Amazon Technologies, Inc.

A9.com, Inc.

Avinash Aghoram Ravichandran

Rahul Bhotika

Rakesh Madhavan Nambiar

Oleg Rybakov

Stefano Soatto

Gurumurthy Swaminathan

Matias Omar Gregorio Benitez

Leo Parker Dirac

Ranju Das

Barath Balasubramanian

Prakash Krishnan

Junxiong Jia

Joaquin Zepeda Salvatierra

Scott Daniel Helmer

Niels Brouwers

Anant Patel

Shaun Ryan James McDowell

Anshul Kumar Jain

Anushri Mainthia

Daniel Bibireata

Pietro Perona

Marshall Friend Tappen

Paulo Ricardo Dos Santos Mendonca

Ajay Kumar Mishra

Hakan Boyraz

Baoyuan Liu

Wei Zhang

Hao Yang

Michael Quang Thai Lam

Zhizhong Li

Tz-Ying Wu

Pashmeen Mistry

Shaun Ryan James Mcdowell

Techniques for class-incremental learning with pretrained machine learning models are described. According to some examples, a pretrained machine learning model that is capable of recognizing objects belonging to a set of object classes is extended to add capability for recognizing additional object classes, without significantly reducing the accuracy of the model for recognizing objects belonging to any of the object classes in the original set of object classes (aka catastrophic forgetting).

Class-incremental learning with pretrained machine learning models

Techniques for feedback-based training may include selecting a scoring machine learning model based at least in part on a test metric, and applying the model on an unlabeled dataset to generate, per dataset item of the unlabeled dataset, a prediction and an importance ranking score for the prediction. Techniques for feedback-based training may further include selecting, based on the importance ranking scores, a result of the application of the model on the unlabeled dataset, providing the result and requesting feedback on the result via a graphical user interface, receiving the feedback via the graphical user interface, adding data from the unlabeled dataset into a training dataset when the feedback indicates a verified result, and retraining the model using the training dataset with the data added from the unlabeled dataset to generate a retrained model.

Feedback-based training for anomaly detection

Techniques for anomaly detection are described. An exemplary method includes receiving one or more requests to train an anomaly detection machine learning model using feedback-based training, the request to indicate one or more of a type of analysis to perform, a model selection indication, and a configuration for a training dataset; training the anomaly detection machine learning model according to the one or more requests using the training data; performing feedback-based training on the trained anomaly detection machine learning model; and using the retrained anomaly detection machine learning model.

Anomaly detection using feedback training

A first set of execution platforms is deployed for a set of operations of a training phase of a machine learning model. Prior to the completion of the training phase, a triggering condition for deployment of a different set of execution platforms is detected. The different set of execution platforms is deployed for a subsequent set of training phase operations.

Dynamically scaled training fleets for machine learning

This disclosure describes automatically selecting and training one or more models for image recognition based upon training and testing (validation) data provided by a user. A service provider network includes a recognition service that may use models to process images and videos to recognize objects in the images and videos, features on the objects in the images and videos, and/or locate objects in the images and videos. The service provider network also includes a model selection and training service that may select one or more modeling techniques based on the objectives of the user and/or the amount of data provided by the user. Based on the selected modeling technique, the model selection and training service selects and trains one or more models for use by the recognition service to process images and videos using the training data. The trained model may be tested and validated using the testing data.

Automated model selection for network-based image recognition service

Techniques for training an embedding using a limited training set are described. In some examples, the embedding is trained by generating a plurality of vectors from a random sample of the limited set of training data classes using a layer of the particular machine learning classification model, randomly selecting samples from the plurality of vectors into a set of samples, computing at least one distance for each sampled class from a center parameter for the class using the set of samples, generating a discrete probability distribution over the classes for a query point based on distances to a center parameter for each of the classes in the embedding space, calculating a loss value for the modified prototypical network, the calculation of the loss value being for a fixed geometry of the embedding space and including a measure of the difference between distributions, and back propagating.

