Amit Aflalo,
Oren Freifeld
arXiv preprint 2026
CPAB transformations are always bi-Lipschitz homeomorphisms but not necessarily diffeomorphisms. This note identifies sufficient regularity conditions that guarantee a CPAB transformation is a diffeomorphism, and shows that they are easy to enforce and hold with probability one for common tessellations.
Amit Aflalo,
Eran Treister,
Chaim Baskin,
Oren Freifeld
NeurIPS 2026
Conference on Neural Information Processing Systems
FG-CPAB separates trajectory integration from parameter-space projection, reducing the gradient cost of CPAB transformations from $\mathcal{O}(dTN)$ to $\mathcal{O}(TN + dC)$. This yields up to $10^4\times$ faster gradients and $500\times$ lower memory, making fine-tessellation diffeomorphisms practical in 2D and 3D.
Amit Aflalo,
Shahaf E. Finder,
Roy Amoyal,
Eran Treister,
Oren Freifeld
TMLR 2026
Transactions on Machine Learning Research
An I/O model of wavelet convolutions (WTConv) guides three algebraic reformulations that enable a fused, I/O-aware implementation, reducing memory traffic by about $2.55\times$ and yielding up to $4.35\times$ faster training at roughly half the peak memory.