Abstract: Local computing resources may be created for machine learning tasks. A kernel of a machine learning development environment implemented as a first container at a host system may execute code to perform a machine learning task that includes a statement that causes a local execution resource to be created at the host system to perform the machine learning task. A second container may be placed as the local execution resource at the host system as a child application at the host system. A network namespace of the first container may be used as the network namespace of the second container, making the second container able to perform network communications with the first container locally on the host system. The second container may perform the machine learning task in accordance with the code.

Abstract: Shared computing resources may be determined for executing actions in an application workflow configuration. An application workflow configuration that describes a workflow that includes actions with respective action code executable to perform the actions may be obtained. Execution of the workflow may be directed according to the application workflow configuration. As part of execution of the workflow, an assignment of two of the actions may be individually determined for a same computing resource. Execution of the respective action code for the two actions may be caused according to the same computing resource.

Abstract: Automated unloading systems and methods may include a preparation subsystem and a transfer subsystem. A package stack may be loaded into the preparation subsystem to open or enable access to individual packages of the stack, and the opened stack may be rotated to an unloading position. Then, conveyance mechanisms of the transfer subsystem may cause automated unloading of individual packages of the opened stack, thereby improving the safety, speed, and efficiency of such package unloading processes.

Abstract: In a data storage system, stored data is secured using a key-based encryption algorithm. The disclosure provides mechanisms for increasing security of the stored data by storing a unique key used to encrypt and/or decrypt the data in a volatile key storage of a backplane, where power to the volatile key storage is routed through a connection that is closed when a corresponding data storage device is connected to the backplane and open when the corresponding data storage device is disconnected from the backplane.

Abstract: A file system manager implemented at a provider network identifies a storage device of a first group of storage devices of a provider network as an initial location of a file system object. Based on an access metric associated with the object, the file system manager initiates a transfer of contents of the object to a second storage device of a different storage device group, without receiving a client request specifying the transfer. In response to an access request received via a file system programmatic interface, contents of the object are provided from the second storage device. Based on a second access metric, the object is transferred back to the first group of storage devices.

Abstract: Disclosed implementations improve data that is used to train a deep neural network to determine a device type or a device operational state at a location based on voltage data at the location. For example, labeled training data for a location may be up-sampled to increase the amount of labeled training data that is available for use in training the DNN by obtaining labeled data from other known locations that are determined to be similar to the location. Likewise, the up-sampled training data may be subsampled during training to remove potentially noisy data from the training data, thereby improving the accuracy of the trained DNN.

Abstract: An intermediate sorted version of a projection of a data set is used to create the projection of the data set. A copy of the data set can be accessed to store an intermediate sorted version and determine a number of portions to the projection of the data set across a number of data stores. Then, distributed execution to create the projection of the data set to be performed from different portions of the number of portions to store at respective ones of the data stores.

Abstract: A method of power-saving and/or proximity-based auto-waking of a mobile device and/or an input device or stylus includes determining, by a mobile device, that motion of an input device has occurred. The method further includes sending, based on the determining of the motion of the input device, a power-on signal to a touch controller circuit and sending a wireless connection beacon using a short-range wireless communication protocol. The method includes receiving acknowledgement data responsive to the wireless connection beacon from the input device and sending, in an instance where a central processing unit (CPU) is in a low-power state, a power-on signal to the CPU. The method further includes receiving, using the short-range wireless communication protocol, input data associated with an interaction of the input device with a display of the mobile device.

Abstract: Technologies directed to calculating an array factor for a phased array antenna are described. One wireless device with a phased array antenna receives inputs and defines a first variable in a first coordinate space and defines a set of amplitudes of a one-dimensional or two-dimensional inverse Fast Fourier Transform (IFFT). Each amplitude is based on the third value, an excitation amplitude and an excitation phase or delay of a respective antenna element. The wireless device determines first array factor values in the first coordinate space by computing the one-dimensional IFFT using the first variable. The wireless device performs a spline or linear interpolation to obtain second array factors at points in the physical coordinate space. The wireless device sends the second array factors to digital beam forming (DBF) circuitry that causes the phased array antenna to steer a radiation beam using the second array factor values.

