search

AshishBora / csgm

Code to reproduce results from the paper: "Compressed Sensing using Generative Models".
MIT License
212 stars 73 forks source link

Question about the loss function #1

Closed xuxy09 closed 7 years ago

xuxy09 commented 7 years ago

I don't understand why you use ||AG(z)-y|| as the loss function instead of using ||G(z)-x|| which makes more sense to me. I intuitively feel that multiplying A with x will lose lots of information. Looking forward to your explanation. Thanks!

AshishBora commented 7 years ago

Because we do not have access to x. We only have access to A and y, and we are trying to reconstruct x.

xuxy09 commented 7 years ago

But y=Ax, if you do not have x, how can you get y?

AshishBora commented 7 years ago

y = Ax is computed outside the estimation procedure. The estimation procedure itself does not have access to x.

Think of it this way : Someone else has x. You give them an A, they compute y = Ax, and give y back to you. Can you find x? This is the problem compressed sensing is trying to solve.