Fig. 7
From: Multi-sample \(\zeta \)-mixup: richer, more realistic synthetic samples from a p-series interpolant
Augmenting a 2D data distribution with non-linear class boundaries (a) with mixup (b) and \(\zeta \)-mixup (c). Notice how \(\zeta \)-mixup generates samples closer to the original data, and this explains why the local intrinsic dimensionality (ID) estimates for \(\zeta \)-mixup (d) may sometimes be lower than the original dataset (e) (Fig. 6): the Fukunaga-Olsen method for local ID estimation using PCA based on nearest-neighbor sampling may yield a more compact distribution for \(\zeta \)-mixup. Conversely, with mixup, a test sample may lie in the vicinity (calculated using k-nearest neighbors; \(k = \{8, 16\}\)) of training samples from classes different from the test image’s correct label, leading to an incorrect prediction (f). This is less likely with \(\zeta \)-mixup (g)