#### Howdy, Stranger!

It looks like you're new here. Sign in or register to get started.

# Neural Network Design Fonts

• Posts: 938
edited February 2016
Can someone explain to me how this is more interesting than the 4 or 5 attempts at aggregation type design that came before? I don’t see the practical or even theoretical value, but I recognize that I may be missing something due to an engineering blind spot.
• Posts: 1,859
Neural networks is a newer buzzword than whatever was used to describe previous attempts.
• Posts: 330
It is a data mining technique - the difference is something like this: let's say you have an uncle who was a chain smoker and died early of lung cancer. And another friend knows another family who smoked a lot and died of neumonia; and somebody else have poor health and coughs a lot, also from smoking. So you know smoking is bad for you, vaguely, from a handful of examples and anecdotes.

Now, if you collect the smoking habits from 50k smokers and how/when they died, you suddenly have a way of actually saying how much one cigarette shorten your life down to the number of minutes.

Samething happens when you draw 5 A's over each other from different typeface, vs drawing 50,000 A's. When you have that amount of data and munch through them, suddenly you have say, degree of serif-ness coming out of the data, in the same way one might quantify how much one cigarette per day shorten your lifespan.

• Posts: 59
There is no magical serifness or proportion one might achieve. In certain styles there are indeed systems. Averaging them gets you no closer to exposing them, rather a surplus of data might as well confuse what is very purely distilled in a single example.

At any rate the proportions are no real secret. Various geometric systems have been devised over the years. Some have been influential, gained a following, and shaped what looks correct in a given style. The rest is up to taste.

The latin typographic alphabet has a written origin. There is no code to crack.
• Posts: 330
edited February 2016
I did not want to go into mathematical jargons. But the average is merely the first principal component. Just as the average shortening of lifespan per cigarette. When you have age, sex, ethnic origin, etc, you can analyze your data into contribution from multi-dimension components. In this case, given sufficient data, one might meaningfully extract parameters which could correspond to serif-ness, weight (in fact the first component would be weight rather than serif-ness), cursiveness, italicness, etc. Just as one might say your lifespan is number of cigarettes adjusted for sex, ethnic origin, age, etc.

https://en.m.wikipedia.org/wiki/Principal_component_analysis

Neural network is one way of doing PCA. There are others.
• Posts: 557
edited February 2016
Can someone explain to me how this is more interesting than...
This thing can recreate a complete typeface having just one letter as input. It's like when a type designer recreate a full typeface from a logo, or from small sample of letters.
I find it very interesting...
• Posts: 165
Can someone explain to me how this is more interesting than...
This thing can recreate a complete typeface having just one letter as input. It's like when a type designer recreate a full typeface from a logo, or from small sample of letters.
I find it very interesting...
• Posts: 1,867
It’s just a tool, so it doesn’t actually “design” fonts, any more than a pen, punch, or Fontlab.
• Posts: 330
That's correct - it does not "design" fonts, but "interpolate" fonts. In the same way that healthy people can expect to live to 80, chain smoker have the first stroke/cancer diagnosis/liver kidney failure around 50, if you do half a pack per week, one might interpolate and predict that you might have your first health scare by 65. That sort of prediction is called  "interpolation".

Well, I guess some can call that "design" in the same sense as one can call the result of prenatal screening "designer babies". It is a conscious decision to select one particular variant out of a range of pre-existed possibilities.

So on the other hand, in the sense of a conscious decision to select from a range of possibilities, it is "design".
• Posts: 1,867
It is a conscious decision to select one particular variant out of a range of pre-existed possibilities.

Like choosing the options on a new car?
Or choosing a color from a palette that someone else has created?
That’s hardly design.

And besides, neural networks aren’t conscious.

• Posts: 330
edited February 2016
Neutral networks are not conscious. As I said, it is a data mining tool - it offers you the choices among the "ranges" of serif-ness, etc.

It is in the same sense that prenatal screening is ,or is not, "designing babies". It is a technique, and quite a powerful one in both cases, which offers some human 'designer' the choices.

There comes a point when you choose options in your car, using exotic wheels, paint it a different color (even using off the shelf paint) that it becomes 'designing' your car.

In fact the mathematics for prenatal screening is about the same.
• Posts: 1,178
It is a conscious decision to select one particular variant out of a range of pre-existed possibilities.

Like choosing the options on a new car?
Or choosing a color from a palette that someone else has created?
That’s hardly design.
In the 21st century, that kind of design dominates.

http://www.printmag.com/interviews/roger-black-typecon-potential/
• Posts: 330
edited February 2016
Some says Apple doesn't "invent" or "design" things - Apple merely put combinations of pre-existed technology together.

So I guess a lot of whether one sees innovation in what some describes as "clever assembly of pre-existing parts", whether it is "font characteristics", "genes", or "colors" / "shapes".

Neural networks as described in the article, was used as a data mining tool - for identifying and isolating those "font characteristics" - and from that, *you* can choose a font outcome with certain combinations of these characteristics. The demo was a bit misleading as it shows the entire range of possibilities. Showing the whole range is not design - picking one out of a range is (maybe).

Just as a similar neural networks analysis on genetic material might identify specific combinations and proportions of genes as "grumpiness" vs "easy-going-ness". So the question is: do you call babies born out of prenatal screening for a "easy disposition" "designer babies"? It is computationally the same process.
• Posts: 2,406