►
From YouTube: Back Seat Driver: 2014 Spring NuPIC Hackathon Demo
Description
Marcus Lewis
A
A
Okay,
hi
guys
I'm
Marcus
and
my
demo
is
going
to
start
wonderfully
so
I'm,
exploring
the
space
of
what
you
can
do
with
event
streams
and
specifically
in
client
software
I
think
that
there
could
be
good
things
here.
If
someone
can
do
it
right,
because
if
you
could
either
reach
a
point
where,
through
anomaly
detection,
you
find
cases
where
you
as
the
product
builder,
agree
that
that's
an
anomaly
like
I
agree.
A
new
pic
good
call
that
shouldn't
be
happening
like
that
would
be
really
cool
to
make
happen.
A
The
other
is
through
prediction,
if
you
could
and
first
of
all
I'm
setting
very
high
expectations
here
for
something
that's
not
going
to
be
anything
close
to
this,
but
this
is
like
the
pie
in
the
sky
kind
of
thing
through
prediction:
it's
very
common
for
a
product
builders,
different
teams
to
have
telemetry
in
place
where
they
see
like
so-and-so,
failed
with
error
code
blah.
But
often
you
can't
do
anything
with
that
or
if
you
try
you'll
go
and
burn
a
week,
trying
to
figure
out
a
priori
I.
A
Why
so-and-so
happened
when
you
could
have
been
like
doing
real
work
through
that
time,
so
the
through
prediction?
If
you
could
predict
that
something's
about
to
occur,
maybe
you
could
do
the
Anita
diagnosis
or
maybe
you
could
swerve
away
and
not
not
reproduce
like
do
something
right
that
so
there's
there
are
a
few
different
things
that
you
could
do
with
this.
So
that's
why
this
is
what
I'm,
looking
at
so
I
just
decided
to
start
with
using
just
windows
with
its
etw
events.
A
Luckily,
I
put
this
on
the
website
and
have
to
do
it
live
from
windows.
So
so
this
is
what
I
started.
I
just
went
with
something
really
basic
I
chose
to
grab
I
ease
you
I
behavior,
and
the
data
ends
up
looking
something
like
this
and
so
I
used
over
time.
I've
built
on
it.
I
added
thread
ideas
later
into
it
as
a
theory
for
improving
it,
so
I'm
going
to
move
on
to
us,
warmed
it,
and
one
thing
I
thought
was
interesting.
A
I'll
move
on
I'll
move
through
this
quickly,
but
I
gave
it
a
thread
IDs
times
stamps
and
the
event
I
used.
The
event
is
just
a
category
which
I
think
that
makes
me
one
a
lot
of
people.
Scalars
categories,
I
think,
is
like
a
whole
new,
a
whole
different
ballgame,
so
that
so
it's
just
differing
concerns
that
you
deal
with.
So
I
thought
it
was
noteworthy
that
through
swarming
it
decided
to
basically
discard
my
time
stamps
and
my
thread
ids
like
well.
A
Thank
you
that
goes
so
I
I
think,
like
eventually
I
would
find
ways
to
use
that
correctly,
because
it
does
get
confused
in
ways
that
I
was
trying
to
solve
by
giving
those
so
I'll.
Just
move
on
to
one
of
the
things
that
was
big
for
me
was
needing
to
be
able
to
make
sense
of
the
data.
So
let's
see
how
okay
cool
I'll
zoom
in
on
this
one,
some
ok
yeah
but
other
other
browsers.
Let
you
zoom
in
on
things
very
nicely
so
anyway.
A
So
I
did
a
few
things
like
added,
color
coding
and
and
these
built-in
bars
that
gave
me
the
ability
to
really
move
around
really
quickly
and
see.
What's
going
on,
and
let
me
reiterate-
I
have
so
many
excel
sheets
here
that,
like
as
good
its
online
is
my
phone
so
so
I
just
then
went
looking
at
things
that
was
my
goal
was
to
just
keep
finding
things
that
surprised
me
and
figure
out
why
those
happen
and
try
to
keep
like
squashing
bugs
until
I
agree
with
it.
A
So
first
of
all,
this
is
how
this
is
just
what
it
ends
up.
Looking
like
it
starts
off
with,
with
it
grabs
events
you
this
this
there's
nothing
surprising
about
this.
It
shows
like
if
first
to
see
something
and
like
I
said
up
here,
I
started
with
like
the
UI
thread,
just
loop,
as
kind
of
training
wheels
of
like
I'm
going
to
get
past
this
so
quickly
and
instead
of
moving
on
to
real
things,
but
first
I
want
to
make
sure
these
don't
confuse
it
so
so
right
away
it
learns.
