►
From YouTube: Angry Bots in AI World: 2014 Spring NuPIC Hackathon Demo
Description
Craig Quiter
A
Cool
okay,
so
yeah.
The
basic
idea
of
this
is
to
actually
recognize
the
image
on
the
top.
These
are
two
eyes,
basically
on
the
top
and
bottom
for
player,
1
and
player,
2
and
player.
1
is
going
to
be
controlled
by
new
pic,
but
actually
I
recognize
the
image
with
a
pre
trained
neural
net
that
does
image
classification.
A
It
did
really
well
on
imagenet
10,000
category
contest,
so
it
basically
takes
this
these
two
eyes
on
top
and
transforms
them
into
a
thousand
numbers
that
represent
features
that
a
team
at
Berkeley
was
able
to
say
whether
something
had
a
cat
or
the
type
of
bird
or
you
know
some
bookcase
or
whatever
category,
and
this
image
net
data
set
and
these
features
actually
are
generalizable.
So
other
teams
have
just
used
this
preacher
a
neural
net
and
won
other
competitions
and
been
able
to
recognize
images,
so
it
works
on
things
other
than
imagenet.
A
Let's
see
so
I
did
like
a
random
bot
on
top
and
I
played
against
it,
and
I
took
the
times
where
the
random
bought
was
successful
in
hitting
me
with
a
shot,
and
I
multiplied
that
experience
by
10
and
ran
that
through
new
pic
and
I
any
time
that
the
bot
was
shot.
I
took
that
out
so
I
only
trained
it
on
the
times
it
was
successful
and
like
Matt,
unfortunately,
I
wasn't
able
to
like
do
something
super
intelligent
here
with
the
bot.
A
A
A
Ok,
so
I'm
controlling
the
guy
and
now
he's
he's
kind
of
settled
on
I'm
just
going
to
not
move
to
the
right,
so
he
either
decides
to
move
to
the
right
really
fast,
the
whole
time
or
just
stay
still
depending
on
what
he
sees,
which
is
kind
of
interesting.
So
I
can
actually
control
everything
except
the
horizontal
direction
on
this
top
bot.
So
I
can
actually
like
go
towards
this
wall
here
and
if
you
look
on
the
bottom
at
the
features
this
this
number
here
is
coming
from.
A
A
What
does
that
vein?
You
in
as
zeros
out
that
yeah,
it's
not
survivable
so
zero
feature
is
just
a
you
know:
it's
a
input
to
an
SVM,
an
image
classification
algorithm
if
it
means
just
a
high-level
feature,
so
it
could
mean
like
a
cat.
You
know
Anna
if
it's,
if
it's
image
or
you
know,
if
a
five
hundredth
feature
in
this
case
feature
189
was
the
one
that
thought
was
was
the
one
new
pic
thought
was
important
to
determine,
which
direction
to
move
next
yeah
I.
A
B
Just
got
several
questions:
I
wanna,
make
sure
I
understand.
What's
going
on
here,
so
you've
taken
up,
you
have
a
thousand
parameters
coming
out
of
this
vision
system.
You
swarmed,
you
picked
one
you're
now
feeding
that
as
a
series
of
data
points
into
new
pic,
you
train
it
on
good
ones,
good
sequences.
How
often
is
that?
How
often
is
that
sampled
10.
B
B
A
B
B
A
A
C
Yeah
just
so
we
were
talking
earlier
I
think
you
know
this
idea
as
well
as
like
the
pong.
The
the
puck
idea
is
really
interesting,
because,
essentially
just
by
watching,
you
play
it's
kind
of
learning
how
to
play
better
and
better.
So
I
think
the
core
of
that
idea
is
a
really
good
idea.
It's
a
very
ambitious
kind
of
gold
to
do
this,
so
I
think
you
know
it's
going
to
take
a
little
while
to
so
work
through
the
kinks
and
things
like
that
and
as.
B
B
C
B
C
B
C
B
Know
you
can
/
sample
and
so,
if
you're
feeding
a
series
of
zeros
or
series
of
any
particular
number
into
CLA
and
you're
trying
to
learn
the
sequences
of
them,
it's
it's.
Basically,
I
think
it's
a
static
variable,
so
I'm
just
curious
down.
There
are
both
zero.
I
guess
or
something
like
that.
Mm-Hmm
may
I
something
to
think
about
too,
because
it
there's
a
tin.
You
higher
data,
isn't
better!
You
want
you
want.
You
want
variability
dinner,
you
can
take
any
data
stream
if
you're
over
sample
it
looks
flat
and
that's
not
good.