Amazon Technologies, Inc.

Calvin Yue-Ren Kuo

Jonathan I Turow

James Christopher Sorenson

Shyam Krishnamoorthy

Alan Conrad Rawcliffe

Mark Edward Rafn

Cheng Pan

Marc John Brooker

Marco Argenti

Weili Zhong McClenahan

William Alexander Stevenson

Kyunghwan Choi

Sunil Mallya Kasaragod

Aran Khanna

Yen-Lin Chen

Jeffrey Tarn

Ambuj Tandon

Eric Jason Brandwine

Samuel John Young

Christopher Andrew Stephens

Rudresh Amin

Chen Yang

Salman Aftab Paracha

Feng Wang

Won Suk Yoo

Dmytro Taran

Shiwei Ni

Naveen Kumar Aitha

Ishwardutt Parulkar

Tristam Kyle MacDonald

Spencer Lyle Voorheis

Jiazhen Chen

Aditya Bhalla

Erik Jacob Sipsma

Robert Battle

Albert Pitera

Mauricio Da Silva Miranda

Yadunandan Parthasarathy

Ankit Masrani

James William Bender

Christoph Saalfeld

Dan Catalin Teodorescu

Sudarshan Anand Chitre

Alexandra Elizabeth Baoboe Lee

Himanshu Jindal

Kadirvelu Rathnavelu

Zheshen Wang

Abdullah Mohammed Jibaly

Therapon Skoteiniotis

Nicolas Webb

Justin Robert Knoepfler

Bryant Cutler

Stuart Olmstead-Wilcox

Harout Sebouh Hedeshian

Georgie Mathews

Simon Poile

Tian Chen

Robert P Cochran

Charles Lawrence Sismondo

Jimi Shah

Dilip Iyer

Xiao Yu

Igor Vodov

Nathan James Schulz

Gabriel Badescu

Jedd Weise

Victor Gails

Shirley Zheng

Jingyu Dong

Jingwei Sun

Haiyang Liu

Tiangang Song

Zhaofeng Zhan

Reda Bouaichi

Frank Miley

Liusu Zeng

Paul Oesterle Malmsten

Shyam Krisnamoorthy

Chung-Ming Chan

Jonathan Allen Jamison

A provider network implements a machine learning deployment service for generating and deploying packages to implement machine learning at connected devices. The service may receive from a client an indication of an inference application, a machine learning framework to be used by the inference application, a machine learning model to be used by the inference application, and an edge device to run the inference application. The service may then generate a package based on the inference application, the machine learning framework, the machine learning model, and a hardware platform of the edge device. To generate the package, the service may optimize the model based on the hardware platform of the edge device and/or the machine learning framework. The service may then deploy the package to the edge device. The edge device then installs the inference application and performs actions based on inference data generated by the machine learning model.

Generating and deploying packages for machine learning at edge devices

A connected device at a client network implements a local data classification service for classifying data based on a data classification service of a remote provider network. The local data classification service receives a request to classify data at one or more data sources of the client network. The request is initiated from a client device of the client network according to a management interface for a data classification service of a remote provider network (e.g., using the same API request used by the remote classification service). The local data classification service obtains at least some of the data from the one or more data sources of the client network. The local data classification service classifies the obtained data according to different types of sensitivity using the data classification engine in the execution environment without the data being exposed outside of a data isolation boundary of the client network.

Local data classification based on a remote service interface

Edge devices of a network collect data. An edge device may determine whether to process the data using a local data processing model or to send the data to a tier device. The tier device may receive the data from the edge device and determine whether to process the data using a higher tier data processing model of the tier device. If the tier device determines to process the data, then the tier device processes the data using the higher tier data processing model, generates a result based on the processing, and sends the result to an endpoint (e.g., back to the edge device, to another tier device, or to a control device). If the tier device determines not to process the data, then the tier device may send the data on to another tier device for processing by another higher tier model.

Model tiering for IoT device clusters

Systems and methods are provided for detecting discrepancies in information that is displayed to a user across different webpages of an online shopping system. User interaction with the online shopping system is simulated in one example to collect data relating to webpages of the online shopping system. The data is analyzed to identify relevant semantic blocks of data and compare the data in those blocks across webpages. In an additional example, source code for the webpages is analyzed using machine-learning systems, which are targeted based on the results of the user interaction simulation to identify problematic source code and correct associated issues.

Detecting and correcting user interface discrepancies

A computing environment is disclosed that receives from devices requests directed toward services accessible in the environment, and that forwards communications from services in the environment to devices registered with the environment. During a registration process at the environment, devices are assigned a device identifier that is used to identify and authenticate each particular device and requests communicated from and to the device via the environment. The computing environment maintains state information for each device that has been registered with the system. As the device interacts with the system, the state information is updated to reflect the changes in the device. When requests to perform functions are received from devices, the computing environment determines for the particular device and the particular function requested what processing needs to be performed by the environment in response to the request.

