►
From YouTube: Paper Review: Catastrophic Forgetting, Context-Dependent Gating and Synaptic Stabilization - 3/31/21
Description
We started this research meeting by responding to a few questions posted on the HTM forum. The HTM Forum is our open source discussion group. It is a great place to ask questions related to Numenta’s work and find interesting projects that people in the community are working on. Join HTM Forum today: https://discourse.numenta.org/
Subutai Ahmad reviews the paper “Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization” by Masse, Grant and Freedman. He first explains his motivations behind reading this paper based on Numenta’s previous work on dendrites and continuous learning. He then highlights the various network architectures simulated in the experiment and the results presented in the paper (i.e. accuracy for each network). Finally, Subutai gives his thoughts and the team discusses the results.
Paper: https://www.pnas.org/content/115/44/E10467
Other paper mentioned:
“Continuous Online Sequence Learning with an Unsupervised Neural Network Model”: https://numenta.com/neuroscience-research/research-publications/papers/continuous-online-sequence-learning-with-an-unsupervised-neural-network-model/
0:00 Answering Questions from HTM Forum
7:12 Paper Review
- - - -
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
B
C
C
C
E
Yeah,
so
this
I'm
sorry,
I
don't
have
a
reference
for
you,
but
I
remember
seeing
there
is
some
work
that
does
exactly
this,
where
they
have.
The
neuron
model
incorporates
the
backward
feedback
as
a
separate
path
through
like
separate
dendrites.
I
forgot
who
did
it?
It
might
have
been
something
related
to
difference,
target
propagation,
but
yeah
this
this.
This
approach
has
been
tried
by
by
some
people
to
varying
degrees.
B
Hey
alex
yeah,
yeah
alexa.
One
of
the
papers
I
know
is
it's
called
towards
deep
learning,
with
segregated
dendrites
by
grief,
lily,
crap
and
richards
back
in
2017
e-life.
So
that's
one
possible.
I
I
probably
butchered
the
first
name
there,
but
yeah.
That's
one
one
paper
I
know
of
that's
in
this.
That
explicitly
looks
at
incorporating
apical
dendrites
for
receiving
back
backprop
information.
E
C
C
A
A
B
A
F
B
B
You
know,
let's
say
in
the
lower
layers
again
that's
what
we've
kind
of
proposed
in
the
past
and
they're
in
relatively
close
proximity
to
the
sensory
input.
That's
maybe
in
higher
layers
and
there's
lots
of
you
know,
connections
back
and
forth
between
the
layers.
It's
not
like
their
grid
cells
are
in
some
really
distant
part
of
the
brain
and
the
sensory
input
is
in
a
completely
different
part
of
the
brain
and
there's
no
way
that
communication
could
get
from
one
place
to
another.
B
So
I'm
wondering
if
there
are
the
assumptions
behind
this
or
some,
if
there's
some
misunderstanding,
or
at
least
I'm
not
understanding
the
intent
of
this
question
correctly
and
in
terms
of
the
hierarchical
topology.
I
think
the
the
topology
is
very
clear
in
the
lower
layers
of
the
hierarchy,
but
as
you
get
higher
and
higher
up,
there's
less
and
less
topology
and
it's
things
tend
to
be
much
coarser
and
and
the
cells
respond
globally
to
a
lot
of
different
types
of
input.
B
So,
as
you
go
from
you
know,
even
in
you
know,
from
v1
up,
you
know
you
have
very
clear
topology
at
the
lower
levels,
but
at
it
there's
almost
there's
very
little.
Topology
each
area
corresponds
to
a
very
large
part
of
the
visual
space,
and
once
you
get
even
up
to
parietal
areas,
there's
almost
no
topology.
As
far
as
I
know,
so
I
think
that.
B
C
B
B
B
Can
you
guys
see
this
powerpoint
slide?
So
this
is
kind
of
a
background.
I
think
a
lot
of
people
are
already
know
this,
but
you
know
in
htms
we
did
continuous
learning
with
with
sparse
representations,
and
this
is
a
slide
from
a
presentation
I
made
late
last
year.
