Abstract: Disclosed are systems and methods to perform neural architecture search (“NAS”) that automatically optimizes for the number of channels to allocate to each layer of a deep neural network. Some implementations include a pairwise slimming that includes a global optimization step. Likewise, in some implementations, a bias toward a region of interest may be applied to channel path selection during training.

Abstract: A request to add tags (e.g., labels, key-value pairs, or metadata) to resources can be digitally signed by the entity making the request, such that the source can be verified and an authorization determination made for each tag. For a request involving multiple services (or entities) that can each add tags, any tag added by a service can be included in the request and digitally signed by that service. Each service processing the request can also digitally sign the request before forwarding, so that each service signs a version of the request, which includes elements signed by other services earlier in the request chain. When the request is received to a tagging service, the service ensures that every tag was digitally signed by the appropriate authorized entity or service, and validates the signatures to ensure that no data was modified or omitted, before adding the tags to the designated resource(s).

Abstract: Devices and methods are provided for remote scheduling of recorded content for digital video recorders. The device may determine a first channel, a first date, and a first time at which a first program will be presented. The device may determine a first recording schedule for a remote digital video recording (DVR) device, wherein the first recording schedule includes the first program at the first channel, the first date, and the first time. The device may receive a synchronization notification that the first recording schedule has been modified at the remote DVR device to include a second program to be recorded. The device may determine a second recording schedule, wherein the second recording schedule includes the second program at the second time. The device may send the second recording schedule to the remote DVR device.

Abstract: An item delivery system can include a carrier positioned on a rail. The rail can be positioned in a delivery vehicle, for example. The carrier can include rigid or semi-rigid sidewall and dividers to form item slots. The item slots can receive one or more items. The carrier can be positioned on the rail and moved around the rail by an advancement device. The advancement device can engage with a portion of the carrier to move the carrier. The carrier can be moved to a position for removal of items from the carrier. The removal position may correspond to an area accessible from an operator area of the delivery vehicle, for example.

Abstract: Methods of video streaming are generally described. In some examples, a camera device periodically captures an image, communicates encoded frame data representing that image to a server, and decodes and stores the previously encoded frame data as a background picture. The server receives the encoded frame data, decodes it, and stores the decoded frame in a buffer for future use. Subsequently, upon initiation of a streaming session, the camera device captures another image and encodes a predicted key frame based on differences between the captured image and the background picture. The camera device sends the predicted key frame to the server, which receives it and reconstructs a facsimile of the captured image utilizing the previously decoded frame stored in the buffer. Methods of acknowledging successfully decoded frames for use in selecting background pictures is also described.

Abstract: Techniques for disabling a sensor are described. In an example, the techniques include receiving input data indicating a request to disable a sensor of the device. Based on the input data, the sensor is disabled by at least decoupling the sensor from a power source and decoupling the sensor from an input terminal of a processor of the device. The techniques also include sending, to a plurality of clients that are implemented on the device and that are configured to receive sensor data generated by the sensor, notification data indicating that the sensor is disabled. Further, the techniques include outputting, by a user interface of the device, an indication that the sensor is disabled.

Abstract: Anomalies can be detected and related mitigation actions can be executed. For example, a computing device receives, from a first source, first data associated with a first process of a payment service in support of transactions of a first plurality of users. The computing device generates first aggregated metrics based on the first data. The computing device determines, by using the first aggregated metrics as first input to a machine learning engine, an anomaly associated with the execution of the first process. The computing device determines, based on the anomaly and a history of anomalies associated with the execution of the first process, an indication of an outage associated with the first process. The computing device causes an execution of a mitigation action based on the indication of the outage. The execution of the mitigation action is in support of transactions of a second plurality of users.

Abstract: Techniques for enabling the identification of anomalous events associated with an object storage service of a cloud provider network using a variational autoencoder model including a pre-trained embedding for selected features of events are described. A variational autoencoder, for example, encodes data into a latent space and reconstructs approximations of the data from an encoding in the latent space. In this context, for example, anomalous events of interest might represent unauthorized or abusive behavior associated with storage resources provided by an object storage service (or in association with other types of computing resources provided by other services of a cloud provider network). Legitimate (or benign) access patterns to an object storage service can be modeled by utilizing observed data plane events stored by an account activity monitoring service. Once trained, the model can be used to identify anomalous events.