Device state management

A technology is described for a device shadowing service that manages device states. An example method may include receiving a state change message associated with a device that a subcomponent of the device has assumed a first state. In response to the state change message, a subcomponent representation associated with a device representation may be updated to indicate the first state received, where the device representation represents the device and the subcomponent representation represent the subcomponent of the device. Thereafter, a request may be received for the first state of the subcomponent of the device. The first state indicated by the subcomponent representation associated with the device representation may be determined, and the first state may be returned in response to the request.

Device subcomponent representations

A provider network implements a service for receiving a request for a deployment from an IoT device and determining whether the IoT device is compatible with a deployment. The service may determine a configuration of the IoT device based on configuration information included in the request for the deployment. The service may then determine whether the configuration of the connected device satisfies configuration requirements for one or more components (e.g., functions) required for the deployment. In some cases, a particular version of a plurality of versions of a function may be compatible with the connected device. If the service determines that the IoT device is compatible with all of the components that are required for the deployment, then the service may generate a package that includes all of the components. The service may then deploy the package to the IoT device.

Determining IoT device compatibility with deployments

Technology is described for a checkpoint service for code function checkpoints associated with instances of code functions executed by a computing hub in a local area network. Subsequent to the execution of a code function being suspended by the computing hub, the computing hub may generate a code function checkpoint describing the execution of the code function. The computing hub may persist the code function checkpoint to storage using a checkpoint service. The computing hub may create the code function checkpoint in response to monitoring metrics associated with the computing hub, such as processor overutilization. The computing hub may resume the execution of the code function at a later time using the code function checkpoint retrieved from the checkpoint service.

Code function checkpoint and restore

Systems and methods are described for implementing a coordinator within a coordinated environment, which environment includes set of coordinated devices managed by the coordinator. The coordinator can be provisioned with a set of tasks, each task corresponding to a segment of code that can be executed by the coordinator, such as to manage the coordinated devices. The coordinator can further be provisioned with event flow information designating a routing of inputs to the coordinator computing device to destinations, such as task executions or coordinated devices. On receiving input, the coordinator can reference the event flow information to pass the input to an appropriate destination.

Localized device coordinator with mutable routing information

A client may send to a provider network a request to classify data at one or more data sources of the client network. The provider network receives the request and transmits the request to a local instance of a network-based data classification service at the client network. The local instance of the network-based data classification service classifies the data at the one or more data sources. The data is not exposed outside of a data isolation boundary associated with the client network during classification of the data by the local instance of the network-based data classification service. The provider network may initially provision the local instance of the network-based data classification service to run on the client network.

Processing requests at a remote service to implement local data classification

A hub device of a network receives data from edge devices and generates a local result. The hub device also sends the data to a remote provider network and receives a result from the remote provider network, wherein the result is based on the data received from the edge devices. The hub device then generates a response based on the local result or the received result. The hub device may determine to correct the local result based on the result received from the remote provider network, and generate the response based on the corrected result. The hub device may generate an initial response before receiving the result from the provider network. For example, the hub device may determine that the confidence level for the local result is above the threshold level and in response, generate the initial response based on the local result.

Split predictions for IoT devices

Systems and methods are described for remotely configuring a coordinator within a coordinated environment, which coordinator can execute code to manage operation of a set of coordinated devices. A client can submit configuration information to a deployment system, including permissions indicating what data resources each set of code should have access to on the coordinator. The deployment system can remotely, and independently of configuration of the coordinator and in a manner that does not conflict with local configuration of the coordinator, determine filesystem permissions that divide access to the data resources among the sets of code. The deployment system can build a directory structure with the permissions and deploy the structure to the coordinator. The coordinator can locally generate directory structure, adopt the permissions, and execute the code to enforce the client-specified permissions.

Remote filesystem permissions management for on-demand code execution

A technology is described for distributed querying of computing hubs for device-generated data. In one example, computing hubs may provide a storage service to devices included in a local device network, and the devices may send device-generated data to the computing hubs to be stored by the storage service. A distributed query may be initiated by identifying the computing hubs that have the device-generated data stored on the computing hubs and sending a message containing query instructions to the computing hubs. The query instructions, when executed on the computing hubs, retrieve the device-generated data from the storage service included on the computing hubs, and the computing hubs may send the device-generated data to a service provider environment.

Distributed querying of computing hubs

Technology is described for providing machine learning (ML) models. A plurality of candidate ML models that are derived from a primary ML model may be generated in a service provider environment. The primary ML model may be associated with a set of parameters and a candidate ML model in the plurality of candidate ML models may be associated with a subset of the parameters associated with the primary ML model. The plurality of candidate ML models may be run against validation data to evaluate performance criteria for the candidate ML models. A performance representation of the candidate ML models with respect to performance results for the candidate ML models may be provided. An ML model may be selected from the performance representation based on the performance results for the candidate ML models.

Transformation of machine learning models for computing hubs

Growing community of inventors

Mercer Island, WA, United States of America

Calvin Yue-Ren Kuo