I
think
where
we
reviewed
a
whole
bunch
of
papers
on
supermass
and
meta
learning
and
oml
and
so
on,
and
I
made
the
point
there
that
when
we
think
about
continuous
learning
in
htms,
we
don't
really
use
the
language
of
gating
and
and
context.
B
So
this
is
the
same
layer
at
different
points
in
time
for
those,
if
you
don't
know
it,
but
if
you
look
at
a
particular,
if
you
look
at
a
layer,
there's
significant
recurrent
connections
back
and
forth
into
the
distal
dendrites,
and
then
what
happens
is
that
that
dendrites
and
well
for
any
given
task,
which
might
be
a
specific
sequence?
You
have
a
very
sparse
sub
network.
That's
specific
to
that
test.
B
So
when
you
have
some
quote-unquote
context,
which
in
the
temporal
memory,
was
the
activity
of
the
cells
in
the
previous
time
step,
that
was
the
context
that
the
dendrites
would
recognize
that
context
and
using
a
winner-take-all
mechanism
or
other
mechanisms,
would
would
select
a
particular
sub-network,
that's
specific
to
that
task,
and
then
all
the
representations
are
extremely
high,
dimensional
and
very
sparse,
and
this
allows
you
to
avoid.
You
know
significant
overlap
between
tasks
there.
So
this
is
so
even
in
our
kind
of
sequence
memory.
B
The
way
we
did
were
able
to
do
continuous
learning.
We
had
all
of
these
connections
there
and
then
based
on
context
that
dendrites
would
choose
specific
subnetworks
to
instantiate,
and
so
the
part
of
what
we've
been
doing
over
the
last
few
months
is
trying
to
see.
Okay,
we
did
this
in
hcm,
we
showed
it
working
in
practical
scenarios.
Can
we
take
these
ideas
and
incorporate
them
into
deep
learning,
and
so
I'm
not
going
to
talk
about
that
work
on
den
rights
right
now.
B
D
B
The
contact
signal
could
come
from
elsewhere.
So
if
you
think,
if
our
in
our
columns
paper
and
the
columns
plus
paper,
for
example,
we
use
the
exact
same
dendritic
mechanisms,
but
there
were
no
recurrent
connections
at
all.
The
context
was
coming
from
a
location
signal
in
that
case,
like
grid
cell
representations.
B
D
B
Yeah,
but
those
that
those
lateral
connections-
that's
what
I
mean
by
these
recurrent
connections.
Those
lateral
connections
are
there
only
in
this
in
the
temporals,
in
the
sequence
version
of
temporal
memory,
when
we
we
could
use
that
same
mechanism
with
zero
lateral
connections,
but
the
connections
onto
the
dendrites
instead
coming
from
the
location
signal
and
the
whole
thing
still
works.
D
B
Yeah,
so
it's
just
it's
more.
The
the
fundamental
idea
is
really
the
fact
that
we're
using
context
of
some
sort
to
select
sub
networks
that
are
going
to
be
instantiated
and
these
sub
networks
extremely
sparse.
To
avoid
overlap
that
that's
really
the
core
of
the
idea
and
then
you
could
say:
okay,
what
is
the
right
contact
signal
and
if
it's
the
previous
activity,
then
you
end
up
with
recurrent
connections
here
onto
the
dendrites
that
that's?
What
gets
you
the
con
context
from
the
past?
B
B
Yeah,
I
might
turn
off
my
camera
because
I'll
otherwise
always
be
looking
down,
and
I
feel
like
that
looks
really
weird
on
that
zoom.
So
I
might
turn
off
the
camera
for
a
bit
okay.
So
this
is
a
paper
we
published
a
while
ago
in
2016
that
specifically
talks
about
continuous
learning
with
these
sparse
networks.
Although
the
language
again
is
not
the
one,
I
just
use
it's
a
different
language.
That's
used
in
here,
that's
like
to
talk
about
it,
okay,
so
let's
get
to
the
paper
itself.
Okay.
B
So
this
is
the
paper
that
I
wanted
to
talk
about,
and
you
can
now
see
why
this
name
paper
name
caught
my
attention.
