Disadvantage of multilayer perceptron over CNN

Question

Suppose I train a binary classifier - a dog or not a dog, images are input, so the question is, let's say I use a multilayer perceptron as a classifier and suppose I train it using images of dogs whose face is in the center of the image, and then if I submit a photo of dog, on which her face is located in the corner, then the perceptron will not be able to correctly classify

But if I use CNN, then it will be able to recognize the dog in this case. Am I right?

Answer 1

Unfortunately this is not so simple. In short the answer is no, there is no guarantee this will happen. While MLP has a high-ish chance of failing, it does not have to, it depends what it ended up learning as discriminating factor. And symmetrically CNNs are not translation invariant (unless by CNN you mean lack of a single linear layer). They can fail if you move an object around, they are just less affected than MLPs.