[Paper Review] Remix: Rebalanced Mixup

Hsin-Ping Chou, Shih-Chieh Chang+3 more|arXiv (Cornell University)|Jul 8, 2020

Imbalanced Data Classification Techniques47 references57 citations

TL;DR

Remix relaxes Mixup by disentangling feature and label mixing factors, labeling mixed samples toward the minority class to better handle imbalanced data, and it improves Mixup-based regularizations across several datasets, especially when combined with re-weighting or re-sampling.

ABSTRACT

Deep image classifiers often perform poorly when training data are heavily class-imbalanced. In this work, we propose a new regularization technique, Remix, that relaxes Mixup's formulation and enables the mixing factors of features and labels to be disentangled. Specifically, when mixing two samples, while features are mixed in the same fashion as Mixup, Remix assigns the label in favor of the minority class by providing a disproportionately higher weight to the minority class. By doing so, the classifier learns to push the decision boundaries towards the majority classes and balance the generalization error between majority and minority classes. We have studied the state-of-the art regularization techniques such as Mixup, Manifold Mixup and CutMix under class-imbalanced regime, and shown that the proposed Remix significantly outperforms these state-of-the-arts and several re-weighting and re-sampling techniques, on the imbalanced datasets constructed by CIFAR-10, CIFAR-100, and CINIC-10. We have also evaluated Remix on a real-world large-scale imbalanced dataset, iNaturalist 2018. The experimental results confirmed that Remix provides consistent and significant improvements over the previous methods.

Motivation & Objective

• Address the problem of deep image classifiers biased by imbalanced training data.
• Introduce Remix to disentangle feature and label mixing factors in Mixup-based regularization.
• Demonstrate Remix's superiority over Mixup, Manifold Mixup, and CutMix on imbalanced benchmarks.
• Show Remix's compatibility and additive benefits when combined with re-weighting or re-sampling techniques.

Proposed method

• Define Remix with separate mixing factors for features and labels: tilde x_RM = lambda_x x_i + (1-lambda_x) x_j and tilde y_RM = lambda_y y_i + (1-lambda_y) y_j.
• Allow lambda_y to depend on class frequencies (n_i, n_j) via a rule involving kappa and tau, enabling label emphasis on minority classes.
• Illustrate that Remix reduces to standard Mixup/Manifold Mixup/CutMix when lambda_y = lambda_x, and describe how Remix can be integrated with existing Mixup-based methods.
• Provide an algorithmic outline (sampling pairs, drawing lambda_x from Beta, computing lambda_y, forming Remix samples, and updating the model with Remix losses).
• Offer guidance on hyper-parameter choices (tau, kappa) and emphasize ease of integration with other imbalance-handling techniques.

Experimental results

Research questions

• RQ1 Does disentangling the feature-label mixing factors improve learning under class imbalance?
• RQ2 How does Remix perform relative to Mixup-based methods (Mixup, Manifold Mixup, CutMix) on imbalanced datasets?
• RQ3 Can Remix augment re-weighting and re-sampling strategies to yield superior results on long-tail and step-imbalance scenarios?
• RQ4 What are the practical hyper-parameter sensitivities (tau, kappa) and integration considerations for Remix?
• RQ5 Is Remix effective on large-scale real-world imbalanced data like iNaturalist 2018?

Key findings

• Remix consistently outperforms vanilla Mixup, Mixup-based methods, and several re-weighting/re-sampling baselines on imbalanced CIFAR-10/100 and CINIC-10.
• Remix achieves strong gains when combined with deferred re-weighting (DRW) or deferred re-sampling (DRS), often yielding the best results among tested configurations.
• The method is especially beneficial under step-imbalance and remains effective under real-world imbalanced data (iNaturalist 2018).
• Remix is compatible with other Mixup-based regularizations (e.g., Manifold Mixup, CutMix) and generally boosts their performance, particularly under severe imbalance (rho = 100).
• Qualitative analyses suggest Remix shifts decision boundaries toward minority classes, creating tighter margins for majorities while improving minority class decisions.