Abstract: Dimension reduction is performed for multimodal embedding searches. A request to perform a search using input data of a modality of a multimodal machine learning model is received. The input data is encoded using the multimodal machine learning model to generate a first embedding that represents the input data. A machine learning dimension reduction technique is applied to the first embedding to generate a second embedding with a reduced number of dimensions. This second embedding is then used to search an index of data objects also generated using both the multimodal machine learning model and the machine learning dimension reduction technique.

Abstract: A traffic controller (TC) circuit monitors an ingress stream being fed to an integrated circuit (IC) component comprising the TC circuit, wherein the IC component is part of a processing pipeline of a computing device. The TC circuit controls input traffic on the ingress stream to processing logic circuitry of the IC component. The TC circuit also controls output traffic on an egress stream being outputted from the IC component. The TC circuit generates an interrupt based on a set of operational metric configurations, and the ingress stream or the egress stream.

Abstract: Techniques for disconnected remote task restarts via situational distributed control are described. A container instance operates in a cluster of container instances that is remote from a control plane. Each agents of the container instances can be configured with a control module that provides purposefully limited control plane functionalities that are activated in specific scenarios. Each agent can be configured, by the control plane, with mappings indicative of other agents—and their currently-executing tasks—that the agent is to serve as a backup control plane for. Upon a failure associated with an agent coupled with a lack of network connectivity from the cluster to the control plane, another agent can become a backup control plane to restart task execution for the affected agent.

Abstract: In some disclosed embodiments, audio data corresponding to an utterance may be received and processed to determine at least a first natural language understanding (NLU) hypothesis and a second NLU hypothesis for the utterance. A determination may be made that the first NLU hypothesis is likely more likely accurate than the second NLU hypothesis, and a device may be caused to present content corresponding to the first NLU hypothesis. The device may further be caused to display a graphical user interface (GUI) element that is selectable to cause execution of an action corresponding to the second NLU hypothesis.

Abstract: Techniques for providing graphic overlays on live videos using a depth machine learning (ML) model are described. According to some examples, a computer-implemented method includes receiving a live stream of a sporting event; determining a mapping of a playing surface of the sporting event depicted in the live stream to a frame of the live stream; generating, by a machine learning model, a depth map indicating a corresponding depth for an individual player of a plurality of players depicted in the frame based at least in part on an input of the mapping and an indication of the plurality of players depicted in the frame; and inserting a graphic into the frame of the live stream based at least in part on the depth map.

Abstract: In response to registration requests, metadata of a plurality of machine learning artifacts are stored. In response to a pipeline definition request, which does not specify resources to be used for the pipeline, a representation of a pipeline comprising nodes corresponding to registered artifacts is stored. A benchmarking operation is conducted to select the types of resources to be used for the nodes, and results of the benchmarking are provided.

Abstract: Input data comprising video and associated audio is processed to generate recap data, such as video with audio that summarizes previous occurrence and entices a viewer. The input data is processed to generate dense text data that includes timestamps, dialogue, descriptions of scenes of the video, and so forth. The input data is also processed to generate summary data, such as scene, episode, and season summaries. A set of one or more machine learning (ML) systems use this as input to generate a narration script. Narration audio is generated from the narration script. The narration script is then used to generate sections of a recap such as an intro, teaser, detail, and outro. Each portion may include narration audio, clips of video, video clips of relevant dialog, background music, audio effects, and so forth. The duration of elements may be constrained based on the presentation of associated narration audio.

Abstract: Disclosed are various embodiments for dynamic deployment optimization for network functions in radio-based networks. In one embodiment, a deployment specification for a network function in a radio-based network is received. A particular node is dynamically selected from a plurality of nodes in a cloud provider network for executing the network function. The network function is deployed for execution on the particular node according to the deployment specification.

Abstract: Domain specialty instructions may be generated for performing text analysis tasks. An input text may be received for performing a text analysis task. A domain specialty may be identified for the input text. Specialty domain identifiers may be inserted as part of generating instructions to perform the text analysis task using a pre-trained large language model fine-tuned to a domain that includes multiple domain specialties. The pre-trained large language model may perform the text analysis task on the input text using the generated instructions. A result of the text analysis tsk performed on the input text may be provided.

Abstract: Systems and methods for integration of speech processing functionality with organization systems are disclosed. For example, a voice interface application may be created to enable a voice interface functionality for devices associated with an organization. Space identifiers of spaces of the organization may be created and associated with the voice interface application. Devices associated with the space identifiers may be enabled for utilizing the voice interface application and may be set up utilizing wireless network identifiers associated with the spaces and/or the organization.

