Hacker News Reader

Comment by layer8 | Hacker News Reader

Isn’t kerning something that would be amenable to be approximately solved by machine learning? I.e. for a “good enough” default kerning? (Ignoring the extra difficulty from the stripes in TFA.)

View on HN

Since it’s an issue of equalizing the perceptual area in the negative space, there’s not really any reason to throw ML at it. Throw actual math at it.

There have been lots of attempts to automate kerning for decades. Most people seem to agree that they're still inadequate and that manual adjustments produce the best results.

Human perception is highly non-linear and largely based on the concepts stored in the brain. Trying to manually approximate that with math is bound to fail once you approach certain fidelity threshold. If you look at the recent color science and spatial perception research, it becomes obvious that it makes total sense to mass profile the perception and throw ML at it. A lot of researchers are still in denial about this, though.

Agree with sibling comments. There's something very slippery and tricky going on with "perceptual area," it's not simple geometry. This is actually an area where I think machine learning has something to offer.

Mrs. Scyzoryk is a typographer, confirms that it’s not just math, but can’t imagine using ML for what is an integral part of developing an „eye” and going through the process.

Just like programming! I can't imagine how we'd teach machines to code.

Yeah and also: new font design is a niche niche niche domain.

It’s not like accelerating font development would give us something new. Most of what most of us need exists at this point.

My initial point exactly

Well, I have some qualifications in typography, a reasonable familiarity with ML techniques, and am fairly good at math (though I can't claim to have won the Putnam), and I can imagine how to apply ML for this task.

I would think that ML is exactly useful in developing an “eye”.

Reply on news.ycombinator.com