So, dumb question...
If you were training some deep learning model...
...should you be trying a few Wavelet transforms on your inputs, and feeding those in to your model, too, to see if your model performs better with wavelet inputs?
If you were training some deep learning model...
...should you be trying a few Wavelet transforms on your inputs, and feeding those in to your model, too, to see if your model performs better with wavelet inputs?
How do you determine if they are good for your application and how do you choose which family of wavelets to apply?
You’ve just described modern machine learning.
Do you know how we discovered X-Rays? Henri Becquerel realized his photographic plates had been darkened after being left in a drawer with uranium sulfate.
Do you know how electric guitar distortion was discovered? Willie Kizart dropped his Fender amp.
Worse things have happened than experimenting by throwing one more transform on your inputs before processing them.
CNNs have gotten so good though it seems a little moot.
The chief advantage of this technique is a) certain regular wavelets have performance benefits over convolutions, especially when talking low level/FPGA/ASIC space b) being not learned can be beneficial, as these is nothing to overfit.
I can see it being handy for embedded/rasppi like applications. In particular, you can festoon a crude face detector with a dozen Haar filters.
End to end training is just soooo convenient though.