►
From YouTube: ECG + HTM - Kentaro Iizuka
Description
2015 HTM Challenge Application submission (3rd place).
B
Hi
I'm
Kent
Erica
I'll
introduce
my
HCM
challenge
project
in
this
project.
I
try
to
detect
animal
spots
of
the
ECG
electrocardiogram
data.
This
project
is
still
in
prototype.
Honestly
speaking,
I
could
not
get
enough
accuracy
of
anomaly
detection
now,
but
I
think
this
project
shows
the
possibility
to
use
HTM
framework
for
bio
sensor
or
medical
application.
B
Okay,
look
into
the
project.
This
is
my
project
structure
gerrae
about
three
parts:
first,
collecting
data
using
ECC
device
with
simple
script:
this
script
correct
data
using
serial
communication.
The
data
saved
to
the
CSV
file.
Second
part
is
the
FFT
fast
Fourier
transform
part.
This
part
extracts
a
characteristic
of
the
ECG
data.
B
B
B
Ok,
this
video
shows
the
normal
ECG
data
learning
reading
epoch.
Almost
an
ahmadi
buddy
at
0.2
to
0.4,
but
unmask
are
going
to
hi
Gregory.
The
high
animal
scores
correspond
to
the
QRS
part
of
the
ECG
data
in
ECG
data.
The
most
peak
and
spike
parts
called
QRS
wave
lubic
tend
to
detect
normal
QRS
wave
as
animal
spot
after
enough
rally.
B
B
B
B
This
is
visualization
of
FFT
combated.
Ecg
data
left
graph
is
a
normal
ECG
data,
light
containing
a
llamar
spot,
but
car
axis
is
time
step
and
horizontal
axis
is
the
frequency
step
card
shows
part
of
its
frequency.
You
can
see
non-uniform
part
in
light
graph.
Actuary
non-uniform
part
shows
animals
in
ECG
data.
B
Okay,
FFT
combated
data
looks
great
let
input
into
the
nupoc
I,
don't
know
what
it's
the
best
way
to
impact
our
into
the
new
peak
I.
Try
three
ideas
for
input
vector
into
the
new
peak
first
idea
is
coordinate.
Encoder
second
is
used
a
lot
of
sky
coder.
Sad
is
specta
encoder
first
I
test
coordinate
encoder.
It
works
good
for
three
to
five
dimension,
but
because,
on
seven
dimension
its
can
learn
but
not
work.
It's
use
huge
size
of
memory,
temp
dimension,
use,
eight
gigabyte
memory
and
my
vector
has
50
or
more
dimensions.
B
B
Inori's
who
are
going
to
high
regularly
I,
don't
know
what
caused
the
problem.
I
had
another
experiment
that
feels
better
encoder.
It's
deleted
only
sent
near
peak,
but
I
want
to
try
this
one.
So
I
put
back
this
encoder
from
past
commit
vector.
Encoder
is
only
one
SCARA
encoder
instance
get
binary
representation
of
each
vector.
Are
you
combine
their
binary
D
presentations
vector
encoder
works.
Good,
their
output
seems
to
be
different
from
a
rot
of
scarring
coda.
For
my
condition,
this
is
vector.
Encoder
output.
You
can
see.
B
Animal
spots
has
high
anomaly
score,
but
it's
not
in
a
high.
It's
only
there
0.1
to
0.15
now
I
continue
to
find
the
best
way
to
encode
time
series
fft
vector
into
the
music.
If
you
are
interested
by
my
project,
I
am
happy.
My
github
repository
include
normal
and
anomalous
ECG
data,
so
you
can
try
to
use
my
project
data
and
code
for
your
experiment.
I
think
my
project
code
is
also
useful
for
the
people
who
wanted
to
run
your
peak
with
time
series
of
Victor.
Please
check
out
that
lead.
C
A
Basically
got
here
is
a
signal,
analysis
problem
and
he's.
Certainly
not
the
first
person
in
our
community
who's
tried
to
solve
a
signal
analysis
problem
with
HTM,
but
he
seems
to
have
tried
the
most
things
and
he's
really
doing
investigative
stuff
here,
Sergey
go
to
the
mic.
If
you
want
to
say
something
else,.
