10_gen_2017

MULTI-SCALE CONTEXT AGGREGATION BY DILATED CONVOLUTIONS

Questions:

[@priba] Que aporta les dilated convolutions que no aconseguíssim amb el algorithme a trous? En una nota del paper diu "Some recent work mistakenly referred to the dilated convolution operator itself as the algorithme a trous. This is incorrect. The algorithme a trous applies a filter at multiple scales to produce a signal decomposition. The algorithm uses dilated convolutions, but is not equivalent to the dilated convolution operator itself." Si l'algorithme a trous ja usa dilated convolutions, entenc que és bo que ens definim aquest operador, però quina novetat hi ha quant a les convolucions?

[@prlz77] Pel que he vist, les dilated les utilitzen per oversampling però no veig que les utilitzin per un model normal de classificació, tindria sentit fer-ho? Tindria sentit aplicar dilated en els canals en comptes de en l'espai?