Abstract: An autonomous mobile device (AMD) moves within in a physical space. The AMD may encounter features in the physical space, such as flooring transitions, or obstacles that may impede movement. Sensors such as an ultrasonic sensor, optical time of flight sensors, inertial measurement unit (IMU), wheel speed sensors, and motor torque sensors acquire sensor data. Based on the sensor data, a location of an obstacle, and characterize whether that obstacle is traversable or non-traversable, is determined. If traversable, the AMD may continue moving past the obstacle. If non-traversable, the location of the obstacle may be added to a map used for route planning. Based on the location of the non-traversable obstacle, the AMD may move to a recovery location away from the obstacle, and resume moving.

Abstract: Systems and methods are described for providing a secure network connection between a serverless instance and an external environment with respect to a serverless execution environment hosting the serverless instance. A secure connection management system may be implemented in a serverless execution environment with at least one host device hosting the one or more function execution environments. The secure connection management system may facilitate establishment and management of a secure network connection between the serverless instance and the external environment in accordance with the connection configuration. The secure connection management system may also be isolated from the function execution environments in a separate execution environment such that the secure connection management system may take action to reduce risk to the external environment against issues with the execution of serverless function code on the serverless instance.

Abstract: This disclosure describes, in part, a mobile apparatus for identifying items within a facility using an omnidirectional imaging system. For instance, the mobile apparatus may include a main frame, a chassis attached to the main frame, a basket that attaches to the chassis in order to prevent the basket from contacting the main frame, and an omnidirectional imaging system providing image data in and around the mobile apparatus. A user may place item(s) within a receptacle of the basket. The mobile apparatus may further include a handlebar module attached to the main frame, the handlebar module including the omnidirectional imaging system and a computing system for identifying items and events in and around the cart.

Abstract: Aspects of the present disclosure enable a system to perform automated risk analysis of source code of applications. The system may include a machine learning model to analyze the source code. The machine learning model may output an application risk value. Further, the application risk value may be used to determine a remedial action. The remedial action may address a risk associated with the application risk value, or allow for publication of the application.

Abstract: The present disclosure generally relates to systems and methods for generating images of an approved identity wearing an article of clothing. The image generation system may generate an image of a person with a blurred face wearing a particular article of clothing based on a received clothing image and estimated attributes from an image depicting the face associated with the approved identity. The image generation system may then deblur the face into a sharp face conditioned on the approved identity based on training of limited data with sharp faces. To generate the final output image, the image generation system integrates the deblurred face into the generated image of the first stage.

Abstract: Techniques for generating a personalization identifier that is usable by a skill to customize output of supplemental content to a user, without the skill being able to determine an identity of the user based on the personalization identifier, are described. A personalization identifier may be generated to be specific to a skill, such that different skills receive different personalization identifiers with respect to the same user. The personalization identifier may be generated by performing a one-way hash of a skill identifier, and a user profile identifier and/or a device identifier. User-perceived latency may be reduced by generating the personalization identifier at least partially in parallel to performing ASR processing and/or NLU processing.

Abstract: Systems and techniques are described for providing video enhancements during sports broadcasts. In various examples, first tracking data representing first respective locations of a first plurality of players on a two-dimensional plane at a first time may be received. First embedding data representing a formation of the first plurality of players at the first time may be generated based at least in part on the first tracking data. A first set of historical plays may be determined based at least in part on searching a first data store using the first embedding data. A first area of the two-dimensional plane corresponding to a play vulnerability may be determined based on respective outcomes of the first set of historical plays. A first graphical overlay may be displayed in association with the first area on a live video feed.

Abstract: Techniques for static script malware detection with semantic code representations are described. A request to perform a static malware detection analysis is received, the request identifying a script comprising code. A semantic representation of the script is generated, the semantic representation including context about statements in the code of the script. A feature space embedding is generated based at least in part on the semantic representation of the script. generating, based at least in part on the feature space embedding, A result indicating whether the script is malicious is generated based at least in part on the feature space embedding. An identification of the script and the result is stored in an entry in a log.

