9 ms·
Recurrent Net Dreams Up Fake Chinese Characters in Vector Format with TensorFlow
- aftbit 11y agoThe more I read about machine learning using neural networks, the more I realize that we've just moved the "decisions" up into the choice of architecture and hyperparameters. This blog post describes a process of tuning the neural network until it generates Kanji that "look right" to the author.
- gertef 11y ago> just That's a small word for major change in programming paradigm
- tel 11y agoOf which there are usually far fewer and of which we can recursively apply ML techniques to fit. So yeah, but it works really well.
- habitue 11y agoWhat's interesting though is that debugging hyperparameters like this is much more at the level of human thought, rather than the nitty-gritty of programming. It's much closer to the Star Trek concept of programming "Computer, create fake chinese characters. These don't look right, stop drawing earlier. Ok, cut out all that are outside a one inch by one inch square."
- sago 11y ago"No, not like that. Why are you doing that? What on earth is causing that? What can I do to change it? I'll throw away the ones that don't work - oh, there are none left. Has it stopped working? Why is it eating 100% CPU and not spitting out any results?" One of the earliest games projects I worked on with learning AI scrapped the learning AI because it had a habit of sometimes, but rarely, causing the characters to run into the corner of a room and stay there until shot. Nobody had the slightest clue how to 'debug' the neural network to solve the intermittent issue. It's all fun and games when things are going well. But like Jurassic Park, when it comes to paying customers, eventually there will be running, and screaming.
- habitue 11y agoYeah, it's definitely not there yet. And quite a few Star Trek episodes center around ambiguities in this kind of "programming".
- hardmaru 11y agoThe problem with this particular algorithm is that I needed to have a lot of data to make it sort of work. I'm not sure how your AI was trained to play that game, sounds like a difficult problem-
- ziyao_w 11y agoPretty cool - some of the characters are actually real Chinese characters.
- trhway 11y ago>suddenly realise they forgot how to write Kanji. I am also guilty of this – even though I read a lot of Chinese and Japanese content in my everyday life, I struggle to write Chinese characters. What we notice is that while we can definitely read and recognise the characters we are able to write, the converse is certainly not true. when written in ink by brush the order of strokes seems to follow very natural flow, yet when it is taught using pencil it looks much less natural.
- zhemao 11y agoIt's not about stroke order. You just forget how it should look. It's like the difference between being able to recognize someone's face and being able to draw it from memory.
- trhway 11y agoThanks. I did feel that my interpretation wasn't enough.
- hardmaru 11y agoIs that why the shape of the fake pen for that Bimoji Kanji Writing NDS video game shaped like a brush?
- joe_the_user 11y agoAbout the reference in article to deep learning producing art: The thing is that what a current net is doing is very simple. It's just extrapolating from a set on a high dimensional feature space. There's no meaning here and so it's hard to generating a bunch of pseudo-data as art. It's fun and interesting when you do it a few times so any semi-random technique like ink blots or fractals but if you produce a steady stream of something like that, it's unoriginal, non-artistic qualities become more obvious.
- anon4 11y agoIf a urinal on a pedestal can be art, I think we've lost the fight when it comes to the word having any meaning.
- joshkpeterson 11y agolol k.
- hardmaru 11y agoWe were actually sort of discussing this a short while back https://twitter.com/hardmaru/status/676067023655337984 https://twitter.com/hardmaru/status/676067023655337984
- patio11 11y agoStroke order is very important to Japanese culture, in a society where the process matters just as much as the end result. Some calligraphists take stroke order very seriously, and will probably explode if they see someone writing a Kanji with incorrect stroke order. One might phrase this "Stroke order is assigned by convention in Japanese; deviating from the conventional order is incorrect, much laik hau Einglish speling iz nat ap four eendevizyuelle tyoise." ("What's the matter? You can sound it out. Pfft, Americans, such rigid traditionalists. It makes sense in the context of their religious views and conservative political tendencies, though." [+]) This avoids unfortunate Man Japanese People They're So Craaaaazy overtones. (FWIW: stroke order is prescriptionist not descriptionist but AFAIK prescriptionism has a virtually hegemonic mindshare among relevant authorities.) n.b. Otherwise this project and post is freaking excellent. [+] You can actually read Japanese-language takes on American culture which are exactly as bad as this Orientialism-in-reverse.
- emmett 11y agoI think a key point about this, which was only clear to me after I learned to write Hanzi in Mandarin, is that if you write the strokes in the wrong order it will not look right at the end. To someone looking at characters for the first time, it's not obvious that the stroke order matters to the reader as spelling does in English. It's not about fetishization of the process over stroke order, it's really obvious to the reader if you wrote the character incorrectly.
- schoen 11y agoI've kind of wondered about how you can tell -- do you know of an example that's online anywhere that makes it clear how it looks different? (... when written with a pen or pencil?)
- yaowenjiaozi 11y agoThe blog post's image on a child's struggle is kinda sorta one example. It shows the pencil at the last stroke of the bottom right character and that stroke order is wrong. Correct stroke order is shown here: https://en.m.wiktionary.org/wiki/%E6%88%90 https://en.m.wiktionary.org/wiki/%E6%88%90 To answer your question, if the image doesn't show the pencil positioned on that last stroke, I'd say that the characters don't look right and may infer but cannot tell for sure that they were written using incorrect stroke orders.
- thrownaway2424 11y agoMaybe this can be used to generate embarrassing nonsense tattoos for idiotic Americans.
- teddyh 11y ago'Twas brillig, and the slithy toves Did gyre and gimble in the wabe; All mimsy were the borogoves, And the mome raths outgrabe.
- niccaluim 11y agoI wonder what RNN-generated fake Latin characters would look like. Sort of a handwriting version of that "what does English sounds like to non-English-speaking people" video. (https://www.youtube.com/watch?v=Vt4Dfa4fOEY https://www.youtube.com/watch?v=Vt4Dfa4fOEY)
- hardmaru 11y agoI reproduced an English handwriting generation experiment in the previou post. http://blog.otoro.net/2015/12/12/handwriting-generation-demo-in-tensorflow/ http://blog.otoro.net/2015/12/12/handwriting-generation-demo...
- JulianMorrison 11y agoMostly like real ones, I expect. We don't have root components to our characters except long and short line, circle, hook, arc, angle, so there aren't many missed opportunities for a longer alphabet. You'd need to run fakeness at the level of words, after teaching it about etymological roots.
- PhasmaFelis 11y agoMy first thought was that, to me, Cyrillic looks how I'd expect fake Latin characters to look. Clearly related, made of the same elements, but I don't recognize most of the characters...
- est 11y agohttp://hanja.naver.com/ http://hanja.naver.com/ Here's a korean site providing many Hanja/Kanji/Hanzi stroke orders
- hardmaru 11y agoCool, I wonder how difficult it would be to extract all of the stroke data. LINE would probably not put the data on github..
- vvpan 11y ago"Dreams up" sounds a lot like modern "handcrafted with love" marketing.