Does g depend on inputs? if not then it can be taken into the weight matrix

yes

thanks

But, the singular vectors of the matrices do change?

so decoupled basis does change?

thanks

Was there a pattern observed whether certain 1D network gradient has a correlation with difficult of the examples ?

Yep, meant difficult to classify

Like images with blur or glare are difficult to classify

Maybe the examples we see now will help

Naive question - In a class incremental setting, when a new class arrives for a particular task - will it create a new pathway or just fine tune the existing pathway for that task?

Interesting, thanks :)

Follow up question - whether the learning dynamics on these pathway structures have been investigated under curriculum learning setting as compared to passing shuffled data samples during training?

In the top plot, since the rates are comparable, why is XOR path completely suppressed?

Great talk, can you comment on how/if this extends to recurrent networks?

Great, talk!

thanks a lot

Thank you

:)

Thank you.