Amazon Technologies, Inc.

Roshani Nagmote

Hagay Lupesko

Sandeep Krishnamurthy

Yao Wang

Vikram Madan

Naveen Mysore Nagendra Swamy

Lai Wei

Jonathan Esterhazy

Ankit Khedia

Vandana Kannan

Anirudh Acharya

Kalyanee Shriram Chendke

Dominic Rajeev Divakaruni

Jiajie Chen

Cheng-Che Lee

Lee Cheng-Che

Ruofei Yu

Techniques are described automatically determining runtime configurations used to execute recurrent neural networks (RNNs) for training or inference. One such configuration involves determining whether to execute an RNN in a looped, or 'rolled,' execution pattern or in a non-looped, or 'unrolled,' execution pattern. Execution of an RNN using a rolled execution pattern generally consumes less memory resources than execution using an unrolled execution pattern, whereas execution of an RNN using an unrolled execution pattern typically executes faster. The configuration choice thus involves a time-memory tradeoff that can significantly affect the performance of the RNN execution. This determination is made automatically by a machine learning (ML) runtime by analyzing various factors such as, for example, a type of RNN being executed, the network structure of the RNN, characteristics of the input data to the RNN, an amount of computing resources available, and so forth.

Automatically determining configurations for executing recurrent neural networks

Features related to systems and methods for automated generation of a machine learning model based in part on a pretrained model are described. The pretrained model is used as a starting point to augment and retrain according to client specifications. The identification of an appropriate pretrained model is based on the client specifications such as model inputs, model outputs, and similarities between the data used to train the models.

Machine learning services with pre-trained models

Techniques for model adaptation are described. For example, a method of receiving a call to provide either a model variant or a model variant profile of a deep learning model, the call including desired performance of the deep learning model, a deep learning model identifier, and current edge device characteristics; comparing the received current edge device characteristics to available model variants and profiles based on the desired performance of the deep learning model to generate or select a model variant or profile, the available model variants and profiles determined by the model identifier; and sending the generated or selected model variant or profile to the edge device to use in inference is detailed.

Model adaptation

Techniques for benchmarking a machine learning model/algorithm are described. For example, in some instances a method includes generating an execution plan for benchmarking of at least one task corresponding to a machine learning model based on an identified machine learning model, identified training data, and at least one objective for the benchmarking job; receiving execution statistics about the execution of the task as a part of the benchmarking job according to the execution plan; and updating the execution plan based at least in part on the received execution statistics of the task.

Benchmarking machine learning models via performance feedback

Growing community of inventors

San Mateo, CA, United States of America

Roshani Nagmote