Chipping away at the "opacity" of ANNs

Are deep neural networks really black-boxes whose inner workings are opaque?

Not quite:

"... for any feed-forward ReLU network, there exists a global linear approximation to a 2-hidden-layer shallow network with a fixed number of units." https://t.co/dDkcVGRyWV pic.twitter.com/belqidsox9

— Chomba Bupe (@ChombaBupe) September 24, 2023