Andrew Sir told in the lecture that Skip connection helps retain information that are destroyed after various steps of convolutions and pooling.
In the last part of his assignment, he says :
Skip connections are used to prevent border pixel information loss and overfitting in U-Net
I am unable to understand the meaning of “border pixel information loss”.
Also, how does skip connections prevent overfitting?
Hi ashish_learns,
Border pixel information is information about the pixels at the border of the original image. Such information can be retained by means of skip connections as explained in the video on U-net architecture intuition.
As for your question about overfitting, my two cents is that whereas short skip-connections increase the chance of overfitting, the long skip-connections in U-net prevent this from happening. On this, see, e.g. this paper.
