Finding the perfect NIRS pipeline

Discrete and dense optimisation are well-studied areas in isolation. Simultaneously optimising over a product of both kinds of spaces is less so. Yet, a large number of real-world phenomena exhibit this dual nature. One of such discrete-dense phenomena, is the optimization of the processing (and analysis) pipeline in fNIRS neuroimaging. For such a problem, we propose a new metric tensor to navigate the discrete-dense space that consequently permits using gradient-based optimization. We illustrate the success of our method in a synthetic fNIRS application, but we claim that the abstract solution remains applicable to other analogous domains.