Abstract: Various embodiments of apparatuses and methods for multi-cast, multiple unicast, and unicast distribution of messages with time synchronized delivery are described. In some embodiments, the disclosed system and methods include a reference timekeeper providing a reference clock to one or more host computing devices. The one or more host computing devices host compute instances, and also contain respective isolated timing hardware outside the control of the compute instances. The isolated timing hardware of the one or more host computing devices then receive respective packets, and obtain the same time to deliver the respective packets. Each isolated timing hardware provides either the packet, or information to access the packet, to its respective destination compute instance subsequent to determining that the same specified time to deliver the packet has occurred. Thus, the respective packets are delivered near simultaneously to the one or more destination compute instances.

Abstract: Scalable garbage collections is performed by a distributed storage system. Garbage collection events are detected by a distributed storage system for different portions of a table. Garbage collection is performed for individual ones of the different portions of the table responsive to detecting the garbage collection events, including identifying one or more versions of a record to reclaim from the different portions of the table based on transaction status information and reclaiming the one or more versions of the record.

Abstract: Share memory communication may be implemented for multiple writers updating a database. Multiple writes may be received at a database management application and written to a back-end of circular buffer by different threads. A storage engine may use a read thread to read from a front-end of the circular buffer to obtain the writes after detecting that a lock indicates that the writes are ready to be read. The storage engine may perform the writes read from the circular buffer to update the database.

Abstract: Systems and methods for graph-based Extract Transform Load (ETL) code generation are disclosed. A natural language request to create ETL code according to a description of the ETL code in the natural language request may be received. A graph representation of the ETL code may be generated that includes nodes corresponding to operations to obtain data from data sources, perform transformations to at least the portion of the data, and store the data in destinations. To create the graph representation, interactions with a generative machine learning model may identify the operations for the ETL code in the description in the natural language request and construct the graph representation that includes the nodes corresponding to the operations along with connections between the nodes identified according to the description in the natural language request. A response to the natural language request may be provided via the interface based on the graph representation.

Abstract: Techniques for improving speculative decoding for improved compute utilization are described. A system may determine a compute budget representing a maximum number of speculative tokens to be verified based on hardware properties, a batch size and an input length. Compute budgets for varying batch sizes and input lengths may be determined offline. At inference time, the compute budget corresponding to a runtime batch size and runtime input lengths may be used to determine the maximum number of tokens that can be verified by a generative model while maintaining efficient compute utilization. The compute budget may be allocated/distributed among the inputs of the batch to efficiently process the inputs in parallel. The allocated budget may be updated for subsequent timesteps. The compute budget may represent the maximum number of tokens that can be verified without the system processing becoming a compute-bound scenario.

Abstract: Systems and methods for a robotic picking assembly system with articulating arms and wheels driven by motors are provided. In one embodiment, the robotic picking assembly may include a picking assembly connected to a robotic arm and having two arms that each may bend and each may have a wheel at the end. Each wheel may be independently driven by a motor. The picking assembly may be positioned over an item in a container (e.g., bin) and the wheels may be positioned on either side of the item and caused to rotate to move the item into the picking assembly. Once the item is moved into the arms of the picking assembly, the arms may be actuated to bend to support the item for transportation. The robotic picking assembly may then move the item to a different location and place the item in the new location.

Abstract: Record-level transaction conflicts may be detected at storage nodes of a database back-end storage service. When a prepare statement is received, a storage node can determine whether another prepare statement acting on at least one of the same records has succeeded. If so, then the storage node can indicate to the database access application that there is a conflict, allowing the database access application to wait to try again or abort the transaction.

Abstract: A system configured to improve noise estimation while performing noise dependent volume control to reduce a sensitivity to voice activity and other non-stationary sounds. For example, a device may adaptively increase a volume level of playback audio based on an amount of ambient noise present in a noisy environment, with an amount of noise compensation determined based on an estimated noise level. To avoid overestimating the noise level during periods of voice activity (e.g., speech), loud transient sounds (e.g., slamming doors, coughing), and/or residual acoustic echo, the device may adaptively slow a noise tracking speed for non-stationary signals. For example, the device may extract features that distinguish between stationary signals and non-stationary signals and calculate a stationarity coefficient to adaptively control noise estimation. Thus, the device may use a normal noise tracking speed for stationary signals and a slower noise tracking speed for non-stationary signals.

