#corruption

Question 1: Single choice

You train a denoising autoencoder (encoder $f$ , decoder $g$ , reconstruction loss $L$ ) with masking noise. Which training setup matches the denoising autoencoder of Vincent et al.?

Train an ordinary autoencoder on clean $x$ , and mask inputs only at test time to check robustness1
Mask part of $x$ to get $\tilde{x}$ and minimise $L(x, g(f(\tilde{x})))$ , with a new mask each time2
Mask a random fraction of $x$ to get $\tilde{x}$ and minimise $L(\tilde{x}, g(f(\tilde{x})))$ on the masked input3
Mask a random fraction of the code $f(x)$ and minimise $L(x, g(\text{masked } f(x)))$ 4

Timed test