Amazon Technologies, Inc.

Fedor Zhdanov

Siddharth Vivek Joshi

Stefano Stefani

Warren Vincent Barkley

Pietro Perona

Sindhu Chejerla

Prateek Sharma

Rahul Sharma

Gurumurthy Swaminathan

Jonathan Hedley

Xiong Zhou

Ragav Venkatesan

James Andrew Trenton Lipscomb

Emanuele Coviello

Anuj Gupta

Benjamin Alexei London

Kari E J Torkkola

Kriti Bharti

Jon I Turow

Jonathan Thomas Greenlee

Alisa V Shinkorenko

Krzysztof Chalupka

Pranav Sachdeva

Sankalp Srivastava

Techniques for zero-shot and few-shot transfer of domain-adapted base networks are described. Multiple machine learning task layers are trained using a shared base feature extractor network. At least one task layer is trained with samples and corresponding labels from a first domain as well as a second domain. At least one other task layer is trained with samples and corresponding labels from only the first domain. Ultimately, the other task layer(s) are adapted to generate labels for the first domain via the base network being weighted based on all trainings.

Zero-shot transfer of domain-adapted base networks

At an artificial intelligence system, training iterations of a first machine learning model are implemented. In a particular iteration, a group of data items are selected from an item collection using active learning, and respective labels selected from a set of tags are obtained for at least some of the items of the group. Using feature processing elements of a different machine learning model, a respective feature set corresponding to individual labeled items is generated in the iteration, and the feature sets are included in a training set used to train the first machine learning model. A trained version of the first machine learning model is stored after a training completion criterion is met.

Artificial intelligence system for media item classification using transfer learning and active learning

A request is received associated with reviewing content. As part of the request, one or more conditions are received and the content is analyzed to identify a first field of interest and a second field of interest. The first field of interest and the second field of interest represent fields of interest associated with the review of the content. At least one of the first field of interest or the second field of interest may not satisfy the one or more conditions and the content, or a portion thereof, may be sent for review.

Determining content to present for human review

A request associated with reviewing content for a field of interest is received. A confidence is determined associated with the content including the field of interest. A machine learning (ML) model determines a first confidence associated with the content includes the field of interest. The field of interest is transmitted for review in instances where the first confidence is less than a confidence threshold. After review, an indication associated with a reviewer reviewing the content and the first confidence associated with the ML model identifying the field of interest is updated to a second confidence.

Adjusting confidence thresholds based on review and ML outputs

Techniques for active learning-based data labeling are described. An active learning-based data labeling service enables a user to build and manage large, high accuracy datasets for use in various machine learning systems. Machine learning may be used to automate annotation and management of the datasets, increasing efficiency of labeling tasks and reducing the time required to perform labeling. Embodiments utilize active learning techniques to reduce the amount of a dataset that requires manual labeling. As subsets of the dataset are labeled, this label data is used to train a model which can then identify additional objects in the dataset without manual intervention. The process may continue iteratively until the model converges. This enables a dataset to be labeled without requiring each item in the dataset to be individually and manually labeled by human labelers.

Active learning loop-based data labeling service

Methods and systems for forecasting demand are described. A method may include determining a demand pattern for each respective item of at least some items of a plurality of items. The method may also include clustering the plurality of items into a plurality of clusters based on the determined demand patterns. The method may further include determining a composite demand pattern for each cluster. The method may additionally include forecasting a demand for an item of the plurality of items based on the composite demand pattern of the cluster to which the item belongs.

Cluster-based demand forecasting procedure

Growing community of inventors

Seattle, WA, United States of America

Fedor Zhdanov