Abstract: Systems and methods are disclosed for generating additional data for one or more frames based, at least in part, on re-synchronization of a transceiver that provides the additional data to one or more additional processors. Systems identity an indication of a trigger that is associated with the one or more frames, cause a set of sensors to generate one or more frames in response to the trigger, generate the additional data, and transmit the additional data to the transceiver in advance of providing the one or more frames.

Abstract: The disclosure describes techniques for enabling an electronic device to determine whether a service set identifier (SSID) associated with a wireless access point (WAP) is associated with a captive portal. If the electronic device determines that the SSID is not associated with a captive portal, the electronic device may determine to refrain from performing a captive-portal check upon establishing a connection with the WAP. For instance, the device may refrain from sending an HTTP Get or other request to a predefined URL and awaiting a predefined response upon determining that the SSID is not associated with a captive portal. In some instances, the electronic device may make this determination by communicating with one or more other electronic devices that store information indicating whether the SSID is associated with a captive portal. In other instances, the electronic device may make this determination by reading data stored in persistent memory.

Abstract: A multitenant solver execution service provides managed infrastructure for defining and solving large-scale optimization problems. In embodiments, the service executes solver jobs on managed compute resources such as virtual machines or containers. The compute resources can be automatically scaled up or down based on client demand and are assigned to solver jobs in a serverless manner. Solver jobs can be initiated based on configured triggers. In embodiments, the service allows users to select from different types of solvers, mix different solvers in a solver job, and translate a model from one solver to another solver. In embodiments, the service provides developer interfaces to, for example, run solver experiments, recommend solver types or solver settings, and suggest model templates. The solver execution service relieves developers from having to manage infrastructure for running optimization solvers and allows developers to easily work with different types of solvers via a unified interface.

Abstract: Techniques are disclosed for generating realistic talking avatar videos from a prompt and a script. A computer system can implement a trained avatar model that includes a diffusion model that can generate a neural texture usable to generate each frame of a talking avatar video. The computer system can generate audio from the script. The script can include text of a narration to be synchronized with a talking avatar. The computer system can generate an avatar expression sequence using the audio. The computer system can receive a prompt characterizing an avatar appearance. Using the prompt and the audio, the computer system can generate the talking avatar video using the trained avatar model taking as inputs the avatar expression sequence and the text prompt. The talking avatar video can include the audio synchronized with a sequence of image frames generated using the neural texture.

Abstract: Pre-trained machine learning models may be adapted using target data sets. A target data set may be received and used to adapt the pre-trained machine learning model. Auxiliary items may be retrieved from an auxiliary data set based on a target item from the target data set. Respective labels for the auxiliary items can then be identified based on target items from the target data set. Those auxiliary items with respective labels that do not match a label of the target item can be retained and used to apply a contrastive loss function to update a backbone of the pre-trained machine learning model. A cross entropy loss function can be applied to update a classifier of the pre-trained machine learning model using the label identified for the target item.

Abstract: Techniques for logical text passage generation and retrieval for retrieval-augmented generation. The techniques involve processing markup language documents to generate logical text passages and their corresponding embeddings. These embeddings are indexed for efficient retrieval. Upon receiving a user utterance, a user query is formed and transformed into an embedding to query the index. Relevant text passages are identified and used to prompt a large language model (LLM), which generates a completion. This completion is then sent as a response to the user. The process effectively bridges user queries with relevant information through advanced embedding and natural language processing techniques, enabling accurate and contextually appropriate interactions within a user-agent dialogue framework.

Abstract: Just-in-time computation of decoder layer features is performed for an early exit decoder of a machine learning model. When generating an inference using an early exit decoder, decoder layer features that were not generated due to a prior early exit are identified. The identified decoder layer features are generated and then provided as input to a decoder layer feature during a decode step.

