►
Description
CEPT SDR Session with Francisco Webber (NuPIC 2013 Fall Hackathon
A
B
Okay
welcome
this
morning,
I
have
had
the
opportunity
yesterday
to
talk
with
many
of
you
on
one
of
the
other
aspects
of
of
our
work,
and
so
I
would
just
kind
of
try
to
make
the
picture
complete
by
now
and
show
you
more
or
less
why
we
have
started
doing
what
we
are
doing
and
how
we
are
doing
it.
B
So,
maybe
to
give
you
a
bit
of
a
context.
I
was
personally
following
jeff's
work
for
many
years
now,
shortly
after
he
published
his
book
and
at
that
time
I
was
deeply
into
trying
to
find
ways
to
better
process,
textual
information
and
stuff
like
that.
That
was
involved
in
a
search
engine
developments
for
various
application
areas.
B
The
latest
one
was
patent
search,
in
fact,
which
is
a
pretty
tough
discipline
in
information
retrieval,
because
you
tend
to
have
a
lot
of
documents
where
even
humans
have
problems
in
reading
it
and
getting
what
it's
talking
about,
and
we
basically
tried
endless
numbers
of
algorithms
methods
to
improve
the
the
the
the
retrieval
of
such
documents
from
repositories.
B
B
B
So
if
you
want
to
improve
this,
if
you
want
to
actually
find
the
five
documents
within
100
million
documents
that
are
actually
important
for
what
you're
doing,
then
you
need
to
invest
a
lot
of
energy
yeah.
So
every
percent,
I
would
say
beyond
the
easy
ones,
is
extremely
expensive
due
to
the
fact
that
you
need
to
do
things
like
annotations.
B
So
our
approach
needed
to
be
completely
different
than
this,
because,
basically
all
the
statistical
approaches
they
have
been
investigated
pretty
deeply
and
we
wanted
to
find
another
approach,
even
if
it's
not
as
good
in
the
beginning,
but
just
to
have
another
approach,
gives
us
another
tool
set
basically
to
to
face
the
problems
that
we
have
there
and
we
started
with
a
pretty,
I
would
say,
straightforward
thinking,
namely
that
language
is
produced
by
humans,
which
means
by
brains
and
language
is
composed
of
symbols.
B
So
it's
not
an
information
that
you
can
understand
if
you
have
not
the
whole
education
and
knowledge
about
the
language,
you
cannot
decipher
it
because
the
the
the
symbols
themselves
don't
tell
you
what
they
are
talking
about,
and
so
that's
exactly
the
problem
the
computer
has
when,
when
he
processes
the
text.
B
So
this
is
one
issue
that
language
and-
and
we
are
just
concerned
about
text,
so
we
are
not
concerned
about
the
speech
recognition
about
ocr
about
all
the
other
disciplines,
so
to
say
that
necessary
to
basic,
basically
bring
in
the
information
from
its
physical
existence
might
be
a
tape
recording
of
somebody
saying
something.
B
So
what
we
focus
on
is
the
actual
text
level,
which
means
we
start
to
reason
at
the
moment
when
we
actually
know
what
word
appears
so
to
say,
and
we
don't
care
how
it
made
it
up
there,
and
the
second
issue
is
that
knowledge
and
the
meaning,
so
to
say
of
words,
is
a
problem
in
the
thought
that
you
don't
in
fact
what
we,
what
we
don't
know
is
how
do
we
represent
the
meaning
in
a
way
that
it
speaks
for
itself
yeah.
B
B
The
same
thing
is
not
true
for
words.
If
you
know
one
word,
it
doesn't
tell
you
how
you
should
read
the
next
one
and
where
the
meaning
is,
and
this
representational
problem
is
in
fact
a
problem
that
is
very
fundamental
to
artificial
intelligence
in
general
yeah.
So
how
do
you
actually
capture
what
we
observe
in
intelligent
behavior?
How
do
you
record
this
and
make
it
happen
somewhere
else?
B
So
what
you
can
do
is
that
you
capture
what
is
called
features
basically
and
you
try
to
create
feature,
sets
that
you
work
on
that
you
process
and
that
you
can
store
in
a
computer,
but
for
the
the
language
case
or
the
text
case
to
be
most
more
precise.
B
The
features
are,
for
example,
in
the
traditional
approach,
just
statistical
values.
The
word
such
and
such
occurred
that
many
times
in
a
text
and
that
many
times
in
a
collection
where
the
text
is
a
part
of-
and
that's
a
very
slim
feature
so
to
say
so.
B
In
order
to
solve
our
problem,
we
need
to
find
ways
in
actually
learning
features
about
text
and
learning
them
in
a
sufficient
amount
and
in
sufficient
quality
that
we
can
reproduce
on
the
quotes.
What
is
meant
by
by
text
or
by
language.
B
So
basically,
what
a
language
does
from
the
point
of
view
of
the
brain
I
would
say
is
that
it
basically
captures
a
set
of
sensorial
information
that
you
got
a
context
of
a
number
of
real
world
things,
and
we
try
to
capture
this
in
order
to
send
it
to
another
brain
and
have
the
other
brain
have
the
same
experience
as
I
had
it.
When
I
created
the
the
the
text
basically
or
the
language
that
describes
it.
So
it's
a
a
way
of
communicating
stimuli.
B
That,
in
principle,
should
go
to
the
brain
and
we
developed
language
so
to
say
to
capture
this
and
to
transfer
it
and
to
allow
another
brain
or
person
to
experience.
Something
without
actually
experiencing
it,
and
this
is
in
fact
an
interesting
improvement.
If
you
compare
this
to
animals
who
are
not
able
to
do
this,
they
lack
this,
I
would
say
social
dimension
and
the
social
processing
of
of
information.
B
What
is
interesting
is
if
you
observe
neuroscience,
or
brain
science
and
the
science
of
of
language
understanding.
So
to
say,
you
see
that
many
of
the
complexities
that
you
encounter
are
very
similar
somehow
and
if
you
go
into
detail,
you
find
out
that
they
actually
come
from
the
same
fundamental
problems,
and
this
basically
shows
nicely
that
brain
and
language
are
extremely
linked,
and
my
idea
was
that
by
better
knowing
how
the
brain
works,
it
will
help
me
in
better
know
how
the
language
thing
works
and
vice
versa.
B
Good,
I
started
thinking
basically
by
doing
a
number
of
assumptions,
so
I
decided
so
to
say
to
take
them
as
given
and
to
try
to
find
a
reasoning
based
on
that,
without
necessarily
looking
for
so
to
say,
the
the
fundamental
truths
of
these
assumptions,
because
it's
the
the
the
field
is
so
complex,
so
that
if
you
want
to
basically
define
all
by
yourself
it
it
is
not
feasible.
B
So
one
assumption
is
that
the
neocortex
is
composed
of
modules,
of
small
structures
that
are
repeated
over
and
over
again
and
that
the
more
of
these
modules
you
have,
the
better
the
more
processing
power
on
the
quotes
you
have.
The
second
assumption
is
that
these
modules
are
designed
according
to
a
building
plan.
B
That
is
the
same
for
all
the
modules,
so
they
are
so
to
say
there
is
one
type
of
module
that
is
available
and
this
one
type
of
mole
of
of
module
has
an
algorithm
that
does
basically
everything
that
is
needed
for
the
whole
structure
as
such,
and
you
see
this
in
in
many
ways,
for
example,
that
the
brain
areas
that
do
vision
don't
look
very
different
from
brain
areas.
That
process,
auditory
information
and
so
on.
B
The
next
assumption
is
that
the
inputs
and
the
outputs
to
the
cla
are
built
of
sdrs,
so
that
there
is
a
a
standard
way
of
inputting
data
in
a
layer
and
a
standard
way
of
getting
this
data
and
by
having
sdrs
at
the
input
and
at
the
output.
It
means
that
you
can
actually
stack
layers
and
feed
the
output
from
one
layer
to
the
input
of
the
next
layer.
B
The
next
assumption
is
that
the
fundamental
input
where
new
information
comes
along
is
always
by
sensors.
They
need
to
be
censored
somewhere
and
they
need
to
transform
the
data
whatever
they
need
to
to
transmit
in
the
format
of
the
sdr
and
to
bring
it
to
the
cortex,
and
the
last
assumption
so
far
is
that
the
organization
of
these
repetitive
structures
is
organized
in
regions,
layers
and,
finally,
in
hierarchies
that
all
communicate
between
each
other
with
sdrs.
B
So
it's
I
I
think
all
of
these
assumptions
are
pretty
obvious
to
you
guys
here.
I
I
just
wanted
to
say
that
I
tried
to
start
with
a
with
a
very
easy
set
of
hypothesis.
From
my
point
of
view
and
to
try
and
find
some
answers,
so,
as
I
said
before,
we
are
focusing
on
a
very
specific.
B
B
So
this
is
a
bit
low,
so
in
principle
you
could
create
a
representation
like
this,
where
you
create,
where
you
get
a
symbolic
input,
symbolic,
because
it
is
text
at
some
point
and
from
the
symbolic
input
we
have
the
the
adaption
so
to
say
of
the
information
until
we
reach
the
word
level
and
from
there
there
is
a
stream
of
word
patterns
that
is
created
and
that
corresponds
to
the
content
of
the
actual
text
or
speech
or
whatever.