Abstract: Techniques implemented by a cloud computing system for providing fallback capacity providers to ensure that infrastructure capacity required to run containerized services is available despite primary capacity providers experiencing a failure. Cloud providers offer container-management services that automate the management and scaling of containerized services of users. The container-management services are supported by capacity providers that manage the computing infrastructure on which the containerized services run (e.g., servers, VMs, etc.). However, there are times when a capacity provider is unable to provision capacity for containerized services, such as due to a large scale failure. Rather than leaving capacity requests unserved, the container-management service designates a fallback capacity provider that acts as a failover for provisioning requested infrastructure capacity.

Abstract: A training system may create and train a machine learning model with knowledge transfer. The knowledge transfer may transfer knowledge that is acquired by another machine learning model that has been previously trained to the machine learning model that is under training. The knowledge transfer may include a combination of representation transfer and instance transfer, the two of which may be performed alternatingly. The instance transfer may further include a filter mechanism to selectively identify instances with a satisfactory performance to implement the knowledge transfer.

Abstract: A namespace monitoring service may track released namespaces such as internet protocol (IP) addresses and manage namespace cooldown pools, available namespace pools, and a registry of released namespaces to detect and mitigate security vulnerabilities that arise from reassignment of namespaces. The namespace monitoring service provides access to the released namespace registry and/or sends a data stream of namespace registry updates. The namespace monitoring service may manage namespace reassignment process and extend the cooldown period of released namespaces or place a hold on available name spaces.

Abstract: Techniques for aligning (merging) knowledge graphs is described. Graph entity alignment is the problem of “joining” two knowledge graphs based on common entities. Most approaches in literature solve this using some annotated seed entity pairs and train a supervised model to rank entities in one knowledge base against another. In some examples, unsupervised alignment approach that uses graph-to-text summaries to encode entities in two or more distinct graphs into the same representation space, encodes those summaries into a common space, and then uses similarity analysis to determine when graph entities should be aligned (merged).

Abstract: Described is a system and method for proactively determining and resolving events having low confidence scores and/or resolving disputed events. For example, when an event, such as a pick of an item from an inventory location within a materials handling facility occurs, the event aspects (e.g., user involved in the event, item involved in the event, action performed) are determined based on information provided from one or more input components (e.g., camera, weight sensor). If each event aspect cannot be determined with a high degree of confidence, the event information is provided to an associate for resolution.

Abstract: Automated container provisioning systems described herein may include one or more pairs of transport belts, one or more guide surfaces, and one or more sensors. The transport belts may engage flaps or other portions of containers, and the guide surfaces may support surfaces or other portions of containers, in order to transport containers to downstream stations within a facility. In addition, the sensors may detect containers during transport, responsive to which various additional operations may be performed with respect to containers, such as sortation, printing, labeling, tracking, packing, accumulation, or others.

Abstract: A resource request is received by a peripheral device from host processing logic. The resource request includes a requested resource size. The peripheral device allocates resource of the peripheral device in response to the resource request. A resource response is sent by the peripheral device to the host processing logic. The resource response includes a location of the allocated resource.

Abstract: A query including a time range can be received. The query can be processed to determine data associated with the time range of the query. Cached data can be analyzed to identify a first block of data including metadata identifying a first timestamp associated with the time range of the query. A communication can be generated to include the first timestamp associated with the first block of data. Based on the communication, a second block of data can be received with metadata identifying a second timestamp associated with the time range of the query. A query result can be generated to include the first block of data and the second block of data.

Abstract: A tensor network decoder accounts for correlated noise between qubits of a two-dimensional code. The tensor network decoder is generated using a graphical noise model for a quantum device that is used (or to be used) to implement the two-dimensional code. For example, an input graphical noise model, such as a hypergraph, may be used to generate a tensor network decoder. Whereas other decoders assume noise is independent and identically distributed (e.g. iid noise), a tensor network decoder accounts for correlated noise not considered in such decoders that assume iid noise.

