[wolfium3]

If you expand a NN to decision trees, the resulting decision tree(s) can take up orders of magnitude more space and take orders of magnitude more time to run than the original NN.

The paper is discussed here: https://www.youtube.com/watch?v=_okxGdHM5b8

[m3047]

Counter argument: For every trainable NN, there exists an equivalent, sparser [edit] trained NN.

https://numenta.com/blog/2019/08/30/case-for-sparsity-in-neu...

https://web.archive.org/web/20211130211241/https://www.bhaut...

https://arxiv.org/abs/1803.03635

[titzer]

So...compressed decision trees?

[sigstoat]

at that point most things are compressed decision trees.

[TeMPOraL]

Possibly related: learning is isomorphic to compression, and the two may be in fact fundamentally the same thing.

[mncharity]

Oh, wow. I'd long thought of my own learning as "when reality surprises, isn't merely what you'd expect, isn't "obvious", then fix that, with emphasis on simplicity and generality". I'd not recognized that as compression. Tnx!