In week 4, lecture Style cost function to compute the correlation of neurons across channels around 2:30 Andrew says the neuron (1 of 9) corresponds to the red channel. The slide shows 9 image patches for each of the 9 neurons.
How do we know that a single neuron is getting maximally activated from the 9 image patches from the same channel (e.g. red)?
Where does this number 9 (9 neurons) come from? Is it from the output (nHxnW) size?
How are we choosing the 9 neurons because there are more than 9 neurons if we consider the output channels?
Prof. Andrew also says, “Let’s say for the sake of arguments, that the red channel corresponds to this neurons so we’re trying to figure out…”
So, this is just assumptions.
As mentioned by Saif,
It is for the purpose of illustrating the concept.
If you are interested in reading further, I recommend referring to the Visualizing and Understanding Convolutional Networks paper, which might help you!
That’s not the point though. Even if it’s an assumption my question is can all of the 9 patches for a single neuron come from a single red channel?