Abstract: An improved data system supports creation of multiple hierarchical database catalogs under a particular account. Users or organizations have the flexibility to organize their data, native or external, into an n-level hierarchy of their choice. There is a root database catalog of the account, and other catalogs (1 or n-levels) are created under it. The data system has extract, transform, and load (ETL) features that allows data to be moved between different data sources and targets.

Abstract: A machine learning network is trained to use biometric input data to generate representation data used to assert identity. The network may be trained to consider particular features of interest. For example, the biometric input data may comprise an image of a user's hand. Particular features may comprise one or more of fine print ridges on the hand, principal lines and creases, wrinkles, veins, and so forth. In one implementation intermediate data produced by a first portion of the network is trained to be representative of the particular feature(s) of interest. The resulting representation data, such as produced by a second portion of the network, may thus be based on these features of interest. In another implementation the image is processed to determine feature data, and the image and feature data are provided as input to the network. A hybrid network may also be used.

Abstract: A cart system for inventory management can include a cart frame including a base, a handle to control movement of the cart frame, a stationary shelf, and a movable shelf disposed above the stationary shelf. The movable shelf is movable in a path of travel vertically overhead of the stationary shelf to change a volume underneath the movable shelf to stack items over each other within the volume and supported by the stationary shelf. A height adjustor can be configured to lower or raise the movable shelf in a vertical direction to an ergonomic height for loading the items and a storage height to store the items while creating sufficient volume below the movable shelf loading the items on the stationary shelf. A cart controller can be configured to receive a height input, and control the height adjustor based on the height input.

Abstract: Systems and methods are provided to round the numbers produced by a systolic array. A rounder, including a plurality of random number generators, can receive a number from the systolic array. Each random number generator may be associated with a random number sequence and may generate a next random number in the random number sequence based on a state value representing a position within the random number sequence. Each random number generator can assume a state based on state value shared by the plurality of random number generators. Each random number generator can cycle through a number of positions within the random number sequence based on the relative position of the random number generator within the plurality of random number generators to generate a random number. The rounder can perform a rounding operation using the generated random number.

Abstract: Provided is a system for facilitating recovery of deleted computing resources in a cloud network environment. A centralized resource recovery service may communicate with a plurality of resource management services that are each configured to create, modify, or delete their respective computing resources such as storage volumes, databases, and compute instances. The resource recovery service may allow creation of immutable resource retention rules that are, under certain conditions, unable to be deleted or modified, even by an administrative user who created such retention rules. Such immutable retention rules may be used to thwart malicious attempts to permanently delete resources by deleting the retention rules that govern such resources. Since such attempts to delete such immutable retention rules would fail, the retention rules would continue to apply and allow deleted resources to be placed in a recoverable state according to the retention rules.

Abstract: A set of security rules for user plane traffic of a subset of user equipment (UE) devices of a radio-based application is received via programmatic interfaces at a service of a cloud provider network. A packet comprising user data associated with a UE device of the subset is received at a network function implementation server of the service. An action indicated in a security rule is implemented with respect to the packet at the server.

Abstract: Methods and systems configured to embed data representing content output by at least one device for downstream content recognition or other downstream processes. For example, a device operates one or more media data embedding components configured to embed information regarding the output media, such as audio or image content. The embedding component is trained to embed information for a subset of known downstream processes with some known processes deliberately held out from the training. Among other benefits, this approach can help reduce over customization of the embedding component and allow more information to preserved by the component for purposes of downstream operations that may yet be configured.

Abstract: Systems and techniques are disclosed for determining discrepancies between conditions and parameters associated with an image. An image processing model may generate output indicating predicted values for an image using the image as input, while a text processing model may generate output indicating corresponding predicted values for an image using textual data associated with the image as input. A comparison of the output data may be performed to determine discrepancies between values. Discrepancies that are sufficiently significant and relevant may be reported for additional analysis.