B
That
is
understood
and
based
on
this,
there
is
some,
let's
say,
reasoning
and
processing
happening,
and
then
the
the
whole
thing
goes
downwards
through
the
same
system
through
the
motor
output
and
we
then
produce
again
symbolic
information
based
on
what
the
result
of
our
processing
about
the
language
has
given,
and
this
layer
of
the
word
sdr
is
what
we
have
tried
to
capture
in
a
technical
fashion,
and
we
call
this
retina,
it's
a
bit
ambiguous.
B
I
know,
but
this
sort
of
happened
while,
while
working
on
this,
the
idea
is
that
the
sept
retina
that
actually
maps
words
into
sdr
word.
Sdr
representations
has
a
very
similar
functionality
if
you
want
as
the
real
retina
that
is
converting
images
of
the
surrounding
into
something
that
the
brain
can
process
and
in
fact
the
retina
itself
is
also
not
only
a
sensor
but
somehow
a
processor
that
pre-processes
the
information
accordingly,
yeah.
B
So
if
we
want
to
represent
words
by
sdrs
in
order
to
be
compatible
and
assuming
that
somewhere
in
the
brain,
this
has
actually
to
happen.
If
we
have
all
the
aspects
of
the
brain
as
I've
listed
before,
we
have
to
see
what
is
needed
from
such
a
word
as
they
are
to
be
a
true
sdr
in
the
cortex
sense.
So
to
say
one
respect.
B
One
aspect
is
that
it
is
a
bit
vector,
so
every
bit
in
this
bit
vector
has
an
actual
semantic
meaning,
and
that's
precisely
what
words
are
trying
to
capture
are
mostly
big
numbers
of
fine
granular
semantic
meanings
of
where
this?
U,
where
this
word
can
be
used
in
what
context
of
other
words
it
can
be
used
and
what
it
represents
in
different
ways
in
different
moments,
even
sometimes,
and
each
of
these
features
is
represented-
let's
say
by
a
bit
in
this
binary
vector
of
the
sdr.
B
So
what
it
needs
to
comply
to
the
sdr
is
the
sparsity.
B
So
we
have
seen
that
in
order
to
properly
process
it
in
the
cortex,
it
has
to
have
a
high
sparsity
in
order
to
keep
so
to
say,
the
the
combinatorial
space
very
large
and
in
fact,
if
you
calculate
what
could
be
possible
so
to
say
in
a
combinatorial
way,
by
using
language,
you
would
have
basically
endless
variations
of
sentences
and
so
on,
and
still
we
just
need
a
very
thin
layer
of
true
sentences,
so
to
say
compared
to
combinatorial
sentences
that
do
not
have
any
sense.
B
A
very
interesting,
a
very
interesting
functionality.
I
would
say
that
you
can
get
by
adding
sdrs
together
and
by
having
this
compositionality.
B
B
This
might
seem
a
bit
complicated,
but,
as
we
found
out,
this
is
a
fundamental
functionality
that
actually
allows
you
to
do
computations
with
the
sdrs,
because
then
they
behave
like
like
a
set
according
to
set
theory,
and
you
can
do
many
operations
according
to
this
set
theoretical
approach
with
sdrs,
and
if
we
manage
to
convert
words
into
these
sdrs,
we
would
be
able
to
do
this
with
words.
B
The
second
most
important
aspect
is
the
similarity,
the
the
way
how
you
could
compute
the
similarity
between
two
representations
and
therefore
the
repres,
the
similarity
between
two
words-
and
this
is
very
hard
to
sort
of
capture,
because
two
different
people
would
have
two
views
on
the
similarity
of
two
words
but
in
what
is
interesting.
B
Is
that
with
all
this
blurriness
that
you
find
there,
we
still
always
basically
find
some
common
ground
and
therefore
we
are
able
actually
to
have
conversations,
because
if
you
would,
if
you,
if
you
would
need
to
clarify
every
word
what
we
actually
mean
before,
we
continue
our
conversation.
B
There
wouldn't
be
any
conversations,
so
the
similarity
is
a
very
subtle
measure
and
we
have
found
out
that
making
this
computable
so
to
say
is
one
of
the
biggest
features
that
we
have
in
using
static
sdr
representation
of
words,
which
I
will
show
later
also,
and
then
I
have
a
last
aspect.
B
You
can't
see
this.
I
tried
to
make
this
in
a
different
color,
because
that's
my
very
personal
approach
to
it.
I
do
personally
think
that
the
topology
in
the
sdrs
is
also
very
important.
So
this
is
so
to
say
very
much
discussed
in
the
last
in
the
last
weeks.
B
I
personally
believe
that
the
sdrs,
as
we
see
them,
resulting
from
words
they
are
very
similar
to
sdrs
from
the
visual
system,
for
example,
where
the
fact
that
a
certain
number
of
features
is
positioned
in
a
certain
place
on
the
sdr
is
part
of
the
whole
information
so
and
has
therefore
be
to
be
considered,
and
this
is
essential,
as
we
will
see
later
on
the
question-
how
to
build
actually
hierarchies
that
process.
This
kind
of
data
yeah.
B
So
I
I
leave
this
for
for
later,
so
that
the
key
on
the
on
the
computing
part
is
so
to
say
that
we
want
to
use
word
vectors
to
actually
be
represented
by
sdrs,
and
we
describe
the
words
by
using
collections
of
features,
as
I
said,
and
built
a
matrix
space.
So
the
matrix
space
is
basically
not
a
space
with
a
coordinate
system,
but
it's
a
space
where
the
measure
is
done
by
a
similarity
of
a
distance
between
two
points.
B
So
the
metric
space
is,
if
you
want
an
endless
environment
that
allows
to
compare
any
two
entities
by
comparing
the
distance
between
the
features
and,
as
we
will
see,
this
works
pretty
well
the
problem
to
do
this
to
actually
create
word
vectors
that
correspond
to
the
meaning
of
words
and
that
comply
with
the
necessities
of
being
sdrs.
B
This
is
pretty
tough,
so
in
many
different
ways
this
has
been
tried
for
many
years.
In
fact,
and
many
different
approaches
have
been
done-
I
mean
the
easiest
approach
would
be
to
create
the
features
manually.
So
you
take
a
list
of
all
the
words
that
you
want
to
use
and
for
each
of
the
words
you
say,
okay,
I
want
to
call
the
color
feature.
I
want
to
call
a
specific
taster
feature
and
then
you
would
have
to
start
and
create
manually
a
dictionary
that
basically
associates
a
number
of
features
to
every
word
problem.
B
There
is,
even
if
you
have
the
manpower
to
do
so.
If
you
give
the
same
so
to
say
task
of
of
of
tagging
words
to
another
group
of
people,
you
would
get
possibly
pretty
different
interpretations,
so
it's
very
hard
to
get
something
objective.
Basically,
that
allows
you
to
make
more
fundamental
claims
based
on
this,
so
the
the
next
try
was
to
find
ways
of
generating
these
these
features,
so
there
was-
or
there
is
recently
published
a
word
to
back
approach
from
google.
B
That
does
this
by
using
the
collocation
information
to
say
for
every
word
in
my
list.
I
look
in
what
documents
this
word
occurs
and
which
words
are
in
two
two
words
before
and
two
words
after
collocated
with
this,
with
the
word
that
I
tried
to
capture,
and
if
you
do
this,
you
can,
in
fact
so
to
say
capture
some
of
the
semantics
of
this
word,
as
you
can
imagine
by
just
looking
at
two
words
before
and
two
words
after
this
is
very
shallow.
B
If
you
further
process
data
that
is
represented
like
this,
you
can
improve
so
to
say
the
unders
understandability
of
text.
Up
to
a
certain
point,
I
mean
it's
it.
It
tends
not
to
be
as
as
good
in
nlp
area
as
the
traditional
approaches,
but
it's
interesting
to
see
that
it
works
at
all
by
using
random
representations.
B
So
we
have
tried
to
develop
from
there
an
approach
that
ensures
that
we
have
a
certain
richness,
semantic
richness
that
represents
every
word.
That
might,
of
course,
not
be
as
rich
as
a
true
human
understanding,
but
at
least
rich
enough
so
to
say
to
get
some
smart
output
from
it.
B
B
And
if
we
have
many
of
these
lists,
because
we
have
looked
at
many
documents
where
this
word
occurs,
we
aggregate
all
these
contexts
to
create
some
sort
of
standardized
amalgamated
context
of
these
lists.
B
And
what
is
interesting
to
note
here
is
that
many
words
have
context
of
a
different
type.
So
to
say
you
will
find
for
the
word
apple.
You
will
find
lists
of
contexts
in
the
computer
field,
for
example,
but
you
will
find
also
about
plants
and
biology,
and
you
will
find
about
music
because
it
has
been
a
record
label
and
so
on.
In
fact,
if
you
look
carefully,
you
will
hardly
find
a
word
that
has
not
many
contexts,
so
it's
not
even
the
exception,
it's
more
or
less
even
the
standard
for
words.
