/sci/ - Science & Math » Thread #13845875

372KiB, 1152x721, 1626365034816.png

View Same Google iqdb SauceNAO

Anonymous Fri 12 Nov 20:16:02 2021 No.13845875 View Reply Original Report

Quoted By: >>13845929

Are there any advantages to Fourier transforming time series data before passing it into a neural network? Are there any architectures that are better for analyzing data in the frequency domain?

Anonymous

Anonymous Fri 12 Nov 2021 20:32:51 No.13845929 Report

Quoted By: >>13845975

>>13845875
I have a sneaking suspicion that neural networks learn to do it on their own.

Anonymous

Anonymous Fri 12 Nov 2021 20:47:10 No.13845975 Report

Quoted By: >>13846259

>>13845929
If that's the case I wonder if that's why adding self attention to convolutional networks shows improvements, since convolution in the time domain is equivalent to multiplication in the frequency domain. I'm thinking that a FT equivalent neural network would consist of a bunch of chained attention maps with two terminating dense layers and an output layer. Would that make sense or am I just retarded?

Anonymous

Anonymous Fri 12 Nov 2021 22:16:19 No.13846259 Report

Quoted By:

>>13845975
imo it sounds too complicated. I would just take the difference of each point in the time series. so you get something like: [-3,+4,-20,+3,+8], then feed that into a transformer.

What kind of data are you working with?

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches