►
Description
2015 HTM Challenge Application Submission (1st place winner).
A
B
So
hi
I'm
Jonathan
Mackenzie-
and
this
is
my
presentation
for
the
new
payment
to
HTM
challenge.
My
presentation
is
on
HTM
models
for
automate
into
the
detection
in
Adelaide
Australia.
So
what
is
automating
to
the
detection?
Firstly,
we
need
to
know
what
congestion
is
and
that's
basically
where
the
capacity
of
the
road
is
exceeded,
and
so
everyone
slows
down.
So
this
can
be
recurrent,
which
you
might
think
of
as
something
like
rush
hour
in
the
morning
when
everyone
goes
to
work
and
Russia
are
on
the
afternoon
what
everyone
goes
home.
B
So
our
goal
is
to
detect
nonrecurrent
condition
and
that's
where
something
like
an
accident.
A
breakdown,
skilled,
low,
burst,
pipe
landslides,
flooding
rights
and
protests
cause
congestion
on
the
road.
So
the
problem
is
basically
that
freeway
autumn
agents
and
detection
is
well
researched
and
there's
lots
of
solutions
for
it.
B
But
we
want
to
try
and
do
the
same
thing
but
on
arterial
roads,
which
is
much
more
difficult
due
to
the
very
nature
of
the
traffic
and
everyone
to
turning
going
different
ways
and
we
use
loop
detectors
which
are
sent
to
the
underneath
the
road
which
accounts
traffic
and
they
come
with
their
own
warts.
So
in
our
solution
we
use
HTM
to
detect
anomalous
traffic
flow.
B
This
is
an
improvement
on
previous
research,
which
mostly
used
simulations
and
supervised
learning
techniques
which
allowed
people
to
easily
set
up
incidents
and
traffic
flow
and
can
really
easily
monitor
everything.
But
we
have
to
use
real-world
data
which
comes
with
all
its
own
problems.
So
the
way
my
system
works
is
we
import
the
data
which
is
provided
by
the
transport
system
center
at
Flinders
University
into
manga
base.
B
The
data
is
analyzed
using
HTM
and
we
reveal
in
a
Python
web
application,
so
we've
currently
got
three
and
a
half
terabytes
of
data
from
seven
years
and
that's
at
five
minute
intervals,
but
for
now
I'm
only
looking
at
two
month
period,
which
is
about
130
intersection
to
170,000
data
points
per
intersection,
and
this
is
private
data.
So
you
don't
have
access
to
it.
B
Sorry,
and
we
also
use
crash
data
to
verify
normally
prediction
and
their
142,000
incidents,
but
in
a
period
we're
looking
at
is
only
attribute
I
own,
this
public
data
and
you
can
get
it
from
da
gsa.gov
today,
you
so
here's
my
readings
data.
Basically,
it's
just
a
collection
of
objects
which
map
the
sensor
to
the
count
of
the
sensor.
B
So
the
analysis
script
uses
HTM
and
there's
an
option
to
do
it
with
one
model
which
takes
in
every
sensor
as
an
input,
and
then
we
get
the
anomaly
score
from
that,
but
they
didn't
work
very
well.
So
we
use
one
model
/
sensor.
So
if
its
intersection
with
24
sensors
that'll,
be
twenty,
twenty
four
models
24
out
different
outputs,
but
still
this
takes
a
long
time.
So
we
use
supercomputer.
B
B
The
results
to
the
database
is
also
a
smoothing
up
option
available
which
we
haven't
used,
but
this
applies
a
median
filter
with
the
window
size
of
your
choosing
to
the
readings
data,
and
you
can
see
down
here
that
this
would
uses
more
processes
then
cause
since
3
meter,
section
3001
uses
20
senses,
so
it
won't
run
as
fast
as
it
could.
So
we
run
it
on
supercomputer,
and
here
you
can
see
supercomputer
q
with
all
the
jobs
lined
up
and
then
speak
computer
has
around
1200
cause.
