What exactly does "channel" refer to in tensorflow/lucid?

3

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.

I guess they're talking about filter. They optimize an image to make a high respond on the filter (channel) – Jérémy Blain – 2018-08-29T08:04:36.117

(I don't read the papers, I don't see them on the github page. Moreover, they don't seem to talk about channel on the github page you've linked) – Jérémy Blain – 2018-08-29T08:06:04.653

Ohh, my bad, I didn't see there are two links ^^" (I edited your post to better see the two links) – Jérémy Blain – 2018-08-29T08:09:49.600

In this image, the channel is 5. There are 5 $3*2$ filters