I attended a talk by Stephane Mallat a few months back about the work his lab is doing on wavelet scattering transforms.
It seems they're making great progress in developing deep neural networks based on them that don't need to be learned.
It's not as hyped as mainstream deep learning methods but I think it holds a lot of promise since it cuts down on learning time, is mostly unsupervised, gives you control over the network's features and is a more theoretically principled way to build networks than just defining the loss and crossing your fingers while it's optimizing as we're doing right now (bit of an hyperbole).
I could go on, it's absolutely fascinating. More info (in english & french): https://edouardoyallon.github.io/thesis.pdf