B
B
This
will
allow
us
also
to
train
it
for
specific
types
of
languages,
so
it
should
be
completely
independent
of
which
language
you
use.
It
should
work
in
english
as
well
as
in
german,
french
or
any
other
language,
but
it
should
also
work
for
languages
in
the
sense
of
let's
say,
medical,
english
or
english
for
chemistry
or
for
aviation
or
whatever.
B
For
our
first
prototype.
We
use
wikipedia
because
we
we
found
that
it
is
more
or
less
a
general
english
approach,
where
we
have
a
well-balanced
importance
of
the
all
the
different
semantic
meanings
and
it
will
be
easier
to
investigate
it,
and
will
it
will
be
easier
for
average
users
who
look
at
our
demos
to
understand
what
what
is
happening
there.
B
So
we
take
the
wikipedia
documents
and
we
process
them
basically
by,
and
we
have
done
this
basically
on
trial
and
error.
We
use
all
the
machinery
that
exists
so
to
say
in
nlp.
We
try
to
apply
it
and
to
basically
get
most
out
of
the
meaning
of
the
documents
as
we
find
them.
B
So
we
do
a
lot
of
annotation
of
filtering
of
preparation
of
segmentation,
of
the
texts
and
in
the
end
we
get
a
collection
of
documents
where
the
most
important
aspect
is
that
every
document
just
talks
more
or
less
about
one
thing,
and
this
one
thing
is
the
word
that
it
is
a
context
of
yeah.
So
this
might
sound
a
bit
blurry,
but
I'm
afraid
that's
as
concrete
as
as
it
as
it
can
be.
So
it's
really.
B
It
has
really
to
be
evaluated
so
to
say
how
you
prepare
such
a
training
couples,
but
we
have
managed
to
automate
this
process.
So
we
don't
need
300
people
to
actually
do
this,
but
we
have
a
set
of
algorithms
that
basically,
at
least
for
languages
like
english,
german
and
french
does
a
good
job
in
the
automated
fashion
and
yeah.
B
Yeah
yeah.
B
And
typically,
what
you
find,
for
example
in
wikipedia
is,
I
don't
know
you
have
henry
ford
and
then
you
have
the
life
of
henry
ford,
the
work
of
henry
ford
and
so
on,
and
we
try
to
segment
this
to
have
always
like
one
one
statement
about
him
in
in
in
one
snippet,
so
to
say:
yeah
and
basically
we
don't
need
to
explain
all
the
words
that
exists
in
in
that
fashion,
because
there
is
not
a
wikipedia
document
for
every
word
that
exists,
but
by
using
so
to
say
a
sufficient
large
number
of
topics.
B
We
end
up
using
all
the
words
that
we
basically
use
by
describing
these
different
topics
and
so
on,
and
that's
by
the
way,
more
or
less
how
we
do
our
language
learning
too.
So
if
we
go
to
school,
we
learn
about
a
certain
number
of
topics,
and
while
we
learn
the
topics,
we
learn
more
and
more
words
to
actually
understand
and
argue
about
these
topics.
B
And
we
build
our
vocabulary
not
based
on
lists
that
we
that
we
learn,
but
by
using
the
words
in
a
description
or
by
talking
about
topics
that
are
in
common
and
by
choosing
the
topics
we
ensure
that
we
have
a
certain
semantic
space
yeah.
So
we
have
history
and
geography,
and
so
everything
that
is
important
in
the
world
basically
becomes
a
topic,
and
then
we
can
decide
for
every
topic.
B
If
we
do
this
with
a
number
of
topics
and
an
interesting
aspect,
maybe
that
that
I
could
add
here,
contrary
to
the
approach,
for
example
behind
word
to
vect,
where
you
need
to
get
larger
and
larger
training
corpora,
which
is,
of
course
not
easy
to
do
for
the
others
than
google.
B
What
we
try
to
use
is
to
to
be
more
and
more
precise
about
the
single
snippet
snippets
that
we
produce
and
to
use
less
and
less
of
these
snippets.
If
you
again
compare
it
how
humans
learn,
learn
language,
you
don't
learn
language
by
reading
5
million
books.
If
you
in
fact
look
how
many
books
you
have
learned
or
read
between
learning,
to
read
at
all.
B
Yeah,
and
so
we
tried
with
our
approach
also
to
rather
go
into
using
less
and
less
training
information,
because
it
has
the
advantage
less
information
needed
less
work
to
prepare
it
less
errors
and
so
on,
and
it
will
also
be
much
easier
to
transfer
this
concept
to
other
languages,
because
I
don't
know,
for
example,
the
thai
wikipedia
is
by
far
not
as
rich
as
the
english
one.
So
if
we
do
need,
I
don't
know
millions
of
wikipedia
documents.
Then
we
would
again
limit
ourselves
to
the
few
languages
where
this
is
available.
B
We
started
now
experimenting
with
other
sources.
The
point
is
that
wikipedia
is
simply
optimal
optimally,
structured
so
to
say
for
the
way
how
we
use
it
and
other
sources
need
a
bit
more
pre-processing
to
do
so,
and
in
fact
we
are
now
investigating
on
how
important
this
encyclopedic
aspect
is,
because
if
it's,
if
we
find
ways
of
not
relying
on
this
enzyclopedic
acid
aspect,
then
we
could
use
other
and
and
more
general
text.
Material,
yeah,.
B
Very
narrow,
it's
the
same
like
dictionaries
here
they
are
good
for
defining
a
word,
but
they
don't
really
give
you
a
context.
Maybe
if,
if
it's
a
dictionary
with
many
examples,
then
you
get
context
again,
but
again
it
would
be
a
be
a
bit
of
a
synthetic
context,
because
that
would
be
kind
of
sentences
stating
the
use
or
the
meaning
of
the
word,
and
we
found
that.
B
Having
really
narrative
descriptions
is
what
makes
a
difference
in
the
end.
F
Have
you
considered,
I
mean
I'm
not
an
expert
in
linguistics,
but
my
passing
understanding
is
that
there
are
theories
that
language
understanding
is
genetically
encoded
in
our
our
brains,
and
so
it's
not
like
purely
a
blank
slate.
Have
you
taken
that
into
consideration.
B
Absolutely,
but
to
tell
you
it's
somewhere,
we
want
to
go
and
not
it's
not
the
starting
point,
it's
more
or
less
a
place
where
we
want
to
to
to
go
by
applying
these
techniques.
The
point
is
that
we
are
actually
not
yet
at
the
level
that
we
want
to
do
language
understanding,
which
is
yet
more
complex.
B
What
we
want
to
start
with
is
to
to
to
understand
words
to
understand
the
meaning
of
the
atoms
of
language,
and
I
will
talk
later
on
the
reason
why
we
want
to
link
our
retina,
in
fact
to
clas
is
because
that
would
be
the
next
step
so
to
say
to
use
the
words
with
a
meaning
to
become
sentences
sentences
with
a
meaning
streams
and
sequences
of
these
patterns
yeah.
So
this
is
just
meant
to
capture
the
meaning,
whatever
this
means
of
single
words
in
a
realistic
fashion.
That's
that's
the
goal
of
this
approach.
E
So
it
I
can
see
how
a
noun
would
work,
because
if
you
have
a
noun
in
text,
then
all
the
words
around
it
are
kind
of
describing
that
noun.
But
for
say
the
adjective
the
adjective
is
there,
but
you're
not
really
describing
the
adjective.
B
It's
not
really
about
describing
it's
fundamentally.
Words
occur
together
when
I
can
formulate
something
that
describes
something
by
using
this
set
of
words.
It's
not
that
I
I'm
not
looking
for
a
description
of
the
word
in
in
in
in
the
pure
sense,
so
to
say,
but
I
want
to
create
realistic
contexts,
and
one
of
the
easiest
and
realistic
context
is
the
explanation
of
a
certain
topic.
For
example,
there
you
find
also
the
adjectives
when
you
explain
again
the
the
word
henry
ford.
B
You
will
find
that
there
is
a
bias
on
which
adjectives
are
used
in
many
contexts
again,
but
the
adjectives
of
of
the
document
about
henry
ford
are
pretty
different
from
the
documents
about
the
naza
yeah,
and
this
is
precisely
what
describes
the
contexts
in
in
which
they
appear
and
for
every
word
in
this
description.
B
All
the
other
words
are
the
context,
so
even
for
the
25th
adjective
in
the
text
about
henry
ford,
all
the
other
texts
is
the
context
for
this
adjective,
and
if
you
have
many
of
these
contexts,
then
you
can
start
and
say
something
about
the
word
about
the
meaning
of
the
word.
So
to
say,
or
at
least
you
can
compare
it
to
other
words.
B
Yeah,
so
the
point
is
that
by
creating
the
retina
corpus,
it's
the
only
place
where
we
try
to
use
all
the
tricks
and
tools
that
the
traditional
nlp
provides
us
with,
even
if
they
are
very
costly.
Even
if
it's
things
like
a
part
of
speech
tagging
the
documents,
which
is
not
an
easy
thing
to
do,
if
you
want
to
do
it
in
a
good
quality,
but
the
point
is:
if
we
do
it
once
in
the
retina,
then
it's
done
after
it's
done.
