►
From YouTube: Clairvaux - Don Tjandra
Description
2015 HTM Challenge Application submission
A
B
Name
of
my
project
is
clever:
it's
a
web
application
that
predicts
on
Celtic
events
using
the
winters
technology
and
what
I
mean
by
that
is
this
weights,
the
HCM,
the
hierarchical
temporal
memory
and
the
data
for
this
application
that
comes
from
the
site.
Ac
led
the
conflict,
location
and
event
data
project,
and
they
they
provide
data
going
back
to
nineteen
nineteen
ninety-seven
describing
our
complex
in
in
the
African
continent.
But
what
I've
actually
used
for
this
for
this
project?
Are
these
monthly
files
and
I've
effluent
a
few
of
them
already
for
this
demo?
B
So
we
can
look
at
one
of
them
as
an
example.
So
here
we
see
that
the
HCM
network
made
over
2,000
correct
predictions
out
of
over
13,000
predictions,
and
in
order
to
get
these
numbers,
you
have
to
be
trained
at
this
many
times,
which
is
way
the
number
of
rows
in
the
file
types
of
training
cycles,
constant,
which
is
10
in
this
case,
and
the
accuracy
is
just
the
collect,
correct
predictions
over
the
total
number
of
predictions.
B
And
in
order
to
look
at
this
in
more
details,
we
have
this
thing
compared
coach
matrix
where
the
rows
are
the
actual
events
and
the
columns
are
to
protect
the
events
at
the
next
step.
As
you
hover
over
each
cell,
you
can
see
the
accounts
as
well
as
well
as
the
build
a
distribution.
So
here
we
see
that
this
is
the
camp
and
that
and
the
probability
of
having
seen
as
having
seen
this
event,
this
is
the
probability
is
the
count
and
the
probability
of
seeing
this
event
at
the
next
step.
B
Now
this
co-occurrence
matrix
tells
us
what
the
work
is
predicting,
but
we
also
want
to
have
some
idea
of
the
accuracy
and
that's
what
this
a
confusion.
Matrix
is
war,
and
here
the
rows
are
the
predicted
events
and
the
columns
are
the
actual
offense
and
the
correct
predictions
are
shown
in
in
Queen
a
lot
diagonal
entries
here.
So
in
this
case
the
duros
here
our
internet
are
basically
a
false
negatives
and
or
the
numbers
in
the
scroll
are
false.
Negatives
and
the
numbers
in
go
ahead
here
in
this
column
are
false.
Positives.
B
So,
as
I
add
more
data,
it's
the
actress's
accuracy,
starting
a
groove
here
at
about
thirty
percent,
and
here
that
we're
at
about
forty
percent.
So
it's
possible
to
see
this
number
go
up
even
higher
as
add
more
data,
fonia
put
just
a
few
files
here
for
this
demo
and
it's
so
it's
also
possible
that
you
know
some
of
these
events
may
be
more
or
less
random.
B
A
So
looks
like
we've:
we've
got
Donnie
on
audio
hello,
Donnie
thanks
for
joining
us,
and
let
me
just
before
going
to
the
panel
so
from
what
I
understand
here,
based
on
all
the
armed
conflict.
Events
in
Africa
Donnie
is
a
feeding
them
in
and
trying
to
predict
what
the
next
type
of
conflict
is
over
time,
so,
basically
using
trying
to
classify
what
what
type
of
category
is
coming
next.
C
C
I
should
say
of
all
the
ones
that
we
saw.
This
was
the
one
that
kind
of
really
opened
my
eyes
about
the
possibilities
of
kind
of
real-time
monitoring,
and
you
know
some
really
innovative
uses
of
this
and
just
to
the
idea
of
monitoring
conflict
in
you
know
globally
and
being
able
to
predict
where
armed
conflict
might
happen.
Next,
I
thought
was
a
really
really
nice
idea.
C
I
think
your
if
I
remember
correctly
you're
a
little
bit
limited
by
kind
of
the
real-time
pneus
of
the
data
that's
available
today,
but
I
think
you
could
imagine
that
in
the
future
this
could
be
really.
You
know,
constantly
streaming
data
and
could
be
a
really
nice
application
for
just
a
monitoring
geopolitical
events.
