By identifying mathematical equivalences in IKPLS computations, you can reduce Y loading calculation from O(KM) to O(M) operations, delivering 2-6× faster model training on CPUs and GPUs without changing results.
This paper accelerates Improved Kernel Partial Least Squares (IKPLS), a fast algorithm for training regression models. The author optimizes two computational bottlenecks: replacing sequential accumulation with direct evaluation for rotation matrices, and discovering that Y loadings can be computed from previously calculated values rather than from scratch.