B
We
can
use
the
retina
to
work
on
any
other
text
and
we
never
will
be
obliged
to
make
a
named
entity
recognition
on
these
other
texts
or
whatever,
but
the
vocabulary
is
already
defined
and
all
the
occurrences
of
the
words
with
their
possible
tags
are
already
defined.
That's
that's
the
approach
so
to
see
yeah.
So
how
does
this?
How
does
this
work?
We,
as
I
said,
we
have
some
a
number
of
source
documents.
B
We
pre-processed
them
in
a
pipeline
and
we
have
in
the
end
documents
that
contains
that
contain
terms
and
out
of
this
term
of
these
documents,
we
create
a
corpus
that
we
then
represent
in
document
vectors,
which
is
basically
every
document
is
described
by
a
vector
that
says
if
a
word
is
present
in
the
document
or
not.
B
So
this
is
very
traditional
so
to
say
in
the
next
step,
we
take
these
document
vectors
and
we
use
a
variation
of
common
networks
to
make
a
2d
map
a
semantic
map
of
these
documents
that
we
have
created.
So
the
result
is
that
every
document
from
my
training
corpus
now
gets
a
coordinate
that
actually
positions
it
on
somewhere
on
this
map.
I
don't
know
if
you
have
seen
this.
There
are
millions
of
ways
of
calculating
these
maps.
What
it
basically
does
is
that
everything,
let's
say
that
talks
about
pets
is
in
a
certain
place.
B
Everything
that
talks
about
cars
is
in
a
different
place
and
all
the
documents
are
arranged
in
a
in
a
way
that
documents
that
are
nearby
are
talking
about
similar
things.
Basically-
and
this
works
pretty
well-
and
we
create
a
map
of
our
training
documents
in
this
step
and
in
the
next
step,
we
need
to
link
the
semantic
that
we
have
captured
by
organizing
the
documents,
which
is
still
relatively
easy.
It
takes
a
lot
of
computation,
but
it's
so
easy
to
understand,
and
now
we
map
all
the
words
to
to
this
map.
B
By
saying
I
have
a
list
of
all
the
words
in
in
my
document
couples.
I
take
the
first
word
and
they
say
in
what
documents
do
you
occur
and
because
every
document
is
positioned
somewhere
else
on
the
map?
I
get
this
representation
of
distributed
pixels
if
you
want
in
my
void
vector.
So
it's
an
extremely
simple
mechanism.
B
It's
sometimes
even
a
bit
scary,
how
simple
it
is
and
in
the
end,
we
so
to
say
post-process
the
the
the
bit
vectors
that
we
get
and
standardize
them
to
become
our
sdrs.
And
then
we
have
created
a
number
of
a
number
of
basic
computational
functions
to
basic,
basically
access
these
sdrs
and
to
combine
them.
B
And
here
again
to
make
the
point
all
we
do
with
conversion
of
words
into
sdrs.
We
call
this
basically
lexical
semantics,
so
this
is
really
just
about
words
about
what
words
mean
what
they
represent,
what
other
words
are
close
and
so
on,
but
we
are
not
talking
about
grammar.
B
We
are
not
talking
about
building
sentences
and
so
on,
because
there,
the
sequence
of
the
words
how
they
occur,
is
of
importance,
and
when
you
hear
sequence
learning
of
patterns,
you
immediately
think,
of
course,
about
cla
and
that's
the
reason
why
we
think
that
linking
up
the
two
systems
might
be
extremely
interesting.
C
C
G
B
Formalize
that
to
get
this,
we
don't
normalize
it
in
general.
What
we
try
to
do
is
to
make
the
training
corpus
that
big,
that
the
less
frequent
words
have
a
sort
of
minimum
representation
and
all
the
higher
densities
we
basically
calculate
them
down
to
whatever
level
we
need
them.
So,
in
the
current
api,
we
compress
that
to
five
percent,
more
or
less
the
thinking
behind
that
was
that
the
cla
wants
to
have
two
percent.
B
So
if
you
want
also
to
throw
away
a
number
of
bits,
then
we
give
you
some
sort
of
three
percent
space,
but
it's
just
a
a
way
of
of
agreement,
so
to
say
and
of
experimentation.
B
In
fact,
we
have
had
recent
results
where
we
worked
on
the
way
how
we
do
this,
we
call
the
specification
of
the
of
the
sdrs,
and
we
have
found
a
new
way
of
doing
this
by
taking
into
account
the
locality
of
the
pixel,
also
not
only
the
the
number
of
pixels
per
position,
but
also
other
many
pixels
around.
Then
it
gets
another
weighting
and
so
on,
and
this
obviously
improves
also
the
the
the
the
discrimination
capacity
of
the
system.
B
H
B
Percent
yeah,
that's
that's
the
way,
that's
the
way
how
we
basically
compute
it.
It's
it's
a
it's
due
to
the
speed,
so
we
want
to
do
this
dynamically
and,
as
we,
the
whole
sdr
computation
internally
functions
with
integers,
and
so
we
did
not
want
to
introduce,
doubles
or
float
and
to
slow
down
everything.
B
And
so
we
sort
of
made
the
steps
and
another
argument
is
atoms,
do
the
same.
They
have
their
quants.
There
is
no
between
no,
it's
just
for
performance
reasons
that
we
decided,
and
this
can
be,
can
be
changed
at
any
point.
Yeah
another.
B
C
Be
two
technical
questions,
so
when
you
create
the
cajon
map,
you
know
there
are
a
lot
of
parameters
you
can
tweak
and
all
of
that-
and
you
know
we
face
this
all
the
time
and
with
different
parameters,
you
can
get
a
different.
C
B
What
we
try
to
have
to
use
up
the
space
best
possible
yeah,
so
we
try
to
map
it
in
a
way
that
we
have
a
sort
of
an
even
distribution
and
in
fact
this
is
reduced
to
the
selection
of
the
features
that
you
take
for
creating
the
map,
which
is
so
to
say
a
tough
thing
to
do
so
in
the
in
the
in
the
first
generation
of
sdrs
that
we
created,
we
have
in
fact
used
again
traditional
tf
idf
stuff
to
calculate
which
are
the
most
important
words
to
do
this
in
the
more
recent
versions.
B
We
have
used
the
older
retina
to
do
this,
because
you
can
also
use
the
retina
to
find
out,
which
are
the
most
important
words
to
reconstruct
the
sdr
of
the
document
and
that
improved.
A
B
The
the
the
semantic
specificity
extremely
so
it's
sort
of
a
boost
trapping
that
you
need
to
do
at
some
point.
You
have
to
decide.
It
should
be
that
way
and
after
having
done
you
kind
of
revisit
the
whole
thing
and
you
refine
it
step
by
step-
and
I
don't
know
where
this
leads,
because
we
are
doing
this
now
only
for
a
short
time.
B
Okay,
so
the
the
colors
are
a
bit
weak,
so
the
first
fundamental
function
that
we
apply
to
these
sdrs
is
similarity.
That's
in
fact,
the
reason
why
we
made
it
was
to
to
be
able
to
calculate
similarity,
and
here
you
have
the
representation
of
cat
and
dog
and
on
the
lower
part,
you
see
the
overlay
of
two
and
normally
you
should
see
the
the
the
blue
and
the
red
dots
separated
and
black
dots
were
a
dot
between
the
two
and
in
general
terms,
I
mean
just
take
this
as
a
symbol.
B
You
have
again
the
areas
that
are
specific
for
specific
aspects
of
the
word,
so
everything
that
is
related
to
home,
family
and
whatever
is
down
here
here-
are
social
aspects,
free
time
aspects,
biology
aspects
and
again
we
don't
know
we
don't
care
to
be
true
what
the
single
dots
mean,
but
what
we
can
say
that
in
the
different
areas
of
the
sdr,
that's
what
I
meant
before
with
the
locality
of
the
representation.
B
You
have
certain
type
of
aspects
that
are
bundled
so,
but
there
are
also
aspects
for
example.
Here
it's
just
blue.
This
is
just
red
that
are
specific
for
the
one
or
the
other,
so
typically
two
words
regardless
how
far
or
close
they
are
together.
The
shares
a
certain
number
of
aspects
and
they
diverge
in
certain.
B
B
H
So
that
I
think
that's
an
interesting
point
because
in
the
so
I
just
got
finished
reading
pinkers
to
language
instinct.
H
In
my
head
that
these
words
are
very
ambiguous
and
how
they're
represented
in
your
brain
there's,
there
are
separate
sdrs
in
your
brain
for
all
the
different
contexts
of
each
word
and
the
way
we're
representing
them
is
is
like
squishing
them
all
together.
So
just
the
fact
that
apples
removes
the
computer
context,
because
it's
plural
but
still
retains
sort
of
the
biological
context,
is
interesting.
H
G
H
B
Well,
it's
very
on
one
level:
you
can
do
it
by
doing
set
arithmetic
with
with
the
sdrs,
but
this
is
only
valid
for
static
representations.
I
can
show.
I
could
even
show
you
right
now,
but
I
need
my
glasses.
A
Here,
for
example,
notice
the
wrong
one.
I
take
this
one,
it's
it's
even
open
already.
