|
@AdamMarblestone | |||||
|
3) my favorite, highly specific internal loss functions pic.twitter.com/X0Zc65LdvV
|
||||||
|
||||||
|
jovo
@neuro_data
|
8. pro |
|
here’s a conjecture: biological learning is never unsupervised. Rather, its goal is to learn representations that are useful for future behavior. @AstroKatie @AToliasLab @danilobzdok
|
||
|
|
||
|
Blake Richards
@tyrell_turing
|
8. pro |
|
I think this is a weird phrasing. You can have an unsupervised learning algorithm that leads to representations that are better for downstream RL or supervised tasks. The purpose of the unsupervised alg is indeed beteer downstream performance, but the alg is still unsupervised.
|
||
|
|
||
|
jovo
@neuro_data
|
9. pro |
|
yes, that is my point that it is weird phrasing. you are a tiger and you just attached a rhino and didn't kill it. now, you are watching the rhino, without "labeled data". why? to build a better model *that enables you to act better in the future attack*.
|
||
|
|
||
|
jovo
@neuro_data
|
9. pro |
|
i guess my point is: calling it unsupervised is a bit misleading, it has a specific goal, and it uses past training data to get better at that goal, not simply "learn a parsimonious model" for its own sake.
|
||
|
|
||
|
Blake Richards
@tyrell_turing
|
9. pro |
|
True, but the reason I would still call it "unsupervised" is because the learning algorithm itself is using only the sensory data, no other rewards or external targets, to determine the parameter updates.
|
||
|
|
||
|
Adam Marblestone
@AdamMarblestone
|
9. pro |
|
You can break this down into a few possibilities
1) something like this, which is not simply optimizing some parsimonious/simple objective like “compression” or “reconstruction”, and whose goal is downstream task performance, but is still unsupervised:
arxiv.org/abs/1804.00222
|
||
|
|
||
|
Adam Marblestone
@AdamMarblestone
|
9. pro |
|
2) unsupervised, e.g., predictive, learning to keep certain representations around, which then happens to make it easier to learn from external rewards
arxiv.org/abs/1803.10760
|
||
|
|
||