Augmentation Invariance Objective

Problem

Return the contrastive objective that trades predictive information against augmentation invariance:

objective = mi_label - lam * mi_aug

where mi_label is the mutual information with the label and mi_aug the information shared between two augmented views.

Hint

Pushing invariance too far can discard label-relevant information.

case 1

in: [2,1,0.5]

expected: 1.5

case 2

in: [1,1,1]

expected: 0

case 3

in: [3,0,10]

expected: 3

case 4

in: [2,2,0.25]

expected: 1.5

Starter codePython

def augmentation_objective(mi_label, mi_aug, lam):
    # Your code here
    pass

4 test cases · run your code against them in the editor