Abstract: A circuit breaker device is used to control flow of electrical power from a first terminal that connects to a power source to a second terminal that connects to a load. A solenoid or other type of actuator within the circuit breaker device is used to operate mechanical switches to control the flow of electrical power. When the solenoid is in a central position, electrical power is able to flow across the device. When the solenoid is moved to a first position, such as in response to a fault, the solenoid activates a trip mechanism that opens a first switch to disconnect the first terminal from the second. The solenoid may also be moved to a second position that opens a spring biased switch comprising electrical contacts, disconnecting the first terminal from the second without opening the first switch or activating the trip mechanism.

Abstract: Techniques for simulating emotional intelligence in language model-based conversational agents through situationally adaptive trait expression are described. A text representation of a recording of a verbal request of a user recorded by a device and a paralinguistic caption indicative of a paralinguistic sentiment of the recording are sent to a user sentiment classifier model, which generates an estimated actual sentiment of the user. A first set of instructions are selected based at least in part on the estimated actual sentiment of the user, and include at least one trait to emphasize in a response to the verbal request. A first prompt to generate the response to the verbal request is sent to a language model, the first prompt including the first set of instructions. The response to the verbal request is received from the language model, and the response is sent to a rendering component for output by the device.

Abstract: A system may use Wi-Fi channel state information (CSI) to configure ultrasonic presence detection (USPD) in a user device. Minor motion events in the vicinity of the user device may occur close to the noise floor of USPD such that noise transients can trigger false presence readings. In contrast, CSI can be highly sensitive to minor movements but may detect motion outside the area of interest; for example, in another room. Accordingly, the system may use CSI to confirm vacancy for purposes of tuning USPD. A first implementation of the system may use CSI to configure conditions for USPD-based state transitions in an adaptive state machine. A second implementation may use statistical methods to calculate parameters for detecting minor motion in USPD data. A third implementation may knowledge distillation techniques to transfer an embedding space from a network that processes CSI to a network that processes ultrasound data.

Abstract: The present disclosure generally relates to privacy-based client device location estimation service. An actual location of a plurality of gateways deployed in the network coverage area is obfuscated into an attributed location of the subregion, where the network coverage area is divided into multiple subregions. Individual gateway of the plurality of gateways receive the modulated signal from at least one client device and measure the strength of the signals of the received signal. The localization service implemented in the network service provider receives the measured strength of the signals and processes the received signal for individual gateway by normalizing the measured strength of the signal with respect to other gateways. The localization service can estimate the location of the client device based on the processed results and the attributed location corresponding to individual gateways.

Abstract: Systems and methods for automatically taring mobile produce scales include receiving sensor data indicating a weight change event associated with a weight sensor, and generating a tare request. In response to the tare request, a difference in weight data associated with the weight sensor before and after the weight change event may be determined. Stability data associated with the weight sensor may also be used to determine whether to automatically tare the weight sensor. If the difference between the weight data is within a threshold associated with the weight change event, the weight sensor may be automatically tared. If the difference between the weight data exceeds the threshold associated with the weight change event, the weight sensor will not be automatically tared. The weight sensor may then be re-tared manually.

Abstract: Approaches presented herein relate to the performing of tasks involving sensitive user data in a multi-tenant resource environment. In particular, various approaches provide for the allocation of resources to perform tasks, such as inferencing tasks, using user-specific data and models deployed in isolated, single tenant environments. A set of base models, code, and artifacts can be maintained and updated as appropriate. When a user requests to perform an inferencing task, these base components can be deployed to a user-specific environment for performing the inferencing task. After the inferencing task has been performed, the model and associated components can be removed from the user-specific environment. In instances where a user might want to further train or fine-tune a model using user-specific data, offline training may be performed to generate a new, user-specific artifact or set of weights, which can subsequently be applied to the base components to effectively obtain a user-specific model.

Abstract: Techniques for evaluating content based on heterogenous features (e.g., numerical and semantic features) are described. Content candidates (e.g., points of interest, events, entertainment, products, etc.) may be retrieved based on a potential relevance to a user input. A system may process the content candidate and the user input to extract numerical and semantic features (e.g., distance, rating, reviews, price, name, category, address, etc.) corresponding to the content candidate. The numerical features may be converted into semantic features (e.g., semantic representations of the numerical features). The semantic representations may be processed by a language model and the numerical features may be additionally processed using one or more encoders to generate semantic and numerical relevance values, respectively, for the content candidates. A learnable weighted sum may be applied to the relevance values to generate a ranking metric, which may be used to rank the content candidates.