Abstract: Pages describing episodes of podcasts or other media entities are constructed by interpreting content of the media entities. A transcript of an episode is determined by one or more natural language understanding techniques and divided into chapters. For each of the chapters, a summary sentence of the chapter and one or more key phrases are determined from the transcript, and participants in the chapter are identified. A summary of the episode is determined from the summary sentences of each of the chapters. A page that describes the episode of the podcast including the summary of the episode, as well as one or more of the key phrases and identities of the participants is generated and provided to prospective listeners to the episode.

Abstract: Based on an analysis of example content items, a proposed initial version of an application to be generated on behalf of a client is determined. At least a portion of a software program implementing the application is generated, and an indication of the software program is provided to the client. Based on input received from the client, additional operations associated with the program are initiated.

Abstract: A neural network accelerator executes instructions to: load a first weight data element of an array of weight data elements from a memory into a systolic array; extract, from the instructions, information indicating a first number of input data elements to be obtained from a first address of the memory and a second number of input data elements to be skipped between adjacent input data elements to be obtained, the first address being based on first coordinates of the first weight data element, and the first and second numbers being based on a stride of a convolution operation; based on the information, obtain first input data elements from the first address of the memory; and control the systolic array to perform first computations based on the first weight data element and the first input data elements to generate first output data elements of an output data array.

Abstract: A system comprising one or more computers implements a synthetic sensor service configured to deploy synthetic sensors to an in-vehicle computing device implementing a synthetic sensor orchestration environment for a vehicle. The synthetic sensor orchestration environment determines a placement decision for a new synthetic sensor to be added to the vehicle based on one or more annotations included in a synthetic sensor package for the new synthetic sensor. The synthetic sensor service and respective synthetic sensor orchestration environments implemented in various types of vehicles provide a consistent way to remotely add additional synthetic sensors and/or other functionality to different types of vehicles after the vehicles have already been put in use by respective owners or operators of the vehicles.

Abstract: An update service of a provider network transmits an update to an edge device by opportunistically selecting one communication channel (or transmission bandwidth) over another (e.g., based on analysis of performance metrics associated with the channel) to transmit different portions of the update to the edge device. For example, the update service may choose the channel with highest available bandwidth or the channel with the lowest failure rates. The update service may receive indications from the edge device that transmission of different update portions was successful. In some cases, the edge device itself may instead perform the selection of one communication channel over another (e.g., based on analysis of channel metrics) to receive different portions of the update at the edge device. The edge device may transmit indications that the edge device successfully received different portions of the update.

Abstract: A materialization configuration request is received via a programmatic interface from a client of a journal-based multi-data-store database. The request indicates a partitioning rule to be used to select, for respective writes indicated in committed transaction entries of a journal, the materialization node at which the writes are to be stored. A control plane component of the database verifies that a set of materialization nodes corresponding to the partitioning rule has been established, and initiates the propagation of writes from the journal to the materialization nodes by respective write appliers.

Abstract: Technologies are disclosed herein for providing asynchronous application store event notifications. An application developer registers with an application store to receive asynchronous push notifications for application store events of interest, such as events relating to in-application purchases, payment events, fulfillment events, application publishing events, or other types of events. When an event for which a notification is to be provided is detected, an application store event message is generated for the event and provided to a notification service. The notification service generates an application store event notification for the event on a topic defined in advance by the application developer. The application store event notification may then be delivered to an endpoint defined by the application developer, such as a hypertext transfer protocol (HTTP) endpoint.

Abstract: Systems and techniques for displaying advertisements on a digital display and gathering impression and view data related to the advertisement that may be used to refine or score the advertisements for greater effectiveness. The impression and view data may be used to identify effective portions of advertisements and subsequently to train a machine learning model to predict impression data for advertisements that may be used to iteratively improve the advertisements.

Abstract: A content delivery service includes a management service for determining an anticipated window of time in which supplemental content can be inserted in the dynamic content stream. The management service will then pre-fetch supplemental content from the third party supplemental content service prior the identification or request for supplemental content to be included in the streaming content. During the streaming of content to individual users via an established channel, as markers for insertion opportunities are identified, the content delivery service will attempt to utilize the pool of pre-fetched supplemental content.

