An attack that recovers training data from GPT-2.http://feedproxy.google.com/~r/blogspot/gJZg/~3/Qm9ni79lltY/privacy-considerations-in-large.html …
Show this thread
-
I’m not super familiar with the field, but doesn’t this also imply these large language models are “memorizing” to a very large extent vs truly “learning”. Can you measure this?
Show this thread
Yes they memorize, but they also have some ability to generalize from what they've memorized, which is what makes them useful. They're not like a dict, more like a locality-sensitive hashtable with a learned distance function, enabling them to interpolate
-
-
That makes sense. I’m curious if you can measure this? Like, if you generate billions of examples, do you get less diversity than you would expect from a model with “true” ability to generate novel text? I recall seeing some research along those lines
-
Measuring generalization is hard, and gets harder the more data the model is trained on. We don't have a good way to do it right now, nor is there much research interest in the topic (but there should be)
- Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.