Table 4 Classification error rates (ERR) on NATURAL
From: Multi-sample \(\zeta \)-mixup: richer, more realistic synthetic samples from a p-series interpolant
Method | CIFAR-10 | CIFAR-100 | F-MNIST | STL-10 | Imagenette | Imagewoof |
|---|---|---|---|---|---|---|
# images (#classes) | 60,000 (10) | 60,000 (10) | 60,000 (10) | 13,000 (10) | 13,394 (10) | 12,954 (10) |
ERM | \(5.48 \pm 0.03\) | \(23.33 \pm 0.09\) | \(6.11 \pm 0.02\) | \(25.74 \pm 0.17\) | \(16.08 \pm 0.15\) | \(30.92 \pm 0.02\) |
mixup | \(4.68 \pm 0.09\) | \(21.85 \pm 0.07\) | \(6.04 \pm 0.20\) | \(25.31 \pm 0.33\) | \(16.20 \pm 0.03\) | \(30.80 \pm 0.04\) |
\(\zeta \)-mixup (\(\gamma =2.4\)) | \(\mathbf{4.42 \pm 0.02}\) + 5.56% | \(21.50 \pm 0.04\) + 1.60% | \(6.04 \pm 0.04\) + 0.00% | \(\mathbf{24.14 \pm 0.10}\) + 4.62% | \(\mathbf{15.16 \pm 0.07}\) + 6.42% | \(30.72 \pm 0.02\) + 0.26% |
\(\zeta \)-mixup (\(\gamma =2.8\)) | \({{\underline{4.67 \pm 0.05}}}\) + 0.21% | \({{\underline{21.35 \pm 0.02}}}\) + 2.29% | \(\mathbf{5.70 \pm 0.07}\) + 5.63% | \({{\underline{24.82 \pm 0.03}}}\) + 1.94% | \({{\underline{15.62 \pm 0.07}}}\) + 3.58% | \(\mathbf{30.21 \pm 0.05}\) + 1.92% |
\(\zeta \)-mixup (\(\gamma =4.0\)) | \(\mathbf{4.42 \pm 0.01}\) + 5.56% | \(\mathbf{21.28 \pm 0.02}\) + 2.61% | \({{\underline{5.89 \pm 0.04}}}\) + 2.48% | \(24.92 \pm 0.22\) + 1.54% | \(15.92 \pm 0.07\) + 1.73% | \({{\underline{30.67 \pm 0.03}}}\) + 0.42% |