Are CNNs rotation invariant and how to cater this

A question I got in an interview

I am happily doing my PhD in Italy but before coming here I was in Sweden doing my PhD which finished after only 2 years due to limited funding. While I was applying for PhD positions, I gave multiple interviews and some of the questions which I remember are as follows:

[embed]Probability based event generation An Interview question I gotthanifbutt.medium.com

[embed]There is a person who jumps and covers 1/2 distance of his last jump, will he reach his target… There is a person who jumps and covers 1/2 distance of his last jump, will he reach his target soon or not? Last of the…thanifbutt.medium.com

One of the questions which I remember is whether CNN’s are rotation invariant or not and how to cater this.

The simple answer to this question is that CNN’s are not rotation invariant by default but the problem can be catered by making the training data capable of handling rotations and one way to do this is by data augmentation. So, if the dataset under consideration contains full 360 degree rotations then the CNN trained using that dataset can be said as rotation invariant otherwise not.

References

[embed]Are CNNs invariant to translation, rotation, and scaling? - PyImageSearch Learn what convolution and cross-correlation are and how the terms are used interchangeably in the deep learning…pyimagesearch.com

That’s it for now. See you later.