Hello and welcome to yet another update to the infamous Japan Map! The latest rounds of updates justify a version revision, not just because of things that happened in the front-end, but also because of some incredible things that occurred in the back-end!

And to cut to the chase: The major new update is that there is a new AI now in place to automatically detect whether a video is based on a location or not. So why is that important you may ask? Well, the thing is that before a lot of videos were placed on spots on the map where they had no business of being.

The videos showcased here have absolutely no business being on the map! The AI removes these automatically now.

Examples for such videos included Unboxings, Anime Talks, Japanese Language Lessons and Reactions. This came about because the location detection AI sometimes triggered on locations being mentioned in the video. Sometimes totally nonsensical entries were also included due to the way in which the grammar was detect. AI is funny like that.

The idea behind the map is to provide valuable information to end users. And having videos show up that are put there in error actually hurts the user experience. Training a new AI that automatically recognized those was therefore given. In the end, an accuracy of around 90% was achieved with a data set of around 2000 videos evaluated by hand.

While I won’t go into too many details on how this works, a couple of things are too interesting to not share. I created this by using SBERT to create embedding vectors of title, description and subtitles. Effectively, this means I created a mathematical representation of the meaning of the aforementioned text.

Out of this, images can then be generated. And now the cool part: On these images, I used something called Convolutions. Effectively, on these images I ran an image recognition AI. The same kind of AI that discerns between images of cats and dogs! Pretty neat right? But yeah, it is pretty standard in the AI field to convert available data into images and then run image recognition algorithms.

The images created by the SBERT vectors look something like this.

To round it all off, there are a number of small changes to the UI that should improve the user experience somewhat. Here they are:

  • Some people have asked for this and I have obliged. You can now hide videos made by yours truly. In case you are interested in seeing a view of what another YouTuber has done or if you simply don’t like seeing my face! You are now able to press a button next to my channel to hide videos from the The Red Value channel
  • The Find Channel search bar will now also give you the option to open the Find Videos prompt. The idea here is to inform you that it is possible to find videos around a nearby location. I felt like a lot of people simply weren’t aware of this feature yet.
  • The Find Channel search bar also lets you choose to show the latest videos. No need to refresh the page to be able to show what everyone else has been up to in the last 7 days.
  • Small little toasts (yes that is what those little popup messages are called), will now give you useful tips on how to use the map and perhaps inform you of new features as you try them out.

All in all, I think the latest improvements really make the map quite a bit more usable. Once again, if you have any suggestions, feel free to get in touch with me! I feel like that the map has now reached a point at which it has become quite presentable and that is why my next steps will be to work on allowing creators to log in and edit their own entries. That is the sneak peek at Version 1.4.



You must be logged in to post a comment.