A
B
B
To
apples
we
see
that
except
of
this
cluster,
it
looks
very
similar
here
and
here
all
these
distributed
clusters
are
very
similar
and
in
fact
I
can
tell
you
what
this
remaining
big
blob
is
here.
It's
it's
about
apple,
the
record
company.
So
if
I
subtract.
B
B
Yeah
so
yeah,
this
is
basically
the
expression
engine.
It's
one
of
the
of
the
two
important
engines
that
we
use
for
manipulating
these
sdrs.
The
one
is
the
similarity
engine.
It's
basically
to
compare
that
and
the
other
one
is
the
expression
engine
that
allows
you
to
do
an
ore
and
sub
and
and
so
on,
with
with.
B
I
don't
like
to
call
it
normalization,
because
it's
not
really,
but
the
re-specification
here
is
done
dynamically
after
the
processing
has
happened,
because
we
want
to
take
as
many
pixels
as
possible
to
do
the
correct
operations
and
only
the
remainer
is
then
specified
to
become
the
two
percent
at
the
five
percent
story:
level:
yeah
yeah
internally
internally.
The
system
always
works
with
with
the
complete
sets,
because
we
don't
need
the
the
sparseness
factor
as
much
as
the
cla
would
need
it.
B
Yeah,
it's
interesting
even
with
various
passwords,
the
still
functions
yeah,
so
it
you
have
really
to
go
and
and
look
for
extremely
rare
stuff,
specifically
due
to
the
wikipedia,
of
course,
and
in
practice,
whenever
you
have,
you
need
words
that
are
too
sparse
to
be
processed
in
a
good
fashion.
You
simply
add
training
material
containing
these
terms,
and
you
add
to
the
pixels,
but
typically
you
want
to
use
a
retina
in
a
specific
environment
and
you
train
it
for
the
environment,
because
the
other
way
around
is
also
true.
B
Industry,
let's
say
I
would
not
be
interested
to
have
all
the
time,
apples
and
pears
from
from
the
computer
business
mess
in
my
in
my
things,
so
I
would
like
to
train
the
way
that
there
is
not
that
there
are
no
pixels
about
apple
computer
industry
related
stuff
yeah,
and
with
that
I
can
improve
the
signal
to
noise
ratio
basically
dramatically.
B
I
mean
I
can
show
you,
there
is
another
another
approach
we
have
used
in
the
second
demo,
so
this
is
basically
just
thought
to
do
this
operations
and
investigate
what
the
outcome
is.
There
is
another
thing
that
basically
tells
you
that
gives
you
the
context
terms
for
a
given
term,
and
it
does
this
by
converting
the
term.
You
said
you
you
have
given
into
an
sdr,
then
it
looks
for
similar
sdr's
and
it
converts
them
back
into
words
and
returns.
B
The
list
of
words,
for
example
the
word
source-
and
I
get
many
different
words
here
and
what
I
also
do
get
is
the
system
can
detect
different
clusters
of
meanings
in
the
representation.
So
here,
for
example,
there
is
a
cluster
that
is
named
water
and
if
I
select
it,
I
get
context
terms
like
water
temperatures,
amount,
carbon
and,
and
so
on
that
are
related
to
the
concept
of
source
and
water.
B
I
get
the
context
terms
like
this
and
if
I
select
computer,
because
there's
also
a
context
between
source
and
computer,
I
get
proprietary
software,
gpl
development
tools
and
so
on
a
whole
other
set
of
context
terms.
G
H
G
B
J
B
Yeah,
so
basically,
but
it's
this
is
just
a
small
change,
because
right
now
it
expects
terms,
and
we
just
have
to
change
it-
to
expect
the
expressions
and
then
it
should
work
yeah.
We
could
even
try
this
again
apple
here.
You
see
the
most
important
context
is
obviously
the
computing
context,
and
here
I
find
the
different.
So
there
is
a
disambiguation
function.
That
might
also
be
interesting
due
to
the
fact
how
simply
how
simple
the
disambiguation
actually
works.
B
B
B
Beetles
then,
I
have
only
two
two
clusters
so
to
stay
left
and
now
we
have
all
the
fruits
and
salads
and
and
so
on.
So
the
same
thing
works
basically
with
the
context
terms,
and
the
nice
thing
is
that
every
one
of
you
who
has
been
working
in
the
nlp
field
knows
how
complicated
it
is
to
do
this
automatic
disintegration
of
terms.
B
Here
we
have
a
very
simple
feedback,
loop,
so
to
say,
feedback.
Subtract
subtraction
allows
us
to
disambiguate
any
word,
so
you
can
type
in
the
the
example
from
from
my
title,
for
example,
the
word
jaguar
and
it
finds
that
jaguar
can
be
a
vehicle
can
be
an
animal,
can
be
a
character,
obviously
from
cartoons
or
whatever,
and
I
can
do
again
the
same
thing
I
can
also
restrict
now.
I
said
I
want
just
nouns
to
be
complex
terms.
B
I
could
restrict
this
to
be,
let's
say
adjectives
and
you
see
the
typical
it
was
to
your
question
before
here.
You
have
typical
adjectives
that
are
related
to
vehicles,
for
example
yeah.
If
I
select
animals,
I
get.
B
B
Yeah,
I
mean
yes,
we
are
finding
them
automatically,
but
with
standard
standard
tooling.
Basically,
we
use
a
stanford
tagger
as
far
as
I
know
to
do
this,
for
example,
but,
as
I
said
before,
we
are
doing
this
we're
doing
this
in
the
sense
of
annotating
the
terms
that
are
in
the
retina,
and
we
do
this
to
avoid
to
annotate
every
single
document
we
want
to
process
later
on,
instead
that
we
create,
for
every
word,
a
part
of
speech
tag.
B
In
fact,
the
part
of
speech
is
also
an
ambiguous
issue,
so
the
very
same
word
might
have
several,
let's
say,
part
of
speech
tags.
What
we
do
with
that
in
the
retina
is
that
we
try
to
find
which
of
the
occurrences
correspond
to
which
correct
part
of
speech
tag.
So
you
will
not
get
a
single
answer
by
saying
I
don't
know,
I
want
only
nouns,
you
might
have
words
that
also
co-appear
as
verbs,
for
example
yeah,
but
the
pattern
that
is
associated
should
be
the
pattern
that
is
true
whenever
it's
used
as
a
noun.
B
So
that's
basically
the
approach
and
the
same
thing
would
be
true,
for
example,
for
doing
an
annotation
of
locations
of
named
entities,
and
so
on.
You
could
basically
restrict
say,
give
me
all
the
context
terms,
but
only
the
terms
that
specify
a
location
or
only
terms
that
are
related
to
a
person
or
whatever
yeah.
So
it
brings
in
a
new
usefulness
so
to
say
for
the
old-fashioned
annotation
job,
to
say
by
being
applicable
to
more
different
areas.
B
Okay,
another
another
practical
use
of
the
of
the
sdrs
that
we
are
doing
is
that
we
are
using
that
to
create
so-called
document
fingerprints.
So
we
decompose
a
document.
In
all
these
words,
every
word
gets
replaced
by
the
fingerprint
of
the
word,
and
then
we
ore
all
those
word
fingerprints
together
and
we
responsify
it
and
in
the
end
we
have
a
fingerprint.
B
We
call
fingerprint,
which
is
an
sdr
for
a
document,
and
the
interesting
thing
is
that
all
the
magic
we
could
do
with
the
words
we
can
now
do
with
the
documents
in
a
very
same
way,
so
we
can
disambiguate
that
we
can
look
for
similar
documents.
In
fact,
that
is
the
core
of
the
search
engine
that
we
are
about
to
build
for
documents
and
just.
C
So
each
document
also
has
a
location
in
this
map.
These
are
not
the
same
documents.
B
B
No,
I
think
you
could.
You
would
not
be
able
to
differentiate
between
a
new
new
listing
document
and
the
training
document.
Okay,
that's
yeah!
If
you
would
look
at
the
whole
collection,
you
would
see
that
it
nicely
fills
up
the
space,
but
that's
the
only
specificity
of
this
yeah
yeah
with
the
document
fingerprints,
as
you
can
imagine,
we
can
do
many
things
on
the
document
level
where
there
is
a
big
business
interest
in
doing
so.
B
B
When
I
do
a
query,
I
don't
formulate
it
by
using
a
query
language,
but
I
take
a
snippet
of
text.
Another
document-
and
I
say
I
want
to
find
all
the
documents
that
are
semantically
closest
to
what
I'm
showing
here
and
then
I
get
the
list.
That
is
ranked
according
to
the
distance
measure
between
my
query
document
and
the
documents
I've
indexed
before,
and
the
nice
thing
is
that
I
can
use
another
feature
that
has
been
in
ir
theoretically
already
for
a
long
time,
but
it
has
been
pretty
complex
to
create
it.
B
I
can
ask
the
user
to
specify
what
kind
of
information
he's
interested
in
so
let's
say
I
have
a
repository
of
technical
scientific
publications,
and
my
user
gives
me
five
documents
about
medical
aspects,
medical
issues
here,
so
he's
a
medical
doctor
and
that's
what
he's
interested
in
he
gives
me.
