Towards Context-Agnostic Learning Using Synthetic Data

TL;DR

This paper introduces a new learning framework that leverages synthetic data to achieve robust, context-agnostic classifiers capable of generalizing well to real-world images, even when trained on minimal synthetic examples.

Contribution

The paper develops a novel risk bound for context-agnostic learning and proposes an algorithm that minimizes bias by sampling independently from data sets, enabling effective training with synthetic data.

Findings

Achieves robust classification performance on standard benchmarks using synthetic data.

Demonstrates good generalization to real-world domains with synthetic training.

Outperforms direct real-world training in terms of robustness to background perturbations.

Abstract

We propose a novel setting for learning, where the input domain is the image of a map defined on the product of two sets, one of which completely determines the labels. We derive a new risk bound for this setting that decomposes into a bias and an error term, and exhibits a surprisingly weak dependence on the true labels. Inspired by these results, we present an algorithm aimed at minimizing the bias term by exploiting the ability to sample from each set independently. We apply our setting to visual classification tasks, where our approach enables us to train classifiers on datasets that consist entirely of a single synthetic example of each class. On several standard benchmarks for real-world image classification, we achieve robust performance in the context-agnostic setting, with good generalization to real world domains, whereas training directly on real world data without our techniques yields classifiers that are brittle to perturbations of the background.