A team is building a denoising autoencoder for handwritten digits: white ink on a black background, with every pixel a value from (black) to (white).
To see what the damage achieves on its own, they make the middle layer as wide as the input: numbers in, in the code, out. A plain autoencoder built like this would simply copy.
They compare four ways of building each training pair. Wherever a scheme makes a random choice, it makes a fresh one every time an image is used.
| scheme | input fed to the encoder | target |
|---|---|---|
| 1 | the digit with every pixel value halved | the clean digit |
| 2 | the digit with a square, placed at random, painted black | the clean digit |
| 3 | the digit flipped left-to-right (every digit, every time) | the clean digit |
| 4 | the digit with a random half of its pixels knocked out (set to black) | that same damaged digit |
Which is the only scheme that forces the network to learn what digits look like?
Select all that apply.