Applications of Independent Component Analysis Terrence Sejnowski Computational Neurobiology Laboratory The Salk Institute.

Applications of Independent Component Analysis

Terrence Sejnowski

Computational Neurobiology LaboratoryComputational Neurobiology LaboratoryThe Salk InstituteThe Salk Institute

PCA finds the directions of maximum variance

ICA finds the directions of maximum

independence

Principle: Maximize Information

• Q:Q: How to extract maximum

information from multiple visual

channels?

Set of 144 ICA filters

• AA: ICA does this -- it maximizes

joint entropy & minimizes

mutual information between output

channels (Bell & Sejnowski, 1995).• ICA produces brain-like visual filters

for natural images.

Example: Audio decomposition

Play Mixtures Play Components

Perform ICA

Mic 1

Mic 2

Mic 3

Mic 4

Terry Scott

Te-Won Tzyy-Ping

ICA Applications

• Sound source separation • Image processing• Sonar target identification• Underwater communications• Wireless communications• Brain wave analysis (EEG) • Brain imaging (fMRI)

Recordings in real environmentsSeparation of Music & Speech

Experiment-Setup:- office room (5m x 4m)- two distant talking mics- 16kHz sampling rate

40cm

60cm

Learning Image Features

Learning Image Features

Automatic Image Segmentation

Barcode Classification

Matrix Linear

Postal

Learned ICA Output Filters

Matrix Postal Linear

Barcode Classification Results

Classifying 4 data sets: linear, postal, matrix, junk

Image De-noising

Filling in missing data

ICA applied to BrainwavesAn EEG recording consists of activity arising from many brain and extra-brain processes

Eye movement

Muscle activity

WHAT ARE THE INDEPENDENT

COMPONENTS OF BRAIN IMAGING?

Measured Signal

Task-related activations Arousal

Physiologic Pulsations

Machine Noise

?

Functional Brain Imaging

• Functional magnetic

resonance imaging (fMRI)

data are noisy and

complex.

I C A C o m p o n e n t T y p e s

S u s t a i n e d t a s k - r e l a t e d

( a )

T r a n s i e n t l yt a s k - r e l a t e d

( b )

S l o w l y - v a r y i n g

( c )

Q u a s i - p e r i o d i c

( d )

A b r u p t h e a dm o v e m e n t

( e )

A c t i v a t e dS u p p r e s s e d

S l o w h e a dm o v e m e n t

( f )

• ICA identifies concurrent

hemodynamic processes.

• Does not require a priori

knowledge of time courses

or spatial distributions.

ICA-2001:http://www.ica2001.org

Contact:[email protected]