Caveats with Stochastic Gradient and Maximum Likelihood Based ICA for EEG

Stochastic gradient (SG) is the most commonly used optimization technique for maximum likelihood based approaches to independent component analysis (ICA). It is in particular the default solver in public implementations of Infomax and variants. Motivated by experimental findings on electroencephalography (EEG) data, we report some caveats which can impact the results and interpretation of neuroscience findings. We investigate issues raised by controlling the step size in gradient updates combined with early stopping conditions, as well as initialization choices which can artificially generate biologically plausible brain sources, so called dipolar sources. We provide experimental evidence that pushing the convergence of Infomax using non stochastic solvers can reduce the number of highly dipolar components and provide a mathematical explanation of this fact. Results are presented on public EEG data.