Newton Method for the ICA Mixture Model Jason A. Palmer 1 Scott Makeig 1 Ken Kreutz-Delgado 2 Bhaskar D. Rao 2 1 Swartz Center for Computational Neuroscience.

Newton Method for theICA Mixture Model

Jason A. Palmer1 Scott Makeig1

Ken Kreutz-Delgado2 Bhaskar D. Rao2

1 Swartz Center for Computational Neuroscience2 Dept of Electrical and Computer EngineeringUniversity of California San Diego, La Jolla, CA

Introduction• Want to model sensor array data with multiple

independent sources — ICA

• Non-stationary source activity — mixture model• Want the adaptation to be computationally

efficient — Newton method

• ICA mixture model• Basic Newton method• Positive definiteness of Hessian when model

source densities are true source densities• Newton for ICA mixture model• Example applications to analysis of EEG

Outline

-10 -5 0 5 10

-10

-5

0

5

10

-10 -5 0 5 10

-10

-5

0

5

10

ICA Mixture Model—toy example• 3 models in two dimensions, 500 points per

model• Newton method converges < 200 iterations,

natural gradient fails to converge, has difficulty on poorly conditioned models

ICA Mixture Model• Want to model observations x(t), t = 1,…,N,

different models “active” at different times• Bayesian linear mixture model, h = 1, . . . , M :

• Conditionally linear given the model, :

• Samples are modeled as independent in time:

• Each source density mixture component has unknown location, scale, and shape:

• Generalizes Gaussian mixture model, more peaked, heavier tails

Source Density Mixture Model

ICA Mixture Model—Invariances • The complete set of parameters to be

estimated is:

h = 1, . . ., M, i = 1, . . ., n, j = 1, . . ., m• Invariances: W row norm/source density scale

and model centers/source density locations:

• Transform gradient (1st derivative) of cost function using inverse Hessian (2nd derivative)

• Cost function is data log likelihood:

• Gradient:

• Natural gradient (positive definite transform):

Basic ICA Newton Method

• Take derivative of (i,j)th element of gradient with respect to (k,l)th element of W :

• This defines a linear transform :

• In matrix form, this is:

Newton Method – Hessian

• To invert: rewrite the Hessian transformation in terms of the source estimates:

• Define , , :

• Want to solve linear equation :

Newton Method – Hessian

Newton Method – Hessian • The Hessian transformation can be simplified

using source independence and zero mean:

• This leads to 2x2 block diagonal form:

• Invert Hessian transformation, evaluate at gradient:

• Leads to the following equations:

• Calculate the Newton direction:

Newton Direction

Positive Definiteness of Hessian• Conditions for positive

definiteness: • Always true for true when model source

densities match true densities:1)

2)

3)

• Similar derivation applies to ICA mixture model:

Newton for ICA Mixture Model

• Convergence is really much faster than natural gradient. Works with step size 1!

• Need correct source density model

20 40 60 80 100 120 140 160 180

-2.03

-2.02

-2.01

-2

-1.99

-1.98

-1.97

Convergence Rates

log likelihood

iterationiteration

Segmentation of EEG experiment trials

trial trial

3 models 4 models

loglikelihood

loglikelihood

iteration iteration

time time

Applications to EEG—Epilepsy

time time

time

loglikelihood

loglikelihooddifferencefrom single model

1 model 5 models

Conclusion• We applied method of Amari, Cardoso and

Laheld, to formulate a Newton method for the ICA mixture model

• Arbitrary source densities modeled with non-gaussian source mixture model

• Non-stationarity modeled with ICA mixture model (multiple mixing matrices learned)

• It works! Newton method is substantially faster (superlinear). Also Newton can converge when Natural Gradient fails

Code• There is Matlab code available!!

– Generate toy mixture model data for testing– Full method implemented: mixture sources,

mixture ICA, Newton

• Extended version of paper in preparation, with derivation of mixture model Newton updates

• Download from:http://sccn.ucsd.edu/~jason

Acknowledgements

• Thanks to Scott Makeig, Howard Poizner, Julie Onton, Ruey-Song Hwang, Rey Ramirez, Diane Whitmer, and Allen Gruber for collecting and consulting on EEG data

• Thanks to Jerry Swartz for founding and providing ongoing support the Swartz Center for Computational Neuroscience

• Thanks for your attention!

Newton for ICA Mixture Model