Questions about clifford_kernel.
xwqtju opened this issue · comments
May I inquire, what is the difference between get_2d_clifford_kernel and these two get_2d_clifford_rotation_kernel? Additionally, is it possible to construct a complete 4D PGA kernel, similar to how get_3d_clifford_kernel is defined?
You might find our papers on Clifford Layers and Geometric Clifford Layers. On a high level, get_2d_clifford_kernel implements the standard Clifford interpretation of CNNs replacing vector multiplication with multi-vector geometric product, while get_2d_clifford_rotation_kernel is the Clifford interpretation of "Quaternion CNNs" and applies a rotational transform. This always seemed to do better in our experiments than the standard Clifford CNN. In the GCAN paper we provide a better understanding of why (better preservation/transformation of geometric properties).
Thx! It seems that a deeper understanding of GCAN requires more time.