It's good that we get visualization tools to explore artificial neural networks: the problem of "transparency" - understanding what an instance of ANN does instead of remaining a "black box", hence "teaching us something" as opposed to "just somehow working" -, is in fact much reduced when the right visualization - e.g. revealing the patterns that the connectional weights come to identify - is applied.
It is a primary goal to understand what causes an emergence of function. ...Which should be necessary to build (and tweak, and hack, and refine, and export) further. (This point of view should be a basic tenet, yet it does not seem to get the proper focus in this realm.)
Neural networks feel like a type of strange magic, like a hack that shouldn’t work but definitely does. This helps reduce that feeling, in the best way possible.
The wisdom of that curly activation can be debated, since in every extreme-point (horizontal) the gradients are zero, and all activations will eventually be trapped or centered there. Histogram plots can be revealing to understand activations.
For me personally, it really helps to be able to see things like how the responses compose to create the outputs.
I know a lot of math-minded people are able to intuitively "get" concepts from symbolic representations, but yeah it definitely makes it easier for me to be able to poke at something while figuring out an unfamiliar topic.
I recall a loner electrical engineer who found a way to put sensors on various parts of a PC motherboard, and had visualization of it. Some things are predictable there, of course, but the main system messaging bus was hours of interest IIR.
It is a primary goal to understand what causes an emergence of function. ...Which should be necessary to build (and tweak, and hack, and refine, and export) further. (This point of view should be a basic tenet, yet it does not seem to get the proper focus in this realm.)