. An input packet is generated and sent to the provider network, which includes the inputs as well as data that is based on the values for the locally predicted states (e.g., a fingerprint or other unique ID). If necessary, a correction packet may be generated at the provider network and sent back to the client.

Abstract: Described is a multiple-camera system and process for re-identifying an agent located in a materials handling facility based on anterior views of agents. An anterior view of a newly detected agent may be partitioned and color signatures generated for each partition. Likewise, stored anterior views of agents (candidate agents) that may potentially be the newly detected agent are partitioned and color signatures generated for each partition. Based on the color signatures, a similarity between the anterior view of the newly detected agent and the candidate agents is determined. The similarity may be used to either determine that the newly detected agent is one of the candidate agents or reduce the set of candidate agents that are considered during a manual review.

Abstract: Systems, methods, and computer-readable media are disclosed for processing input data to determine an entity such as a product, service, user profile, etc. referenced in or otherwise relevant to a semantic context of the input data. Information related to the entity may be provided as an information package (e.g., a card) that is shareable as part of an electronic message. The card may include a representation of a network resource identifier that identifies a network resource, a network location of the network resource, and an access mechanism for accessing a representation (e.g. a product detail page) of the network resource. The network resource identifier may include one or more tags or tokens that identify an electronic messaging application provider and/or a user such as a sender or recipient of an electronic message that includes the card so as to enable compensating the provider and/or the user for a purchase of a product or service to which the card relates.

Abstract: Devices, systems, and methods are provided for user profile management across multiple devices. A method may include determining, by a first device, that the first device is signed into a customer account; determining that the customer account includes a first user profile and a second user profile, the first user profile having first identity data, and the second user profile having second identity data; determining that third identity data for the first user profile is available for presentation, the third identity data based on a user interaction with a second device; determining that the first user profile is enrolled on the first device; presenting the first identity data using the first device; receiving a second user request to access the first user profile; and presenting, using the first device, content based on the second user request and a user preference associated with the first user profile.

Abstract: Systems and methods are described for providing auxiliary functions in an on-demand code execution system in a manner that enables efficient execution of code. A user may generate a task on the system by submitting code. The system may determine the auxiliary functions that the submitted code may require when executed on the system, and may provide these auxiliary functions by provisioning or configuring sidecar virtualized execution environments that work in conjunction with the main virtualized execution environment executing the submitted code. Sidecar virtualized execution environments may be identified and obtained from a library of preconfigured sidecar virtualized execution environments, or a sidecar agent that provides the auxiliary function may be identified from a library, and then a virtualized execution environment may be provisioned with the agent and/or configured to work in conjunction with the main virtualized execution environment.

Abstract: Techniques for multi-stream transactional event processing under ACID semantics in a distributed log-based append-only datastore are described. A transaction coordinator writes events that are part of a transaction to a transaction buffer, where the events can be made visible to clients involved in the transaction while other clients are not aware. Upon committing the transaction, an optimistic concurrency control based technique is utilized to attempt to obtain locks on all events involved in the transaction across one or multiple shards by one or multiple designated writer nodes. When all involved writer nodes indicate that they are able to commit their events, the transaction can be committed.

Abstract: This disclosure describes, in part, systems for enabling facilities to customize the layouts of items located within the facilities. For instance, an entity may be associated with multiple facilities. The facilities may be located at various geographic locations, include various retail layouts/sizes, sell various items, and/or the like. As such, it may be difficult to generate planograms that can be used by more than one of the facilities. Therefore, for different facilities, the entity may determine (1) locations of inventory locations within the facilities, (2) categories of items to include at the inventory locations, (3) identifiers of items associated with the categories, (4) quantities of items to include at the inventory locations, and/or the like. The facilities can then use these determinations to determine the actual items, as well as the locations for the actual items, at the inventory locations.

Abstract: Systems and methods are described for controlled migration of workloads between cell systems arranged in a cell-based architecture. Each cell system can implement a portion of an overall workload of the architecture, which may represent a network-accessible service. The isolation provided by cells may prevent widespread problems at the service, but may also conceal errors that might occur should workloads between the cells be redistributed. Such redistribution is often forced at inopportune moments, such as when a cell of the service has already failed. Systems and methods described herein enable detection of such errors by repeatedly migrating portions of workloads between cells. The system can monitor health information during or between migrations to ensure continued health of the service. If the service appears to be unhealthy after a migration, future migrations can be halted to enable a cause of the unhealthy state to be identified and rectified.

