Mis on mixup ja cutmix? Kuidas need mõjutavad regulatsiooni?

Mixup and CutMix are data augmentation methods used to improve the generalization ability of models, especially in computer vision tasks.

• Mixup involves linearly mixing two images and their labels. For example, a new image is created as a weighted sum of two original images, and the label as a weighted sum of their labels. This helps the model learn from intermediate examples, reducing overfitting.

• CutMix instead of mixing the entire image, cuts a rectangular region from one image and inserts it into another, with the label formed proportionally to the area of the inserted region. This preserves local features and promotes more stable training.

Both methods serve as regularizers, as they force the model to be less confident and more resistant to noise, improving generalization and reducing overfitting.