Abstract: Disclosed are systems, methods, and apparatus to provide a distributed domain authority (“DA”) for management of DA certificates, client host end-entity certificates, and DA key pairs for secure communication within a provider network. In accordance with the disclosed implementations, each of a plurality of hardware security modules (“HSM”) performs the role of DA and issues end-entity certificates to individual HSMs and client hosts communicating with the HSMs. The private key of a DA key pair never leaves the HSM set in plaintext form and the DA key pair and its corresponding DA certificate are rotated periodically (e.g., daily, weekly) as part of a domain state rotation.

Abstract: An audio enabled device comprising a microphone array and loudspeaker can detect user respiration for sleep analysis if enabled by a user. The loudspeaker emits ultrasound audio signals which are captured by the microphone(s) and processed with respect to a reference signal to determine when the user's chest is moving and thus detecting the user's respiration. The device may employ a line-fitting and/or other techniques to determine respiration data from the detected movement. This information may be used, along with other movement data, acoustic data, image data, etc. to determine data regarding the user's sleep. Such information may be made available to the user through other device(s) associated with the user's profile.

Abstract: Techniques are generally described for controllable image-to-video generation. In various examples, a first image representing at least a first object may be received. First input data including a selection of the first object in the first image for animation may be received. Second input data including at least a first bounding box indicating a target location of the first object may be received. A latent diffusion text-to-image model and the first image may be used to generate a first plurality of visual tokens. One or more first grounding tokens may be generated representing a location of the first bounding box. The latent diffusion text-to-image model may be used to generate a video animating the first object based on the first plurality of visual tokens and the one or more first grounding tokens.

Abstract: An electric vehicle supply equipment (EVSE) pedestal is provided. The EVSE includes a top shroud to which an EVSE may be mounted. The shroud may also be configured to be universal and receive any number of different types of EVSEs. The EVSE pedestal includes one or more side panels that are provided at an angle with respect to a top surface of the EVSE pedestal. This results in the EVSE pedestal comprising a “trestle” or “pyramid” shape, which provides for increased structural rigidity and provides the EVSE pedestal the ability to withstand environmental conditions, such as high-velocity wind. One or more of the side panels may also be removable to, among other reasons, allow for a user to access the interior of the EVSE pedestal, which may contain some of the electrical connections associated with the EVSE.

Abstract: A content moderator which includes an artificial intelligence model is iteratively enhanced. An enhancement iteration comprises generating synthesized content units based on analysis of content units which were erroneously labeled by a first version of the moderator, and transforming the first version to a second version by training the first version on the synthesized content units. From among labeled content items generated by the second version, additional erroneously labeled content items are identified for analysis in a subsequent iteration.

Abstract: An encapsulation packet, formatted according to an encapsulation protocol, is obtained at a virtualization manager of a server from a packet processing intermediary. The virtualization manager determines that the encapsulation packet includes an encapsulated packet directed to a virtualized run-time environment which was launched at the server. In response to a determination, by the virtualization manager, that the virtualized run-time environment is unable to process the encapsulated packet, the virtualization manager transmits a message to the packet processing intermediary, indicating that the virtualized run-time environment is unable to process the encapsulated packet.

Abstract: A multi-appliance logical connection is created between a transit gateway (TGW) of a cloud computing environment and a group of network addresses of a first virtual private cloud (VPC). As addresses of the group are assigned to software-defined appliances, respective communication sessions of a routing information exchange protocol are established between the appliances and the TGW. The TGW routes packets associated with a second VPC to next-hop targets indicated in routing information received via the sessions.

Abstract: Technologies directed to forming and using a wireless mesh network (e.g., compliant with the IEEE 802.11s standard in an existing home wireless network for channel state information (CSI) sensing) are described. One method includes receiving a request to create a wireless local area network (WLAN) mesh network between wireless devices of a home wireless network. The method receives, from the devices, information about a home AP device and the other neighboring devices. The method calculates, using the information, a channel score for available channels for the WLAN mesh network. The method sends commands to the devices to form the WLAN mesh network using a channel having a highest channel score. The method receives, from the devices, channel state information (CSI) data representing channel properties of a mesh link of the WLAN mesh network.