Abstract: Systems and methods are provided for machine learning-based rendering of a clothed human with a realistic 3D appearance by virtually draping one or more garments or items of clothing on a 3D human body model. The machine learning model may be trained to drape a garment on a 3D body mesh using training data that includes a variety 3D body meshes reflecting a variety of different body types. The machine learning model may include an encoder trained to extract body features from an input 3D mesh, and a decoder network trained to drape the garment on the input 3D mesh based at least in part on spectral decomposition of a mesh associated with the garment. The trained machine learning model may then be used to drape the garment or a variation of the garment on a new input body mesh.

Abstract: A method of particle-based threat scanning may include obtaining a sample from a sample source, generating a plurality of particles from the sample, wherein each particle from the plurality of particles is an array of unique bytes generated based on one or more particle properties, and determining whether the sample is associated with a known threat by comparing the plurality of particles to particle threat signatures in a threat database.

Abstract: Disclosed are various embodiments for performing user account migration between platforms. In one embodiment, a client device is configured to obtain an update to a first application. The update causes the client device to obtain a second application. A first user account on a first platform associated with the first application is migrated to a second user account on a second platform associated with the second application. The client device launches the second application for access to the second user account after the first application is launched.

Abstract: Disclosed herein are techniques for implementing a large fully-connected layer in an artificial neural network. The large fully-connected layer is grouped into multiple fully-connected subnetworks. Each fully-connected subnetwork is configured to classify an object into an unknown class or a class in a subset of target classes. If the object is classified as the unknown class by a fully-connected subnetwork, a next fully-connected subnetwork may be used to further classify the object. In some embodiments, the fully-connected layer is grouped based on a ranking of target classes.

Abstract: A neural network is trained to process input data and generate a classification value that characterizes the input with respect to an ordered continuum of classes. For example, the input data may comprise an image and the classification value may be indicative of a quality of the image. The ordered continuum of classes may represent classes of quality of the image ranging from “worst”, “bad”, “normal”, “good”, to “best”. During training, loss values are determined using an ordered classification loss function. The ordered classification loss function maintains monotonicity in the loss values that corresponds to placement in the continuum. For example, the classification value for a “bad” image will be less than the classification value indicative of a “best” image. The classification value may be used for subsequent processing. For example, biometric input data may be required to have a minimum classification value for further processing.

Abstract: Described are systems, methods, and apparatus that detect keywords in one or more speech segments to authenticate that the speech is generated by the speaker as part of an intentional enrollment by the speaker into a service. For example, as a speech segment is received as part of an enrollment process, the speech segment may be converted into a log melspectrogram and the log melspectrogram may be processed using a machine learning model to determine if an expected keyword is represented by the log melspectrogram. If the keyword is detected, it may be determined that the speech output by the speaker is output as part of an intentional enrollment process.

Abstract: At an artificial intelligence system, one or more classifier training iterations are performed until a training completion criterion is met. A particular iteration comprises obtaining, via an interactive interface, asynchronously with respect to the start of the iteration, class labels for data items identified in a previous iteration as candidates for labeling feedback. The particular iteration also comprises identifying, based on an analysis of classification predictions generated using classifiers trained using class labels obtained via the interface, another set of data items as candidates for labeling feedback. After the training criterion is met, a classifier trained using labels obtained during the iterations is stored.

Abstract: A journaling file system may implement remote durable logging. Updates to a file system may be received, and log records describing the updates may be stored in a locally-accessible file system change log. The update may then be acknowledged as committed. The log records may then be sent to be stored in a network-based data store in a remote version of the file system change log. Once it may be determined that the log records are stored in the remote version, storage space for the log records in the local file system change log may be reclaimed. Various types of restoration and duplication techniques may be implemented based on the remote version of the change log to restore a file system at an originating device or to duplicate the file system at a different device.

Abstract: Techniques for a server-based reconnection of disconnected devices in a network are described. In an example, a computer system receives, from a second device, first data of a first device. The first data indicates that the first device has disconnected from the network due a change in the password of the network. The computer system determines that the first device is authorized to connect to the network. Based on the first device being authorized, the computer system transmits a request for a token. The computer system compares the token to a token stored by the computer system. In response to the tokens matching, the computer system transmits second data to the first device via the second device. The second data includes an updated password of the network. The first device uses the second data to reestablish a connection with the network.