So
it's
kind
of
funny.
The
way
I
found
out
about
this
paper
is,
I
use
mendeley
to
keep
track
of
my
papers
and
and
to
read
my
papers
and
they
have
so.
They
know
what
papers
I'm
reading
and
they
have
a
email
that
gets
sent
out
every
week
or
so
that
they
suggest
new
papers
to
read,
based
on
what
you've
been
reading
and
I've
often
found
really
good
papers
for
that.
B
It's
a
pretty
interesting
kind
of
service,
and
so
this
came
up
my
feed
a
couple
weeks
ago.
I
think
I
posted
on
slack
and
then
quran
looked
at
it
and,
and
he
and
I
talked
about
it
and
then,
as
he
talked
about
it
with
me,
it
became
clear
that
it's
a
really
interesting
paper
and
sorry.
I
spent
you
know
time
reading
it,
and
so
I
have
to
say
this:
it's
a
very
well
written
paper.
B
It's
like
really
clearly
written
the
the
hypotheses
are
very
clear
and
they
go
through
it
step
by
step
and
as
I'm
reading
it
it's
like.
I
have
questions
about.
Oh
what
about
this,
and
then
you
know.
Lo
and
behold.
The
next
paragraph
would
talk
about
exactly
that
thing,
so
I
felt
for
me
anyway,
it
was
reading
it.
It
was
a
very
kind
of
logically
logical
paper
and
very
clearly
written.
B
So
I
really
like
that
part
of
it,
so
this
appeared
in
pnas
our
proceeding
of
the
national
proceedings
of
the
national
academy
of
sciences,
and
let
me
just
highlight
first
the
abstract.
You
can
see
this
highlighted
area
here
and
you
can
see
I'll.
Just
read
that
part:
it
says
here
drawing
inspiration
from
algorithms
that
are
believed
to
be
implemented
in
vivo
by
even
vivo.
They
just
mean
in
brains,
running
live
in
vivo.
We
propose
a
complementary
method,
adding
a
context.
B
So
the
core
of
what
they're
looking
at
is
very
similar
to
what
I
what
I
just
described,
and
they
test
different
aspects
of
this,
and
then
they
show
how
it
impacts,
continual
learning,
the
other
piece
that
they
do
is
they
include
something
called
synaptic
stabilization.
So
this
is
basically
elastic
weight,
consolidation
or
synaptic
intelligence.
There
are
two
methods
that
we've
discussed
in
the
past:
I'm
not
going
to
I'm
not
planning
on
going
into
those
in
much
detail.
I
think
we've
done
a
journal
club
of
that.
B
If
you
do
elastic
weight,
consolidation
or
synaptic
intelligence,
then
you
can
alleviate
some
of
the
problems
of
catastrophic,
forgetting
and
very
briefly,
what
happens
with
those
two
methods
is
that
any
given
synapse
tries
to
kind
of
handle
multiple
tasks?
Well,
so
their
insight
was
with
backprop.
If
you
learn
on
a
new
task,
then
there's
nothing
preventing
the
weight
from
moving
completely
away
from
the
old
tasks
right.
So
you
can.
If
you
learn
on
on
task
one
and
the
weights
become
adjusted
to
task
one.
B
Now,
you
don't
see
anything
more
from
task
one,
and
you
go
to
task
two.
Then,
when
you're
learning
task
two
the
weights,
the
way
backdrop
works.
It's
just
gonna
try
to
minimize
tasks
too
the
loss
for
task
two,
so
the
weights
can
move
completely
away
from
what's
good
for
task
one
and
in
ewc
and
synaptic
intelligence.
B
They
try
to
have
synapses
that
are
good
for
kind
of
multiple
tasks
at
a
time
or
if
it's
really
already
well
tuned
to
a
task,
they
won't
move
those
weights
as
much
and
they'll
try
to
use
other
weights
that
are
not
as
important
for
that
task.
So
there's
a
couple
of
different
ways
of
of
doing
it,
and-
and
you
can
already
see
what
the
issue
might
be-
that
well,
as
you
have
more
and
more
tasks,
you
can't
have
the
same
weight
sort
of
serving
too
many
different
masters
right.
B
It's
and
so
that's
what
they
mean
by
this
statement
here
it
is
uncertain
whether
synaptic
stabilization
alone
can
support
continual
learning
across
large
numbers
of
tasks.
So
it's
going
to
be
really
hard
for
you
get
to
hundreds
of
tasks
or
thousands
of
tasks
for
a
single
synapse
to
really
serve
well
across
all
of
these
tasks.
B
So
that's
that's
here
and
the
idea
here
is
they're
going
to
invoke
sub
networks
and
then
only
the
synapses
or
weights
within
that
set
network
have
to
learn
different
tasks.
So
if
you
have
completely
non
overlapping
sub
networks,
then-
and
you
have
let's
say-
you
have
10
sub
networks
and
100
tests-
then
any
sub
network
only
has
to
really
learn
10
tasks
right.
That's
the
basic
kind
of
thing
hypothesis
that
they're
looking
at
okay.
B
B
This
is
sort
of
remarkably
similar
to
the
numbers
that
we're
using
in
our
dendrite
networks
and
the
numbers
that
we
used
in
an
htm
in
terms
of
the
number
of
mini
columns.
You
know
we
have
2048
and
our
old
htm
things
and
at
quran
and
in
our
dendrite
networks,
we're
also
using
2048.
So
this
is
remarkably
similar
in
terms
of
numbers,
so
that's
the
basic
network
they
use,
and
then
they
train
it
with
basic
backdrop,
and
they
can
also
train
it
with
these
synaptic
stabilization
methods
that
I
discussed.
B
F
B
So
that's
sort
of
this
b
here
this
network
is
given
a
one
hot
encoded
signal,
this
listing
here
which
encodes
the
task
id
and
it's
just
fed
in
and
there's
a
layer
of
weights
here.
So
it's
just
another
input
into
the
hidden
layers,
so
each
of
the
hidden
layers
get
the
input
and
raw
input
and
also
gets
the
the
task
id
as
input,
and
it
can
just
learn
those
okay.
So
that's
this.
That's
a
scenario
that
they
test.
B
What,
if
you
just
given
the
context
signal
as
another
input,
and
you
just
treat
that
as
a
feed
forward
system
and
then
in
network
in
panel
c?
What
you
see
is
what,
if
you
had
completely
distinct
sub
networks
for
each
task,
id
okay.
So
let's
say
here:
this
is
an
example
that
shows
you
know
five
five
tasks.
B
So
let's
say
you
had
a
one
hot
encoded
vector
that
that
had
five
different
components
and
then
for
each
task.
It's
going
to
only
allow
a
specific
sub
network
to
become
active.
So
that
would
be
you
know
like
here
right
so
for
this
particular
task.
Only
those
activities
would
be
allowed
to
propagate
everything
else.
Would
be
set
to
zero.
B
Okay,
so
you
have
completely
distinct
non-overlapping
networks,
so
this
is
what
they
call
the
split
network
scenario
and
then,
in
panel
d
erase
this
stuff,
so
in
panel
d
they
have
what
they
call
context
dependent
gating-
and
here
you
have
again,
you
have
your
one
hot
encoded
input
specifying
the
task.
You
have
your
regular
input
and
for
each
task
here
you
have
a
specific
subset,
a
random
subset
of
the
hidden
units
that
are
that
are
allowed
to
propagate
okay.
So
what
happens?
B
E
B
B
It's
just
used
to
select
which
units
become
active,
and
this
is
also
very
similar
to
what
quran
tribes
and
the
ins
in
c
it's
his
original
hard-coded
thing
had
very
distinct
sub-networks
that
were
not
overlapping,
whereas
in
d
two
different
you
know
you
can
have
overlapping
sub
networks.
So
a
given
task,
you
know
so
maybe
for
task.
B
F
B
B
Yeah,
so
so
the
the
first
result
is
that
if
you
just
do
standard
backdrop
and
you
train
on
tasks
sequentially
on
on
permutation
mnist,
then
after
100
tasks
you
get
to
about
19.1
percent,
maybe
I'll
just
switch
to
the
the
charts
yeah
that
might
be
easier.
Oh
okay,
yeah!
They
don't
have
a
chart
on
that
sorry
yeah!
So
the
baseline.
Is
that
standard
backdrop
without
anything
special
gets
to
19.1
after
100
permutations.
B
So
here's
the
different
you
can
see
the
figures
here
with
the
different
results.
So
if
you
look
at
panel
a
they
try
both
ewc
and
synaptic
intelligence
and
going
up
to
100
tests,
and
so
ewc
gets
up
to
somewhere
around.
I
don't
know
like
72
percent
after
100
tasks
and
synaptic
intelligence
is
better.
It's
like
I
don't
know
like
83
or
something
like
that
here.
B
Then,
if
you
look
at
panel
b
here,
the
dotted
lines
show
what
happens
if
you
add
that
context
signal
as
input
into
the
network,
so
you're
still
doing
the
the
synaptic
stabilization
thing,
so
you
still
have
either
ewc
plus
context
or
synaptic
intelligence
plus
context,
and
you
can
see
that
you
know
the
accuracy
is
improved.
If
you
add
the
task
id
as
as
a
feed
forward
input
into
the
system,
okay
and
that
with
synaptic
intelligence,
plus
the
context
signal
you
get
to
roughly
90
accuracy.
B
Okay,
is
that
pretty
clear
and
then
panels
see
what
what
is
panel
c.
Oh
panel
c
is
okay.
Now
suppose
you
have
those
split
networks,
okay,
so
every
task
id
would
have
a
completely
distinct
subset
of
the
network's
network
active
and
then
you
and
you
have
all
of
these
various
combinations
you
can
have
okay,
so
ewc
plus
context.
B
B
B
They're
non-overlapping
exactly
and
then
the
last
one
panel
d.
It
shows
what
they
call
their
context,
dependent
gating
stuff
here
and
that's
where
for
each
task
id,
you
choose
a
random
subset
of
the
networks
of
the
network
to
become
active
and
you
allow
overlap
in
there
and
that's
where
they
got
the
best
results.
That's
here,
so
they
get
up
to
about
95.
B
F
B
B
Just
they
just
split
it
evenly,
they
didn't
really
try
different
variations
of
that.
You
know,
you
know
with
permeated
mnist.
The
tests
are
so
similar
that
they
may
not
be
much
advantages
in
tweaking
that,
so
maybe
that's
why
they
didn't
do
it,
because
each
task
is
basically
identical
to
the
previous
task.
B
Okay,
with
with
permuted
on
this
I
mean
so
this
panel
d
is
kind
of
interesting,
so
so
clearly
with
allowing
overlapping
subnetworks,
they
got
the
best
result.
So
that's
good
these.
These
two
lines
here
are
the
same
as
these
two.
So
there's
just
showing
the
you
know,
the
impact
of
synaptic
intelligence
in
ewc
on
its
own.
What's
really
weird
to
me,
is,
is
the
black
solid
line,
so
that
is
just
doing
context,
dependent
gating
without
any
synaptic
stabilization,
but
allowing
redundant
c,
and
you
can
see
it
does
pretty
poorly.
B
B
I
mean
it's
better
than
regular
backprop,
but
it's
it's
quite
a
bit
worse
than
ewc
or
si
on
its
own.
But
somehow,
when
you
do,
you
know,
let's
say
synaptic
intelligence
plus
context,
dependent
gating.
That
gives
you
the
the
very
best
results
up
here,
so
that
was
that
was
kind
of
interesting,
maybe
if
they
had
done
some
sort
of
learning
of
the
masks
like
you're
suggesting
kevin,
maybe
this
the
black
solid
line
could
have
been
better.
B
B
No,
no,
it's
very
different
from
dropout.
So
let's
look
at
some
of
these
numbers
and-
and
I
think
the
numbers
are
interesting-
so
what
they
did,
what
these
charts
show
is
getting
80
of
the
hidden
units.
So
what
that
means
is
that
their
sparsity
is
80,
that
80
of
the
hidden
units
are
set
to
zero
for
a
given
task.
B
So
it's
you
know
there
are
2
000,
choose
20
possible
sub
networks
here
and-
and
so
this
is
a
really
huge
number,
and
so
you
can
see
that
you
know
you.
The
chance
of
a
significant
overlap
between
any
two
sub
networks
is
going
to
be
small.
So
it's
not!
It's
not
a
noise
as
such
they're
explicitly
saying
for
task.
B
F
F
B
Yeah
and
the
other
thing
other
you
know
so
dropout
is
ran
completely
random.
It
doesn't
look
at
the
input
or
the
task
at
all
and,
of
course,
and
the
other
thing
is
dropout
is
only
used
during
learning
during
inference.
There
is
no
dropout,
whereas
here
it's
learning
an
inference.
It's
the
same,
so
it's
quite
a
pretty
different
from
dropbox.
B
Yeah
yeah,
so
I
looked
at
their
appendix
where
they
actually
tested
the
sparsities
a
little
bit
more.
So
they
mentioned
that
briefly
here
that
they
tested
different
levels
of
sparsity,
so
I
looked
at
their
appendix
and
what
they
have
here
is
the
accuracy
they
look
at
the
drop-off
in
accuracy
as
a
function
of
sparsity.
B
So
if
you
just
do,
if
you
just
get
50
of
the
units
you
end
up
here,
you
know
so
you
know
50
sparsity,
yeah
still
still,
you
know,
keep
keep
keep
in
mind
that
the
access
here
is
down
to
90.
It's
a
pretty
it's
scaled
to
90
from
90
to
100.
So
it's
still
pretty
good.
It's
still
well
above
back
prop
or
you
know,
even
synaptic
intelligence,
but
as
you
increase
the
sparsity,
you
know
this
is
increasing
the
sparsity.
B
B
H
H
Less
less,
if
you
have
less
neurons
left
you
it's
not
enough
neurons
to
classify
anymore,
because
you
can
see
if
you
start
at
zero,
it
gets
worse
and
the
like.
The
starting
performance
is
better.
The
more
you
have
available,
of
course,
to
classify,
but
it
decreases
very
little.
So
I
guess
it
could
go
on
a
bit
at
least.
B
Yeah
exactly
so
so
yeah,
so
I
I.
I
think
that
is
that's
going
to
be
right.
You
know,
as
you
get
higher,
when
you
get
too
sparse,
accuracy
is
going
to
fall
off,
but
given
that
it's
kept
increasing
here,
I
think
for
this
purpose
of
a
paper.
You
would
want
to
keep
going
and
see
where
the
falloff
is
right.
So
I
would
have
liked
to
see
a
chart
where
you
just
look
at
the
accuracy
after
100
tasks-
and
maybe
you
look
at
you-
know
sparsity
here
and
the
you
know
the
accuracy
here
and
yeah.
B
B
D
B
E
D
G
In
that
case,
going
back
to
what
you
were
saying
earlier
about
the
87
percent,
that
means
that
if
they
have
2000
hidden
units
in
each
layer,
that's
260
of
them
would
be
used
for
a
particular
sub
network.
And
then,
if
you
have
relu,
that
means
roughly
130
are
getting
through.
So
so
that's
well.
I
guess
I
don't
know
if
that,
how
that
compares
to
tasks
like
permeated
amnest,
but
that
still
seems
pretty
wide.
They
should
be
able
to
go
past.
87.
B
E
F
E
B
Would
be
universal
exactly
and
that
you
know
I
talked
about
the
dimensionality
stuff
in
a
previous
meeting
but
yeah
this
is,
as
you
increase
the
dimensionality.
You
should
be
able
to
increase
the
sparsity
and
I
think
even
the
absolute
number
of
units
that
are
allowed
to
be
on
can
decrease
as
you
increase
dimensionality.
B
B
B
Okay,
so
that's
that's
kind
of
their
basic
result.
They
also
tried
it
on
imagenet,
which
I'll
explain
in
a
second
this,
the
second
set
of
things
they
do
I'll
go
much
more
quickly,
they're
trying
to
say,
okay.