Let's
say
five
documents
that
basically
he
likes.
B
I
take
the
documents
and
make
a
document
fingerprint
of
that
and
instead
of
ranking
the
result
list
after
the
search
against
the
distance
measure.
To
my
query,
I
can
rank
it
to,
according
to
the
distance
measure,
to
the
fingerprint
of
the
actual
user
profile
so
to
say,
and
this
allows
basically
to
make
document
searching
using
two
strategies,
namely
by
refining
the
definition
of
what
I'm
interested
in
and
by
defining.
What
I'm
looking
for,
but
just
to
to
show
you
that
this
functionality
is
already
so
to
say.
B
We
get
this
already
from
the
static
sdrs,
and
all
of
that
I'm
just
saying
to
put
the
emphasis
on
that.
We
believe
that
there
is
a
lot
of
information
captured
in
the
sdr
and
that
the
cla
should
feel
well
in
learning
about
this
information.
D
Well,
so
that
would
be
probably
a
good
way
to
determine
how
well
the
system
will
work
in
any
other
fields.
It
would
be
very
easy
to
assess
them.
B
There
is
another
way
of
doing
I
mean
just
to
evaluate
so
to
say
the
contents.
That
would
be
one
aspect
to
become
more
and
more
specific,
with
the
training
training
coupos.
Another
way
is
to
change
the
resolution
of
of
the
actual
sdr.
B
I
mean
we
have
chosen
to
use
128
by
128
because
it
somewhere
in
the
middle
of
we
are
able
to
actually
compute
that
because
it
takes
quite
some
time
to
process
the
retina
to
in
this
resolution,
and
it
makes
the
use
of
the
of
the
word
sdrs
more
or
less
easy,
so
the
bigger
they
become.
B
Everything
gets
slower
basically,
and
we
try
to
find
the
minimum
resolution
that
is
necessary
to
get
a
certain
quality
out.
But
if
you
have
a
very
broad
field,
where
you
need
a
lot
of
resolution,
then
you
would
need
maybe
to
go
to
five,
twelve,
five,
five,
twelve
or
even
larger
yeah,
so
we
basically
scale
whenever
the
processing
becomes
cheaper.
B
We
can
scale
this
accordingly
yeah,
maybe
just
I
would
just
like
to
make
a
short
interruption
here
and
ask
my
colleague
daniel
to
come
over
and
just
oh,
it's
another
question.
B
B
Okay,
yeah
so
daniel,
who
came
with
me
from
vienna.
They
are
the
the
first
users
of
of
our
of
our
word
sdrs
and
his
company
phase.
Six
works
in
the
e-learning
area
and
what
they
do
is
that
they
provide
english
learning
functionality
so
to
say,
and
they
have
to
face
a
certain
number
of
problems
so
to
say,
and
he
wants
to
generate
or
create
the
next
generation
of
e-learning
systems.
B
And
we
are
now
working
together
for
some
months
and
they
use
the
sdrs,
which
I
find
very
interesting
to
individualize
the
training
environment
for
a
specific
language
learner.
And
maybe
you
just
give
us
a
short
overview
on
sure
on
what
you
and
what
you're
doing
that.
Try
to
find.
L
Okay,
well,
hello:
my
name
is
daniel.
Like
francisco
introduced
me
phase,
six
is
the
company
we've
been
providing
educational
software
in
germany,
mostly
for
about
10
years,
and
we're
now
looking
to
make
the
software
smarter?
Our
ultimate
goal
is
to
create
a
digital
tutor
which
fulfills,
who
fulfills
many
of
the
features
that
our
human
tutor
will
fulfill
and
our
principal
aim.
L
Our
principal
goal
is
to
personalize
education,
because
every
student
learns
differently
learns
at
a
different
pace,
has
different
strengths
and
weaknesses,
and
the
school
system
simply
does
not
provide
the
kind
of
individual
and
personal
service
or
tutoring
that
the
student
would
need
to
fulfill
his
full
potential,
and
so
the
situation
in
education
is
that
most
kids
do
not
live
up
to
their
potential
and
get
frustrated
about
learning
about
school,
and
so
I
think
technology
can
really
help
a
great
deal
with
personalizing
education
and
we're
specifically
focused
on
english.
Let
me
get
my
water
quickly.
L
We're
specifically
focused
on
english
because
it
turns
out
that
if
you
want
to
have
a
high
quality
product
and
if
you
want
to
really
personalize
along
many
dimensions
and
not
just
superficially,
you
need
to
go
deep
in
the
vertical.
You
need
to
have
a
lot
of
knowledge
about
the
actual
subject
that
you
want
to
teach,
and
so
we
we
picked
english
language
as
a
very
as
a
subject.
L
A
lot
of
people
around
the
world
are
learning
and
which
is
also
a
vehicle
to
other
things,
because
language
is
the
vehicle
to
knowledge,
and
so,
if
we
want
to
personalize
the
learning
and
teaching
of
english
for
every
individual
student,
there
are
several
aspects
of
things
that
need
to
be
personalized.
And
I'm
going
to
talk
to
about
two
of
them
where
the
fingerprints,
the
semantic
fingerprints
or
sdrs,
we
call
them
here
are
actually
extremely
useful.
L
Content
to
show
the
student
exercises
which
actually
engage
the
student,
because
they're
interested
in
the
topic
and
every
student
would
be
interested
in
different
topics,
and
so
we
we're
going
to
use
the
similarities
in
order
to
find
the
type
of
content
that
the
individual
student
would
like
to
engage
with,
because
countless
reach.
Research
and
studies
show
that
if
the
student
is
engaged,
they
learn
many
times
more
and
they
remember
many
times
more
than
when
they
are
bored
or
simply
not
engaged
and
not
interested.
L
G
We'll
go
to,
let's
see
what
is
it
so
many
things
that
we
don't
need,
chico,
I'm
looking.
L
And
I'm
going
to
make
use
of
something
that
francesco
here
showed
us
earlier,
which
is
the
capability
to
do
sentence
level?
Well,
you
talked
about
document
level
sdrs,
and
that
means
we
can
also
do
sentence
level
sdrs
one.
Second,
there
we
go,
I'm
going
to
put
here
a
the
whole
sentence.
That
includes
the
word
jaguar,
so
we're
sticking
to
the
example
jaguar
we
have
here.
The
maya
also
includes
some
mentions
of
the
jaguar
in
their
mythology
and
what
the
system
created.
Let
me
scroll
up.
L
A
little
bit
is
a
fingerprint
that
contains
the
semantics
of
the
whole
sentence.
L
What
this
tells
us,
the
jaguar
in
this
sentence
very
likely
means
the
animal
and
not
the
vehicle,
and
it
turns
out
that
semantic
context
on
the
sentence
level
is
a
very
important
determiner
of
the
actual
dictionary
definition
of
the
word
to
disambiguate
it.
So
we
have.
This
is
something
which,
for
us
is
quite
intuitive,
because
we
also
capture
as
humans
the
semantic
context
of
the
sentence
in
order
to
guess
what
does
jaguar
actually
mean
here
now.
Why
is
word
sense,
disinfiguration,
so
important
for
individualization
of
education?
L
Well,
if
you
think
about
it
like
francisco
said
every
word,
almost
every
word
has
many
many
meanings,
especially
simple
words.
If
you
think
of
the
word
party,
it
can
mean
the
fun
party
where
you
go
to
celebrate,
or
it
can
be.
The
political
party
and
the
student
who
learns
english
would
learn
those
two
different
senses
at
very
different
stages
of
his
education.
L
So
I
have
here
a
little
demo.
I
don't
need
this
one
one.
Second,
where
is
the
new
one
that
oh
here
we
go?
So
let's
take
the
sentence
to
fire
a
worker
and
the
the
word
that
we
want
to
disambiguate
is
fire.
There
can
be.
You
know
you
can
fire
an
employee,
you
can
fire
a
gun,
you
can
fire
ceramics,
you
can
fire
somebody
with
enthusiasm
and
it
turns
out
that
the
semantic
context
is
not
the
only
factor
that
we
need
to
consider
in
order
to
disambiguate
it.
L
But
it's
one
important
one,
and
what
we
do
here
is
to
map
it
back
to
dictionary
definitions.
We
have
a
strategic
collaboration
with
oxford
university
press.
They
believe
in
our
new
approach.
It's
basically
the
next
generation
of
english
learning
and
so
we're
using
their
dictionaries
as
as
kind
of
information
sources.
In
order
to
understand
what
are
the
properties
of
different
word
senses
sentence
level.
Semantics
is
one
of
them.
We
are
using
the
sdrs
for
that,
but
others
are
grammar
structure.
L
For
example,
you
fire
missiles
were
fired
at
the
enemy,
so
something
was
fired
at
somebody
is
a
different
grammar
structure.
Then
you
fire
an
employee
or
you
fire
somebody
with
enthusiasm
yeah,
so
you
have
different
grammar
frames
or
verb
frames
surrounding
the
head
word
that
you
want
to
disambiguate,
which
you
also
need
to
consider.
