Organizing the world of fonts with AI | Hacker News Reader

Organizing the world of fonts with AI

medium.com

48 pointsbrudgers19 comments

This is really fun and interesting work, but it also goes to show how diluted the term/buzzword "AI" really is.

This is barely even machine learning (what exactly is being "learned" here?). 5 years ago this would have been "dimension reduction and clustering".

AI has always been a much more general term in Computer Science than in popular use. Look at historic papers to see.

What is being learned is an embedding (lower-dimension representation) that generalises well enough to group "similar" fonts together.

Both dimensionality reduction and clustering are important parts of machine learning: even if some of the dimensionality reduction might be 100% mathematical, most of the recent ones involve statistics and some engineering, hence being included under the ML umbrella.

T-SNE doesn't learn an embedding or generalise at all. There is no way to add new points to his plot without running the entire T-SNE algorithm again from scratch.

Though who even cares if it is AI or not, it's new and looks useful.

Dimension reduction and clustering have long been the focus of AI/ML. Reference, Bishops book an early 2000's classic, Pattern Recognition and Machine Learning, chapter 9.1 is k-means clustering.

People started using machine learning as a term to differentiate the goal from actual intelligence. They're not the same. Early work in AI was generally about achieving goals (aka A*, blocks world, etc), which is not the same thing as dimensionality reduction or clustering.

Russell and Norvig (1995) include machine learning as one of the four elements needed to pass the Turing Test. They cite Gold's Identification of Language in the Limit from 1967 as a laying the groundwork for machine learning applied to natural language processing.

http://stpk.cs.rtu.lv/sites/all/files/stpk/materiali/mi/arti...

http://web.mit.edu/~6.863/www/spring2009/readings/gold67limi...

Metallurgy is an important aspect of high end bicycle design. But you would never argue that metallurgy is bicycle design.

https://en.wikipedia.org/wiki/AI_effect

One of the properties of artificial intelligence is that in the rear view mirror it is always unimpressive "standard programming practice." Searle's Chinese Room [1] probably describes the problem well enough -- once people are familiar with the mechanism that solved a difficult problem for humans, the machine around the mechanism no longer seems intelligent.

In the early days of artificial intelligence, the general problem solver's unguided search of an exponential solution space was cutting edge (and GPS was non-exhaustive despite searching a combinatorial explosion of potential solutions). A decade the A* algorithm [2] that's in beginning game tutorials was cutting edge research and today, Eliza's pattern matching against natural language powers the 'bot revolution'.

I suspect that in a few years people will say "That's a convolutional neural network, not AI."

[1]: https://en.wikipedia.org/wiki/Chinese_room

[2]: https://en.wikipedia.org/wiki/A*_search_algorithm

I'd say right now a convolutions neural net by itself isn't AI. That's a how not a what.

When I think of AI, I think of hard AI, which includes the ability to learn new things, extrapolate, integrate across senses, and generally reason like an animal (ideally human) brain would.

It's currently unknown what will achieve (if anything) true AI. Maybe a NN? Maybe a really fancy use of A* search? Some crazy rules engine with a trillion rules a la Searle's Chinese room? It doesn't really matter how in the end as long as you achieve the what (AI).

This typography post is not even remotely close to it, nor is wit.ai, or generally almost all things hyped as AI. It's really frustrating to me having actually focused on studying AI and neuroscience to see these labels just haphazardly applied to literally anything. I'm fine with the title of machine learning, as that appropriately sets expectations. At least in the post the author used ML even if nothing was learned because that's closer to what occurred.

I read Searle's Minds brains and logic before the commercial world wide web in a philosophy of the mind course. While the main thrust of his argument certainly made an impression, the most intellectually interesting part was the tidbit about thermostats. A few years later during a phase in my life where I was buying philosophy books, I picked up a copy...maybe in part because the bit about thermostats has utility in internet arguments if for no reason other than offering a sneak attack from behind and in part because it is a Humeian stance on the limits of human knowledge and I admire Hume.

One day, within the last decade and probably about six years ago, I realized that the source of the thermostat tidbit is McCarthy. Which smells a bit relevant here given the claim to define what is and isn't 'true AI'.

Machines as simple as thermostats can be said to have beliefs, and having beliefs seems to be a characteristic of most machines capable of problem solving performance. However, the machines mankind has so far found it useful to construct rarely have beliefs about beliefs, although such beliefs will be needed by computer programs that reason about what knowledge they lack and where to get it. Mental qualities peculiar to human-like motivational structures , such as love and hate, will not be required for intelligent behavior, but we could probably program computers to exhibit them if we wanted to, because our common sense notions about them translate readily into certain program and data structures. Still other mental qualities, e.g. humor and appreciation of beauty, seem much harder to model.

Yesterday morning it was hot here in my office and so I turned on the ceiling fan: it was not so hot as to close up the house and fire up the AC. My human reasoning was exactly like that of a thermostat. One of the things about Humeian rigor is that one does not get to pick and choose the comfortable path when thinking deeply. The other thing about Humeian rigor is it's still OK to head down to the pub for a pint at the end of the day.

Anyway, 'true AI' is always over the horizon. Once we see how the parts are attached the technology is no longer so sophisticated as to be indistinguishable from magic.

Personally I don't believe it's just over the horizon. I think we'll continue to have advancements in new areas like autonomous driving that will push the envelope. Perhaps we'll one day be able to assemble everything and create Minksy's society of minds, but that path is very different from the human brain and how it works. However, it's something that is far more reasonable to create by human engineering rather than relying on (or trying to reverse engineer) the magic soup of chemicals and physics that is biology.

I had seen a similar research project by people from Microsoft. They had mapped fonts onto a higher dimensional data. Then you can generate new fonts on the fly by specifying a new point and getting a font for that.

https://github.com/erikbern/deep-fonts

I'm surprised it was all driven solely by rendering a single word "handgloves". The diagram hints that it might have even been an all-caps "HANDGLOVES". I wonder if the results would have been better with a longer phrase that included other characters, or some mix of lower/upper case.

Very nice and promising thing. I would suggest performing a second iteration with the following improvements:

- classify on all characters (not only one word)

- add some more fonts, at least those most popular ones. I tried with a few very popular (i'd say classic) fonts, but none was found

I would imagine most designers would prefer to see something like "RSOgfcq" if they're font surfing. An 'A' conveys so little about a typeface.

Agree that "a" is not helpful, but "RSOgfcq" (or an actual word) is too long to use in such a map.

Doing that forces you to focus on the trees instead of the forest. If you use an actual English word then you can compare gestalts, which is typically more important. That allows you to understand the emotional impact, etc.

Reply on news.ycombinator.com