PowerPoint Presentation

unknownlippsΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

67 εμφανίσεις

DECODING
COGNITIVE STATES
FROM BRAIN IMAGES

By

Rahul E

Chittibabu N


Under the guidance of

Dr
.
Krithika

Venkataramani


Introduction


Goal: To train classifiers that could
decode subject's
cognitive
state based on his
brain FMRIs


Specifically, whether the subject is seeing
a picture
or reading
sentence


Brain FMRIs indirectly represents
neural activity
in the brain as a
3D
image


Classifiers
we trained for
:


GNB
classifier (86% accuracy)


KNN classifier (90% accuracy)


SVM classifier (85% accuracy)



A note on Brain FMRIs


FMRI :functional Magnetic
Resonance Imaging



it
measures the ratio of
oxygenated haemoglobin
to
deoxygenated haemoglobin in
the
blood



Oxygen
content is directly
proportional to brain activity


High resolution, Typically
contains ~
15000 voxels



Temporal response
are usually
smeared over
several seconds

Dataset


StarPlus fMRI data by
CMU



6 subjects
, around 54
trials per
subject



Each trial lasted for about 27
seconds



Subject is shown an
image,
followed by
a sentence then asked
to
decide whether
the sentence correctly
represents the
image



FMRIs
where recorded at a rate of 2
scans per
second

Timeline

Prior Approaches


Learning Algorithm


f
: fmri
-
sequence[I
1
,I
2
….I
n
]


cognitive state


I
1
,I
2
….
I
n
are Images the taken
during a contiguous time
interval


Cognitive state


seeing a picture/ sentence


Gaussian
Naive Bayes (
GNB)


Feature independence


Support
Vector Machine (SVM
)



L
inear
kernel Support Vector
Machine


k
Nearest
Neighbour (kNN)


Euclidean distance metric


considering values of 1, 3,5 and 7 for k

Prior Approaches (cont..)


Dimensionality reduction


Average



average
all voxels in an ROI


ActiveAvg(n)


For
each
ROI


select the n
most active
voxels


average
the selected voxels


Active(n)


select the n most
active voxels in the whole brain.

Our approach



Two stage Dimensionality reduction


Choose most active m voxels


Choose the most differentiating n voxels (out of m)

Dimensionality reduction


Discard
those voxels, which can’t be
used to distinguish between pictures
and
sentences



For each voxel, create a “sentence”
and “picture” distribution



Do t
-
testing between distributions



Choose those voxels which gives
maximum t
-
testing value


Classification


For classification we tried three classifiers


GNB classifier


Dependencies between voxel activities

TAN
tree (won’t work)



assumed independencies between voxel
activities


SVM Classifier


Linear kernel support vector machine


KNN


Distance metric

Euclidean distance


K=1,2,3 and 5

Results

Method

Published

approach

Our

approach

GNB

82

86

SVM

89

90

1NN

78

85

3NN

82

85

5NN

82

82


Accuracies are average accuracies over all single subjects
classifier


Hold Out Cross
-
validation (70% training data, 30% test
data)

Comparison

Active voxels vs. accuracy

Classifier for all the subjects

Classifier

Accuracy

GNB

77

SVM

80

1NN

82

3NN

77

5NN

82

Conclusion


Obtained best accuracy
of 90% for the
SVM
classifier


TAN tree is not a good approach


We can extend
our work to
Syntactic
ambiguity study

References


Mitchell, T.M. and Hutchinson, R. and
Niculescu
, R.S. and
Pereira, F. and Wang, X. and Just
, M
. and Newman, S.
Learning
to decode cognitive states from brain image
,
Machine
Learning
, 2004


Springer



X. Wang, R. Hutchinson, and T. M. Mitchell.
Training fMRI
Classifiers to Detect
Cognitive States
across Multiple Human
Subjects
, Neural Information Processing Systems 2003
.
December
2003


Friedman
, N. and Geiger, D. and
Goldszmidt
, M.
Bayesian
network classifiers
,
Machine learning
, 1997


Springer



http://
www
-
2.cs.cmu.edu/afs/cs.cmu.edu/project/theo
-
81/www/


http://en.wikipedia.org/wiki/Functional_magnetic_resonance_i
maging