►
From YouTube: Famous: 2014 Spring NuPIC Hackathon Demo
Description
Francisco Webber, Soeren, and Erik from CEPT.
A
So
my
name
is
eric
and
this
is
siren
and
we
are
both
from
sept
and
in
this
hack
we
tried
to
combine
the
sub-retina
and
the
cla
and
people
who
have
been
at
the
last
hackathon.
No,
probably
the
fox
eats
rodent,
and
we
tried
to
extend
it
a
little
bit.
Our
main
goal
here
was
to
get
more
exposure
to
the
cla
to
get
to
know
it
better
to
understand
it
better
and
also
to
test
drive
a
new
set
of
retinas
that
we
have
developed,
which
have
some
special
properties
and
so
for
the
first
experiment.
A
A
Physicist
physicist,
you
have
like
the
profession
itself
and
then
you
have
like
exposure
of
physicists
in
movies,
where
crazy
physicist
wants
to
build
a
doomsday
machine.
You
have
historical
aspects,
you
have
university
teaching
aspects,
so
they
are
very
complex
and
very
spread
out
fingerprints,
for
example.
A
This
would
be
the
semantic
fingerprint
for
physicist
and
you
can
see
if
we
visualize
it
that
it's
very
much
spread
out
over
the
semantic
map
of
our
semantic
fingerprints.
So
the
semantic
fingerprints
that
we
use
they
have
topology.
A
And
if
bits
are
more
in
the
upper
right
corner,
then
they
will
relate
somehow.
You
can't
say
exactly
what
it
is,
for
example,
the
big
blob
on
the
right.
That
seems
to
be
like
the
physics
or
physicist
area
and
from
the
pattern
of
the
bits
you
can
see
that
this
is
very
spread
out
and
applies
to
many
different
contexts.
A
A
So-
and
these
are
all
persons
that
the
cla
has
never
seen
before,
you
can
see
that
up
there
yeah,
I'm
sure.
So
it's
like
katy
perry,
robert
oppenheimer,
don
henley,
and
it's
quite
good.
A
A
So,
let's
introduce
something
a
second
verb,
so
here
we
have
just
four
sentences:
that's
all
and
we
say
eugene
weekner
like
mathematics,
most
compounding
like
mathematics,
so
they're,
both
physicists
and
then
we
take
katy
perry
like
fans
and
then
sinatra
like
fans,
and
then
we
and
then
we
run
this
again.
The
likes.
A
Yes,
so
I
it
has
like
the
corpus
of
the
input,
data
of
all
the
singers
and
physicists,
but
on
top
of
that
now
we
place
these
four
sentences,
with
the
expression
of
both
going
paulie
likes.
Mathematics
and
katy
perry
likes
fans.
A
A
Do
you
think
without
changing
the
training
set,
we
could
see
what
barack
obama
would
be?
Yes,
we
can
try
to
then
afterwards,
but
let
me
go
to
the
next
step
because.
B
A
B
A
Now
we
have
actors
like
mathematics
and
tom
cruise
like
mathematics
and
initially
I
was
a
little
bit
disappointed,
but
then
I
was
thinking
it
didn't,
have
a
lot
of
exposure
when
you
think
about
how
little
or
few
sentences
the
cla
saw
and
in
my
mind
of
course,
actors
are
more
closely
related
to
singers
than
to
physicists,
but
for
the
cla.
Maybe
that's
not
true.
F
It's
really
not
that
it's
really
related
to
which
people
right.
Isn't
it
more
like
it's,
which
it's
never
seen.
Physicist
is
the
first
word
here
right,
so
I
mean
you're
really
saying
if
I
give
it
actors
it
has
to
it
has
to
it
has
to
be
related
to
the
names
you
saw
before
right.
There
has
to
be
an
overlap
between
actors
and
the
in
the
names
you
know,
nancy
sinatra
or
whatever.
A
F
Yeah,
but
for
it
to
do
the
right
thing
here,
the
the
word:
the
sdr
for
actors
has
to
overlap
substantially
with
the
sdrs
for
the
people
who
there
are
act
or
musicians
actually,
and
that's
interesting.
I
don't
know
how
much
you'd
get
an
overlap.
There
would
be
interesting.
Did
you
say
singers
like
because,
like
nancy
sinatra
and
katy
perry
are
singers.
D
B
B
D
The
thing
is
when
you,
when
you
see
the
demo
that
you've
just
done
the
problem
is
that
human
beings,
looking
at
what
you've
just
been
doing,
go
well.
