Amazon Technologies, Inc.

Google Inc.

Mu Li

Alexander Johannes Smola

Yida Wang

Zhi Chen

Leyuan Wang

Eugene Jon Shekita

Edo Liberty

Hao Yu

Yizhi Liu

George Karypis

Trishul Amit Madhukar Chilimbi

Shuai Zheng

Bor-Yiing Su

Yao Wang

Amr Ahmed

Zhen Zhang

Xingjian Shi

Tatsuya Arai

Abhi Vinayaka Sharma

Ryan Matthew Brand

Yohei Nakayama

Nick Dean Erickson

Hang Zhang

Justin Chiu

Haichen Shen

Jonas Mueller

Tuning result records indicating tuned schedules for machine learning tuning tasks are stored at a data store accessed from a tuning service. A given schedule indicates at least an order in which sub-operations of a tuning task are to be executed. In response to determining that the data store does not include a result record whose tuning task meets a similarity criterion to a tuning task determined from a tuning request, performance predictions for a set of candidate schedules are obtained using a performance prediction model, without running code which executes the candidate schedules. In a response to the tuning request, executable code corresponding to a preferred schedule identified from the candidate schedules using the performance predictions is included.

Tuned executable code service for machine learning models

A resource set which includes multiple servers with a respective plurality of training computing devices is identified for training a machine learning model. The resource set is subdivided into partition groups, such that each partition group can store a respective replica of state information of the model. The model is trained using the partition groups. The training comprises a multi-stage gathering of a portion of the state information at training computing devices of a particular partition group. Different types of communication channels between training computing devices are used in respective stages of the gathering, including inter-server communication channels in one stage and an intra-server communication channel during another stage. A trained version of the model is stored.

Distributed training of machine learning models

Techniques for code-free automated machine learning (ML) are described. Users can train high-quality ML models and pipelines without necessarily needing to write code by providing a training dataset to a code-free machine learning service. The service may deploy an ML orchestration function and a storage location on behalf of a user. When a modification is made to the storage bucket, such as by the user providing a training dataset, the orchestration function is invoked and can automatically initiate an AutoML process using at least the training data to train multiple ML model variants. The resultant ML model(s) and associated metrics can be provided to the user, deployed behind an endpoint, and/or used to generate inferences.

Code-free automated machine learning

Techniques for optimizing and deploying convolutional neural network (CNN) machine learning models for inference using integrated graphics processing units are described. A model compilation system optimizes CNN models using optimized vision-specific operators as well as both graph-level tuning and tensor-level tuning to explore the optimization space for achieving heightened performance. The model compilation system may also implement a heuristic-based two-stage technique for falling back certain operators of CNN models to use CPUs when needed or otherwise beneficial.

Unified optimization for convolutional neural network model inference on integrated graphics processing units

Techniques for determining and utilizing optimal aggregation schedules are described are described. A deep machine learning model can be trained using multiple processing elements implemented in one or multiple computing devices and that are interconnected using one or multiple types of links. An optimal aggregation schedule for such arbitrary topologies can be determined automatically. The determination may include solving a linear program on the spanning tree polytope. The optimal aggregation schedule can be utilized by the multiple processing elements to train the deep machine learning model.

Optimal message scheduling for aggregation

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for handling faults in a distributed key-value storage system. One of the methods includes receiving an indication that a machine storing a primary replica of a first replication chain is inactive, in response to receiving the indication, generating a concatenated replica comprising a first replica of the first replication chain and a second replica of a second replication chain, the second replication chain comprising replicas of a second key segment, the second key segment being adjacent to the first key segment in the multiple key segments of the plurality of keys, and providing, to another machine in the ordered sequence of machines, a notification of availability of the concatenated replica.

Fault tolerant distributed key-value storage

Growing community of inventors

Union City, CA, United States of America

Mu Li