My understanding is that the convolutional layer of a convolutional neural network has four dimensions: input_channels, filter_height, filter_width, number_of_filters. Furthermore, it is my understanding that each new filter just gets convoluted over ALL of the input_channels (or feature/activation maps from the previous layer).
HOWEVER, the graphic below from CS231 shows each filter (in red) being applied to a SINGLE CHANNEL, rather than the same filter being used across channels. This seems to indicate that there is a separate filter for EACH channel (in this case I'm assuming they're the three color channels of an input image, but the same would apply for all input channels).
This is confusing - is there a different unique filter for each input channel?
The above image seems contradictory to an excerpt from O'reilly's "Fundamentals of Deep Learning":
"...filters don't just operate on a single feature map. They operate on the entire volume of feature maps that have been generated at a particular layer...As a result, feature maps must be able to operate over volumes, not just areas"
...Also, it is my understanding that these images below are indicating a THE SAME filter is just convolved over all three input channels (contradictory to what's shown in the CS231 graphic above):