1
$\begingroup$

See this example: convnet quiz Udacity. How to get from input depth = 3 to output depth = 8? My assumption: In this example we have 8 filter (kernels) and each of them slides over the 3 inputs. So in total we have 24 convolutions. That would give me a depth of 24? So how to reduce to eight?

Update: I found this mapping table by Yann. LeNet5, see page 8 Table1. However the question is, whether this table is still used in the same fashion as in early convnets or today we might use a different mapping sheme? E.g. just sum 3 filtered maps into one of the 8 output maps?

$\endgroup$
0
$\begingroup$

After the input layer, depth is simply the number of filters. From the official Stanford course materials:

"The depth of the output volume is a hyperparameter: it corresponds to the number of filters we would like to use."

Source: http://cs231n.github.io/convolutional-networks/

$\endgroup$
  • $\begingroup$ See my update. Your answer cannot be correct. $\endgroup$ – Randy Welt Jul 26 '16 at 16:05
  • $\begingroup$ @RandyWelt From the official Stanford class materials on GitHub: "First, the depth of the output volume is a hyperparameter: it corresponds to the number of filters we would like to use." Source: cs231n.github.io/convolutional-networks $\endgroup$ – Ryan Zotti Jul 26 '16 at 17:15
  • 1
    $\begingroup$ @RandyWelt You're right though that the table in the paper you reference seems to contradict this. There are many subtly different flavors of CNNs, and while there is no one "right" way to treat filters, there is one widely used way (for now, at least), which is what I provided in my original answer. In some cases I've have seen connections that will skip an entire layer, but again, that's not widely adopted. $\endgroup$ – Ryan Zotti Jul 26 '16 at 17:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.