Fake it ’til you make it — все що вам потрібно знати про ML
upd: розходимось, нема на шо тут дивитися
Тут під’їхала новина:
Beginning in 2016, according to one source, Ring provided its Ukraine-based research and development team virtually unfettered access to a folder on Amazon’s S3 cloud storage service that contained every video created by every Ring camera around the world. This would amount to an enormous list of highly sensitive files that could be easily browsed and viewed. Downloading and sharing these customer video files would have required little more than a click.
This real-time crime-fighting requires more than raw video — it requires the ability to make sense, quickly and at a vast scale, of what’s actually happening in these household video streams. Is that a dog or your husband? Is that a burglar or a tree? Ring’s software has for years struggled with these fundamentals of object recognition. According to the most recent Information report, „Users routinely complained to customer support about receiving alerts when nothing noteworthy was happening at their front door; instead, the system seemed to be detecting a car driving by on the street or a leaf falling from a tree in the front yard.”
To jump-start the process, Ring used its Ukrainian „data operators” as a crutch for its lackluster artificial intelligence efforts, manually tagging and labeling objects in a given video as part of a „training” process to teach software with the hope that it might be able to detect such things on its own in the near future. This process is still apparently underway years later: Ring Labs, the name of the Ukrainian operation, is still employing people as data operators, according to LinkedIn, and posting job listings for vacant video-tagging gigs: „You must be able to recognize and tag all moving objects in the video correctly with high accuracy,” reads one job ad. „Be ready for rapid changes in tasks in the same way as be ready for long monotonous work.”
tl;dr: Ring винаймають операторів для ручного розпізнавання образів. Машин льорнінг та імедж рекогнішн працюють погано.
Є хтось звідти щоб підтвердити або спростувати цю інфу?