1. Detec%ng
Emo%on
from
EEG
Signals
Using
the
Emo%ve
Epoc
Device
Rafael
Ramirez
Zacharias
Vamvakousis
Universitat
Pompeu
Fabra
Barcelona,
Spain
Presented
by:
Álvaro
Barbosa
University
of
Saint
Joseph
Macau
SAR,
China
Brain
Informa%cs
2012
2. Mo%va%on
• Study
of
emo%ons
in
human-‐computer
interac%on
has
increased
in
recent
years
• Growing
need
for
computer
applica%ons
capable
of
detec%ng
users’
emo%onal
state
• Facial
and
voice
informa%on
– can
be
consciously
controlled
and
modified
– interpreta%on
is
oSen
subjec%ve
•
Here,
we
use
EEG-‐based
emo%on
detec%on
3. Contrib%ons
• Method
for
EEG-‐based
emo%on
detec%on
• Use
of
low-‐cost
technology
-‐>
Emo%v
EPOC
headset
• We
do
not
rely
in
subject
self-‐reported
emo%onal
states
(as
most
previous
work
do)
• Instead,
we
use
a
library
of
emo%on-‐annotated
sounds
(IADS
Lib
-‐
hp://csea.phhp.ufl.edu/media/iadsmessage.html)
5. Data
Collec%on
• 6
healthy
subjects
(mean
age
=
30);
listening
to
12
(5-‐10s
long)
emo%on-‐annotated
sounds
(IADS
Lib)
• Emo%v
EPOC
headset
-‐
14
data-‐collec%ng
electrodes
(AF3,
F7,
F3,
FC5,
T7,
P7,
O1,
O2,
P8,
T8,
FC6,
F4,
F8
and
AF4)
and
2
reference
electrodes
6. Feature
Extrac%on
• Alpha
(8-‐12Hz)
and
Beta
(12-‐30Hz)
bands
are
par%cular
bands
of
interest
in
emo%on
research
for
both
valence
and
arousal
• We
apply
bandpass
filtering
for
extrac%ng
alpha
and
beta
frequency
bands
• EEG
signal
in
four
loca%ons
in
the
prefrontal
cortex:
AF3,
AF4,
F3
and
F4
• Arousal
=
a(AF3+AF4+F3+F4)/b(AF3+AF4+F3+F4)
• valence
=
aF4
/bF4
−
aF3
/bF3
7. Classifica%on
Learning
Task
• Detect
emo%onal
state
of
mind
of
a
person
based
on
observed
EEG
data
• We
approach
this
problem
as
a
two
2-‐class
classifica%on
problem
– high/low
arousal
– posi%ve/nega%ve
valence
ArousalClassif
ier
(
EEGdata([
t,
t
+c]))
→
{high,
low}
ValenceClassifier
(
EEGdata([
t,
t
+c]))
→
{posi%ve,
nega%ve}
c=1s
and
with
increments
of
t
of
0.0625s
12. Results
(3)
• Results
indicate
that
the
EEG
data
contains
sufficient
info
to
dis%nguish
between
high/low
arousal
and
posiFve/negaFve
valence
states
• Machine
learning
methods
are
capable
of
learning
the
EGG
paerns
that
dis%nguish
these
states
• Different
accuracies
among
different
subjects
• For
a
subject,
similar
accuracies
with
different
learning
method
13. Results
(4)
• Inter-‐subjects
accuracy
differences
may
be
due
to
– different
degrees
of
emo%onal
response
between
different
individuals,
or
– amount
of
noise
for
different
subjects.
• Anyway,
there
exists
considerable
varia%on
in
EEG
responses
among
different
subjects
14. Conclusion
• Low-‐cost
emo%on
detec%on
system
• no
self-‐assessment
informa%on
about
the
emo%onal
states
by
the
subjects
• linear
discriminant
analysis
and
support
vector
machines
classifica%on
• Classifiers
able
to
discriminate
between
high-‐
low
arousal
and
posi%ve-‐nega%ve
valence
15. Future
work
• Improve
classifica%on
accuracy
– Systema%cally
exploring
different
feature
extrac%on
methods
and
learning
methods
• Incorpora%ng
self-‐assessment
informa%on
would
very
likely
also
improve
the
accuracies
of
the
classifiers