When error is back propagated in a convolution layer in CNN, do we need to average the error deltas ? As a 5x5 filter mask may traverse a pixel more than once (except some of the boundary pixels).

asked Oct 24 '13 at 00:10

gul's gravatar image

gul
30334

I am using stochastic gradient method (no batch processing).

(Oct 24 '13 at 00:15) gul

pushing error deltas from a layer with feature maps of size 9x9 to a layer with features maps of size 13x13, the error deltas may be computed following number of times

count13by13[169] ={1,2,3,4,5,5,5,5,5,4,3,2,1, 2,4,6,8,10,10,10,10,10,8,6,4,2, 3,6,9,12,15,15,15,15,15,12,9,6,3, 4,8,12,16,20,20,20,20,20,16,12,8,4, 5,10,15,20,25,25,25,25,25,20,15,10,5, 5,10,15,20,25,25,25,25,25,20,15,10,5, 5,10,15,20,25,25,25,25,25,20,15,10,5, 5,10,15,20,25,25,25,25,25,20,15,10,5, 5,10,15,20,25,25,25,25,25,20,15,10,5, 4,8,12,16,20,20,20,20,20,16,12,8,4, 3,6,9,12,15,15,15,15,15,12,9,6,3, 2,4,6,8,10,10,10,10,10,8,6,4,2, 1,2,3,4,5,5,5,5,5,4,3,2,1};

do we need to average the error gradient accordingly ?

(Oct 24 '13 at 19:37) gul

One Answer:

No, they are added together.

Jake Bouvrie from MIT has some good notes for implementing CNNs. It's targeted at matlab but shouldn't be too hard to adapt to other languages since the descriptions are quite thorough.

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.70.1419&rep=rep1&type=pdf

answered Oct 28 '13 at 07:42

Henry%20Robbins%20Gouk's gravatar image

Henry Robbins Gouk
312

Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.