1. Suppose we have an image of 3x32x32. This means the image has three channels (possibly RGB) each of dimensions 32x32. If we apply one convolutional filter of 3x3 to this image then the filter will have dimensions 3x3x3 because the image has three channels. We apply each channel of the convolutional filter to the corresponding channel in the image. As an example consider an image below of 2x5x5, in other words 2 channels each of 5x5, and one convolutional filter of dimensions 2x3x3. N 3 7 2 1 co 4 10 4 2 -1 -1 1 -19-18 ? -1 1 .1 4 5 8 9 2 ? ? ? 5 1 -1 6 7 2 1 ? ? ? -4 3 ? 6 9 5 0 4 ? 2 ? 1 7 4 8 9 ? ? ? -1 1 1 15 15 ? 4 6 3 1 0 1 1 1 1 ? ? ? 0 0 4 2 6 2 132 -1 1 1 ? ? ? 1 8 1 1 3 7 0 What about the activation? Do we activate before or after averaging? Usually it is after averaging. In the above case the activation would give -1 and -1 in the final output of the convolution