A
It
gets
the
hang
of
it.
So
I
did
other
things
where
I
added
red
multiple
browser
tabs
open,
and
it
doesn't
like
that
and
that's
why
I
gave
it
thread
IDs,
but
that's
not
in
this
picture,
but
it
doesn't
matters
on
this
picture.
The
swarm
still
in
this
case
decides
to
get
rid
of
that
IDs.
So
I
think
that
means
I
need
a
different
way
to
put
it
into
it.
So
I'll
move
on
to
the
thing
that
really
surprised
me
was,
I
call
this
late
into
it.
A
A
Is
it
just
expects
something
like
it
is
expected
like
every
so
often
something
weird
happens
and
we're
expecting
that
and
it
because
something
we
didn't
happen,
that
was
the
anomaly,
but
it
doesn't
have
a
name
for
the
thing
that
it
because
the
because
the
classifier
is
still
out
putting
that
it
expects
that
it
expects
the
exact
event
that
happens,
but
here's
one
encoding
of
that
event.
Here's
another
dramatically
different
one
and
a
lot
of
these
values
I
did
check.
Our
are
also
is
a
subset
for
the
most
part
of
this.
A
A
As
far
as
fool
for
one
I
kind
of
suspect,
I'm
doing
I'm
giving
events
to
new
pic
incorrectly,
because
other
people
in
the
room
would
definitely
have
insights
on
this,
but
basically
I'm
giving
if
you
just
give
the
raw
event
stream
and
if
I
event,
it's
like
you're
kind
of
giving
a
diff
of
sorts
rather
than
giving
like
the
picture
or
giving
snapshots.
So
I
suspect
like
that.
A
B
C
B
B
B
A
B
B
A
C
A
C
D
There
is
a
difference
between
how
the
anomaly
score
is
calculated
and
how
the
eventual
predictions
are
calculated.
The
nominee
score
is
done
in
the
column
space
and
then
the
predictions
go
through
a
classifier,
and
then
we,
you
know
so
they're
small
differences
there.
So
perhaps
that's
has
something
to
do
with
it.
I'm,
not
exactly
sure.
What's
going
on,
okay.
B
Just
just
one
question
because
I
think
what's
happening
here
is
that
there
do
you
have
one
model.
That's
basically
listening
to
a
different
kind
of
structure
done
like
grok
listens
to,
because
what
Gras
places
too
is
a
timestamp
and
one
scalar
value
right
right
and
what
you're
giving
it
is
much
more
structured
information
right
simultaneously
right,
and
it
just
might
be
that
the
spatial
and
temporal
structure
of
the
data
that
you're
given
it
is
just
to
two
disjoint.
It
yeah.
A
B
That
do
you
get.
Some
of
the
fields
are
changing
drastically
because
you're
using
category
data,
for
example,
for
a
lot
of
the
values
and
what
happens
is-
is
that
that
just
throws
off
a
whole
lot
of
the
spatial
pooling
and
basically
confuses
the
thing.
And
then
that
causes
bursting,
because
it
thinks
that
sequence
has
been
interrupted.
B
But
it's
because
if
you
have
like
deep
structure
in
the
semantics
of
the
field
structure
that
you're
feeding
in
then
you
have
to
do
a
lot
of
work
to
get
something
that
when
things
are
not
anomalous,
that
you
get
everything
going
green
all
the
time
and
that
it
detects
real
anomalies.
But
what
you
could
be
looking
out.
B
There
is
just
that,
there's
like
an
alphabet
change
or
that,
like
a
new
piece
of
vocabularies,
being
introduced
that
just
that
particular
combination
of
fields
has
never
been
seen
before
by
the
the
training
process
and
that's
what
generating
these
their
false.
Positive,
a
anomaly
detection.
But
it's
actually
because
of
the
dimensionality
of
the
structure
of
the
day
that
you're
giving
it.
E
A
E
Then
the
data
was
really
this
part
like
it
was
like
basically
he's
built
on
those
timeouts
that
you
have
no
idea.
If
there's
think
they're
not,
and
then
you
have
multiple
chaps
and
then
the
thread
I
mean
the
your
browser
is
going
to
be
slow,
sometimes
and
speed
up
again.
So
it's
you,
you
should
look
at
the
analogies
and
the
data
behind
their
clicks
or
or
what
you.
Why
is
happening
with
the
analytics
behind
it,
I
mean
d,
analytics
I
mean
the
behavior
of
like
conversion
or
anything
like
that.