Abstract: A mechanical design for a braking system of a delivery device is described. The braking system includes one or more brake modules. Each brake module is configured to engage a different wheel of the delivery device to stop movement of the wheel when power to the brake module is less than a threshold. Each brake module includes: an electromagnet; a magnetic plate; a pawl configured to engage a wheel to stop movement of the wheel; a solenoid comprising a solenoid pin, the solenoid pin having a first end coupled to the magnetic plate and having a second end coupled to the pawl, and a manual override mechanism that is configured to allow movement of the wheel when the power to the brake module is less than the threshold. The brake module is configured to disable the manual override mechanism when the power to the brake module is greater than or equal to a threshold.

Abstract: Techniques for a Predictive Connection Manager Service (PCMS) to predict when client applications will send service requests to backend services, and proactively establishes connections, caches data, or takes other actions, to reduce latencies between receipt of and response to these service requests. The PCMS analyzes historical usage data for the client applications to identify usage patterns, and uses those usage patterns to proactively scale resources to handle service requests. The PCMS can be implemented as a pass-through proxy for client applications to reduce frictions for managing how users interact with backend services. For instance, the PCMS can install client-side drivers such that updates or patches for the drivers need only be installed on the PCMS rather than on each client device.

Abstract: Techniques for generating a dataset from a knowledge graph are described. An exemplary method includes receiving a request to generate a dataset from a knowledge graph to be stored in the storage; generating a dataset comprising a plurality of mention-concept pairs from the knowledge graph according to the request based one or more of a synonym-based and graph-based evaluation of the knowledge graph and a custom ontology for the knowledge graph; and storing the generated dataset in the storage.

Abstract: Systems and methods for implementing lockless, leaderless transaction protocols for scalable applications such as distributed data stores are disclosed. Applications may be implemented using multiple nodes individually implementing all or a subset of lockless, leaderless protocols. A transaction, such as a write to a data object, may be received at a node, the operation including a user-provided key. A composite key may be generated using a monotonic sequence number, the composite key globally unique across the application. A new data object is created based on data provided and once the object is created, the object is immutable. Other operations are implemented such that no data object may be written, modified, or deleted by more than one operation, thus eliminating contention and allowing for lock-free operation. Leaderless operation, as well as replication using erasure coding and redundant data paths, further minimize latency of the protocol.

Abstract: Input data, such as audio and/or video data, may be captured from a first room, for example via microphones and/or cameras within the first room. A first quantity of people within the first room may be determined based at least in part on the input data. An alert may be provided when the first quantity of people exceeds a threshold quantity of people. Additionally, locations of people within the room may also be detected based at least in part on the input data. A first proximity of a first person in the room to a second person in the room may be determined. An alert may also be provided when the first proximity is less than a threshold proximity.

Abstract: Techniques for applying a perceptual quantizer (PQ) electro-optical transfer function to optimize dithering and adaptive quantization for non-perceptual quantizer (e.g., SDR) color spaces are described. According to some embodiments, a computer-implemented method includes receiving a request to encode a video comprising a frame having pixels with non-perceptual quantizer code values of luminance at a content delivery service, determining, by the content delivery service, perceptual quantizer code values for the frame based at least in part on the non-perceptual quantizer code values, determining, by the content delivery service, one or more encoding parameters based at least in part on the perceptual quantizer code values, encoding, by the content delivery service, the frame based at least in part on the one or more encoding parameters into an encoded frame, and transmitting the encoded frame from the content delivery service to a viewer device.

Abstract: Systems and methods for device control by a natural language processing system are disclosed. A user may desire to utilize a voice-enabled device to associate an accessory device with a hub device without having to utilize third-party software associated with the accessory device and/or the hub device. The user may provide a user utterance to associate the accessory device with the hub device. Audio data corresponding to the user utterance may be analyzed and utilized to generate and send directive data to a third-party remote system to transition the hub device to a join mode. Upon association completion, audio may be output confirming that the association has been established successfully.

Abstract: Techniques are described for assisting users with the process of modernizing software applications from a source software architectural pattern (e.g., a model—view—controller (MVC) pattern, a multitier architecture, etc.) to a target architectural pattern (e.g., microservices, event-driven architecture, etc.). A decomposition analyzer receives as input artifacts associated with a software application under analysis and an architectural pattern that a user desires to use for a modernized version of the software application. The decomposition analyzer generates, based on the application artifacts, a graph-based model of the software application, where the model includes nodes representing application components (e.g., classes) and edges representing dependency relationships among the components. The decomposition analyzer further uses machine learning techniques and other types of analyses to identify recommended groupings of nodes from the graph-based model for decomposition.

