Hacker News new | comments | show | ask | jobs | submit login

I’m not sure, I think I would maybe break classes into multiple labels, but that becomes even more finicky to train. At the end of the day, there are many more things that are not hotdogs, than things that are hotdogs, so you do have to provide more examples of the not hotdogs to train something from scratch properly — I don’t see a way around it.

Honestly I think the biggest gains would be to go back to a beefier, pre-trained architecture like Inception, and see if I can quantize it to a size that’s manageable, especially if paired with CoreML on device. You’d get the accuracy that comes from big models, but in a package that runs well on mobile.




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: