►
From YouTube: Niels Leadholm on Hierarchical Feature Binding and Robust Machine Vision - September 16th, 2020
Description
Niels Leadholm, a visiting researcher, discusses his (recently de-anonymized) PhD research on hierarchical feature binding and robust machine vision. He first explores the issue of robust machine vision and his motivation in developing a deep-learning neural network architecture using a biologically-inspired approach. Many AI systems nowadays are vulnerable to adversarial examples. Niels explains how the characteristics of “feature binding,” which happens in a primate’s brain, can be implemented in machine learning systems to enhance robustness.
If you want to follow Niels’ work, you can follow him on Twitter (@neuro_AI).
Interested in being a visiting researcher at Numenta? Apply to our Visiting Scholar Program here: https://numenta.com/company/careers-and-team/careers/visiting-scholar-program/
- - - - -
Numenta is leading the new era of machine intelligence. Our deep experience in theoretical neuroscience research has led to tremendous discoveries on how the brain works. We have developed a framework called the Thousand Brains Theory of Intelligence that will be fundamental to advancing the state of artificial intelligence and machine learning. By applying this theory to existing deep learning systems, we are addressing today’s bottlenecks while enabling tomorrow’s applications.
Subscribe to our News Digest for the latest news about neuroscience and artificial intelligence:
https://tinyurl.com/NumentaNewsDigest
Subscribe to our Newsletter for the latest Numenta updates:
https://tinyurl.com/NumentaNewsletter
Our Social Media:
https://twitter.com/Numenta
https://www.facebook.com/OfficialNumenta
https://www.linkedin.com/company/numenta
Our Open Source Resources:
https://github.com/numenta
https://discourse.numenta.org/
Our Website:
https://numenta.com/
A
Then,
how
do
we
develop
this
kind
of
unified
representation
of,
say
a
blue
triangle
localized
in
space,
rather
than
kind
of
you
know,
disembodied
free
floating
features,
and
this
can
also
be
thought
of
in
a
hierarchical
sense.
So
you
know
how
do
we
understand
that
it's
a
particular
vertical
and
horizontal
bar
again
located
in
space
that
are
forming
the
representation
of
a
t
and
and
and
how
do
we
have
access
to
that
information?
A
A
Then
that
can
inform
parsing
and
structural
description
of
the
next
layer
of
processing.
Encoding
is
about
capturing.
What
features
are
actually
part
of
this
object
in
in
kind
of
an
object,
centered
frame
of
reference,
and
so-
and
this
is
what
I'm
kind
of
talking
about
today
and
in
particular
kind
of
capturing
this
idea
that
this
vertical
bar
is
causally
part
of
this
high-level
kind
of
abstract,
neuron
and
so
just
kind
of
to
give
some
kind
of
images
to
maybe
help
with
the
kind
of
intuition
of
that.
So
with
capsule
networks.
A
B
A
A
Yeah,
that's
what
I'm
going
to
try
and
cover
kind
of
towards
the
end
after
I've.
So
I
I
was
going
to
touch
on
kind
of
how
exactly
I've
implemented
this
in
deep
neural
networks,
because
I
think
kind
of
seeing
explicitly
what
the
representations
look
like
in
the
network
will
make
everything
a
bit
more
concrete
and
in
particular
the
kind
of
how
this
this
relates
to
yeah
grid
cell
cortical
grid
cells
and
displacement
cells,
but
yeah.
A
So
the
kind
of
the
original
proposal
for
a
house
might
be
implemented,
builds
on
this
idea
of
polychrony,
which
is
you
can
think
of
kind
of
as
similar
to
synchrony,
where
you
have
kind
of
spacer
temporal
patterns,
but
that
these
are
asynchronous
but
time
locked
so
just
to
kind
of
unpack
that
a
bit.
If
you
have
these
kind
of
presynaptic,
neurons,
b,
c
and
d
that
are
driving
activity
in
a
and
e,
you
kind
of
often
there's
been
or
at
least
traditionally
many
years
ago.
A
A
The
idea
with
polycrine
is
to
actually
make
use
of
the
fact
that
one
has
these
kind
of
non-uniform
conduction
delays
such
that
the
same
neurons
can
fire.
The
identity
is
the
same
and
sorry
so
along
the
x-axis.
Here
is
time
and
along
the
y-axis,
the
identity
of
the
neurons,
but
depending
on
the
particular
pattern
with
which
they
fire,
the
signals
will
arrive
at,
for
example,
different
post-synaptic
cells
and
therefore
encode
different
things.
A
Yeah,
so
you
know
it's
a
great
question.
I
mean
when,
when
this
paper
came
out,
it
was
kind
of
very
fashionable
this
idea
of
kind
of
neurons,
encoding
information,
and
you
know
sinfire
chains
and
fire
braids,
that
sort
of
thing,
and
then
there
was
this
paper
cortical
songs
which
you
know
turned
out
they'd,
which
kind
of
claimed
some
experimental
evidence
of
this,
but
then
they've
kind
of
done,
some
statistically
not
kind
of
great
practice,
and
so
for
a
while.
There
hasn't
been
as
much
interest
in
this.
A
If
you
were
to
assume
that
this
is
actually
how
the
the
kind
of
brain
encodes
information,
you
would
need
to
record
from
an
enormous
number
of
cells
in
parallel,
but
even
if
you
have
those
recordings,
so
even
detecting
these
kind
of
patterns
in
silico
is
challenging,
and
it's
only
over
the
last
couple
years,
that
kind
of
robust
statistical
methods
have
been
developed
to
both
kind
of
align,
the
spike
times
as
well
as
actually
detect
these
kind
of
patterns.
From
from
the
statistics
of
the
activity,
I.
D
Mean
it's
a
very
challenging
thing
to
assume
neurons
could
do
this.
It
reminds
me
of
a
bit
of
you
know:
there's
a
group
of
people
who
argue
that
the
inter-spike
intervals
are
encoding
information,
so
the
the
time
between
individual
spikes
in
a
neuron.
It's
not
a
rate
coding,
it's
not
a
binary
coding,
but
it's
a
timing
between
interspikes
and
they're
compelling
the
argument
for
is
it's
really
full
of
information.
D
You
can
pack
a
lot
of
information,
which
is
true,
but
I
don't
think
it's
actually
happening
in
the
brain
just
because
it's
an
attractive,
mathematical
concept.
So
this
kind
of
reminds
me
of
that.
Like
hey,
this
thing
could
really
pack
a
lot
of
information,
but
it
doesn't
doesn't
seem
the
kind
of
things
that
neurons
could
actually
do
in
the
brain.
A
Yeah
so
yeah,
so
I
mean
I
fully
accept
that
it's
it's
possible
that
the
the
brain
can't
do
this
and
so
kind
of
the
the
approach
being
taken
here
is
so
you
know
like
the
core
computation,
which
I'll
kind
of
touch
on
a
bit
more.
Is
this
idea
of
coincidence
detection
as
a
means
of
encoding
causal
relations
between
features
in
the
hierarchy,
and
so
there
are
different
ways
that
this
could
be
implemented
in
the
brain.
A
So
a
kind
of
a
much
slower
mechanism
might
be
the
kind
of
the
firing
rate
dynamics
that
neurons
take
are
are
kind
of
continuous
as
well,
and
so
those
can
kind
of
summate
in
in
certain
ways.
So
so
it's
possible
that
that
something
else
is
actually
what's
implementing
this
and
that's
why,
for
the
work,
I've
been
doing,
I've
been
kind
of
trying
to
abstract
that
away,
but
but
this
is
at
least
one.
B
Right,
I
remember
there
was
this
guy
lokendra
shastri
in
berkeley
who
was
working
on
this
like
using
synchrony
as
a
way
to
do
bindings
and
relational
information.
He's
been
working
on
it
since
the
80s
and
he
seemed
to
be
always
in
like
one
corner
or
working
decades.
In
the
same
thing,
I
don't
know
if
this
is
related
to
his
work
at
all
or
not.
A
There's
a
lot
of
literature
on
synchrony
for
binding,
but
that's
quite
a
different,
although
kind
of
polycrine
synchrony
seem
quite
similar
and
then
obviously
I
do
have
some
similarities.
Synchrony
for
binding
is
very
different
idea,
because
that's
essentially
just
saying
that
you
know
in
the
t
and
the
l
example,
all
of
the
neurons
that
are
firing
for
the
t
are
firing.
A
Actually,
you
have
neurons
that
are
explicitly
encoding
this
these
binding
relations
and
therefore
can
be
read
out
by
higher
layers,
so
yeah
so
I'll?
I
can.
I
can
come
back
to
this
slide,
but
but
so
kind
of
returning
to
this.
This
toy
example,
if
say,
we've
got
a
t
and
an
l.
So
the
t
is
at
location,
a
b
is
a
location.
A
A
A
And
then
you
would
have
another
one
if
it
was
this
neuron
that
drove
the
t.
But-
and
you
know
the
desire,
then,
is
in
this
case
when
you
present
both
the
t
and
the
l,
it's
the
binding
neuron,
that's
encoding
the
relation
between
these
two
that's
active
and
not
the
other
one,
but
because
all
four
of
these
neurons
are
coactive.
A
So
the
kind
of
possible
approach
to
take
then
is
to
bring
in
this
idea
of
polychromy
and
in
particular
that
if
you
have
these
conduction
delays,
then
this
kind
of
third
neuron
that's
introduced
is
serving
as
a
coincidence,
detector,
for
this
causal
relation
and
in
particular,
is
active
if
and
only
if
it
was
actually
this
vertical
bar
neuron.
That
drove
this
one
because-
and
hopefully
this
animation
will
work
because
the
signals
will
propagate
this.
B
A
It
may
seem
kind
of
awfully
convenient
to
have
the
delays
to
kind
of
enable
this
relationship,
but,
interestingly
stpp,
if
you
start
with
random
delays
and
sorry
stp
spike
time,
dependent
plasticity
is
the
kind
of
main
form
of
plasticity
that
we're
aware
of
in
the
in
the
brain
it.
It
leads
to
exactly
these
kind
of
timings
and
kind
of
enabling
this.
These
representations
also
requires
top-down,
lateral,
connectivity
and
then
conduction
delays,
so
all
kind
of
elements
from
neuroscience.
D
A
No,
it's
a
it's
a
great
point,
and
so
the
kind
of
first
thing
to
say
is
obviously
the
way
I've
been
showing.
This
is
just
kind
of
a
toy
diagram
for
intuition.
In
reality,
each
of
these
features
would
be
encoded
in
a
in
a
distributed
manner.
But,
more
importantly,
you
know
these.
These
representations
that
the
binding
is
actually
occurring,
for,
at
least
in
my
view,
are
you
know,
they're
not
going
to
be
some
sort
of
grandmother
cell
it
may
be.
A
You
know
that
we
have
a
limited
number
of
kind
of
high
level
objects
consistent
with
kind
of
you
know
the
the
human
ability
to
represent
in
parallel.
You
know
between
four
to
seven
objects
at
the
same
time.
So
so
something
like
that,
in
which
case
the
the
order
of
as
in
the
magnitude
of
binding
neurons
that
you
would
need
is,
is
on
the
order
of
the
number
of
neurons
low-level
neurons
that
you
would.
B
A
The
delays
don't
change,
they
are
initialized
randomly
and
then
essentially
the
delays
that
naturally
through
stdp,
the
ones
that
coincidentally
arrive,
will
be
strengthened
and
the
others
will
be
pruned.
So
you
just
need
to
start
with
a
distribution
of
random
delays:
okay,
okay
and
then
yeah.
This
is
just
showing
you
know
again,
just
for
the
sake
of
intuition
just
some
example:
kind
of
png's
that
that
might
kind
of
encode.
This
kind
of
information
and
yeah
like
like
stdrs,
you
know,
activities
sparse
and
distributed
and
then
yeah.
A
So
yes,
so
so
there's
nothing
that
actually
adjusts
the
delays.
The
idea
is
that
you
just
start
with
a
random
distribution
of
many
delays
and
because
of
the
number
of
synapses
that
the
neurons
share
and
the
kind
of
sorry
the
number
of
neurons
in
the
network.
They
there's
kind
of
inevitably
going
to
be
some
some
neuron
that
will
receive
coincident
activity
from
these
two
neurons.
D
I
think,
as
long
as
you
as
long
as
you
don't
expect
the
number
three
neurons
to
detect
a
lot
of
things
at
once.
They
only
have
to
detect
a
few
then
and
then
you're
willing
to
take
a
distributed
representation
of
the
three
layers,
and
I
guess
it
would
work
it's
a
little
bit
like
our
spatial
pooler.
Without
without
learning
about
training
it
works.
D
A
Yeah,
if
if
there
was
just
no
no
neuron
that
received
the
activity,
then
then
yeah.
That
would
be
the
case,
but
I
mean
the
other
thing
is
so
spiking
neurons
are
kind
of
interesting,
because
all
kinds
of
things
can
happen.
I
mean
the
the
timing
with
which
this
neuron
spikes
as
well.
So
let's
say
it
does
receive
this.
A
A
Right
right,
so
so
yeah,
so
I
mean
ultimately,
so
I
should
say
you
know
this
this.
This
is
a
theory,
but
it
hasn't
been
kind
of
demonstrated
conclusively
in
with
kind
of
empirical
simulations
that
these
kind
of
binding,
neurons
reliably
emerge.
So
yeah
I
mean,
as
you
point
out
kevin
it
may
be
that
you
find
okay.
Well,
you
know,
maybe
maybe
some
sometimes
binding
relations
never
get
encoded,
and
and
what
is
the
kind
of
parameter
space
that
one
needs
in
order
to
ensure
that
that
is
the
case.
A
A
But
but
then
the
kind
of
idea
here
is
that,
rather
than
kind
of
border
ownership
being
just
a
limited
phenomenon,
these
kind
of
kind
of
hierarchical,
binding
representations
exist
that
kind
of
multiple
levels
of
the
hierarchy.
So
you
know
whether
it's
contours
and
how
we
represent
those-
or
you
know
a
feature
like
an
eye.
B
A
Great,
so
so
yeah
and
just
going
back
to
this
toy
example
again
of
vertical
bars,
driving
some
sort
of
kind
of
invariant
representation
and
the
kind
of
ideas.
How
can
this
yeah
be
encoded
in
a
deep
learning
system
and
so
in
the
kind
of
what
I
was
discussing
before
there
was
kind
of
this
importance
put
on
things
like
spiking
and
conduction
delays
and
that
sort
of
thing,
but
kind
of
from
an
algorithmic
point
of
view.
A
Ultimately,
what
this
binding
representation
doing
is
doing
is
encoding
a
causal
relation
between
this
little
feature
and
this
high
level
feature
and
and
saying
that
it
was
this
representation
that
was
responsible
for
its
activity
and
not
this
one
and
so
to
to
kind
of
do
this.
In
a
cnn,
I
thought
I'd
just
briefly
kind
of
go
over
again,
just
the
convolutional
neural
networks,
so
yeah
you
kind
of
have
the
lower
level
representation,
whether
that's
pixel
space
or
a
particular
layer.
A
And
so,
if
you
kind
of
look
at
say
a
lynette
five
type
architecture,
where
this
two
convolutions
two
max
boolean,
you
know
you
might
imagine
that.
Okay.
So
this
in
the
max
pool
layer,
this
it's
kind
of
encoding,
contour
a
is
somewhere
in
the
top
half
of
the
image
ways
in
the
layer
below,
as
well
as
other
representations
that
has
this
kind
of
contour
a's
at
position
x,
so
some
more
spatial
specificity
so
to
to
capture
which
neuron
in
this
layer
was
causally
responsible
for
the
max
fold.
A
Representation
is
very
straightforward
in
the
case
of
max
pooling,
and
it's
actually
an
operation
that
already
exists,
which
is
known
as
uncooling.
An
unpulley
is
normally
used
in
the
decoder
part
of
a
an
autoencoder
to
enable
reconstruction,
but
here
it's
being
used
in
the
feed
forward
part
of
the
network,
and
so
the
idea
is,
if
you
kind
of
think
back
to
these
representations.
A
This
t
is
a
more
invariant
kind
of
t,
neuron
say.
Whereas
this
one
has
a
smaller
receptive
field
and
because
of
that
smaller
receptive
field
and
because
of
the
higher
dimension
of
the
space
in
which
it's
embedded,
it
actually
encodes.
Spatial
information
in
in
this
case
that
the
kind
of
t
or
whatever
it
was
was
in
you
know
the
top
left
corner.
A
A
more
kind
of
interesting
question,
then,
is
how
to
capture
which
low-level
features
kind
of
simpler
features
were
responsible
for
driving
this
more
abstract
one.
So
more
kind
of
like
the
vertical
bar
that
drove
the
invariant
t
and
so
to
do
this.
I
kind
of
introduced
this
operation
called
kind
of
gradient
unpooling,
which
in
flavor
is
very
similar.
A
A
So
instead
to
kind
of
again
emphasize
this
idea
of
importance.
I
have
this
hyperparameter
gamma,
which
is
essentially
the
proportion
of
the
largest
gradients
that
I
take,
and
those
largest
gradients.
Their
location
is
then
used
to
generate
that
same
boolean
mask
that
I
mentioned
before,
and
so
yeah.
A
This
information
is
up
projected,
which
I
can
show
yeah
a
bit
more
kind
of
maybe
clearly
here,
where
the
kind
of
the
max
pool
representations
are
are
still
there.
So
you
know
this
still
recognizes
the
importance
of
these
kind
of
abstract,
invariant
representations,
but
alongside
them
literally
just
concatenated
are
the
unpooling
and
grading
unfollowing
representations,
and
so
you
know
you
might
imagine
that's
something
like
contour
a
somewhere
in
top
half
contour
a
is
at
position
x.
A
A
A
The
the
kind
of
nice
thing
about
that
is,
you
can
actually
yeah
take
a
an
image,
a
kind
of
high
dimensional
input
and
rapidly
and
in
parallel,
extract
both
kind
of
features,
but
also
some
degree
of
kind
of
spatial
information
about
those
features,
and
that
I
think
that
you
know
that
could
be
the
kind
of
representation
that
could
then
be
fed
into
these
kind
of
cortical
grid
cells.
As
kind
of
an
initial
kind
of
location-based
representation
to
kind
of,
then
then
enable
further
processing.
A
B
A
A
B
A
Decision
to
make
is
kind
of
how
many
of
these
are
you
going
to
include
yeah
in
practice,
I
found
not
just
one
unfully
and
one
grading
gradient
pulling
is
sufficient
for
mnist,
whereas
I
used
two
of
each
for
bgg
on
super
10.,
and
that
was
a
I
think,
a
15
team
player.
I
can't
actually
remember
now
yeah
network
yeah,
but
because
I
mean
I
I
think
super
ty.
You
mentioned
that
marcus.
A
Maybe
you
were
looking
at
applying
capsule
networks
to
extract
some
kind
of
spatial
information
for
kind
of
feeding
into
the
kind
of
cortical
grid
cells.
I'm
wondering
if
this
could
be
kind
of
yeah.
I
mean,
as
I've
mentioned
kind
of
these
ideas,
in
my
view,
are
complementary
to
what
capsule
networks
are
doing,
but
then
the
advantage
is.
They
are
very
fast
and
easy
to
train.
B
A
C
A
A
But
but
I
guess
just
just
as
a
teaser
to
hopefully
get
people
interested
it
does,
it
does
make
a
big
difference.
It
does
seem
to
result
in
yeah
I
mean
yeah.
I
don't
particularly
like
the
kind
of
the
way
state
of
the
art
is
measured
in
in
a
lot
of
machine
learning
papers,
so
I'm
hesitant
to
make
that
claim,
but
but
it
certainly
leads
to
robustness
across
a
huge
number
of
different
adversarial
attacks.
B
A
A
It's
similar
in
that
yeah
you
end
up
with
sparsity
and
and
that
I
guess
you
know
the
k
winner,
if
it's
just
based
on
the
activity
in
this
layer,
is
still
capturing
some
idea
of
kind
of
importance
and
actually
at
kind
of
one
point.
What
I
originally
wanted
to
do
was
rather
than
just
using
the
gradients,
was
to
multiply
these
kind
of
gradient
values
that
extracted
by
the
activity
and
then
apply.
A
This
kind
of
gamma
kind
of
operation
say:
okay,
which
of
these
were
the
most
important,
but
that's
actually
tensorflow
doesn't
like
that.
Because
of
this
particular
grading
operation,
that's
not
implemented,
I
was
able
to
do
it
in
pie
torch,
but
in
practice
I
didn't
see
a
huge
difference
in
terms
of
the
the
kind
of
adversarial
robustness.
A
And
so
yeah,
so
that's
again,
maybe
one
of
the
the
reasons
to
plug
for
the
kind
of
biological
implementation
of
this
is
that
you
know
that
does
it
kind
of
in
parallel
using
local
information
very
rapidly.
Whereas
you
know
here,
I've
had
to
implement
a
non-local
algorithm
to
kind
of
artificially
try
and
capture
this
causal
relationship.