Not a dumb question at all. There has been some research on using a wavelet scattering transform as a feature extractor instead of learned convolutions in the f

Not a dumb question at all. There has been some research on using a wavelet scattering transform as a feature extractor instead of learned convolutions in the first layers of deep neural networks. It works, but isn't as good, _given enough data_. For low-data regimes it makes sense to do this.