Abstract: A diffusion model is provided with a reference image and a text prompt, and generates output images having content specified in the prompt and style characteristics represented in the image. The reference image is encoded to generate multiple image embeddings. The prompt is encoded to generate a text embedding that is separate from the image embeddings. After each denoising step of the diffusion model, the text embedding is used to perform a first cross-attention operation on the denoising output, and the image embeddings are used to perform a second cross-attention operation. During earlier denoising steps, the second cross-attention operation is associated with a low weight parameter, causing the operation to have a smaller effect on the output image. During later denoising steps, a high weight parameter is used. In some cases, the input image for the model may be generated by performing a diffusion process using the reference image.

Abstract: Techniques are described herein for image-based determination of container fullness. An example method can include a computing system receiving, from an image-capturing device, a first two-dimensional image of a container including a first item. The computing system can determine, using a machine learning model, three dimensional information based at least in part on the two-dimensional image, the three dimensional information indicating a fullness of the container. The computing system can compare the fullness of the container to a target fullness of the container. The computing system can determine whether the fullness of the container exceeds the target fullness based at least in part on the comparison.

Abstract: Methods and systems of using an approximate duplicate event filter are disclosed. Event processing systems of a service provider network such as a security service may receive a large volume of events to analyze. Events that have previously occurred can be given a lower priority for analysis than new events. Events that are similar to previous events can likely also be given a lower priority for analysis. These non-exact match events can be detected by comparing token representations of new events to token representations of previous events. Events may be considered according to the context in which the events occur.

Abstract: A control plane server of a network-accessible service of a cloud provider network obtains an indication from a client that security artifacts to be used for establishing connections among constituent services of an application are to be obtained automatically by the service. The control plane server transmits, to a first agent established at a first resource at which a first constituent service of the application runs, a set of security artifacts obtained from an artifact source and assigned to the first constituent service by the control plane server. The set of artifacts is used to establish a connection between the first agent and a second agent at a second execution resource at which a second constituent service of the application runs. Messages between the constituent services are sent using the connection.

Abstract: Systems and methods are provided for swift restoration of a network connection and communication session between computing systems. A network connection between two computing systems may be established. When the network connection is interrupted, one computing system may send a reestablishment connection to the other computing system, and the network connection may be reestablished based on the reestablishment communication without establishing a second network connection.

Abstract: Disclosed are various embodiments for a simulation framework that facilitates the integration and management of multiple services in a software-in-the-loop simulated environment. A simulation runtime engine can deploy multiple services which may have dependencies upon one another in a simulated environment. Network service calls sent from one simulated services to another service may be intercepted to ensure that the network service call is sent to the simulated instance of the service instead of the actual non-simulated instance of the service. In addition, system time calls are intercepted to ensure that the requested service is provided the simulation time.

Abstract: A large language model (LLM) may be used to generate an overall entity representation of an entity using an input that includes individual representations based on graph data, text data, and image data associated with the entity. Graph data that represents characteristics of the entity and types of relationships between the entity and other entities is used to generate a graph representation. Text associated with the entity is used to generate a text representation. Image data associated with the entity is used to generate an image representation. These representations are used to generate an input to the LLM, which is trained to generate an entity representation based on the input. The entity representation may be used by other models, such as to determine entities having similar or differing characteristics.

Abstract: Techniques for updating a model are described. An example includes determining a proper subset of weights of a first trained machine learning (ML) model to change during an optimization procedure; updating only the proper subset of weights of a second ML model on a first source dataset to generate an optimized ML model; and storing the optimized ML model.

Abstract: An automated conversation intermediary performs several operations during a conversation between end users and chatbots. Prior to adding natural language input from an end user to a chatbot prompt, the intermediary annotates the input to indicate a role of the end user, using hidden markers for the annotation that are not presented to the end user during the conversation. The intermediary receives output generated by the chatbot, which also includes hidden markers in addition to natural language to be presented to the end user, and removes the hidden markers before providing the natural language portion of the output to the end user.

