Start-up is the first K-drama that I’ve ever watched. It’s about young people with entrepreneurial spirit juggle with life and love as they try to keep their startup alive.
I was intrigued primarily because of its name. I got hooked. I eagerly wait for the weekends for 2 new episodes to come out. Story-telling and music is just something else!
The best thing about the show is that they don’t shy away from details. Unlike most Hollywood flicks where you see people typing really fast and speaking gibberish, they actually refer to proper ML techniques and code.
This post is about the apps that they go on to build in the show. I write down possible techniques to build the same on our own.
Object detection and localization
Samsan Tech (Nam Do-San’s startup) won the CODA competition on image recognition.
Handwriting font generation
Won In-Jae’s team builds this in the hackathon organized by Sandbox, a startup accelerator. I think this is fantastic. I remember one of my friends talking about building this using GANs. He actually built a proof-of-concept of this which led to a new project within the organization.
Generating fonts is pretty expensive. If we could use ML to do this, it would be really cool!
Handwriting Forgery classification
Samsan Tech builds this in the same hackathon by Sandbox, where they go head to head with Won In-Jae’s team. One of the judges suggests using the font generated by In-Jae’s team to be passed as an input to the Forgery Detection model. The model actually fails to recognize the AI-generated font as fake. Nam Do-San (the guy in the photo above) figures out an issue with his model later and fixes it.
Noongil - an app that describes images in natural language
Technique: Image captioning (CNN + LSTM) + Speech to Text(Wavenet?)
This was presented on Demo day at Sandbox. Every startup has to demo their app after 6 months where they’ll be judged by investors from all over the world. The stake is high and they need to secure funding to keep their business alive.
I love this one because their motivation was to help the blind navigate through life. Nam Do-san is motivated to build a startup around this because his girlfriend’s grandma was on the verge of losing her vision. Really touching!
Guardian AI - detect suspicious actions from live CCTV footage and inform authorities of the same
Tehnique: CNN + LSTM
In-Jae presented this app on Demo day. She mentioned that she didn’t need funding as she already had sales that could sustain the business.
I think this is similar to Noongil with a different application in mind. This was intended to be used by enterprises/governments for monitoring via CCTV footages. The classifier will also be trained to localize and describe any actions that need to be reported. Noongil was designed to be used from a smartphone, which had to sacrifice accuracy for battery-life/performance.
Imagine having this app run on your security cameras outside your house that automatically triggers an alarm/informs the police on seeing suspicious activities like a break-in. What if a massive crowd starts gathering on the streets outside for riots? What if some people are being robbed in the middle of the night? I think this would be a really cool application for law-enforcement agencies.