Decomposing event related EEG using Parallel Factor
Morten Mørup
Informatics and Mathematical Modeling Intelligent Signal Processing
Technical University of Denmark
Outline
Non-negativity constrained PARAFAC
Application of PARAFAC to the EEG
(Harshman & Carrol and Chang 1970)
Alternating Least Squares (ALS)
ALS corresponds to maximizing the likelihood of a Gaussian
Consequently, ALS assumes normal distributed noise.
Gradient descent
Especially good for cost functions without analytical solution.
Let C be the cost function, then update the parameters
according to:
Why imposing Non-negativity constraints
Most PARAFAC algorithms known to have problems of degeneration among the factors
Degeneration result of factors counteracting each other.
Some solutions:
Sparseness/regularization constraints i.e. c
1||A||
2+c
2||B||
2+c
3||S||
2Orthogonality constraints, i.e. A
TA=I
Non negativity constraint on all modalities
(if data is positive and factor components considered purely additive)
How to impose non-negativity constraints
Active set algorithm (Bro & Jong, 1997)
Iteratively optimizes cost function until no variables are negative.
Gradient descent with positive updates
Update parameters so they remain in the positive domain.
Among various other methods
Non-negative matrix factorization (NMF)
Generalization to PARAFAC
(Lee & Seung 2001)
Electroencephalography (EEG)
EEG measures electrical potential at the scalp arising
primarily from synchronous neuronal activity of pyramidal cells in the brain.
Event related potential (ERP) is EEG measurements
time locked to a stimulus event
History of PARAFAC and EEG
Harshman (1970) (Suggested its use on EEG)
Möcks (1988) (Topographic Component Analysis) ERP of (channel x time x subject)
Field and Graupe (1991)
ERP of (channel x time x subject)
Miwakeichi et al. (2004)
EEG of (channel x time x frequency)
Mørup et al. (2005)
ERP of (channel x time x frequency x subject x condition)
time
time
frequency
Wavelet transform
Complex Morlet wavelet - Real part - Complex part
Absolute value of wavelet coefficient
Captures frequency changes through time
time
channel
subjects
Möcks (1988)
Field & Graupe (1991)
time
frequency
channel
Miwakeichi (2004)
PARAFAC Assumption:
Same signal having
Various strength in each subject mixed in the channels.
PARAFAC Assumption:
Same Frequency signature present to various
degree in time mixed in the channels.
The Vector strength
Vectors coherent, i.e. correlated Vectors incoherent, i.e. uncorrelated
Vector strength a measure of coherence
Visual Paradigm
(Herrmann et al. 2004)
Expected result: Coherence around 30-80 Hz, 100 ms,
stronger in Objects having LTM representation.
Inter trial phase coherence (ITPC)
time frequency
channel
Mørup et al.
(article in press, NeuroImage 2005)
subject
Condition
n
e e
n Xe c f t t f c t X
f c ITPC
1 1
, ,
, ) ,
, , (
Parafac Assumption: Same Frequency signature present to various degree in time, mixed in the channels and present to different degree in each condition and each subject.
Factor components only additive (non-negativity constraint) ITPC normal distributed - proven by bootstrapping.
The ITPC is the vector strength over trials (epochs)
Proof of normality of ITPC
Bootstrapping:
Randomly select
Data from the epochs to form new datasets (each epoch might be
represented 0, 1 or several times in the datasets).
Calculate the ITPC of each of these datasets.
Evaluate the distribution of these ITPC’s.
Coherent region Incoherent region
Test of difference between conditions over subjects
Time Frequency Channel
F-test value
K
k S s
K k
K S k t f c I k s t f c ITPC
K t f c I k t f c I S t
f c Z
1 1
2 1
2
/ , , , ) , , , , (
1 / ) , , ( ) , , , ( )
, , (
K k
S s S s
k s t f c KS ITPC
t f c I
k s t f c S ITPC k t f c I
1 1 1
, , , 1 ,
, ,
, , , 1 ,
, , ,
Mørup et al.
(article in press, NeuroImage 2005)
5-way analysis
Mørup et al.
(article in press, NeuroImage 2005)
Time-frequency decomposition of ITPC
Time-frequency Subject condition
Channel
Pull paradigm - 6 subjects, 2 condition.
Even trials: Right hand was pulled by a weight
Odd trials: Left hand was pulled by a weight.
References
Bro, R., Jong, S. D., 1997. A fast non-negativity-constrained least squares algorithm. Journal of Chemometrics 11, 393-401.
Carrol, J. D., Chang, J., 1970. Analysis of individual differences in multidimensional scaling via an N.way generalization of 'Eckart- Young' decomposition. Psychometrika 35, 283-319.
Field, Aaron S.; Graupe, Daniel “Topographich Component (Parallel Factor) analysis of Multichannel Evoked Potentials: Practical Issues in Trilinear Spatiotemporal Decomposition” Brain Topographa, Vol. 3, Nr. 4, 1991
Harshman, R. A., 1970. Foundation of the PARAFAC procedure: models and conditions for an 'explanatory' multi-modal factor analysis.
UCLA Work. Pap. Phon. 16, 1-84.
Herrmann, Christoph S; Lenz, Daniel; Junge, Stefanie ; Busch, Niko A; Maess, Burkhard “Memory-matches evoke human gamma- responses” BMC Neuroscience 2004, 5:13
Lee, D. D., Seung, H. S., 2001. Algorithms for non-negative matrix factorization. Advances in Neural information processing 13,
Miwakeichi, F., Martinez-Montes, E., Valdes-Sosa, P. A., Nishiyama, N., Mizuhara, H., Yamaguchi, Y., 2004. Decomposing EE data into space-time-frequency components using Parallel Factor Analysis. Neuroimage 22, 1035-1045.
Möcks, J., 1988. Decomposing event-related potentials: a new topographic components model. Biol. Psychol. 26, 199-215.
Mørup, M., Hansen, L. K., Herrmann, C. S., Parnas, J., Arfred, S. M., 2005. Parallel Factor Analysis as an exploratory tool for wavelet transformed event-related EEG. NeuroImage Article in press,