B
And
then
you
can
reveal
it
using
the
and
your
web
application.
So
here
you
can
see
the
lovely
city
of
Adelaide,
and
this
is
a
central
business
district,
which
is
the
only
area
that
a
traffic
for
so
we
can
click
on
section
and
kind
of
button
to
see
the
data
for
the
intersection.
So
here
we've
got
intersection,
3083
to
signalize
t-junction
and
here's
all
the
readings
on
sensor
56-
and
here
we
have
the
anomaly
scores
in
blue,
normally
likelihood
in
red
incidents
in
green
and
the
orange
dot
indicate
that
a
sensor
has
exceeded
the
threshold.
B
B
So
certainly
here
there
was
a
significant
dip
due
to
this
incident
here
and
if
we
mouse
over,
we
can
see
that
it
was
caused
by
inattention
and
was
a
rear
end
that
cost
two
thousand
dollars
and
down
here
we
can
see
on
the
map
exactly
where
it
occurred
for
the
going
north,
but
it
did
think
this
bit
of
traffic
over
here
was
anomalous
on
that
was
on
center,
848,
etc.
We
can
click
the
button
to
change
the
sensor.
B
So
here
we
have
a
list
of
incidents
with
the
date
and
which
innocent,
which
intersection
a
decoder
and
if
any
of
its
senses
exceeded
2.99
threshold,
then
they'll
be
shown
here
with
the
sensor.
We
can
filter
to
only
show
ones
which
have
exceeded
and
will
click
on
this
one
here,
30
40,
and
so
here
we
can
see
it
was
an
accident
at
that
time
resume
in
we
can
see,
but
there
was
a
dip
in
traffic
at
that
time
and
that
did
come
up
as
an
anomaly
on
these
sensors
5672
natey.
A
First
off,
as
someone
who
has
built
traffic
analysis,
applications
with
new
pic
before
I
just
want
to
say
that
the
amount
of
software
it
takes
to
do
what
Jonathan
did
is
pretty
significant,
so
impressive,
just
in
the
maturity
of
the
software
and
a
short
period
of
time,
Thank
You
Jonathan
for
your
work.
Judging
panel.
C
B
Would
throw
one
giant
model
that
took
in
every
sensor
as
an
input?
It
could
only
give
one
output
and
that
was
I,
wonder
normally
output,
so
you
can
get
anomaly
/
in
two
percent,
so
rather,
and
so
it
was
difficult
to
tell
which
sensor
was
actually
acting
up
because
I
said
intersect.
Any
incident
on
one
line
might
not
slow
the
traffic
in
another
in
a
significant
way,
and
so
you
really
need
to
detect
that.
B
So
if
new
pic
had
outputs
on
multiple
outputs
for
anomaly
scores
that
have
a
good,
but
yet
it
didn't
really
work,
because
the
score
it
produced
was
pretty
much
just
very
low
the
whole
time,
no
matter
what
and
I'm
not
sure
if
that's
doodle,
that
I
didn't
put
enough
data
into
the
model,
but
that's
just
what
I
found
so
I
pretty
quickly
switched
to
one
model.
/
sensor,
Thanks.
D
B
I'd
say
fairly
confident
because
I
think
that
the
damage
caused
is
a
pretty
good
indicator,
the
severity
incident
and
so
the
length
that
it
would
take
to
clean
up,
etc,
but
I
think
it's
fairly
accurate,
because
people
are
required
by
law
to
report
any
accident
on
the
road
to
the
police.
I
mean
I
think
they
should
have
a
data
entry
person
entering
it,
but
there's
never
a
way
to
verify
back,
because
it's
done
at
a
time.
Chang
right.
D
E
Occurred
to
me
that
you
know
just
my
casual
observation
as
a
driver
that
sometimes
accents
they
quickly
go
to
the
side
and
doesn't
slow
down
anybody,
and
sometimes
it
just
backs
up
everything
so
that
the
correlation
between
an
accident
and
anomaly
in
traffic
flow
is
not
always
so
great
and
I
didn't
know.
