FAST CONVERGENT FACTORIAL LEARNING OF THE LOW­DIMENSIONAL INDEPENDENT MANIFOLDS IN OPTICAL IMAGING DATA

Penio S. Penev, http://venezia.rockefeller.edu/, http://camelot.mssm.edu/
PenevPS@IEEE.org, kaplane@rockvax.rockefeller.edu

In many functional­imaging scenarios, it is a challenge to sep­ arate the response to stimulation from the other, presumably in­ dependent, sources that contribute to the image formation. When the brain is optically imaged, the typical variabilities of some of these sources force the data to lie close to a low­dimensional, non­ linear manifold. When an initial probability model is derived by the Karhunen­Lo‘eve Transform (KLT) of the data, and some fac­ tors of this manifold happen to be accessibly embedded in suitably chosen KLT subspaces, vector quantization has been used to char­ acterize this embedding as the locus of maximum likelihood of the data, and to derive an improved probability model, in which the factors---the dynamics on this locus and away from it---are esti­ mated independently. Here we show that such a description can serve as the starting point for a convergent procedure that alterna­ tively refines the estimates of the embedding of, and the dynamics on, the manifold. Further, we show that even a very crude initial estimate, from a heavily mixed subspace, is sufficient for conver­ gence in a small number of steps. This opens the possibility of hierarchical semi­blind separation of the independent sources in optical imaging data, even when their contributions are nonlinear.