So
this
was
a
really
nice
innovative
application.
C
Did
you
get
a
sense
of
whether
and
it's
obviously
a
pretty
random
thing
as
well?
Did
you
get
a
sense
of
whether
forty
percent
is
a
good
number
and
a
second
follow-on
question
is
new?
Pic
can
predict
sort
of
probability
distributions
or
events
I,
wonder
if
you
try
to
look
at
that
to
see
if
there
was
some
way
you
could
give
a
confidence
estimate
to
the
prediction,
so
it
may
be.
Some
types
of
sequences
are
really
random,
but
there
are
some
types
where
you
can
pretty
confidently
predict
that
this
next
conflict
is
going
to
happen.
C
B
B
The
second
was
that
the
battle,
no
change
of
territory,
which
is
basically
like
a
billion
we're
until
the
contested
location,
doesn't
change
in
the
in
the
third
one
is
the
violence
against
civilians
and
so
that
htn
it
can
predict
the
rights
and
protests
really
well
because
just
because
it
has
a
high
count
in
the
data
but
the
other
times
they
have
low
counts.
So
that's
why
that?
B
So,
that's
why
it's
you
know
it's
not
getting
accurate
predictions
for
them,
so
I,
basically,
I
basically
tried
to
running
locally
with
monthly
data
on
my
local
machine,
starting
from
January
up
till
about
October,
and
what
I
can
see
is
that
at
the
beginning,
it's
the
HTM
is
sort
of
like
making
random
cases
in
the
in
the
co-occurrence
matrix.
You
know
it's
showing
on
us,
colored
cells
in
and
all
the
other
columns,
but
as
a
as
I
progressed.
B
Further
up
in
time,
so
when
I'm,
looking
at
the
October's
data,
for
example,
it's
really
showing
just
cost
college
cells,
just
just
in
one
column,
in
the
vice
and
protest
column,
which
means
that
it's
a
getting
more
confident
but
I'm
predicting
that
event,
but
not
so
much
about
the
other
defense,
because
it's
just
because
there's
not
enough
data
for
them
yet.
But
then
this
it's
also
starting
to
start
it's
time
to
learn
them
to
protect,
but
the
second.
If
the
second
most
common
event,
which
is
that
battle,
no
change
of
their
authority.
B
D
B
D
B
Yeah,
it's
I
mean
that
they
are
similar
in
some
ways.
I
mean
they're
human
janitor
defense.
They
have
a
timestamp
and
location,
so
it
I
think
it
would
be
possible
to
defeat
the
data
and
and
see
how
how
it
would
work
and
while
the
visualization
itself
completes
not
going
to
change
that
much
it's
just
a
you
know
the
matrix
of
categories,
so
I
haven't
looked
at
that
the
climatic
data
I
guess
their
categories.
They
have
different
categories
for
the
different
time
types
in
so
I
think
it
is
possible
to
use
this.
D
B
Okay,
yeah
well
I
like
right
now
that
the
code
is
it's
hard
coding,
a
few
things
like
the
encoders,
it's
hard
coded
for
the
specific
fields,
but
I
think
with
a
little
bit
more
like
it
is
possible
to
make
something.
That's
that's
much
more
flexible,
that's
so
it
doesn't
work
on.
So
it's
not
really
hardcore
only
for
the
specific
data
set,
but
it
could
potentially
work
with
other
data
sets.
B
E
You
I
thought
this
was
a
really
fascinating
kind
of
application.
I'm
curious,
I
couldn't
make
out
the
specific
names
associated
with
the
rows
and
columns,
but
my
question
is
in
any
sense:
do
you
view
this
as
as
actionable?
Okay?
So
let's
say
you
to
go
back
to
the
air
traffic
thing
to
you.
This
is
predictive
that
you
see
a
strong
signal
that
some
next
event
will
happen.
Is
that
interesting
or
actionable.
B
Me
so
I'm,
not
really
a
you
know
like
an
early,
an
expert
on
national
security,
for
example,
but
I
think
it
may
be
of
use
for
it
for
like
a
government
organization
or
someone
involved
in
national
security,
for
example,
to
if
they
were
once
but
right
now.
The
the
event
that
its
most
confident
on
is
the
the
riots
and
protests.
