Achieving Robustness in the Wild via Adversarial Mixing with Disentangled Representations

Abstract

Recent research has made the surprising finding that state-of-the-art deep learning models sometimes fail to generalize to small variations of the input. Adversarial training has been shown to be an effective approach to overcome this problem. However, its application has been limited to enforcing invariance to analytically defined transformations like p -norm bounded perturbations. Such perturbations do not necessarily cover plausible real-world variations that preserve the semantics of the input (such as a change in lighting conditions). In this paper, we propose a novel approach to express and formalize robustness to these kinds of real-world transformations of the input. The two key ideas underlying our formulation are (1) leveraging disentangled representations of the input to define different factors of variations, and (2) generating new input images by adversarially composing the representations of different images. We use a StyleGAN model to demonstrate the efficacy of this framework. Specifically, we leverage the disentangled latent representations computed by a StyleGAN model to generate perturbations of an image that are similar to real-world variations (like adding make-up, or changing the skin-tone of a person) and train models to be invariant to these perturbations. Extensive experiments show that our method improves generalization and reduces the effect of spurious correlations.

Out-of-Distribution Generalization via Risk Extrapolation (REx)

“…Improving the generalization of deep learning models has become a major research topic, with many different threads of research including Bayesian deep learning (Neal, 1996;Gal, 2016), adversarial (Engstrom et al, 2019;Jacobsen et al, 2018) and non-adversarial (Hendrycks & Dietterich, 2019;Yin et al, 2019) robustness, causality (Arjovsky et al, 2019), and other works aimed at distinguishing statistical features from semantic features (Gowal et al, 2019;Geirhos et al, 2018). While neural networks often exhibit superhuman generalization performance on the training distribution, they can be extremely sensitive to minute changes in distribution (Su et al, 2019;Engstrom et al, 2017; In this work, we consider out-of-distribution (OoD) generalization, where a model must generalize to new distributions at test time without seeing any training data from them.…”

Extrapolatable Relational Reasoning With Comparators in Low-Dimensional Manifolds

“…However, the projection of test data tends to escape from this sub-manifold into the region where the comparator function is never trained on, resulting in incorrect prediction. [44,8], adversarially learn representations that are domain-invariant [28,1], disentangling representations to separate functional variables with spurious correlations [19,15], and constructing models with innate causal inference graphs to reduce dependence on spurious correlations [3,7]. Our work aligns more with the line of works on discovering inductive-bias that improves generalisation.…”

Learning perturbation sets for robust machine learning

“…Generative modeling and adversarial robustness Relevant to our work is that which combines aspects of generative modeling with adversarial examples [Xiao et al, 2018a, Sharif et al, 2019, Bhattad et al, 2020, most of which either use p perturbations, run user studies to define the perturbation set, or simply do not restrict the adversary at all. Gowal et al [2019] trained a StyleGAN to disentangle real-world perturbations when no perturbation information is known in advance. However the resulting perturbation set relies on a stochastic approximation, and it is not immediately obvious what this set will ultimately capture.…”