DeepLearning5.3, MaxPooling as inductive bias for images
From Wulfram Gerstner
views
comments
From Wulfram Gerstner
MaxPooling enforces an inductive bias towards local translation invariance: it does not matter where exactly the edge occurs within a small neighborhood. Several convolutional layers of MaxPooling can give rise to global translation invariance: a cup is a cup, and it does not matter where it is placed in the image.