So
I'm
not
sure
like
exactly
what
what
kind
of
actions
one
would
take
when
when
they
see
that
you
know
and
that
that
this
effect
is
predicted
to
happen
on
the
next
day.
B
But
I
think
for
some
of
the
other
events
that
are
more
extreme
like
the
battles
where
them
some
fatalities
are
involved.
I
think
it
could
be
a
something
that
that
could
be
used
to
do
for
someone
to
make
some
decisions.
To
example,
the
either
like
find
a
way
to
defend
that
event
from
happening
or
maybe
take
take
some
actions
to
maybe
move
some
resources
plan
to
remove
some
people
ground
to
them.
Minimize
the
damage,
for
example,
or.
B
B
Yeah
because
some
of
these
offense
have
vitality,
so
I
mean,
for
example,
rice
and
protest.
Probably
is
not
gonna
have
any
fatalities,
but
some
of
them
boys,
team
ones
like
the
battles
and
and
the
cotton
catholic
events
that
they
do
have
some
fatalities
then,
and
the
fatality
czar
actually
in
the
theater
itself.
But
I
didn't
you
use
that
in
this
case,
for
classification.
F
It's
a
lovely,
ambitious
application.
You
know,
I'm
not
convinced
that
the
actual
the
their
patterns
and
the
data
that
you
could
act
upon.
You
know
it's.
It's
it's
on
one
extreme
year,
you're
trying
to
do
something
like
predicting
earthquakes.
Well,
you
can
classify
them.
You
can
say
the
percentage.
There
are
different
types
and
but
you
know
if
it's
not
it
may
might
be
there.
I
don't
know,
but
you
know
you
also
might
have
just
any
predicting
the
you
know.
The
first
order,
zeroth-order
prediction
of
of
likelihood
so
well.
F
F
Think
it's
I
love
the
idea,
I'm,
not
sure
that
if
the
data
itself
is
sufficient
to
make
a
good
prediction
on
these
things,
and
but
you
know
that's
what
I
would
be
thinking
everything
like
what
other,
what
other
things
that
are
like
this,
that
might
be
I
get
more
data
that
might
be
more
predictable
and
so
in
sequential
it's
just
that.
You
know
it's
an
unproven
hypothesis.
Union
you
end
up,
will
predicting
the
most
likely
events.
If,
especially
if,
if
there
really
is
a
structure,
isn't
there
isn't
temple
structure
in
the
data?
F
G
In
a
way,
it's
an
interesting
experiment
to
see
if
there
is
a
pattern
and
something
you
wouldn't
immediately
think
there
is
a
pattern
for
I
think
most
I
personally
I
would
think
that
conflicts
are
somewhat
random,
like
you
could
I
guess
you're
making
the
argument
that
conflict
baguettes
conflict.
Therefore,
there
is
some
sort
of
a
pattern
there
for
the
conflict.
What
happened
protest.
G
C
C
A
F
Rare
the
event
is
a
harder,
it
is
to
predict
it.
That's.
Why
use
the
earthquake
example?
You
know
and
again
you
may
say:
I
like
hey
they're,
predicting
earthquakes
here
in
Northern
California.
We
know
it,
but
you
know
when
so
we
don't
do
anything
about
it.
You
know
a
put
some
water
in
your
basement
type
of
thing.
I
was
running.
You
know
you
could
also
just
we're
bringing
something
we
should
probably
keep
going,
but
that
you
know
there
could
be
other
things
that
are
less
than
armed
conflict.
F
There
might
be
things
like
food
shortages
or
transportation,
forages
or
other
things
that
might
be
more
frequent,
perhaps
less
significant,
but
more
predictable
that
he
saw
it
just
diseases
but
thinks
it
might
happen
on
a
fairly
regular
basis
that
you
could
track
and
and
then
they
might
be,
the
more
often
they
occur,
the
more
actionable
they
become
as
well.
So
because.
F
We
ran
through
this
very
similar
issues
when
we
were
looking
at
credit
card
fraud.
You
know
you
have
lots
and
lots
of
data
and
then
there's
there's
a
fraud
that
might
occur,
but
it's
it's
too
rare
an
event,
sometimes
to
make
it
a
predictable
data
stream
link
in
that
regard,
right
anyway,
but
very,
very
cool.