The general KAN framework uses learnable activation functions on the edges of the computational graph followed by summation on nodes ... functions are replaced by grids of re-weighted sine functions ...