Abstract: Methods for providing power to electronic components (e.g., servers in a datacenter, or other devices) can include providing AC power received from an external source to the devices. During periods of high demand, when costs of external AC power exceed a threshold, or when availability of external AC power is decreased, distributed backup power supply systems can be utilized to temporarily offset or to reverse consumption of external AC power. Such distributed backup power systems can be periodically recharged during periods of low power demand, and thus consistently retain adequate charge for backup use in the event of power loss.

Abstract: Techniques are described for providing managed virtual computer networks that have a configured logical network topology with virtual networking devices, such as by a network-accessible configurable network service, with corresponding networking functionality provided for communications between multiple computing nodes of the virtual computer network by emulating functionality that would be provided by the virtual networking devices if they were physically present.

Abstract: An application (or process) may have an amount of steady state work to perform per unit time, as well as one or more mechanisms for doing a lower quality job of that work in the event the application falls behind. Approaches presented herein can utilize a clock monitor that enables the application to determine whether a clock loss was encountered that was due to an external source, and is of an amount of time that may be naturally recoverable by the application. If so, the application can enter into a mode of operation wherein the activation of one or more recovery mechanisms is postponed for a period of time to provide the application time to recover. If, after the period of such mode operation, the application has not recovered from the real time clock loss, then the recovery mechanism(s) can be activated as appropriate.

Abstract: Techniques for facilitating returns of ordered items include receiving, by a control device, a first code associated with an order, determining that the code corresponds to a valid order, and granting access to a controlled access area where a user may examine the ordered items. If a second code associated with a particular item is received by the control device, a determination is made that the code corresponds to an item associated with the order, and access to a container within the controlled access area is granted. Once an item is placed in the container, sensors in the container determine characteristics of the item. If the determined characteristics correspond to item data that indicates the expected characteristics of the item, data indicative of a return of the item is generated.

Abstract: Techniques for tuning parameters for machine learning models are described. Different values for a parameter are tested to determine the value that results in an optimized model. A parameter value may be selected for testing using a search algorithm based on how the model performs with respect to other values for the parameter. Different values may be tested until a stopping criterion (such as time for testing, number of trials, amount of enhancement in performance, etc.) is met. In some embodiments, the techniques may be used to determine parameter values for natural language processing models.

Abstract: A machine learning algorithm, such as a random forest regressor, can be trained using a set of annotated data objects to estimate the risk or business value for an object. The feature contributions for each data object can be analyzed and a representation generated that clusters data objects by feature contributions. Any clustering of data objects with incorrect scores in the visualization can be indicative of gaps in the regressor training. Adjustments to the inputs can be made, and the regressor retrained, to eliminate clustering of errors for similar feature contributions. Correcting the risk score estimations can ensure that the appropriate security policies and permissions are applied to each data object.

Abstract: Disclosed are various embodiments for integrating multiple access technologies in a radio-based network. In one embodiment, first data is received from a first wireless device connected to a radio-based network via a first access technology. Second data is received from a second wireless device connected to the radio-based network via a second access technology. The first access technology and the second access technology use a different physical layer and a different media access control layer but use a network layer that is in common. The first data and the second data are processed at the network layer in the radio-based network using a common processing element.

Abstract: Devices and methods are provided for using virtual fulfillment center locations to facilitate delivery of items. A device may receive a virtual location associated with transportation of an item to a first location. The device may determine a geographic zone associated with the virtual location. The device may determine a first delivery time associated with delivering the item from the first location, and determine a second delivery time associated with delivering the item from the virtual location. The device may receive a purchase request for the item, from a second device, wherein the purchase request is associated with delivery to a second location. The device may determine an estimated delivery date to the delivery location based on the first delivery time or the second delivery time. The device may send, to the second device, an indication of the estimated delivery date.

Abstract: Various embodiments of an automated adaptive bitrate (“ABR”) encoding system and method that encodes an input video into a set of encoded video streams for ABR delivery, also known as an ABR ladder. The automated ABR encoding system encodes the input video into an initial number of speculative encodings, based on respectively determined resolution and quality level settings for the individual speculative encodings. It then estimates quality scores for a plurality of potential output encodings based on respective output statistics of the speculative encodings. It selects a plurality of better quality output encodings from the plurality of potential output encodings based on the quality scores. It then encodes the input video a second time into the plurality of better quality output encodings. The automated ABR encoding system selects one or more output encodings from the plurality of better quality output encodings to include in the ABR ladder.

