Tensorflow/Lucid is able to visualize what a "channel" of a layer of a neural network (image recognition, Inception-v1) responds to. Even after studying the tutorial, the source code, the three research papers on lucid and comments by the authors on Hacker News, I'm still not clear on how "channels" are supposed to be defined and individuated. Can somebody shed some light on this? Thank you.
The channel they are talking about is the depth of the layer L.
In this image, the channel is 5. There are 5 $3*2$ filters
They optimize an image from noise, to better respond to the filter F, in the layer L. Then you obtain this kind of image, and you can try to interpret the prupose of that filter (it learns to detect eyes, faces, wheels, and so on)
EDIT: To be more precise, you visualise the channel only if you optimise images for every filters, otherwise, you obtain a filter visualisation (if you do the optimisation one time, for one filter)
They call it "channel", because in a colored image, you have 3 dimensions : witdh, height, and channel (for color), as layers which have also 3 dimensions.