Abstract: Log records are collected from resources used for various constituent services of a distributed application during an execution of the application. An aggregated event record sequence corresponding to the execution is generated from the log records, with individual event records indicating occurrences of events for which expected responses of the constituent services are indicated in correctness checking rule sets associated with the application. An action responsive to a violation of a correctness checking rule set is initiated after the violation is detected based on analysis of the event records.

Abstract: Disclosed are various embodiments for an interactive user interface for predictive simulation results that are derived from the automatic sampling of a simulation space. In some embodiments, a system is configured to receive a manifest and an input data object for executing a simulation project on a simulator. The system can generate sample inputs for a sample space of the simulation project based on a sampling space criteria for the simulation project. A simulation of the simulation project is executed to generate simulation outputs based on the input data object. A machine learning model can be trained based at least in part on simulation outputs. The system can be configured to generate a user interface that displays a predictive simulator output for the simulation project based on querying the machine learning model with an unsimulated input.

Abstract: Devices and techniques are generally described for compression of machine learning models. In some examples, a first weight value wi of a machine learning model is determined. The first weight value wi may be associated with a first region comprising an interval of numbers. A first loss for the first weight value wi may be determined using a first loss function comprising a weighting term comprising a sinusoid function with an argument of ??rwi. ? may be selected such that a maximum value of the weighting term for the first region is of a format that is compatible with first machine learning accelerator. A gradient of the first loss function may be determined. A second weight value wi may be determined using the gradient and the first loss. The second weight value wi may be stored in a first memory.

Abstract: An apparatus for use at an inventory location stowing items comprises an array of sensor elements. Each sensor element comprises a plurality of conductive elements that may be driven as a self-capacitance sensor. A set of one or more sensor elements may be associated with a particular lane. Based on the changes in self-capacitance values and the association, the lane a user interacted with may be determined. The determination of the set of sensor elements associated with a particular lane may be determined during a setup process. During operation, the conductive elements may be driven to compensate for changes in the self-capacitance resulting from temperature changes resulting from changes in ambient temperature or the items. Compensated capacitance data may be used to determine interaction data, count of items in the lane, and so forth. Sensor elements at one or both ends of a lane may be used to disambiguate interactions.

Abstract: A system for storing data includes a rack, one or more data storage modules coupled to the rack, and one or more data control modules coupled to the rack. The data storage modules may include a chassis, two or more backplanes coupled to the chassis, and one or more mass storage devices (for example, hard disk drives) coupled to the backplanes. The data control modules may access the mass storage devices in the data storage modules.

Abstract: Techniques for utilizing a cloud-side link module and a provider substrate extension (PSE) link module to secure a communications channel between a cloud provider network and a provider substrate extension are described. A PSE link module receives a request originated by a compute instance in the PSE that is destined to a destination within the cloud provider network. The request was encrypted using an encryption scheme that encrypts traffic of a virtual private cloud that the compute instance operates in. The PSE link module decrypts the first request, encrypts it using a separate encryption scheme, and transmits it via a secure tunnel to a second link module in the cloud provider network. The second link module can decrypt the encrypted traffic and cause it to be validated before it is passed on via the cloud provider's substrate network to be processed.

Abstract: Devices, systems, and methods for routing network requests using signature-based authentication may include receiving, by a load balancer, a network request including a destination uniform resource indicator (URI) and a signature header; verifying, by the load balancer, a certificate for a domain name of the destination URI; initiating, by the load balancer, a serverless instance based on the verifying; allocating, by the serverless instance, a region for the network request; copying, by the serverless instance, the signature header of the network request; retrieving, by the serverless instance, a first address for an intended destination address of the network request; resolving, by the serverless instance, the destination URI to the intended destination address, wherein the intended destination address is different than the first address; and sending, by the serverless instance, content of the network request and the signature header to the intended destination address without modifying the signature header

Abstract: Systems and methods are disclosed for automated container stacking stations. In one embodiment, an example system may include a housing having a first side and a second side, a lifting component configured to lift a first container from a first position to a second position, and to lower the first container from the second position to a third position, such that the first container is stacked on a second container, and a container centering component having a first container engagement member and a second container engagement member, the container centering component configured to align the first container in a predetermined position.

Abstract: Database recovery of independently replicated shards may be performed. A recovery event to make a mirror location of shards of a database that is replicated independently and asynchronously from a primary location of the database is detected. To make the mirror location a new read-write location of the database according to the detected recovery event, a recovery point may be identified based on a gossip protocol that shares progress across the different shards in the mirror location that is used to access the database as the new read-write location.