L
L
The
word
census
integration
extremely
important
in
order
to
serve
up
those
sentences
and
examples
to
the
student
which
contain
the
word
in
the
sense
that
the
student
is
supposed
to
know
in
on
his
level
of
learning
and
not
something
which
will
only
supposed
to
show
up
a
few
years
later.
Traditionally,
this
has
been
the
turf
of
publishers.
Publishers
spent,
spend
educational
publishers,
textbook
publishers,
elt
english
language
teaching,
publishers,
spend
considerable
resources
and
and
pay
very
expensive
editors.
L
In
order
to
do
just
that,
in
order
to
write
books
and
texts
which
introduce
the
which
manage
the
progression,
it's
called
progression,
so
the
progression
of
which
word
senses.
Do
you
introduce
at
which
stage
and
then
selecting
the
content
for
that
so
far,
there
has
not
been
a
way
to
automate
this
and,
I
think,
we're.
Basically,
we
can
come
very
close
to
automating
this
and
revolutionizing
a
little
bit
the
publishing
industry.
L
By
doing
so
so
this
was
the
one
use
case
for
the
fingerprints,
and
we
we've
seen
this
with
the
example
of
jaguar
and
the
second
one
that
I
talked
about
was
to
get
content
which
is
interesting
for
the
student,
and
we
can.
I
will
demo
this
right
now
we
have
a
corpus
with
one
billion
sentences.
L
This
corpus
is
composed
of
various
sources.
It
includes
the
complete
english
wikipedia.
It
includes
200
000
books
from
literature.
It
includes
the
british
national
corpus.
It
includes
every
kind
of
corpus
that
we
could
get
our
hands
on
and
that
we're
allowed
to
quote,
and
so
we
have
one
billion
example,
sentences
from
which
we
can
generate
the
content
and
and
select
the
content
which
is
appropriate
for
the
student
in
order
to
generate
exercises.
I'm
going
to
show
this,
so
let's
take
any
word
that
the
student
would
need
to
learn.
L
L
It's
thinking,
and
so
it
just
went
through
one
billion
example
sentences,
and
it
gave
me
any
number
I
just
I
put
100
here
as
the
number
to
return.
It
can
give
me
any
number
of
sentences
which
include
the
word
achieve,
which
is
the
word
that
the
student
needs
to
learn
and
talk
about
a
certain
subject
in
municipal
elections
for
local
government
boards.
The
labor
candidates
were
achieving
a
significant
share
of
the
vote.
They
can
apply
other
other
methods
to
filter
the
resulting
sentences
even
more
and
now.
If
we
put
this
back
into
the
similarity
engine.
L
B
L
L
So
we
have
apple
sub
computer.
I
can
put
the
term
here
the
term
two
and
then
I
get
the
similarity
score,
which
is
we're
using
yeah
apis
to
to
do
that
and
I'm
happy
to
share
them.
Our
programmer
basically
created
a
whole
set
of
sept
apis
and
integrated
them
into
google
spreadsheet,
because
then
you
can
do
overviews
and
computations
and
you
can
combine
things
and
you
can
do
all
kinds
of
things
to
play
with
it
and
to
evaluate
the
results.
L
E
A
L
Was
the
showcase
for
your
technology?
Okay,
great.
H
A
A
B
Our
sdrs
have
been
created
in
fact,
are
modeled
along
the
the
development
so
to
say
of
of
the
cla
theories
in
in
over
the
last
years.
B
So
it
makes
perfectly
sense
now
to
sort
of
use
both
functionalities
together,
as
I
said,
the
sdrs
by
it
by
themselves
allow
to
do
a
certain
number
of
static
understanding
from
language,
but
as
soon
as
we
want
to
go
to
the
so
to
say
that
the
grammar
level,
if
you
want,
we
need
a
backend
that
actually
manages
the
sequences,
how
they
occur
in
real
text
and
so
on.
B
So
we
want
to
create
so
to
say
or
extend
the
functionality
by
being
able
to
learn
sequences
and
just
in
theory,
we
should
be
able
to
generate
text,
so
we
have
had
the
very
first
example
of
generated
sdrs
according
to
the
training
that
happened
in
in
the
cla
and
with
when
we
have
seen
that
so
to
say
what
comes
out
does
in
fact
make
sense
compared
to
the
other,
to
the
other
meanings
that
we
have
seen
there.
B
So
I
think
that
by
doing
we
can
use
several
so
to
say
sdr
outputs
on
the
quotes
that
occur.
If
we
use
the
prediction
functionality,
we
can
basically
do
what
I
would
say
the
the
perceptive
talk.
It's
basically
understanding
what
a
certain
text
is
about
by
guessing
what
the
next
word,
basically
in
a
sentence
would
be
and
to
use
this
functionality
to
give
an
alarm
as
soon
as
the
interpretation
of
what
this
sentence
could
mean
differs
from
what
actually
was
said.
B
B
I'm
pretty
sure
that
the
match
will
be
pretty
good
and
you
can
give
the
the
speech
recognition
system
good
hints
on
what
the
actual
word
was,
that
it
was
hearing
so-
and
I
think
this
is
very
similar
to
the
example
of
two
persons
standing
in
the
discotheque
with
a
lot
of
noise,
but
they
still
can
follow
each
other's
conversation
because
they
actively
interpret
what
they
would
expect
to
be
said.
Next,.
B
B
If
you
do
that,
let's
say
400
answers
and
you
count
how
many
times
it
was
right.
You
get
a
percentage
on
how
good
the
the
algorithm
works.
If
you
try
to
do
this
with
data
from
let's
say
images,
vision,
videos
or
so
you
have
the
problem
that
you
have
to
prepare
the
data
you
have
to
tag
it.
You
have
to
create
a
collection
of
let's
say
pictures
where
you
have
manually
identified
things
that
you
would
expect
your
algorithm
to
detect.
B
If
you
do
this
with,
I
don't
know
the
the
heating
and
temperature
measures.
This
is
very
hard
to
create
measuring
sets
using
this
this,
this
sort
of
data-
and
I
think
that
using
language
to
create
a
measuring
corpus
is
much
more
intuitive.
It
allows
you
to
so
to
say,
see
much
faster
if
the
there
hasn't
has
been
an
improvement
or
not,
and
in
our
case
we
have
even
by
using
the
sdrs,
we
have
even
a
way
of
numerically
measuring
this
phenomenon,
so
we
have
a
reference
text.
B
We
pass
it
through
the
retina,
we
feed
it
into
a
cla
and
we
look
what
we
have
predicted.
So,
every
time
we
get
the
prediction,
we
compare
the
actual
word
that
there
is
that
that
should
have
been
found
and
the
word
that
was
predicted
and
we
can
use
the
api
by
computing,
the
the
the
euclidean
distance
between
the
two
sdrs
and
we
immediately
get
a
numeric
value.
B
That
tells
us
how
good
the
system
performed
and,
as
we
have
already
sort
of
started,
doing
such
a
collection
with
with
the
kids
tales,
which
is
a
small
collection
but
which
I
think
would
be
a
very
good
first
step
in
setting
up
a
setup
like
this,
and
this
would
basically
allow
you
to
measure
any
variant
or
any
improvement
that
you
might
add
to
the
to
the
cla
code
and
immediately
measure.
If
it's
actually
improving
or
not.
B
B
I
think
that
text
data
is
very
well
suited
for
that,
and
maybe
a
more
philosophical
approach
to
this.
If
you
want
to
test
the
cortical
algorithm
using
some
real
world
physical,
so
to
say
data,
you
basically
have
to
evaluate
the
whole
embodiment
of
the
system
if
you
want,
which
makes
it
very
complicated.
B
If
we
just
limit
ourselves
to
this
level
of
the
word
where
it
appears
so
to
say
in
the
human
brain,
we
don't
need
to
take
into
account
the
whole
embodiment
part
in
the
measurement
loop,
but
we
can
stay
so
to
say
in
brain
to
brain,
to
evaluate
it,
and
I
think
this
could
be
very
efficient,
and
this
is
just
the
last
aspect
that
is
very
important
to
me.
B
B
If
you
sort
of
link
every
input,
potentially
every
every
sensor
bit
to
every
input,
there
is
no.
B
In
my
understanding
at
least,
there
is
no
obvious
hierarchy
that
you
can
find
there,
but
the
fact
that
we
have
in
the
lower
corner-
let's
say
aspects
as
we
have
seen
about
leisure,
for
example,
by
saying
that
in
the
next
level
there
is
one
bit
that
corresponds
to
these
nine
potential
representations
of
leisure
aspect
makes
a
lot
of
sense
and
all
the
the
the
the
thoughts
behind
the
the
hierarchical
structure
and
so
on
make
sense
again
in
in
this
context
so
yeah
this,
but
it
I
think
this
has
to
be
found
out
experimentally.
B
But
I
have
a
very
strong
belief
that
we
could
use
this
sort
of
trivial
hierarchy
hierarchy
that
comes
from
the
nature
of
the
data
that
we
that
we
work
with
with
the
sdrs
okay.
So
any
any
more
questions.
J
I
am
francisco
okay,
if
you,
if
you
include
the
the
the
depth
of
the
columns
in
in
the
first
layer
of
cla,
then
you're
actually
getting
that
already,
because
you
have
60
000
outputs
generated
from
16k
inputs
and
they
they.