Prototypical network algorithms for few-shot learning

Embodiments described herein are directed to allowing manipulation of visual attributes of a query image while preserving the visual attributes of a query image. A query image can be received and analyzed using a trained network to determine a set of items whose images demonstrate visual similarity to the query image across a plurality of visual attributes. Visual attributes of the query image may be manipulated to allow a user to search for items that incorporate the desired manipulated visual attributes while preserving the visual attributes of the query image. Content for at least a determined number of highest ranked, or most similar, items related to the modified visual attributes can then be provided.

Visual similarity and attribute manipulation using deep neural networks

Various embodiments provide for visual similarity based search techniques that certain desirable visual attributes of one or more items to search for items having similar visual attributes. In order to create an electronic catalog of items that is searchable by parts-based visual attributes, the visual attributes are identified and corresponding feature vectors are extracted from the image data of each item. Thus, feature values of parts-based visual attributes of items in the electronic catalog can be determined and used to select or rank the items in response to a search query based on desirable visual attributes. To conduct a search, a user may define desirable visual attributes of one or more items. The feature vectors of the desirable visual attributes are determined and used to query the electronic catalog of items, in which items having visual attributes of similar feature vectors are selected and returned as search results.

Parts-based visual similarity search

Approaches introduce a pre-processing and post-processing framework to a neural network-based approach to identify items represented in an image. For example, a classifier that is trained on several categories can be provided. An image that includes a representation of an item of interest is obtained. Rotated versions of the image are generated and each of a subset of the rotated images is analyzed to determine a probability that a respective image includes an instance of a particular category. The probabilities can be used to determine a probability distribution of output category data, and the data can be analyzed to select an image of the rotated versions of the image. Thereafter, a categorization tree can then be utilized, whereby for the item of interest represented the image, the category of the item can be determined. The determined category can be provided to an item retrieval algorithm to determine primary content for the item of interest. This information also can be used to determine recommendations, advertising, or other supplemental content, within a specific category, to be displayed with the primary content.

Object recognition

Image classification and related imaging tasks performed using machine learning tools may be accelerated by using one or more of such tools to associate an image with a cluster of such labels or categories, and then to select one of the labels or categories of the cluster as associated with the image. The clusters of labels or categories may comprise labels that are mutually confused for one another, e.g., two or more labels or categories that have been identified as associated with a single image. By defining clusters of labels or categories, and configuring a machine learning tool to associate an image with one of the clusters, processes for identifying labels or categories associated with images may be accelerated because computations associated with labels or categories not included in the cluster may be omitted.

Cluster-trained machine learning for image processing

A computing device can be configured to analyze information, such as frames captured in a video by a camera in the computing device, to determine locations of objects in captured frames using a scene-based tracking approach without individually having to track the identified objects across the captured frames. The computing device can track scenes, a global planar surface, across newly captured frames and the changes to (or transformation) the scene can be used to determine updated locations for objects that were identified in previously captured frames. Changes to the scene between frames can be measured using various techniques for estimating homographies. An updated location for the particular object in the currently captured frame can be determined by adjusting the location of the object, as determined in the previously captured frame, with respect to the transformation of the scene between the previously captured frame and the currently captured frame.

Approaches for scene-based object tracking

An object recognition system may recognize an object in a query image by matching the image to one or more images in a database. The database may include images corresponding to multiple viewpoints of a particular device. Key points of the query image are compared to key points in the database images. Database images with many overlapping key points to the query image are selected as potential matches. The geometry of objects in the potential matches is verified to the geometry of the object in the query image to determine if the overlapping key points have a similar geographic relationship to each other across images. Objects in geometrically verified database images may be selected as potentially matching objects to the object in the query image. When a potential matching image is found, the system may confirm the match by performing matching with a second image of the object.

Growing community of inventors

Seattle, WA, United States of America

Avinash Aghoram Ravichandran