Two column

Left

We have a number of n-dimensional points; Normally, we identify each one of them using n numbers (e.g., the point (0,1) in R^2 is identified using 2 numbers)

Right

But, what if we could give n-dimensional points a lower dimensional (<n) identifier?

This only make sense if...

This only make sense if the n-dim points are mostly within a finite subset of their n-dim space, which allows them to be mapped onto a lower (<n)-dim space.

We could use a hash map, and it’d be perfectly able to map an n-dim point to a (<n)-dim point, but would it be able to map the (<n)-dim point back to its n-dim version? Nope.

An Autoencoder, however, can map back and forth between an n-dim point and a (<n)-dim one.

It in fact can do all of the following stuff that a hash map just can’t,

  1. map an n-dim point, A, to a (<n)-dim that then gets mapped back into another n-dim point, B.
    • This is useful if u want information in A that’s not in B (usually noise) to be ignored in the mapping-back.


Example


Related Notes

ThumbnailNoteType

Sources