Of
course,
that's
going
to
happen
because,
of
course,
wolfgang
paulie
likes
mathematics
right
but
you've,
just
given
a
a
computer
system
with
a
some
sort
of
memory,
lookup
thing
that
produces
sdrs
the
text
of
wolfgang
paoli.
C
B
F
A
A
F
Know
the
answer
to
these
questions
is
it's
not
whether
you
get
the
right
answers,
whether
you
get
a
good
answer
and-
and
we
have
to
remember
that
so
it's
kind
of
hard
to
know
until
you
see
some
of
the
other
paul
tunis.
But
albert
einstein
was
a
very,
very
famous
guy
and
he
didn't
shy
away
from
his
publicity
and
you
know-
and
he
was
vetted
by
kings
and
queens,
and
so
you
know
he
could
definitely
be
up
in
the
fans
category.
You
never
really
know
we
can't.
F
E
Very
cool,
very
that's,
really,
really
cool
question.
The
detailed
question
is
the
sparsity
level
coming
out
of
the
sept
retina.
Is
that
have
been
normalized
in
this
api?
So
is
that.
A
E
F
Well,
so
we
we're
discussing
this
earlier
in
the
week
and
I
you
know
our
original
actual
reaction
was
well.
You
have
to
get
it
normalized
to
certain
percentage,
but
then
super
tight
pointed
out
that
maybe
not,
and
then
we
thought
about
a
little
bit
more
and
we
said
yeah
given
it
may
not
actually
care
too
much.
We
thought
it
would,
but
maybe
not
so
it's
more
might
be
more
impossible.
C
Yeah,
maybe
yeah.
What
will
happen
is
if,
if
let's
say
physicists
had
or
if
fans
had
you
know
very
high
density,
like
five
percent
and
mathematics
had
very
low
density
of
one
percent,
and
you
know
it
might
kind
of
predict
fans,
but
even
kind
of
predict
fans
might
overload
what
mathematics
is.
So
you
have
to
be
a
little
careful.
C
G
Yeah,
so
I
just
wanted
to
say
that
I
was
really
thrilled
by
this
I
mean
it
is
definitely
not
a
scientific
analysis
by
by
looking
at
those
results,
but
still
it
sort
of
behaves.
How
one
would
expect
it
intuitively.
So
you
don't
need
to
sort
of
think
about
this
for
two
hours
to
come
up
with
some
sort
of
plausible
internal
explanation.
G
But
the
really
fascinating
thing
is
that
we
are
looking
here
at
things
like
10
examples
of
this
three
additional
examples
of
that,
and
it
already
looks
so
to
say
very
interesting
and
it
seems
to
sort
of
cross-link
aspects
that
they
did
not
see
linked
together
before,
and
I
think
that
this
has
to
do
with
the
fact
that
we
so
to
say
have
a
very
rich
semantic
data
that
we
are
working
with.
I
think
that
is
what
is
the
most
to
me,
the
most
important
outcome.
G
So
if
one
would
try
to
just
make
a
simulator
of
this
simulation
using
traditional
nlp
things,
you
would
need
like
1000
times
the
amount
of
data,
and
it
probably
would
take
hours
to
actually
pipe
the
whole
data
through
and
and
again.
It
would
not
give
like
a
clear
statement
like
this,
so
I
find
this
really
unexpected.
Even.
E
You
said
you're
hooked
into
cla.
I
think
this
is
gets
me
really
excited
about
sept
in
reverse,
because
a
lot
of
these
sequence
learning
things
we
use
numbers
and
things
like
that,
where
they
don't
have
that
much
semantic
overlap
between
the
different
patterns.
But
with
word
with
you
know,
language,
like
you
guys,
have
done.
There's
a
lot
of
semantic
overlap
between
things.
So
there's
a
lot
of
scope
for
generalizability,
which
there
isn't
in
a
lot
of
other
kinds
of
problems,
and
so
there's
you
get
a
lot
of
interesting
behavior
here
due
to
that
generalizability.
A
Yes
and
one
thing
that
I
forgot
to
mention
so
in
order
to
get
this
running,
for
example,
I
said
at
the
beginning
that
we
physics
or
physicists
is
a
very
distributed
semantic
fingerprint.
