Not the author, but I don't think visualizing neural networks is useful for understanding the general concept. I think it's useful for understanding a specific

Not the author, but I don't think visualizing neural networks is useful for understanding the general concept. I think it's useful for understanding a specific network. For example, if you see the weights change extremely quickly, that means your learning rate is probably too high. If you run a network 3 times and see completely different sets of weights, that tells you your model is highly unstable (not necessarily a bad thing.) It can also give you a sense of whether you have a lot of features that matter or just a few.