1
$\begingroup$

I understand that as a network learns about an output with regards to an input, weights are updated according to how wrong the guess was for that node. So, over time, the weights move in the "direction" towards the correct value.

Is it possible to use a separate neural network that takes as input the weights of the first network while it trains to try approximating that "direction" and, in effect, pushing the weights in that direction faster?

$\endgroup$
2
$\begingroup$

That is similar to a chicken-and-egg problem; If the second network could figure out the direction and learn faster, why couldn't the first network do that too and learn even faster.

However, I would recommend looking up Actor Critic Methods. Here there are two networks but one (the critic) evaluates the other (the actor) to head it in the right direction.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.