Style F 36 by 48

1
Inferring individual perceptual experience from
MEG Robust statistics approach Andrey
Zhdanov1,4, Talma Hendler1,2, Leslie
Ungerleider3, Nathan Intrator4 1Functional Brain
Imaging Unit, Tel Aviv Sourasky Medical
Center 2Psychology Department and faculty of
Medicine Tel Aviv University3Lab of Brain and
Cognition, NIMH, NIH, Bethesda, USA 4School of
Computer Science, Tel Aviv University
Introduction
Classifier Construction
Inferring subjective experience from functional
brain signals is one of the major goals of brain
research. It has been attempted in several
branches of neuroscience such as EEG-based
brain-computer interfaces and fMRI data analysis.
Here we introduce inference methodology that is
based on machine learning techniques and apply it
to infer perceived stimulus in magnetoencephalogra
m recordings.
By assuming a timepoint-wise correspondence among
the segmented signals and assigning each signal a
target label (face or house) we formulate the
stimulus inference problem as a classical
high-dimensional machine learning problem. We
then address the problem by using a regularized
Fisher Linear Discriminant analysis (see above).
Our investigation focuses on 1. Selecting the
optimal subset of MEG channels and timepoints 2.
Selecting the optimal value of the regularization
parameter ? 3. Investigating the relation between
those values, in particular the structure of
?-time space. 4. Obtaining and interpreting the
resulting MEG sensor weight maps.
Fisher Linear Discriminant
Fisher separation measure The Fisher separation
measure d between two sets of scalars x1, x2, ,
xn and y1, y2, , ym is given by where µx
and µy are means and sx and sy are standard
deviations of the two sets. Linear Discriminant
Analysis (LDA) Given two sets of vectors X
x1, x2, , xn and Y y1, y2, , ym we search
for a direction p that maximizes the Fisher
separation of the projections of X and Y on p
px1, px2, , pxn and py1, py2, , pym.
The solution is given by where S Sx Sy is a
sum of covariance matrices for X and Y and µx, µy
vector means of X and Y.
Results
A
B
C
Curse of dimensionality and regularization In
many practical applications the dimension of the
data is larger than the number of samples. In
such cases the classical LDA problem is ill-posed
the covariance matrix S from (ii) is singular.
In such cases some regularization of the problem
is required. One commonly used regularization
method is to replace S with regularized matrix
S where ? is a regularization parameter and I
is an identity matrix. Optimal value of ? is
usually estimated from the data using
cross-validation.
(A) Prediction accuracy of the regularized LDA
classifier for each of the 10 subjects using
signals from 274 sensors at a single timepoint.
Error bars denote 1-std wide margin around the
average error. (B) Error rate as a function of
regularization parameter for subject ZK. Solid
blue line denotes the average error rate
estimated by cross-validation, dotted lines mark
1-std wide margin. The vertical line marks the
minimum of the smoothed error rate (red line).
Three plots below show the distribution of sensor
weights corresponding to different values of the
regularization parameter. (C) Error rate as a
function of number of sensors used for classifier
construction (for single timepoint) for subject
TE. For comparison, red line depicts the error
rate obtained by using all 274 sensors. Dotted
lines mark 1-std wide margin. All error
estimates were obtained by 100-fold
cross-validation.
Experimental Setup

• Mixed experiment design
• Visual categories were presented as events of
  either face or house.
• Emotional valence of face was presented in 40
  sec epochs of neutral or fear.10 sec blank
  separated between epochs.
• The presentation duration of each category
  within epochs was a replay of prior perceptual
  experiences in a binocular rivalry paradigm for
  face competing with simultaneously presented
  house (Tong et al, 1998). Altogether the
  experiment lasted 15 minutes.
• Data acquisition and preprocessing
• 274-channel CTF MEG recording system
• Signal sampled at 60Hz
• Signals segmented into intervals -0.3, 1 sec
  around the timepoint of stimulus switch.

Conclusions
Linear combination of MEG sensor signals created
using regularized Fisher Discriminant Analysis
was shown to be useful for inferring subjective
experience. In addition, stable spatiotemporal
patterns of discriminating weights that were
discovered might produce an insight into the
neurophysiologic mechanisms of perception.
Optimization on spatio/temporal sensors provided
a significant prediction results, suggesting more
succinct interpretation of brain activity. We
expect that further development of this technique
may provide the means to detect not only
externally-driven but also internally-driven
events in in-vivo brain signals.
Neutral epoch




Negative epoch




References
40 s
10 s
Tong, F. et al. (1998) Binocular Rivalry and
Visual Awareness in Human Extrastriate Cortex.
Neuron 21, 753-759. Friedman, J.H. (1989)
Regularized Discriminant Analysis. Journal of the
American Statistical Association 84,
165-175 Acknowledgments Dr. David Papo, Dr. Tom
Holroyd, Mariam Eapen, Yonatan Weintraub, Israel
Science Foundation, Bikura Foundation
zhdanova_at_post.tau.ac.il