Multivariate brain signals appear often as realizations of nonstationary processes that can be modeled as a linear mixture of latent processes. These latent processes tend to include both stationary and nonstationary components. Stationary subspace analysis (SSA) aims at factorizing the observed multivariate signal into stationary and nonstationary components.
Existing SSA methods involve a non-convex optimization problem that becomes computationally intractable in higher dimensions. We discuss computational strategies to overcome this issue. First, we investigate a community structure based SSA strategy that breaks down the problem into smaller dimensions. Second, we propose a momentum based gradient descent approach to speed up the convergence of the optimization process. Third, to deal with multiple solutions that may arise from multiple epochs we consider clustering techniques involving canonical angles between subspaces, which can be used as a measure of distance between the recovered stationary space and the true stationary space. We apply these techniques to analyze local field potentials of rats to assess the impact of an induced stroke on the brain.
|