Why
is
this
the
case?
How
come
having
this
context
dependent
networks?
Why
is
it
so
much
better
than
just
using
you
know,
let's
say
synaptic
intelligence
on
its
own
right
and
what
they
looked
at.
B
Is
you
can
look
at
how
important
each
synapse
is
to
the
to
the
all
the
tasks
and
you
can
plot
the
the
importance
and
this
importance
comes
out
of
this
synaptic
intelligence
loss
function
pretty
directly.
Let's
see,
if
I
can,
you
know
the
the
synaptic
intelligence
stuff.
You
know
you
have
your
regular
loss
and
then
you
add
this
other
loss
here
where
you
look
at
the
weights
before
and
from
the
previous
task
and
the
current
test,
and
you
compute
this
omega
and.
B
So
they
have
a
measure
of
how
important
a
particular
synapsis
and
what
they
hypothesized
is
that
you
want
to
have
a
whole
bunch
of
synapses.
That
are
not
really
that
important,
so
you
can
allocate
to
new
tasks,
and
so
if
every
synapse
is
really
important
in
the
tasks
you've
already
seen,
then
it's
going
to
be
really
hard
to
learn
new
things
without
without
some
loss
and
accuracy,
and
so
they
plotted.
B
B
B
It's
going
to
change
very
little
because
it
was
important
for
previous
tasks,
so
si
on
its
own
is
going
to
prefer
this
stuff
on
the
on
the
on
the
left
here,
and
so
that's
why
they're
saying
synaptic
intelligence
plus
context,
dependent
gating,
is
allows
you
to
get
so
much
higher
accuracy.
That's
their
the
way.
They're
good.
B
Basically,
they
did
a
very
a
smaller
version
of
imagenet,
so
they
had
imagenet
images
that
were
downscaled
to
32
by
32.
So
it's
you
know,
presumably
for
speed
things
and
they
split
mns
into
100
from
the
thousand
classes.
They
created
a
hundred
tasks.
Each
task
had
10
different
labels,
so
they
they
took,
they
just
split
up
the
thousand
classes
into
into
100
tasks
of
10
10
classes
each
and
then
they
they
they
try.
B
Two
different
architectures
one
is
a
multi-headed
architecture,
so
the
multi-headed
architecture
had
a
thousand
oops
a
thousand
output
units,
and
they,
you
know,
for
each
task.
10
of
the
output
units
were
assigned
to
the
category
for
that
task,
and
so
they
would
basically
set
990
output
units
to
zero
at
any
for
any
particular
task
and
only
allow
the
other
10
to
to
learn.
B
They
created
a
baseline
maximum
accuracy.
You
can
get
is
about
56.5
and
that's
if
you
essentially
look
at
just
train
for
each
task,
just
train
those
10
categories.
Look
at
the
accuracy
then
reset
all
the
weights
train.
Another
another
task.
Look
at
the
accuracy
reset
the
weights
train,
another
task,
and
that
will
give
you
the
kind
of
the
maximum
oops.
Let
me
get
rid
of
my
notes
here
that
will
give
you
the
maximum
possible
accuracy
on
this
network.
B
And
they
found
this
that,
if
you
just
do
synaptic
stabilization,
that
gets
you
so
that
the
maximum
accuracy
is
56.5
and
with
just
synaptic
stabilization
you
can
get
up
to
about
55.
So
basically
they
concluded.
This
was
not
a
hard
enough
task,
and
you
know
si
on
its
own
is
able
to
solve
this
without
catastrophic
forgetting
so
then
they
did
a
single
head
architecture
which
is
similar
to
what
happened
in
permeated
amnest.
So
here
the
output
unit
only
consists
of
10
units
and
so
for
any
given
task.
B
B
D
B
B
B
B
B
B
Okay,
so
so
pretty
nice
results
pretty
impressive
results
that,
with
this
they
were
able
to
pretty
much
recover
all
the
accuracy
on
this
imagenet
task
and
the
kind
of
the
last
thing
here
I
made
a
list
of
kind
of
some
of
the
issues
that
I
had
so
overall,
I
really
like
the
paper
again,
it's
very
clearly
written.
