latent space

Machine learning borrowed the word space from us, and I would like to check what it did with it.

A model turns everything it has seen into points. Not pictures, not words: coordinates, in a space with hundreds or thousands of axes, none of which anyone chose or can name. Things that behave alike end up near each other. That is the entire trick, and it is a spatial argument made almost entirely by people who do not think of themselves as spatial.

Near means similar. Direction means something even when nobody can say what. There are neighbourhoods, there are dense parts and thin parts, and the thin parts are exactly where a model will invent something with total confidence.

I want to be careful not to oversell the metaphor. A latent space has no floor, no gravity and no up. You cannot stand in it and it does not care about your body, which is precisely what makes it useful and precisely why I would not hand it a room and walk away.

Still. We are the discipline that has spent three thousand years on what near means. Somebody from here should probably be in that room.