Abstract: Systems, methods, and computer-readable media are disclosed for robotic picking assemblies with different concurrent flow rates. In one embodiment, an example system may include a picking assembly coupled to a vacuum system. The picking assembly may include a first suction cup assembly having a first flow rate, the first suction cup assembly including a first suction cup and a first mesh filter disposed in the first suction cup, a second suction cup assembly having a second flow rate, the second suction cup assembly including a second suction cup and a second mesh filter disposed in the second suction cup, and a third suction cup assembly having the second flow rate, the third suction cup assembly including a third suction cup and a third mesh filter disposed in the third suction cup. The picking assembly may grasp items based on the first flow rate and the second flow rate.

Abstract: Data for training a machine learning algorithm to detect airborne objects is generated by capturing background images using a camera aboard an aerial vehicle, and generating a trajectory of a synthetic object, such another aerial vehicle, in three-dimensional space. The trajectory is projected into an image plane of the camera, determined from a pose of the camera calculated using inertial measurement unit data captured by the aerial vehicle. Images of the synthetic object may be rendered based on locations of the trajectory in the image plane at specific times. Pixel-accurate locations for the rendered images may be determined by calculating a homography from consecutive images captured using the camera, and adjusting locations of the trajectory using the homography. The rendered images may be blended into the images captured by the camera at such locations, and used to train a machine learning algorithm.

Abstract: A method of implementing a network-enabled secure door lock, comprising obtaining measurements of an environment associated with a door from a variety of sensor types; generating, based at least in part on the measurements, a set of inputs to a machine-learning model; inputting the set of inputs into the machine learning model to determine a status of the door; generating a message that indicates the status of the door; and transmitting the message to a user device.

Abstract: Speech processing tasks may be allocated at least partly to a local device (e.g., user computing device that receives spoken words) and at least partly to a remote device to determine one or more user commands or tasks to be performed by the local device. The remote device may be used to process speech that the local device could not process or understand, or for other reasons, such as for error checking. The local device may then execute or begin to execute locally determined tasks to reduce user-perceived latency. Meanwhile, the entire media input, or a portion thereof, may be sent to the remote device to process speech, verify the tasks and/or identify other user commands in the media input (or portion thereof).

Abstract: Systems and methods are disclosed for hierarchical categorical and sub-categorical topic modeling allowing, in response to a query in natural language, a set of results to be determined which are both semantically relevant to the user and diverse, by containing information complementary or adjacent to that of the query. Such a paradigm permits exploration and discovery of new topics and ideas in large collections of documents. In some embodiments, one or more non-negative matrix factorization (“NMF”) algorithms are applied in determining a hierarchical topic model including the semantically-related categories and sub-categories. The dataset may include authorized social media data collection, and machine learning techniques can optimize the generation of the topic model and/or the search results.

Abstract: A network infrastructure modernization service described herein may allow a customer to migrate the underlying network infrastructure topology and configurations of the on-premises environment onto the cloud environment. By collecting the network infrastructure configurations and/or runtime metrics, generating a network model based on the configurations and/or runtime metrics, and applying modernization rules to the network model to generate a network infrastructure template, the network infrastructure modernization service described in the present application enables the customers to be migrate their on-premises applications as well as the underlying network infrastructure onto the cloud environment.

Abstract: Described herein is a neural network accelerator (NNA) with a decompression unit that can be configured to perform multiple types of decompression. The decompression may include a separate subunit for each decompression type. The subunits can be coupled to form a pipeline in which partially decompressed results generated by one subunit are input for further decompression by another subunit. Depending on which types of compression were applied to incoming data, any number of the subunits may be used to produce a decompressed output. In some embodiments, the decompression unit is configured to decompress data that has been compressed using a zero value compression scheme, a shared value compression scheme, or both. The NNA can also include a compression unit implemented in a manner similar to that of the decompression unit.

Abstract: Techniques for providing and implementing a single skill associated with custom functionality and system-provided functionality are described. The skill may be used to invoke functionality in response to a user input without requiring a user remember exact formulations to cause the functionality to be performed. The skill may be associated with more than one domain. For example, the skill may be associated with custom sample user inputs (corresponding to the custom functionality) that correspond to a custom domain while the skill may also be associated with system-provided sample user inputs (corresponding to the system-provided functionality) associated with a non-custom domain.