Multi-sample -mixup: richer, more realistic synthetic samples from a p-series interpolant

Area of Science:

• Deep Learning
• Computer Vision
• Machine Learning

Background:

• Data augmentation is crucial for deep learning model regularization.
• Mixup, a popular method, generates synthetic data via convex combinations but can create unrealistic samples with incorrect labels.
• Existing methods struggle with off-manifold data generation.

Purpose of the Study:

• To introduce Zeta-mixup, a generalized mixup technique for improved data augmentation.
• To address limitations of existing mixup methods, such as off-manifold sampling and incorrect labels.
• To enhance the realism, diversity, and label accuracy of synthetic training data.

Main Methods:

• Propose Zeta-mixup, a generalization of mixup allowing convex combinations of multiple samples using a p-series interpolant.
• Investigate the preservation of intrinsic data dimensionality.
• Implement and evaluate Zeta-mixup against baseline methods.

Main Results:

• Zeta-mixup generates more realistic and diverse outputs compared to mixup.
• The method better preserves the intrinsic dimensionality of datasets, crucial for generalizable models.
• Zeta-mixup implementation is faster than mixup and outperforms mixup, CutMix, and traditional augmentation on 26 diverse image datasets.

Conclusions:

• Zeta-mixup offers provably desirable properties for data augmentation in deep learning.
• The technique enhances model generalizability and performance in image classification.
• Zeta-mixup represents a significant advancement over existing mixup-based augmentation strategies.