Abstract: A search language is provided for writing rules that allow semantic and syntactic checking of source code to find potential errors in a target code. The language allows for wildcards (also called placeholders), which provide a more expansive matching of search terms. A compiler is described that converts a search language query into a semantic graph representation. The target code being searched can also be converted to a target graph representation. A comparison can then be performed between the semantic graph representation and the target graph representation to determine if the semantic graph representation is contained within the target graph representation. If so, then a semantic match is found in the target source code. More specifically, semantic matching can be realized by checking homomorphism against the target graph representation. In one example, the semantic matching can include a comparison of nodes and edges connecting the nodes.

Abstract: A first-in-first-out (FIFO) device includes a plurality of FIFO stages coupled in series. Each FIFO stage includes a data register representing a FIFO entry, a valid register storing a data valid signal to provide a valid signal to indicate whether the FIFO entry is storing valid data, and an input multiplexer to provide data to a FIFO stage. Upon receiving a push signal to store input data into the FIFO device, a FIFO stage of the FIFO device determines whether to update the data register of the FIFO stage based on respective valid signals from the FIFO stage, and neighboring FIFO stages of the FIFO stage. Upon receiving a pop signal to provide output data from the FIFO device, the FIFO stage determines whether to provide data stored in the data register of the FIFO stage or the input data of the FIFO device to the preceding FIFO stage.

Abstract: Automated scaling-related operations may be performed dynamically during execution of a spatial simulation. A spatial partition may be locally reassigned, based on application workload information, from a first application to a second application on the same worker. A quantity of applications on a worker may also be changed during execution of a spatial simulation. A parent spatial partition may be split into child spatial partitions, and child partitions may also be merged back into a common parent partition. Indications of partition splits and merges on each of a plurality of workers may be reported to the plurality of workers. A spatial partition may also be remotely reassigned from a first worker to a second worker, such as based on worker-level resource consumption information and partition information. A quantity of workers that are used to implement a spatial simulation may also be changed during execution of the spatial simulation.

Abstract: Systems and methods are described for using multiple threshold maps for color dithering. An example method may comprise determining a tertiary RGB color. The example method may comprise selecting, using data representing the tertiary RGB color, a dithering threshold map indicating a pattern of alternating threshold values. The example method may comprise selecting, based on the tertiary RGB color and the dithering threshold map, a tertiary color value set comprising a red value, a green value, and a blue value. The example method may comprise mapping, based on the dithering threshold map, the alternating threshold values to a pixel set of a portion of an RGB display. The pixel set may include one or more of a red pixel, green pixel, and/or blue pixel. The example method may comprise activating, based on the tertiary color value set and the dithering threshold map, at least in part, the pixel set.

Abstract: Systems and methods are disclosed for shuttle height monitoring systems for shuttles powered by linear synchronous motors. An example system may include a track, a first sensor disposed adjacent to the track, a second sensor disposed adjacent to the track, a third sensor disposed adjacent to the second sensor, and a controller. The controller may be configured to determine, using the third sensor, that the first shuttle is adjacent to the first sensor, determine, using the first sensor, a first distance between a first lower portion of the first shuttle and the first sensor, determine, using the second sensor, that the second shuttle is adjacent to the first sensor, and determine, using the first sensor, a second distance between a second lower portion of the second shuttle and the first sensor.

Abstract: In various examples, systems and methods for stability prediction for 3D robotic bin packing are described. A set of contact points for a first item at a first candidate placement position may be determined. A first contact point having a minimum value along a first coordinate dimension and a second contact point having a maximum value along the first coordinate dimension may be determined. A third contact point that is a maximum distance between a line between the first contact point and the second contact point may be determined. A determination may be made whether a centroid of the first item is within a triangle defined by the first contact point, the second contact point, and the third contact point. A stability of the first item at the first candidate placement position may be determined based on the location of the centroid.

Abstract: A system analyzes network traffic at the network edge in a Point-of-Presence (POP) to prevent denial of service attacks, such a reflection attack, and filters such attack traffic. The POP can analyze a source identifier of the network traffic, such as a source prefix, and compare the source identifier to known source identifiers from the same geographic region as the POP. If the source identifier is outside of the POPs geographic region, then the network traffic is rate limited, such as by blocking the network traffic or reducing the network traffic to a desired threshold. To further ensure that the traffic is routed to the appropriate POP, the POP edge location announces anycast addresses for receiving the network traffic. The geographic region can be defined as a city, a country, sub-regions within a country, a portion of a continent (e.g., southeast region of Asia), a continent, etc.

