I'm in the middle of watching the show Start-Up on Netflix. I just finished episode 5 (some spoilers). Some background: Seo Dal-mi and Do-san are our main characters. They and Do-san's machine-learning-developer buddies are all on the same team at a hackathon. At the end of the hackathon, each team presents their product and business plan to the judges and an audience, and the top 5 teams (out of 40 - so this is pretty competitive) will be admitted to Sandbox, a fictional Korean Y Combinator. Seo Dal-mi has spent years working at crappy part-time jobs and entered the competition because she was inspired by Do-san's "company" to start her own. Do-san and his friends technically do have a company, and are very good at getting deep neural nets to attain state-of-the-art accuracy on classification tasks, but are terrible at every other aspect of business, and their company is currently falling apart. Dal-mi's sister In-jae is also in the competition, but the two have a terrible rivalry, went 15 years without speaking to each other, and mostly despise each other. In-jae forms a different team and, out of spite, does her project using the same dataset that Seo Dal-mi's team chose: a computerized collection of handwriting images from a bank.

What follows is the best machine learning dramatic face-off I've ever seen in anything. Dal-mi's team has made an excellent forgery detector: it takes handwritten text as input, and outputs "forgery" or "not a forgery" as output, with super-high accuracy. In-jae's team, meanwhile, has made a font-generator: it takes some sample handwriting as input, and outputs a full custom font-set, that you can then use for whatever digital application you want, like design. The presentations are going straightforwardly, until a judge asks... what if we put some writing from this font-generator into the forgery-detector? Would it identify it as a forgery? The two teams go on stage to set it up, In-jae's machine-fabricated font is fed into Dal-mi's fabrication-detector, and everything is on the line: the fate of their projects, and whether they get into Sandbox, all depends on what the detector network outputs.

This is the best. I've heard of this type of contest before, in real life: the contest between email spammers and spam filters; between doctored photos and videos and the ability to detect the doctoring; between auto-generated text...

george.e.hovey@gmail.com

george.e.hovey@gmail.com