KAN network

[2404.19756] KAN: Kolmogorov-Arnold Networks

KANs have a fully-connected structure similar to MLPs

but instead of fixed activation functions on nodes (neurons), they have learnable activation functions on edges (weights).

This is the original KAN. n -> 2n -> 1

Then the author generalized the KAN network, make it deeper and flexible.

Why? why just stack.

No further explaining, but SIMPLY stacking.

Grid

This is kind like concept of resolution, a computation complexity and fidelity trade-off.

Activation function

It seems that the activation function choosing has no specific reason (as if it fit the previous assumption.), or I do not aware - I just skimmed the paper.