stokesj / ewc Goto Github PK

As I understand it, the fisher weights of x for weights are the squares of the gradients d/dw (log x). (Let me know if this is incorrect.)

In my own implementations I use the loss term for x, but in other implementations I see the model prediction used instead. I can't remember why I used this, although I suspect it's because I needed a value >0 (for which raw predictions of a Q-network would not work).

Should the Fisher matrix be based on the raw output rather than the error? (In practice, it seems to work based on the error, which makes sense, although I suspect basing it on the output would work better.)

I am not sure i understand correctly.

it looks like you are compute fisher matrix for current task with samples after the training a current task. am i right?

Then, when moving to third task... does fisher diagonal compute for all previous tasks? or just compute previous task only? I think if the former, EWC is useless because we do not want to see previous. right?

i also implement the EWC using tensorflow. but, i am not sure that i am right. so i am referring your code :D. Thank you for your advanced work!

p.s why do we need to compute fisher diagonal per example? why not per batch?