Abstract: Techniques are generally described for dynamic band steering for wireless mesh networks. An example method includes receiving a request from a client device to connect to a wireless network, wherein connecting to the wireless network includes a band steering process and determining that the client device is present in a database of client devices. The example method further includes determining that the request from the client device was received within a pre-determined window of time, wherein the window of time is based on a timestamp of a last known activity of the client device and disabling, based on determining that the client device is present in the database and determining that the request was received within the pre-determined window of time, the band steering process. The example method also includes executing a wireless network connection routine based on the request.

Abstract: A configuration hierarchy may be used to update a storage engine configuration. A storage engine configuration manager may determine a storage engine configuration for a specific storage engine using a specified hierarchy of storage engine configurations that generate a composite of features from across different tiers of the hierarchy. A storage engine can receive the storage engine configuration and apply the configuration to update execution of the storage engine without restarting the storage engine.

Abstract: An address having a set of address bits is received to access a memory device having a plurality of memory channels. The set of address bits may include an atomic address that indexes to an atomic data transfer unit, and an address offset that indexes into an offset in the atomic data transfer unit. At least a part of the atomic address is provided as an input to a hash function that provides a one-to-one mapping of a unique output for each unique input. A memory index is derived based on an output of the hash function. A memory address formed by concatenating the memory index with the address offset is provided for accessing the memory device.

Abstract: Security checkpoint access history data and video may be analyzed to determine physical relationships between checkpoints, such as distances and access dependencies between checkpoints. The physical relationships may be used to determine remaining defense ratings for checkpoints relative to a protected asset. The remaining defense ratings may be used to prioritize security incidents and provide corresponding security alerts. The physical relationships may also be used to issue instructions for caching identity verification data in a local checkpoint data store. The checkpoint access history data may also be used to generate an access model for one or more identities and to detect an inconsistent access behavior that is inconsistent with the access model, thereby indicating a security incident may have occurred.

Abstract: An indication of a graph data set is obtained at a machine learning service. At a server identified for training a model using the graph data set, a plurality of mini-batch training iterations is conducted. A given mini-batch iteration comprises execution of a group of asynchronous tasks including a graph sampling task run at a CPU, and a model parameter learning task run at least in part at a training accelerator. At least one task of a given iteration is run in parallel with a task of another iteration. A trained version of the model is stored.

Abstract: Implementing machine learning generated item scores to guide the handling of items is described herein. The method comprising receiving, from a first user device, first data indicating that a first item and a second item are to be arranged in a first container; determining for the first item a first item score and for the second item a second item score, the first item score and the second item score being generated by a machine learning model and indicating that the first item is to be arranged in a first container prior to the second item being arranged in the first container; and causing, based at least in part on the first item score and the second item score, the first item to be arranged in the first container prior to the second item.

Abstract: Scalable recovery of a database is performed by a distributed data store. Storage nodes can monitor connection status of database access applications that have attached respective portions of a table stored at the storage nodes to provide access to the table. A connection failure of a database access application can be detected. Transactions associated with the failed database access application can be identified and determination of transactions that are not committed made. The storage nodes can remove versions of records of the table associated with the transaction as part of recovering the respective portions of the table from the connection failure of the database access application.

Abstract: The techniques described herein relate to generating a portrait aspect ratio video from a landscape aspect ratio video. The techniques can include a content editing engine identifying a first frame of a first shot of a landscape video and determining a plurality of objects in the first frame. The content editing engine can identify a primary subject of the plurality of objects based at least in part on the first frame. The content editing engine can generate a first set of bounding boxes for the primary subject. The content editing engine can generate a second set of bounding boxes for the primary subject based at least in part on the first set of bounding boxes. The second set of bounding boxes can have a portrait aspect ratio. The content editing engine can generate a portrait video based at least in part on the second set of bounding boxes.

Abstract: Techniques for a service provider network to automatically mitigate malicious activity by an internet protocol (IP) address in a cloud computing environment are discussed herein. A system can detect malicious data hosted by the IP address and determine whether or not to allow subsequent data exchanges by the IP address. In some examples, the system can validate an impact to a resource, a service, a network, or other entity of a service provider in the cloud computing environment. The system can determine whether to block the IP address based on the validated impact to the resource, the service, the network, or the entity. The system can determine an action for a particular IP address to reduce or eliminate exploitation, disruption, and/or data theft, by a threat actor hosting a botnet attack.