So
the
one
trick
that
we
did,
which
differs
from
fluent,
is
that
when
we
enter
the
third
fingerprint
we
take
the
and
between
the
person
and
physicist
or
singer,
because
we
want
to
condense
the
fingerprint
only
to
concentrate
on
the
profession
part
so
that
we
can
say
okay.
A
C
So
what
do
you
do
with
when
you
fit
in
that
example,
where
you
fed
in
tom
cruise?
How
did
you
know
to
in
that
case,
you
would
want
to
end
it
with
actor.
B
A
I
just
always
end
like
the
profession
with
the
person,
but
this
is
just
like
to
help
me
within
the
heck
to
create
a
more
focused
fingerprint,
which
only
represents
the
profession.
E
A
With
more
context
and
longer
learning
time
that
this
would
happen
automatically,
but
it
was
just
like
a
shortcut.
I
think.
D
G
Maybe
maybe
this
was
not
clear:
it
was
not
about
to
find
out
what
a
profession
a
person
is.
It
was
taking
one
person
where
we
know
it's
a
physicist,
and
we
take
this
person
as
an
example
fingerprint
to
end
it
with
the
name
of
the
profession,
to
get
only
those
points
that
are
really
relevant
to
the
profession
and
not
for
physicists.
That
is
more
precise,
it's
only
so
to
say
physicist
as
a
profession.
That
would
be
the
one
and-
and
you
can
achieve
this
by
taking
a
random
person.
G
A
Yeah
we
do
this
quite
often
also
like
we
call
this
synthetic
fingerprint,
so
you
want
to
represent
something.
For
example,
you
want
to
have
a
fingerprint
which
represents
computers,
so
we
just
take
like
a
couple
of
computers
and
we
and
them-
and
this
is
then
the
essence
of
computers-
they
might
have
different
bits,
but
when
we
end
them
then
essentially
this
looks
like
a
synthetic
fingerprint
representing
a
certain
class
exact
concept.
D
So
I
think
the
nice
thing
about
this
and
the
reason
why
the
people
on
the
kind
of
new
pic
and
jumento
side
of
things
are
asking
so
many
questions
about.
This
is
because
this
is
a
classic
example
of,
like
obviously,
there's
something
really
interesting
going
on
here,
but
none
of
us
know
which
bits
of
that
are
happening
on
the
sept
side
and
which
bits
of
it
are
happening
on
the
cla
site
right
and
that's
exactly
the
way
that
the
brain
works
so
that
it's
some
of
it
happens
on
the
above.
D
D
D
F
F
I
mean
this
gets
back
to
the
varying.
This
gets
back
to
the
varying
levels
of
sparsity
in
the
in
your
sdrs,
and
I
was
wondering
if
we,
if
we
brought
them
all
down
to
the
same
two
percent
through
other
means,
would
it
might
have
worked
better.
A
Yeah
it
might
have
been,
but-
and
this
is
more
to
also
like
what
training
data
went
into
the
retina,
because
usually
when
you
try
to
specify
without
looking
at
the
topology,
then
the
strongest
context
will
win.
And
that
might
have
been
like
this
exactly
this
context.
But
it
might
have
been
also
like
the
movie
physicist,
the
movies
context
or.
A
B
D
F
B
D
D
It
only
meant
fruit
right
exactly
so
yeah,
so
the
thing
is
that
people,
our
age
now
are
confused
by
some
of
these
things
that
sept
has
produced
by
just
analyzing
the
statistics
of
the
meaning
of
words,
and
so,
for
example,
albert
einstein
to
a
kid
in
popular
culture.
Now
is
not
a
physicist,
but
he's
a
guy
who
sticks
out
his
tongue
in
a
black
and
white
picture.
F
For
example,
it
was
in
some
apple
heads,
so
yeah
we
probably
need
to
get
going
here.
But
what's
this
gonna
point
about
the
part
of
what
sept
has
talked
about
don
doing-
and
I
don't
speak
for
them
but
is
obviously
this
is
this
is
what
the
people
who
wrote
wikipedia
entries.
F
Think
of
albert
einstein
and
that's
biased,
but
the
whole
point
here
is:
you
can
build
the
different
sets
of
finger
word
fingerprints
by
different
corpuses.
So
you
you
put
this
through
pubmed.
You
get
one
thing
you
put
it
through.
You
know
physics,
letters,
you
get
another
thing,
you
threw
it
through
the
patent
office,
you
get
another
thing
so
and
that's
exactly
what
you
want
and
that's
part
of
the
power
and
you
can
sell
those
differences
and
make
some
more
money.