In
episode
of
Leave
made
my
question
here:
we're
using
the
the
traffic
incident
reports
just
to
as
label
data,
or
is
that
somehow
the
algol
of
the
system?
What
was
the
goal?
So
what
would
you?
B
The
ultimate
goal
is
to
have
it
actually
identify
an
accident
that
that
causes
congestion
because,
as
you
say,
if
there's
a
small
way
around
and
they
pull
over,
that
doesn't
really
cause
congestion.
You
really
want
the
ones
that
do
cause
congestion
because
of
the
more
severe
ones
but
I
think
in
my
application
it
was
mostly
just
for
now
to
search
for
anomalous,
traffic
and
I.
Think
we
found
a
fair
bit
of
that,
but
yeah
you.
D
F
B
The
point-
nine
nine
four
point:
nine
nine
nine
was
provided
to
me
by
at
taylor.
I
think
I
think
I
was
a
standard
threshold
that
that's
used
in
new
pic
projects.
Good
you
can
take
the
logarithm
of
the
anomaly
likelihood
and
then
that
will
show
you
it
because
it
really
only
peaks
what
it
actually
thinks
as
anomalous
behavior
going
on.
So
that's
what
I
was
told
was
best
practice.
Okay,.
G
Question
so
you
trained
the
model
with
your
supercomputer.
Did
you
also
run
swarming
to
said
the
parameters
for
these
models?
I.
B
Did
run
a
swarm,
but
it
wasn't,
it
didn't
produce
results
very
different
from
that,
provided
in
the
standard
models
and
I
think
there
was
a
video
made
by
Matt,
Taylor
and
Scott
Perry
on
you
should
swamp
for
anomaly,
detection
and
I
think
the
answer
was
basically
no.
So
it
was
a
waste
of
three
hours
me.
G
I'm
just
curious,
if
maybe
one
of
the
reason
why
sometimes
it
doesn't
detect
an
animal,
he
is
a
min
max
problem
or
something
like
that.
If,
for
this
particular
intersection,
maybe
the
like
I'm,
is
it
like
the
number
of
its
incidence?
Maybe
maybe
there
is
this
issue
and
that's
why
it
doesn't
detect
the
enemy.
You
see
what
I
mean
if
the
the
maximum
in
your
model,
pram
is
said
that
some
value,
but
you
end
up
with
something
that's
much
higher.
It's
kind
of
inquiry.
B
G
B
Maximum
number
of
vehicles
through
a
sensor
I,
was
provided
by
the
team
here
at
Flinders,
was
that
200
vehicles
is
pretty
much
the
maximum
number
of
vehicles.
You're
gonna
get
ever
because
that's
like
fully
because
it's
in
five
minute
intervals,
so
that
that's
pretty
much
constant
high,
very
high
speed
traffic
over
a
single
sensor.
Ok,.
H
H
B
Frankly,
that
I
can't
really
tell
which
one's
its
best
up,
because
I
think
it
it
found
it
didn't
separate
end
of
like
12
p.m.
midnight
very
well
from
other
day
other
times,
and
it
was
a
huge
gap
in
the
data
for
about
three
days.
I
think
that
it
thought
was
very
anomalous,
but
I'm
not
really
quite
sure
about
how
which
fits
its
best
at
its
have
to
look
through
all
the
incidents
that
thought
were
highly
anomalous.
Yeah.
H
It's
very
challenging
with
all
of
this
data
to
figure
that
out.
If
but
that's
yeah,
I
think
sort
of
piggybacking
on
Marian's
thing.
It
might
be
interesting
to
do.
You
know
offline
as
kind
of
a
sweep
through
your
parameters,
and
you
know
there
may
be
a
way
to
kind
of
improve
the
overall
system
a
little
bit
yeah.