Can one make a one-to-one correspondence of such numbers to a copyrighted work of art?

The way these types of networks function results in the generation of wholly unique, never before seen images. I guarantee you will never find a single copyrighted work of art that is generated by a model like this.

It seems then that the closest analog is the network is a derived work then.

Arguably it's a derived work in the same way that a person developing a taste for art after looking at thousands of images and then going and painting their own original creation is.

To be fair... that is also an apt description of a 2D image.

Sure, but in the case of the copyrighted work, we're talking about a specific combination and sequence of numbers which represent that 2-D image. The numbers here representing the weights that make up the model have nothing to do with those.

Yes, I agree that there is an argument to be made here, just not on the basis that the model is 'just a matrix of numbers.' Of course, when we're talking about this many weights, one must wonder exactly how much of the input images is actually pretty much directly encoded in the resulting network.