Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

For an ML model, a sofa with a tiger pattern might be a tiger, if in its training dataset tiger stripes always means tiger.

It does not have common sense.



For small models, yes. But for the kind of massive multimodal models getting trained these days, the concept of "a pattern on an object" will exist. It doesn't need to have seen a tiger sofa before. But tiger + sofa + patterns on an object is enough to not run from the tiger print.




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

Search: