***
**W5: Equations 10 through 15 are badly formatted and hard to read. It is also unclear what the individual contribution of the four cross-entropy terms are, and a suitable way to choose $\alpha$ and $\beta$.**
**A5:** According to your suggestions, we have reformatted Equations 10 -15 to clarify the presentation. In addition, we have included the following contents in subsection 3.3.1 in page 6 to explain the individual contributions.
“Notice that $L_{ce}(\psi_W, y)$ is the standard cross-entropy loss. $L_{ace}(\psi_W, \mathbb{Y})$ loss facilitates the backbone to provide a stable feature map at the early stage so that it speeds up convergence. Target loss $L_{amce}(\psi_W, \mathbb{Y})$ aims to learn task-relevant information in the generated mixed samples. $L_{mce}(\psi_W, y_{mix})$ facilitates the capture of task-relevant information in the original mixed samples.”
Also, we have conducted ablation experiments by gradually adding the regularization losses and computed their accuracy. The experimental results are listed in Table 5 in page 9 and Table 8 in Appendix at page 16.
Regarding the optimization of hyperparameters, existing mixup augmentation approaches [1-10] usually conduct comparable experiments on CIFAR-100 , Tiny-ImageNet, ImageNet-1K, CUB-200-2011, FGVC-Aircrafts, and Standford-Cars, as detailed in A.1 in the Appendix. However, these original datasets include only two subsets: either training set and validation set or training set and test set. In these works [1-8], they treat the validation set as test set. So, they do not employ a validation set in the experiments and all models are trained on the training set and stopped with predefined epochs. Then, the accuracy on the test set or validation set are reported for comparison. Therefore, for fair comparison, we follow their experimental settings and do not use the validation set to choose the best checkpoint. Instead, similar to works SuperMix and AutoMix, we provide the sensitivity of the hyperparameters N, λ, α, and β with the dataset (as shown in Figure 6) in the ablation experiments in page 9 . More experimental results are included in Tables 9 and 10 in the Appendix at page 16.
We do think, nonetheless, that your suggestions are very valuable . In the future, we will split, for each dataset, the current training set, into an actual training set and a validation set, while keeping the test set unchanged. The hyperparameters will be optimized on the validation set. Alternatively, we will optimize the hyperparameters by cross-validation on the training dataset.
***
**W6: Section 4.2 mentions the proposed method has the lowest calibration error, but there is no table showing the ECE of other baselines. Fig.4. shows the ECE of only the proposed method.**
**A6:** In the previous version, the ECE of the tested approaches were shown in the corresponding sub-figs in Figure 4. According to your suggestion, we show, in the revised vision, the ECE in Table 6 in Appendix in page 14 and Figure 4 in page 8.
|Classifiers|MixUp|CutMix|FMix|GridMix|PuzzleMix|AutoMix|AdAutoMix|
|---|:--:|:--:|:--:|:--:|:--:|:--:|:--:|
|ResNet-18|15.3|4.4|8.9|6.5|3.7|3.4|3.2(-0.2)|
|Swin-Tiny Transformer|13.4|10.1|9.2|9.3|16.7|10.5|9.2(-0.0)|
***
**W7: Typos and minor mistakes: “to facility representation” Mxiup → mixup bad formatting in eq 5 notation in eq 6, 7 is unclear. is the ∗ scalar multiplication? what is meant by “inherent meaning of images” - this sounds slightly unscientific, this should be explained in a bit more detail.**
**A7:** Thanks for pointing this out. We have corrected the “to facility representation” to “to facilitate representation”. The “Mxiup ” is changed to “mixup”. Eq.5 have been reformatted. Also, we have provided the explanation that the notation ∗ denotes scalar multiplication in page 4 in the revised version.
In our work, the inherent meaning of images is related to the the semantic meaning of the images. For example, if there is cat in an image, the label of cat is the semantic meaning. If we fully mask the regions associated with cat, there is no pattern for classification. We have added the explanation in the following sentence at page 6.
“It is possible, therefore, that the inherent meaning of images (i.e. the semantic meaning of the images) collapse”