Deep Learning
What are the two actors in a convolution operation?
Convolutional filter and a slice of an input matrix.
Card 1 of 259. Answer: Convolutional filter and a slice of an input matrix.
Convolutional filter and a slice of an input matrix.
Card 1 of 259. Answer: Convolutional filter and a slice of an input matrix.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
The filter, or kernel, and the patch of the input it covers. The filter is a small grid of learned weights, such as 3 × 3; the patch is the same-sized slice of input beneath it. Their element-wise products are summed, plus a bias, into one output value.
A vertical-edge filter with every row [1, 0, −1], on a patch with every row [10, 10, 0], gives 10 + 10 + 10 = 30, a strong edge response; a flat patch gives 0. Sliding the filter fills the feature map: a 5 × 5 input, a 3 × 3 filter, stride 1 and no padding give a 3 × 3 map, as (5 − 3)/1 + 1 = 3.
On an RGB image the patch is 3 × 3 × 3, so a filter has 27 weights plus a bias, and 64 filters have 1,792 parameters at any image size, since every patch shares them. Libraries compute cross-correlation, without flipping the kernel, which is harmless because the weights are learned.