What is the use of advanced convolution?

Question

I refer to Layered Context Aggregation using expandable consoles .

• A 2x2 core has holes in it, so that it becomes a 3x3 core.
• A 3x3 core has holes in it, so that it becomes a 5x5 core.
• The above assumes, of course, interval 1.

I can clearly see that it allows you to effectively use 4 parameters, but has a susceptible field of 3x3 and 9 parameters, but has a susceptible field of 5x5.

Is the case of extended convolution just saving on parameters, benefiting from a larger susceptible field and thereby saving memory and computation?

Answer 1

TL; DR

• (. Convolutions)

• , , .

• , .

• Strided Convolutions - , .

@Rahul WaveNet, 2.1 Dilated Causal Convolution. . :

• (a) 1- 3x3. , 3x3.
• (b) 2- 3x3. - , 0. , 5x5 9 0, . 7x7, 3x3. ( , , ).
• (c) 4- 3- . 9x9 9 0. () , 7x7 , , , 7x7 .

, :

• 3 33 1, 7x7 . 1515 . .
• 3 33 , , 15x15 , , . , - , . .

Answer 2

, , , , :

• / . , , .
• . , 1D , , .

WaveNet, , .

, .