Abstract: Testing recommendations for an updated software build can be generated. For example, a computer system can determine a classification of a modification of a component of a software build into a category based on a text file that indicates the modification of the component. The computer system can also determine a recognition of an entity type for the component of the software build based on the text file. The computer system can generate, based at least in part on the classification and the recognition, a recommendation indicating one or more tests for use in testing the modification to the software build.

Abstract: Computer-implemented techniques for autonomic control plane scaling in a container orchestration cluster. The techniques include a method that involves managing a container orchestration cluster's control plane. When a control plane instance is underutilized, it handles control plane requests and monitors its resource usage. If it becomes overutilized (hot state), it identifies another instance with spare resources (cold state) and redirects control plane traffic to it to maintain performance and availability.

Abstract: Techniques for implementing cross-domain identity trust using trusted token issuers are described. An identity provider is registered with an identity service of a cloud provider network in association with an organization or user. Thereafter, tokens generated by the identity provider can be provided to the identity service as part of token exchanges to obtain new tokens useful for accessing resources managed or provided by applications, such as services, within the cloud provider network. The token exchange process can include use of administrator-provided mapping information indicating how to associate user identity information from an identity provider token with user identity information in a directory utilized by the identity service, whereby the new tokens include identity information that is easily utilized by receiving applications for making access control determinations and performing logging for auditability.

Abstract: Systems and methods are provided for a scalable load balancer to enable client certificate access control using mutual transport layer security authentication comprising hardware configured to receive trust store files including a certificate authority (CA) bundle and a certificate revocation list (CRL), generate a single trust store file based on the trust store files, distribute a snapshot of the single trust store file to one or more nodes of the scalable load balancer, verify, using the snapshot, a client certificate from a client device by searching for a trusted CA from the CA bundle in the snapshot, determine whether the CRL contained in the snapshot includes the client certificate, and upon verifying the client certificate and determining that the CRL does not contain the client certificate, transmit connection data from the client device to an application target.

Abstract: A system is provided for determining privacy controls for output including sensitive data. A user may subscribe to receive an output in the future based on the occurrence of an event. The system may determine when the event is occurred triggering the output, and determine that the output includes outputting sensitive data. The system may determine output data that does not include the sensitive data, send the output data to a device, and may request the user to provide an authentication input to receive the sensitive data.

Abstract: Described are systems and methods for detecting sleep-related breathing events during a sleep session of a user based on sensor data from a wearable device, such as a wrist-worn wearable device. For example, disclosed implementations detect biometric changes of the user that may be indicative of sleep-related breathing events and determine, based on features determined from those biometric changes, that a sleep-related breathing event has occurred. In addition, the disclosed implementations provide a presentation and explanation of the determined sleep-related breathing events.

Abstract: A technique to compute statistics of data elements include serially inputting the data elements into a compute channel. The compute channel can generate a first running mean and a first running variance associated with data elements of the vector having odd sequence indices, and a second running mean and a second running variance associated with data elements of the vector having even sequence indices. Subsequent to serially inputting data elements into the compute channel, the first running mean and the second running mean are aggregated to generate a mean associated with the data elements of the vector, and the first running variance and the second running variance are aggregated to generate a variance associated with the data elements of the vector.

Abstract: Techniques for compressing a neural network model by mixing compression ratios (sparsity patterns) are described. The weight tensor of a neural network model is divided into weight groups. The pruning cost of compressing the weight values according to a compression ratio is determined for each weight group, and a pruning cost distribution for the compression ratio is generated from the pruning costs of the weight groups. A cost threshold can then be selected from the pruning cost distribution, and weight groups having a pruning cost below the selected cost threshold are compressed according to the compression ratio. The remaining weight groups can be compressed using one or more less aggressive compression ratios. The cost threshold can be adjusted to tune the overall sparsity and accuracy of the compressed neural network.