This one is a little bit weird, I am not sure it is really useful. However, the concept is pretty fun:
Reshape and sort models a and b in the order of model c's weights
Take the fourrier transform of both resulting distributions
Merge the frequency components according to a linear filter (see below)
Turn the merged frequencies back into a weights distribution
Reorder and reshape according to c's weights order
The filter used in the spectral domain can be parameterized here: https://www.desmos.com/calculator/wsikh5hmyf
Everything in red is the ratio of the frequencies of model b that contribute to the result. x=0 is the lowest frequency component and x=1 is the highest.
When beta = 0, alpha is a hard ratio of the frequency components. a takes all the frequencies below alpha and b takes all of the frequencies above. In this case you could say a contributes the low frequencies and b the high frequencies.
This one is a little bit weird, I am not sure it is really useful. However, the concept is pretty fun:
a
andb
in the order of modelc
's weightsc
's weights orderThe filter used in the spectral domain can be parameterized here: https://www.desmos.com/calculator/wsikh5hmyf Everything in red is the ratio of the frequencies of model
b
that contribute to the result. x=0 is the lowest frequency component and x=1 is the highest.When beta = 0, alpha is a hard ratio of the frequency components.
a
takes all the frequencies below alpha andb
takes all of the frequencies above. In this case you could saya
contributes the low frequencies andb
the high frequencies.