Re:Transformers struggle with generalizing tasks beyond pre-...
by haltist on Tuesday November 07, 2023 @12:09PM
(#38179536)
Neural networks approximate the statistical distribution of whatever they are trained on. There is no reason to expect that any neural network will perform well on out of distribution tasks, e.g. multi-digit addition and multiplication.
[ Parent
| on HN ]
Re:Transformers struggle with generalizing tasks beyond pre-...
by ctoth on Tuesday November 07, 2023 @12:12PM
(#38179577)
This is (probably) wrong, see:
https://gwern.net/scaling-hypothesis And specifically the part where it discusses addition under the heading Blessings of Scale.
Re:Transformers struggle with generalizing tasks beyond pre-...
by haltist on Tuesday November 07, 2023 @12:16PM
(#38179636)
So why can't GPT do arithmetic? At what scale do you suppose it will manage to add numbers?
Re:Transformers struggle with generalizing tasks beyond pre-...
by og_kalu on Tuesday November 07, 2023 @12:34PM
(#38179943)
GPT-4 can do arithmetic just fine. In fact, despite being severely hobbled by Tokenization (https://arxiv.org/abs/2110.09485), it can perform arithmetic better than you could without a dedicated tool or pad.
Re:Transformers struggle with generalizing tasks beyond pre-...
by Der_Einzige on Tuesday November 07, 2023 @12:42PM
(#38180060)
Besides your answer on tokenization being right on the mark (gwern even has articles discussing this exact issue - https://gwern.net/gpt-3#bpes), that linked paper is extremely powerful for theoretical thinking about this problem. I literally just left a comment claiming that almost all of human thought is encoded in LLMs, leading almost all tasks to be considered interpolation.
If this is true, I suppose based on what we know of the curse of dimensionality it makes sense. Very thought-provoking work.
Re:Transformers struggle with generalizing tasks beyond pre-...
by og_kalu on Tuesday November 07, 2023 @12:49PM
(#38180193)
Damn I linked the wrong paper lol.
Meant to link this. More pertinent to my point on Tokenization https://arxiv.org/abs/2310.02989