Is CNN sensitive to biases?

Question

Neural networks

Let's say I'm training a convolutional neural network that recognizes dogs, and let's say I've trained this network with photos of dogs with all the dogs in the center of the photo.

And now I want to classify a photo of a dog, in which the dog itself is in the corner, so will this neural network cope?

PS this is about the condition that global pooling is not used before the fully connected layer

Answer 1

If in a peasant way on the fingers, then it is not so sensitive, the core of the convolution passes through all corners. But in order to smooth it out, they also use the augmentation technique so that the dogs visit the corners too.