Robust learning on ISIC 20181, based on Learning with Noisy Labels via Sparse Regularization (ICCV 2021)2.
-
Noise Corruption: a clean label y is flipped into a random noisy version $y ̃
$with probability $ η(x,y ̃) = p(y ̃|y, x)$-
Symmetric: with equal probability
$η/6$ -
Asymmetric: NV↔MEL, BCC↔BKL, VASC↔DF, AKIEC↔DF
-
$η = 0, 0.1, 0.4$
-
-
Loss Function
a) Focal Loss $$ \mathrm{FL}\left(p,\ y\right)=-\sum_{t=1}^{n}{y_t\left(1-p_t\right)^\gamma\log{\left(p_t\right)}} $$
where
$γ=0.5$ b) Focal Loss + SR $$ \mathrm{FLSR}\left(p,\ y\right)=-\sum_{t=1}^{n}{y_t\left(1-p_t\right)^\gamma\log{\left(p_t\right)}}+\lambda |p_t|_p^p $$
where
$p=\frac{exp{\left(\frac{z_i}{\tau}\right)}}{\sum_{j=1}^{n}exp{\left(\frac{z_j}{\tau}\right)}},\ \tau=0.5,\lambda_t=5・ 1.005^{t/1}$ c) GCE3 $$ \mathrm{GCEL}\left(p,y\right)=\sum_{t=1}^{n}{y_t\frac{\left(1-p_t^q\right)}{q}} $$
where
$p=0.7$
AUC | |||||
---|---|---|---|---|---|
Noise Type | None | Asymmetric | Symmetric | Asymmetric | Symmetric |
𝜂 | 0 | 0.1 | 0.4 | ||
Focal Loss | 0.9856 | 0.9766 | 0.9691 | 0.9235 | 0.9224 |
Focal Loss + SR | 0.9854 | 0.9778 | 0.9725 | 0.9443 | 0.9517 |
GCE | 0.9837 | 0.9804 | 0.9789 | 0.9256 | 0.9674 |
ACC | |||||
---|---|---|---|---|---|
Noise Type | None | Asymmetric | Symmetric | Asymmetric | Symmetric |
𝜂 | 0 | 0.1 | 0.4 | ||
Focal Loss | 0.855 | 0.817 | 0.803 | 0.573 (0.691) | 0.683 (0.782) |
Focal Loss + SR | 0.853 | 0.821 | 0.829 | 0.700 (0.720) | 0.784 (0.776) |
GCE | 0.851 | 0.826 | 0.827 | 0.605 (0.702) | 0.782(0.798) |
(): the best model on validation set
Loss Function
Noise Type
The robustness of aforementioned models under adversarial attacks is also tested.
Trained with sparse regularization, the model is relatively robust against perturbations.
Training Method | Accuracy | |||
---|---|---|---|---|
Clean | Gaussian | FGSM | PGD | |
FL | 0.855 | 0.675 | 0.287 | 0.000 |
FL+SR | 0.853 | 0.745 | 0.309 | 0.000 |
GCE | 0.851 | 0.698 | 0.243 | 0.000 |
While trained on dataset containing hand-crafted noisy labels, the model gets higher accuracy under adversarial attacks.
Noisy Labels | Accuracy | |||
---|---|---|---|---|
Clean | Gaussian | FGSM | PGD | |
Clean | 0.853 | 0.745 | 0.309 | 0.000 |
0.1 Asymmetric | 0.826 | 0.673 | 0.230 | 0.000 |
0.1 Symmetric | 0.829 | 0.726 | 0.381 | 0.000 |
0.4 Asymmetric | 0.700 | 0.599 | 0.291 | 0.000 |
0.4 Symmetric | 0.784 | 0.716 | 0.417 | 0.000 |
Footnotes
-
HAM10000 Dataset: (c) by ViDIR Group, Department of Dermatology, Medical University of Vienna; https://doi.org/10.1038/sdata.2018.161 ↩
-
X. Zhou, X. Liu, C. Wang, D. Zhai, J. Jiang and X. Ji, "Learning with Noisy Labels via Sparse Regularization," 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 72-81, doi: 10.1109/ICCV48922.2021.00014. ↩
-
Zhang, Zhilu, and Mert Sabuncu. ‘Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels’. In Advances in Neural Information Processing Systems, Vol. 31. Curran Associates, Inc., 2018. https://proceedings.neurips.cc/paper/2018/hash/f2925f97bc13ad2852a7a551802feea0-Abstract.html. ↩