Skip Connections in U-Net

Still Confused why do we use skip connections in U-Net?

It would be worth listening again to what Prof Ng says about this. He’s the best at explaining everything. But my recollection of what he says and what it says in the notebook is that the point of U-Net is to reconstruct a new version of the original image with the same geometry, but with different contents. You want each original pixel converted to a “label” value that identifies the object type that the pixel is a part of, right? So the point of the skip connections is that they make it easier for the algorithm to reconstruct the geometry of the original image. In the “downsampling” path, it works like a classifier where you lose the geometry but gain knowledge about the objects. Then in the “upsampling” path, you need to reconstruct everything, which is where the information from the skip connections is helpful.

But as I said at the beginning, don’t take my word for this. Listen again to what Prof Ng says with the above ideas in mind and I bet it will make more sense this time.

Thanks for everything really !