Question

Assume you are using Logistic Regression for classyfing images of cats and dogs.

What happens if we randomly permute the pixels in each image (with the same permutation) before we train the classifier?

Will we get a classifier that is better, worse, or the same than if we used the raw data?

Give a short explanation (at most three sentences).

HINT: The location of pixels relative to each other seem to hold some kind of information. Does a random permutation of all pixels position affect this locality? Does the model we use exploit pixel locality?

Answer 1

The performance will be worsen for sure, since Logistic regression tries to learn the decision boundary based on the relative position by learning weights. if the pixels are permuted randomly before training this relative information stored will be lost however if you try to permute all the images with some sort of same logic than this information will be retained and model might be able to learn the decision boudry even after permutation, though while trying to classify new image we need to permute the subject image with the same permutaion logic that we have used during training. Pixel locality will be exploited if we use different permutaion logic for different images and should retain if we are using the same permutation logic on all the images used for training.