What
they
do
is
add
in
the
sequence
information.
So
they
would.
They
would
be
recognizing
phrases,
combinations
of
words
and
sentences.
So.
G
B
Yeah
I
mean
the
point
is
that
I
think
already
here
in
the
in
the
beginning,
to
use
the
spatial
puller
to
decide
for
himself
which
of
the
nine
bits
is
obviously
the
sub
pattern
in
this
area
that
occurs
more
frequently
than
other
sub
patterns.
B
I
think
this
it
would
make
sense
to
use
also
the
spatial
polar
in
the
in
the
first
level,
because
this
is
in
reality
it's
much
bigger.
It's
not
nine
pixels,
it's
much
bigger
than
that,
the
the
areas
that
you
have
and
in
the
experiment.
As
far
as
I
know,
you
have
bypassed
the
the
spatial
polar
in
the
first
step
and
in
principle
this
contradicts
the
concepts
that
you
should
use
the
spatial
puller,
basically
as
a
part
of
every
layer.
B
As
far
as
I
understood
it-
and
I
think
that
by
it
is
that
the
spatial
pooling
that
it
makes
in
this
sector
means
something
different
than
the
spatial
pooling
that
it
would
do
for
the
whole
sdr,
because
in
the
locality
it
basically
just
says
which
of
the
sub
variations
are
more
frequent
for
representing
a
feature
within,
let's
say
laser
than
any
others,
and
it
would
already
be
able
to
put
some
of
the
learning
at
this
very
low
end
and
potentially
prevent
confusion
on
a
higher
level.
That's,
but
it's
just
a
a
guessing.
C
B
B
Nobody
would
have
guessed
before.
It
was
measured
that
the
different
angles
of
the
of
the
edges
that
you
see
in
the
first
visual
layer
end
up
in
differentiating
faces,
for
example,
yeah
yeah.
So
I
think
that
it
would.
There
would
be
very
interesting
to
to
play
around
with
this
aspect
and
to
find
out
how
it
can
be
used.
D
If
you
extend
that
algorithm
to
make
a
mapping
into
3d
dimension
space,
how
would
that
affect
the
efficiency
of
this
approach?.
D
D
B
I
mean
to
tell
you
the
truth:
it
is
already
extremely
costly
to
compute
2d
map
in
of
this
size,
so
to
do
that
in
three
dimensions.
This
is.
D
For
example,
you
scanned
millions
of
words
very.
D
B
Yeah,
but
I
think,
even
if
it's
narrow,
even
if
you
I
don't
know
you,
take
a
hundred
words
out
of
the
legal
system,
you
still
need-
let's
say
five
thousand
words
of
common
english-
to
embed
them,
to
make
reasonable
sentences
and
to
have
those
five
thousand
words
mean
something
reasonable
in
the
representation
you
need
a
certain
size,
yeah.
B
B
I
B
Yeah
this
this
was
in
fact
another
example
to
use
translation
software,
so
you
could
do,
I
would
say,
a
dump
translation
system
that
basically
just
tries
to
map
one
one
word
to
the
other.
That
would
be
doable.
I
would
say
just
with
the
sdrs
themselves.
B
This
could
be
useful
to
say
I
just
need
to
find
out
what
this
webpage
is
about.
I
don't
want
to
know
exactly
what
is
written.
I
want
to
know
if
it's
about
chemistry
or
about
aviation,
but
if
you
want
to
do
real
translation
in
a
human
sense,
then
you
would
actually
need
a
sequence
learner
coming
afterwards
and
that
could
be.
I
You
could
create
a
similar
document
clustering,
so
so
you
would
actually
be
able
to
map
the
document
clusters
onto
each
other
from
one
language
to
another.
You
have
you
with
your
2d
representation
of
all
the
documents.
I
mean
what.
B
I
mean,
I'm
pretty
sure
it
ends
up
in
enormous
calculations
in
the
end.
But
it's
I
think
that
would
be
the
approach
that
is
worth
trying.
Yeah
and
I
I
believe
I
mean
we
have
to
investigate
the
behavior
of
the
sdrs
with
the
clas
further
to
sort
of
make
a
better
strategic
planning
for
that.
But
I
think
just
out
of
gut
feeling
this
would
be
an
approach
that
makes
sense.
E
So
if
you,
if
you
gave
it
to
wikipedias
in
different
languages,
would
the
article
about
dog
look
the
same
like
visually
in
german
as
it
did
in
english,
so
that
you
would
know
if.
B
B
If
you
would
have
a
lot
of
money
and
if
you
would
say
okay,
I
take
10
000
english
wikipedia
documents
and
I
have
the
best
professional
translation
units
translated
exactly
it
should
be
very
similar
yeah
and
that
could
be
another
approach
of
creating
translation,
potentially.
C
B
It's
it's
very
sensitive,
so
if
you
add
to
a
collection
of
let's
say
a
million
documents,
you
add
five
documents
and
you
recreate
the
semantic
map,
which
is
a
competitive
learning
algorithm.
The
map
could
look
completely
different,
but
really
as
as
if
you
would
have
chosen
a
complete
other
set
of
documents.
So
this
is
practically
it's
an
iterative
algorithm.
So
if
you
change,
I
don't
know
two
degrees
in
the
first
step,
the
100
step
might
end
up
completely
differently,
but
for
our
for
our
use,
the
actual
topology
of
the
map
is
not
relevant.
B
The
only
thing
that
is
relevant
is
that
we
map
all
our
words
to
the
same
map
yeah,
it's
the
same,
as
I
mean
it's
in
in
in
humans.
It's
the
same.
We
have
a
completely
different
understanding,
I'm
pretty
sure
about
every
word
that
we
know,
but
still
we
share
sufficient
contexts
that
are
sufficiently
similar
so
that
we
can
engage
in
a
conversation.
If
we
would
not
share
any
context,
we
we
could
just
talk
to
each
other
without
transferring
any
meaning
yeah.
B
G
J
Get
it
so
you
can
get
a
fingerprint
for
any
string
effectively.
Okay
and-
and
as
I
was
thinking
that
I
realized
that
the
the
sdr
will
have
a
kind
of
an
invariance
that
comes
from
the
sentence,
so
the
so
it's
as
you
build
up
the
sentence.
If
you
imagine
that
there's
a
higher
level,
that's
looking
at
the
sentence,
it
will
actually
have
an
sdr
that
would
correspond
to
that
sentence
right
and
that,
as
each
word
was
added
that
that
word.
G
J
And
then,
eventually
for
the
full
documents,
this
could
be
a
way
of
measuring
by
the
way,
and
that's
the
the
sdr
of
the
of
this
of
the
sequence
so
far,
and
that
would,
for
example,
what
you
could
do
is
you
could
connect
up
the
the
sent
the
sequence
so
far
learned
and
get
that
the
retina
for
that
back
from
the
sdr
for
that
back
from
set
so
to
compare.
J
As
they
are,
what
would
happen
would
be
if
you
were
talking
about
politics
or
if
you
started
talking
about
jaguar-
and
you
know
we
said
the
jaguar
and
then
dot
dot
dot
then
it
would
be.
It
would
contain
fingerprints
for
animals
as
well
as
fingerprints
for
cars,
but
then,
as
the
next
word
appears,
that
would
improve
so
there's
a
there's
kind
of
an
invariance
that
that
changes
throughout
the
sentence
in
the
brain
is
obviously
helping
us
to
disambiguate,
because
the
previous
sentence
was
about
was
about.
G
J
So
I
think
that
the
sentence
level
sdrs
would
would
be
invariant
across
the
across
the
conversation
so
far
at
a
very
high
level,
yeah
and
that
that
would
be
transmitted
possibly
downwards
and
help
to
assist
the
upwards
disambiguation
of
the
different
components
of
an
sdr
for
jaggers,
for
example.
So
it's
quite
possible
that
we're
boosting
the
we're
boost
we
would
be
boosting
the
recognition
of
animals
if
we
had
previously
been
talking
about
animals
yeah.
B
Something
like
that
is
probably
the
mechanism
that
drives
us
humans
to
create
things
like
ontologies,
for
example,
is
because
we
want
to
group
stuff
together
and
the
the
the
interesting
thing
with
ontologies
is
that
it's
pretty
easy
to
agree
with
others
on
a
certain
ontology.
If
I
actually
explain
it
to
them,
and
also
in
in
the
moment
when
we
try
to
teach
somebody
something
we
tend
if
they
are
grown
up,
I
mean
if
they
are
young,
we
give
them
examples
of
context.
B
We
distribute
context
for
different
words
to
explain,
but
people
who
have
already
a
certain
understanding
of
language.
We
rather
transfer
a
part
of
the
ontology
tree,
and
we
explain
the
first
and
the
part
of
the
ontology
tree
and
the
person
then
can
itself
regroup
the
the
sub
patterns
into
it.
But
again,
this
will
be
interesting
to
see
how
the
how
the
cla
handles
this
and
what
we
can
create
by
fingerprinting
the
different
parts
and
comparing
it
to
how
the
cla
handles
it.
B
H
Do
otherwise
be
here
for
the
demo
right
now.