ICA Projects at CNL
Te-Won Lee, Michael Lewicki, Tony Bell(at Interval Research) Terry Sejnowski
Scott Makeig, Tzyy-Ping Jung, Martin McKeown Colin Humphries, Terry Sejnowski
Definitions of terms -- EEG, MEG, ERP, ERF
Electromagnetic fields associated with brain processes and recorded outside the head produce electroencephalographic (EEG) and magnetoencephalographic (MEG) data. Averages of EEG epochs time-locked to a set of experimental events of interest are called event-related potentials (ERPs). Similar magnetic averages are known as event-related fields or ERFs.
Suitability for ICA decomposition
In the EEG/MEG frequency range (roughly 0.1-100 Hz) the mixing of brain fields at the scalp electrodes is basically linear. Although skull attenuates EEG signals strongly and "smears" (low-pass filters) them spatially, this does not affect the linear relation between potential in the brain and potential at the scalp. Fields propagate to the sensors (electrodes or SQUID coils) through volume conduction without significant delays. This makes EEG and MEG data suited to linear decomposition via ICA. A number of "frequently asked questions" about the application of ICA to averaged or spontaneous EEG/MEG data are answered in Frequently Asked Questions about ICA applied to EEG/MEG data.
First Applications
The ICA algorithm of Bell & Sejnowski was first applied to EEG and ERP data in Makeig S, Bell AJ, Jung T-P, and Sejnowski TJ, "Independent component analysis of electroencephalographic data." Advances in Neural Information Processing Systems 8, 145-151,1996. This paper demonstrated the successful decomposition of 14-channel ERP data consisting of only 624 data points. Further details have now been published in a PNAS paper on ICA applied to ERP data. Preprint html and Postscript versions of this paper are also available for review and download from this site.
Toolbox
A Matlab toolbox for EEG/MEG analysis using ICA is also available for download. The toolbox consists of scripts for ICA decomposition and plotting of results, together with general-purpose EEG plotting and computational routines. A demo script (icademo) illustrates application of the ICA routines to both synthetic and actual ERP data.
http://www.cnl.salk.edu/~scott/ica.html
View summary of
recent changes to the toolbox.
Martin McKeown, Tzyy-Ping Jung, Scott Makeig, Terry Sejnowski
fMRI data
fMRI data is a complicated mixture of different sources of variability: cardiac and respiratory pulsations, subtle head movements, task-related activity changes and machine noise. Changes related to the performance of psychomotor tasks may constitute as little as 10-15% of the variance of the Blood Oxygen Level Dependent (BOLD) contrast signal in a 1.5T magnet, so extracting the small task-related changes from the measured signal is difficult.
ICA decomposition of fMRI data
ICA, in the manner applied to ERP and EEG (see above), is inappropriate for fMRI analysis because the number of "channels" (i.e. voxels) greatly exceeds the number of time points in a typical fMRI experiment. In 1997, it was first proposed to look for spatially independent patterns of activity in fMRI data [ref]. This assumes that the spatial distributions associated with each of the above sources of variability are independent, and that the contributions from each spatial pattern sum linearly to represent the data. The time courses associated with the different spatial patterns can potentially be correlated, allowing for the detection of spatial patterns whose time courses are transiently task-related (TTR) as well as consistently task-related (CTR). The criteria of spatial independence appears to be a powerful way to separate task-related activations from other sources of variability making up the BOLD signal, as explained in Frequently Asked Questions about ICA applied to fMRI data.
Further details have been published in a PNAS paper of ICA applied to fMRI data (which can be downloaded) and a Human Brain Mapping paper.
Marni Bartlett, Terry Sejnowski
In a task such as face recognition, much of the important information
may be contained in the high-order relationships among the image pixels.
Some success has been attained using data-driven face representations based
on principal component analysis, such as "Eigenfaces" (Turk & Pentland,
1991) and "Holons" (Cottrell & Metcalfe, 1991). Principal component
analysis (PCA) is based on the second-order statistics of the image set,
and does not address high-order statistical dependencies such as the
relationships among three or more pixels. Independent component analysis
(ICA) is a generalization of PCA which separates the high-order moments of
the input in addition to the second-order moments. We developed image
representations based on the independent components of the face images and
compared them to a PCA representation for face recognition.
ICA was performed on the face images under two different
architectures. The first architecture provided a set of statistically
independent basis images for the faces that can be viewed as a set of
independent facial features. These ICA basis images were spatially
local, unlike the PCA basis vectors. The representation consisted of the
coefficients for the linear combination of basis images that comprised each
face image. The second architecture produced independent coding variables
(coefficients). This provided a factorial face code, in which the
probability of any combination of features can be obtained from the product
of their individual probabilities. The distributions of these coefficents
were sparse and highly kurtotic. Classification was performed using nearest
neighbor, with similarity measured as the cosine of the angle between
representation vectors. Both ICA representations were superior to the PCA
representation for recognizing faces across sessions, changes in
expression, and changes in pose.
Papers
on face image analysis using ICA by Marian Stewart Bartlett.