►
From YouTube: Duck Duck Moose [DEMO #5] (2015 Spring NuPIC Hackathon)
Description
Frank Carey of the NYC Bots and Brains Meetup works with ImageNet and Torch to create SDRs representing features as input for NuPIC.
http://nupic2015spring.challengepost.com/submissions/37839-duck-duck-moose
A
It
says:
shame
leap
shamelessly
plug
our
meetup
group,
BOTS
and
brains,
and
if
you're
interested
in
sort
of
this
cross
between
neuroscience
and
a
I
highly
recommend,
you
check
it
out
so
I've
been
interested
in
AI
for
a
while
and
I've
been
teaching
myself
a
lot
of
the
deep
learning
stuff
and
so
I
was
curious.
How
I
could
sort
of
bastardize
both
these
algorithms
and
sort
of
glue
them
together
and
sort
of
a
contrived
example
that
I
came
up
with?
Was
a
duck
duck
goose
game?
Everyone
remembers,
duck
duck
goose
as
a
kid
right.
A
You
walk
around
the
circle.
You're
patting
people
on
the
head,
duck
Duck,
Duck
and
eventually
hit
two
gates,
and
it's
like
it's
like
a
an
anomaly
right.
So
what
I
wanted
to
do
was
a
new
pic
cannot
far
as
I
know,
really
handle
image
data
it's
just
much
too
large
and
there's
not
a
hierarchical
representation,
but
deep
learning
actually
is
really
great
at
getting
the
Thunder
lying
features
and
the
underlying
hierarchy
of
the
image
itself,
and
so
what
I
decided
to
do
is
try
to
see
if
I
could
use
the
here
at
coming.
A
If
see,
if
I
could
use
the
deep
learning
network
to
as
a
feature
detector,
that
I
could
use
to
actually
generate
the
SDR
automatically
for
the
images,
so
I
used
a
set
of
images
from
imagenet
and
if
you
haven't
heard
about
imagenet,
this
is
what
everybody
is
crushing
right
now
with
some
of
the
algorithms
lately
and
what
they
do
is
they
basically
have
labeled
data?
They
have
a
picture
of
a
duck
in
this
case,
and
then
they
have
some
ones
labeled
it
with
duck.
A
But
in
this
data
set
there's
only
a
thousand
categories,
so
they
had
to
like
be
selective
and,
interestingly
enough.
There's
not
duck
is
not
one
of
the
categories
in
this
data
set.
They've
got
all
kinds
of
weird
other
strange.
You
no
exotic
birds,
but
they
don't
have
duck.
They
have
goose,
but
they
don't
have
a
duck.
So
I
thought
a
lot
of
when
you
get
these
results
back.
I
want
to
show
you
that
real
quick
when
you
get
the
results
back
they're
effectively
probabilities.
So
even
though
it's
not
a
dud,
they
don't
have
duck.
A
It
gives
suggestions
that
are
similar
to
duck
right.
It
gives
like
a
lot
of
the
other
birds.
It's
going
to
take
the
man,
let's
scroll
up
here,
and
so
you
can
see
it's
really
dark
one
second,
so
these
are
some
of
the
categories
that
come
back
right
so
for
this
one.
Its
first
guess
was
see
here
for
this
one.
Its
first
guess
was
that
seventy-three
percent
chance
that
that
was
a
goose
whatever.
That
was
right,
then
the
second
best
gas
was
twenty
one
percent.
A
Here
I
wanted
to
go
further
than
this
actually
remove
the
classifier
and
actually
get
at
some
of
those
underlying
features,
but
just
kind
of
ran
out
of
time,
and
so
the
you
know,
the
sort
of
dumb
idea
of
the
game
is
that
you're
playing
duck
duck
goose
and,
as
you
click
this,
you
can
see
these
pitaah
ten
or
as
let
you
know
some
of
them
like
goose,
were
actually
so
popular.
So
there's
the
SDR
right
I
tried
to
format
it
like
something
like
you
guys
recognize.
A
So
in
this
case
it
was
here's
three
categories
for
this
image,
right
that
it
said
these
were
the
top
three
guesses
for
that
image
right
and
only
of
the
thousand.
You
know
in
the
data
set
categories,
and
so
some
of
these
have
ten.
Some
of
these
have
one,
but
most
I'll
have
something-
and
so
you
know
I'm
all
sort
of
I'm
very
ready
to
move
forward
with
it.
But
it's
I
want
to
dig
into
those
underlying
features,
actually
get
this
much
more
much
more
less,
sparse,
I
guess
you'd
say
so.
B
A
They
seem
like
neural
networks
or
deep
now,
deep
learning
networks
are
not
really
great.
They
have
some
temple
components,
they're
ready
now
with
our
current
Nets,
but
the
sort
of
long-term
and
they're
very
you
know,
compute,
heavy
and
so
I'm
thinking
that
we
could
sort
of
get
the
best
of
both
worlds
with
deep
learning.
That's
a
really
great
at
is
that
sort
of
encoding,
so
we've
all
been
doing
me
and
cook,
like
you,
know,
who's
like
trying
to
figure
out.
How
do
I
encode
my
data
today
right?
We
were
all
doing
that.
A
Well,
the
deep
learning
stuff
is
a
really
good
at
at
least
with
certain
types
of
day,
like
images
of
actually
generating
that
sparse
representation
for
us.
So
my
interest
is
like
how
I
can
actually
take
that
data
and
then
kind
of
glue
these
two
algorithms
together.
So
we
started
that
process,
but
you
know
a
little
more
to
go.