Abstract: In an embodiment, a request associated with a digital asset is received over a first channel. Connection information for a second channel is sent over the first channel. The first channel and the second channel are synchronized. End user input related to the digital asset and context information of the digital asset is received. The end user input is classified into at least one of a first class or a second class. An instruction that includes an indication of the first class and relevant metadata to cause a compute device to execute the instruction to be sent over the first channel and in response to the end user input being classified as the first class. A message to cause the compute device to output content of the message is sent over the second channel and in response to the end user input being classified as the second class.

Abstract: Inferences may be obtained to handle access requests at a non-relational database system. An access request may be received at a non-relational database system. The non-relational database system may determine that the access request uses a machine learning model to complete the access request. The non-relational database system may cause an inference to be generated using data items for the access request as input to the machine learning model. The access request may be completed using the generated inference.

Abstract: Phonetic similarity search may be performed for performing custom automatic speech recognition. A ranked list of possible texts for speech in audio data generated by a general speech recognition machine learning model may be obtained. Unique tokens from different texts in the ranked list may be determined. Respective vectors of phonetic features for the unique tokens may be determined. A similarity search to generate a set of candidate entities using the respective vectors of phonetic features in a phonetic feature index generated for a custom set of entities to be included in a speech recognition. Similarity scores for the candidate entities paired with the tokens may be generated and then used to determine a candidate entity to replace a token in one of the texts to provide as a text of speech in the audio data.

Abstract: A system provides an application programming interface to initiate a request to provide predictions based on a dataset. In response to the request, the system analyzes the dataset and selects, based on the analysis, one or more types of predictive models to use to make predictions. The system trains the predictive models using the dataset, and uses the predictive models to make predictions. In response to an additional dataset, the system replaces one or more of the predictive models using different model types, based on analysis of the additional dataset.

Abstract: A circuit breaker device is used to control flow of electrical power from a first terminal connected to a power source to a second terminal connected to a load. A first electromagnet in the device is used to activate a trip mechanism and open a switch to disconnect the first terminal from the second, such as in response to a fault. The fault may be detected using dedicated circuitry or using one or more sensors. Once tripped, a manual reset is required. A second electromagnet is used to open the switch a smaller distance, without activating the trip mechanism, and to subsequently close the switch, in response to control commands. A latching mechanism may be used to retain the second electromagnet in a position that causes the switch to remain open until a subsequent command is received.

Abstract: A biometric identification system uses images acquired using different modalities. A suitability prediction module (SPM) is a machine learning model trained to accept an input image and provide a predicted distance value (PDV) as output. If the PDV is less than a threshold value, the input image may be deemed suitable for recognition. The SPM is trained using training data comprising first modality images and associated distance values. The training data is based on first modality images that are of unknown suitability and second modality images associated with the same identity that are presumed suitable. The second modality images are processed to determine a set of representation data in a representation space, from which a centroid in a representation space is calculated. The first modality image is processed to determine first representation data, and a distance in the representation space is calculated between the first representation data and the centroid.

Abstract: Described is a multiple-camera system and process for detecting, tracking, and re-verifying agents within a materials handling facility. In one implementation, a plurality of feature vectors may be generated for an agent and maintained as an agent model representative of the agent. When the object being tracked as the agent is to be re-verified, feature vectors representative of the object are generated and stored as a probe agent model. Feature vectors of the probe agent model are compared with corresponding feature vectors of candidate agent models for agents located in the materials handling facility. Based on the similarity scores, the agent may be re-verified, it may be determined that identifiers used for objects tracked as representative of the agents have been flipped, and/or to determine that tracking of the object representing the agent has been dropped.

Abstract: Disclosed are systems and methods for training deep neural networks (“DNNs”) to increase the detection robustness of those DNNs to adversarial attacks through the use of non-uniform perturbations of training data that consider feature characteristics of features of the training data that are used to train the DNN. Features of training data are perturbed in a non-uniform manner based at least in part on the correlations of the different features and/or the importance of those features in detection of malicious data by the DNN.