Apple Inc.

Ashish Shrivastava

Cuneyt Oncel Tuzel

Russell Y Webb

Tomas Jon Pfister

Shahab Kaynama

Joshua M Susskind

Seyed Moosavi-Dezfooli

A method includes obtaining a time-series of training samples that include one or more states, a ground truth value, an output value produced in the presence of the one or more states, and an actual error value that is defined as a difference between the ground truth value and the output value. The method also includes training a machine learning model using the time-series of training samples such that the machine learning model is configured to determine a condition-dependent error distribution for a current time step based on simulated states for the current time step.

Conditional error models

A generative network may be learned in an adversarial setting with a goal of modifying synthetic data such that a discriminative network may not be able to reliably tell the difference between refined synthetic data and real data. The generative network and discriminative network may work together to learn how to produce more realistic synthetic data with reduced computational cost. The generative network may iteratively learn a function that synthetic data with a goal of generating refined synthetic data that is more difficult for the discriminative network to differentiate from real data, while the discriminative network may be configured to iteratively learn a function that classifies data as either synthetic or real. Over multiple iterations, the generative network may learn to refine the synthetic data to produce refined synthetic data on which other machine learning models may be trained.

Generating more realistic synthetic data with adversarial nets

A neural network is trained to defend against adversarial attacks, such as by preparing an input image for classification by a neural network where the input image includes a noise-based perturbation. The input image is divided into source patches. Replacement patches are selected for the source patches by searching a patch library for candidate patches available for replacing ones of those source patches, such as based on sizes of those source patches. A denoised image reconstructed from a number of replacement patches is then output to the neural network for classification. The denoised image may be produced based on reconstruction errors determined for individual candidate patches identified from the patch library. Alternatively, the denoised image may be selected from amongst a number of candidate denoised images. A set of training images is used to construct the patch library, such as based on salient data within patches of those training images.

Growing community of inventors

Santa Clara, CA, United States of America

Ashish Shrivastava