D
Discussion
of
this,
it's
very
sad
to
me
that
the
answer
to
the
encoding
program-
a
problem-
is
right
on
that
picture.
Ecg
data
actually
has
discrete
time
steps
and
there
are
several
features
you
can
measure
about
the
time
step.
So
you
measure
the
amplitude
and
the
durations
of
the
different
phases
of
a
single
beat.
If
you
just
use
like
those
four
scalar
encoders,
it
would
actually
work
very
well.
D
F
A
F
Look
to
me
like
what
you
know
signal
in-house
is
a
lot
of
stuff
in
this
case,
he's
got
a
repeating
waveform
and
you
and
and
you
want
to
find
out
what
a
DBAs,
and
so
this
is
the
kind
of
thing
that
the
scale
encoder
really
ought
to
learn
and
you
might
be
able
to
use
the
reset.
So
you
might
been
do
various
other
things
to
make
it
so
like
hey.
F
Take
that
peak
no
I
said
to
peak
that
cure,
as
peak
doesn't
always
occur,
so
I
I,
don't
know
enough
about
it,
but
I
felt
I
thought
they
shouldn't
strive
more
on
that,
because
I
think
I
think
this
is
the
kind
of
thing
you
could
really
learn.
These
are
like
subtle
little
melodies
and
you
can
learn
them.
You.
G
I'm,
making
no
two
surprises
using
a
50.
Do
is
just
a
pattern
for
a
couple
of
problems
where
you
have
noisy
oscillating
signals
like
sound
eg
ECG,
where
it
seems
that
all
these
projects,
independently
they've
sort
of
converged
to
taking
the
a
50
and
then
including
that
I,
don't
just
like
this.
Whatever
this
one
seems
different
than
me,.
F
F
A
Don't
believe
he
encoded
time,
I
think
it
was
just
a
sequence.
However:
I
mean
one
of
his
problems.
He
ran
into
it.
Just
a
scalar
encoder
was
that
he
had
different
people
who
had
different
ECG
patterns,
and
so
he
couldn't
learn
like
a
typical,
healthy
heartbeat
from
a
sample
of
a
bunch
of
different.
F
F
E
H
F
F
Come
in
with
certain
having
been
involved
this,
although
we
come
to
certain
biases
and
my
assumption
was
generally-
you
want
to
do-
monitor
lots
and
lots
of
individual
sources
and
build
also
out
to
individual
models.
My
assumption
I
made
god
I-
must
have
misunderstood
him.
I
thought
he
was.
You
know
they're
trying
to
mana,
what's
normal
for
a
person
and
see
if
a
change
occurs
like
I'm,
never
saying,
but
maybe
that's
not
what
he
was
in,
that
would
that
would
have
been
a
very
I
think,
a
very
doable
project,
yeah
yeah.
A
C
A
F
Well
to
me
this
was
actually
one
of
the
easier
ones
it
should
have
worked,
because
it's
such
a
regular
thing,
but
again
for
an
individual.
It
would
be
regular
and
then
deviations
right.
So
that's
what
you
wanted.
I
think
this
is
a
great
one
because
other
things
like
you
know,
EEG
data,
it's
just
all
over
the
map.
Well,.
A
H
H
Feels
like
we're
very
close
to
solving
it
and
I
also
kind
of
felt
that
maybe
he
gave
up
a
little
too
quickly
on
the
scalar
encoder,
and
maybe
that
would
have
worked.
You
know,
maybe
something
specific
could
work.
I
also
think
an
FFT
based
approach
could
work,
but
that's
harder.
You
have
to
figure
out
how
to
set
up
the
HTM
properly
for
that,
but
the
FFT
based
approach
would
be
very
general
to
lots
of
different.
F
H
F
E
A
question
about
the
scalar
encoder:
can
it
adapt
to
a
variable
frequency
input
like
to
sink
in
if
you
were
really
looking
at
these?
As
a
doctor,
you'd
say:
here's
this
bike,
here's
this
bike
here,
there's
this
bike,
not
just
n
seconds
later,
but
this
spike
matched
up
to
that
spike.
This
bike
match
up
to
that.
Even
if
the
spikes
came
at
variable
time
apart
can
the
scalar
encoder
kind
of
auto
sync
onto
itself
to
adjust,
adapt
it's
a
different
to
variable
frequency
data,
I.