I'm curious, have you been bitten by representational comparisons before?
-
-
-
let's say I've seen those who've been bitten without realizing they have.
- Još 3 druga odgovora
Novi razgovor -
-
-
What if you take avg of the representation of an ensemble of NNs tho?
-
what is the correct way to average them?
- Još 1 odgovor
Novi razgovor -
-
-
Isn't this a byproduct of linearity of CCA/CKA, their inability to properly disentangle the representations across different networks? We have found recently http://arxiv.org/abs/1905.11141 that representations of much more complex networks (VGG/ResNet) align with a more powerful metric
-
My view isn't that CCA is 'wrong' per se, far from it. Rather we should use comparison techniques of varying power: Orthogonal - eg Proscrustes, we didn't do Linear - eg CCA Topological - eg we used fixed points and turned them into graphs for isomorphism
- Još 2 druga odgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.