Table 2 A brief comparison of existing mixing-based data augmentation methods summarizing their key idea, the space in which the interpolation is performed, the number of hyperparameters in the method, and the number of samples used for mixing to generate 1 new sample
From: Multi-sample \(\zeta \)-mixup: richer, more realistic synthetic samples from a p-series interpolant
Method | Key idea | Interpolation space | Number of hyperparameters | Involves additional optimization | Number of samples mixed |
|---|---|---|---|---|---|
SamplePairing [38] | Linear interpolation of pairs of images with a ratio \(\lambda = 0.5\); use labels of the first image | Input | 0 | ✗ | 2 |
Between-Class Learning [37] | Linear interpolation of pairs of images from different classes and their labels | Input | 0 | ✗ | 2 |
mixup [36] | Linear interpolation of pairs of samples and their labels | Input | 1 \((\alpha )\) | ✗ | 2 |
CutMix [39] | Paste a rectangular patch from one image onto another; mix labels proportionally | Input | 3 \((r_x, r_y, \lambda )\) | ✗ | 2 |
GridMix [40] | Paste a grid-based region from one image onto another; assign a mixed label and grid-based labels | Input | 2 (N, p) | ✗ | 2 |
Manifold Mixup [41] | Linear interpolation of latent representations and their labels | Latent | 1 \((\alpha , {\mathcal {S}})\) | ✗ | 2 |
MixFeat [42] | Linear interpolation of latent representations only | Latent | 1 \((\sigma )\) | ✗ | 2 |
AdaMixUp [25] | Train an additional network to learn mixing policy from data | Input | 0 | ✓ | 2 |
AutoMix [44] | Bi-level optimization for mixed sample generation and mixup classification | Input | 3 \((\alpha , l, m)\) | ✓ | 2 |
OptTransMix, AutoMix [43] | Optimization using optimal transport (OptTransMix) in input space or DNNs (AutoMix) in latent space for barycenter learning | Input/Latent | 2 \((n, \sigma )\) | ✓ | 2 |
SuperMix [99] | Iterative optimization-based salient masks for mixing | Input | 5 \((\alpha , \kappa , k, \sigma , \lambda _s)\) | ✓ | 3 |
Co-Mixup [96] | Iterative optimization-based mixing to maximize data saliency and encourage submodular diversity | Input | 6 \((\alpha , \beta , \gamma , \eta , \tau , \omega )\) | ✓ | 4 |
\(\zeta \)-mixup (Ours) | p-series-weighted convex combination of entire mini-batch of samples and their labels | Input | 1 \((\gamma )\) | ✗ | \(m (\ge 2)\) |