Abstract: Techniques for detecting user activity and repetitions (rep) counting are described. A system may dynamically determine a threshold value for counting reps. The threshold value may be based on an activity predicted by the system using sensor data. The threshold value may be based on a rep rate. Changes in the threshold value may enable more accurate rep counting when a user is performing an activity that involves taking reps, such as, walking, running, rowing bicycling, etc., and may further cease counting of reps when a user is performing an activity that does not involve taking repetitions but may result in sensor data that triggers rep counting.

Abstract: An index handler determines, with respect to a key to be inserted into an index, that a candidate destination node of the index meets a split criterion. The index handler generates and embeds a deferred split descriptor comprising an identifier of a new node within the destination node. Before an insert-completed indication is provided, the destination node is written to a back-end data store without acquiring a lock and without writing out the new node to the back-end data store. During the traversal of the index, the index handler identifies another deferred split descriptor indicating a second new node. After providing the indication that the key was successfully inserted, the index handler writes the second new node to the back-end data store.

Abstract: Devices and techniques are generally described for adaptive targeting for voice notifications. In various examples, first data representing a predicted likelihood that a first user will interact with first content within a predefined amount of time may be received. A first set of features including features related to past voice notifications sent to the first user may be determined. A second set of features including features related to interaction with the first content when past voice notifications were sent may be received. A first machine learning model may generate a prediction that a voice notification will increase a probability that the first user interacts with the first content based on the first data, the first set of features, and the second set of features. Audio data comprising the voice notification may be sent to a first device associated with the first content.

Abstract: Described is a system and method for presenting event information to a user and, if necessary, obtaining confirmation of different aspects (user, item, action) of the event. In some implementations, an event includes a user, an action, and an item. For example, an event may include a user picking an item from an inventory location, a user placing an item into a tote associated with the user, etc. if the aspects of the event cannot be determined with a high enough degree of confidence, a user interface may be generated and sent to the user requesting confirmation of one or more of the aspects of the event.

Abstract: Methods, systems, and computer-readable media for automated threat modeling using application relationships are disclosed. A graph is determined that includes nodes and edges. At least a portion of the nodes represent software components, and at least a portion of the edges represent relationships between software components. An event is received, and a sub-graph associated with the event is determined. The event is indicative of a change to one or more of the nodes or edges in the graph. Threat modeling is performed on the sub-graph using one or more analyzers. The one or more analyzers determine whether the sub-graph is in compliance with one or more policies.

Abstract: Methods, systems, and computer-readable media for relevance-based schema matching for targeted catalog enrichment are disclosed. A catalog management system determines textual descriptors associated with a category of items in a catalog. The system determines relevant attributes for the category based (at least in part) on analysis of the textual descriptors. The relevant attributes are selected from a larger set of candidate attributes for the category. The system modifies the catalog based (at least in part) on schema matching for a plurality of items in the category. The items are associated with descriptive terms from a plurality of data sources and expressed according to a plurality of source schemas. For an individual item, the schema matching determines a correspondence between one or more of the descriptive terms in one of the source schemas and a corresponding attribute in a target schema comprising the relevant attributes.

Abstract: Systems and methods are provided for evaluation of communication paths through networks to determine whether communication is permitted across one or more internal network boundaries. The analysis may be used to determine whether a node in one isolated network (e.g., VPC, VPN, client on-premise network, etc.) is able to communicate with a node in another isolated network across region and/or segment boundaries. The automated analysis can allow users (e.g., network administrators) to see what high-level policies (e.g., Cloud WAN policies written in a declarative language) are interfering with or permitting communication between the nodes.

Abstract: A video packaging and origination service can include one or more encoder components that receive content for encoding and transmitting to requesting entities. During the operation of the encoder components, a management service associated with the video packaging and origination service can receive information related to updates or modifications to the encoder components. Responsive to the information, the management service can instantiate and configure an updated encoder. Once the updated encoder component is instantiated and configured, both the updated and target encoder components receive content for streaming, encode content and generating encoding content. Additionally, the updated encoder component can initiate a notification to the target encoder component that a handover will be occurring and begin a negotiation process identifying a handover event.