An autoencoder has been trained on thousands of handwritten 1s, with a code of just two numbers. The 1s in the training set vary in two ways — how thick the stroke is, and how far it leans — and nobody told the model about either.
Three training images you have studied closely encode to:
| image | code |
|---|---|
| — a thin, upright 1 | |
| — a thick, upright 1 | |
| — a thin, leaning 1 |
is with only the thickness changed. is with only the lean changed.
Training codes fill this whole region densely, and the space is well-behaved here: moving a code by a given step, in a given direction, changes the decoded 1 in the same way wherever you start from.
A new image arrives and encodes to .
What should you expect to look like?
Select all that apply.