Independent Component Analysis for functional Magnetic resonance imaging, why it works even though it can’t possibly work.

Ingrid Daubechies
Princeton University

In functional magnetic resonance imaging (fmri) one observes (indirectly) a noisy and smoothened version F (v, t) of the neuron activity at point v in the brain and at tone t, typically while the person (whose brain is being imaged) performs some type of task. The hope is to understand brain function for complex tasks by analyzing F. Unfortunately, the brain does many other things at the same time as working on the task, and it is necessary to separate F into several components. Often the task-related component is identified by correlation with the tone course of the task. In some applications one wants to sidestep this correlation, and so-called “blind source separation” is called for. One method that has been proposed, with varying success, is independent component analysis (ICA).

The talk will review what ICA is, where it comes from, why it can’t possibly work for fmri even though it does (sort of), and discuss where this leads us.

Back to MGA Workshop V: Math Analysis and Multiscale Geometric Analysis