Publication

DAIR: Data Augmented Invariant Regularization

Uncertainty and Robustness in Deep Learning Workshop at ICML


Abstract

A fundamental problem in machine learning is to learn representations that are invariant to certain transformations. For example, image representations are desired to be invariant to translation, rotation, changes in color, or background; natural language representations ought to be invariant to named entities. Naturally, data augmentations are a simple yet powerful way to address such invariance. However, such data augmentations requiring either additional data collection or careful engineering to capture all invariances. In this paper, we argue that a simple yet effective additional loss, called Data Augmented Invariant Regularization (DAIR), could improve the performance even further. DAIR promotes additional invariance on top of data augmentations at little marginal cost, and is consistent with any learning model. We empirically evaluate the performance of DAIR on two vision tasks, Colored MNIST and Rotated MNIST, and demonstrate that it provides non-trivial gains beyond data augmentation, outperforming invariant risk minimization.

Related Publications

All Publications

Interspeech - October 12, 2021

LiRA: Learning Visual Speech Representations from Audio through Self-supervision

Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Björn W. Schuller, Maja Pantic

ICML - July 18, 2021

Latency-Aware Neural Architecture Search with Multi-Objective Bayesian Optimization

David Eriksson, Pierce I-Jen Chuang, Samuel Daulton, Peng Xia, Akshat Shrivastava, Arun Babu, Shicong Zhao, Ahmed Aly, Ganesh Venkatesh, Maximilian Balandat

IEEE Transactions on Image Processing Journal - March 9, 2021

Inspirational Adversarial Image Generation

Baptiste Rozière, Morgane Rivière, Olivier Teytaud, Jérémy Rapin, Yann LeCun, Camille Couprie

ICML - July 12, 2020

Lookahead-Bounded Q-Learning

Ibrahim El Shar, Daniel Jiang

To help personalize content, tailor and measure ads, and provide a safer experience, we use cookies. By clicking or navigating the site, you agree to allow our collection of information on and off Facebook through cookies. Learn more, including about available controls: Cookies Policy