Please enable JavaScript.
Coggle requires JavaScript to display documents.
NN_Conv_Layers (Pooling (with mask 2x2 (stride=1=default
overlap in…
NN_Conv_Layers
Pooling
-
typical-filter: max
mean, L^2 norm,...
other benefit: detect feature
invariant to scale/orientation change
generalize the lower-level-info
help object detection
we may add (zero-) padding
(to left-bottom) for programming
reason but using RLU, practically
drop these dummy padding
stride=conv-mask-shift
during operation
stride[2,2], shift-2 in both x,y
direction
with mask 2x2
-
stride 2,2
no-overlap
output-size-reduced
-
-