I
encourage
you
to
read
it.
If
you
haven't
read
it,
they
explain
everything
and
very
logically,
and
they
test
a
lot
of
different
conditions.
B
So
some
of
the
issues
I
had
you
know
here,
you
have,
you
know
the
task
id
ahead
of
time.
So
that's
given
to
you,
you
know
it's
not
like
you're
learning
the
context
from
the
input
or
some
other
thing
you
you're,
given
the
the
task
id
the
masks
are
hard-coded
like
kevin
mentioned,
there's
no
concept
of
learning
these
masks.
B
You
know
they
didn't
explicitly
mention
this,
but
the
weights
are
all
dense,
so
there's
no
sparsity
in
the
weights.
This
is
purely
looking
at
activation.
Sparsity-
and
I
said
here,
xdg
by
itself
did
not
help
much,
but
I
guess
in
the
imagenet
one
it
did
better,
but
really
this,
but
the
bottom
line
is
this:
this
they
really
required
synaptic
intelligence
plus
context-dependent
gating
to
to
get
the
best
results.
B
I
think
it's
very
close
to
the
supermass
paper
super
mess
paper.
Also,
you
had
you
know
pre-determined
random
masks
for
every
task,
although
they
allowed
sort
of
mixtures
of
sub
networks
and
in
supermass
you
didn't
train
the
weights
themselves,
which
was
kind
of
one
of
the
unique
things
about
supermass.
But
you
know
conceptually
it's
very
simple:
it's
very
close
in
the
space
of
papers.
B
It's
somewhat
similar
to
animal
because
we're
doing
this
multiplying
the
activations
by
this
by
a
context
signal
animal
is
better
because
you
learn
the
context
signal
through
meta
learning,
but
I
actually
like
this
paper
better
because
they
train
the
full
network,
whereas
an
animal,
they
only
train
the
output
layer.
So
it's
not
to
me:
that's
not
really
continual
learning
here,
they're
training,
the
full
network
and
they're
just
using
the
contact
signal
to
determine
which
parts
get
trained.
B
You
know
we
talked
about
the
sparsity
percentages.
I
thought
that
was
interesting.
You
know
they
should
be
able
to
go
much
faster,
particularly
if
they
go
higher
dimensionality,
but
it's
good
that
they
at
least
looked
at
it.
I
thought
that
thai
there
was
a
bunch
of
stuff
about
neuroscience
in
the
beginning,
but
I
thought
the
tie
to
neuroscience
was
a
bit
loose.
B
H
B
H
F
B
B
B
H
Yeah
that
actually
is
another
side
question
I
had
so.
The
catastrophic
forgetting
for
condition
c
was
actually
a
lot
weaker
than
I
had
expected,
so
they
just
split
it
into
five
sub
networks,
but
every
network
still
has
to
solve
20
tasks.
So
I
was
a
bit
surprised
that
the
accuracy
stayed
so
high
compared
to
the
baseline
of
using
one
network
like
just
dividing
it
by
five.
F
A
B
D
B
B
H
I
guess
it's
when
they
also
apply
synaptic
intelligence.
Additionally,
maybe
that
helps
a
lot,
but
I
guess
that
brings
me
to
my
main
question,
which
is
those
plots
where
they
show
the
synaptic
importance
and
that
there's
lower
importance
yeah
and
do
they
have
this
also
for
the
split
network,
because
that
would
be
interesting
if
actually,
this
effect
is
coming
from
the
overlap
or
if
it's
just
coming
from
splitting
the
network.
B
D
D
F
B
G
D
G
And
then
so,
each
of
them
is
only
learning
20
tasks
each.
So
then
I
guess
a
fair
comparison
point
would
be
the
final
accuracy
in
panel
c
versus
where
the
curves
get
to
after
20
tasks
in
panel
d,
especially
since
you're,
including
the
context
dependent
gating
and
there
it's
up
at
beyond
95
percent,
but
they're
in
panel
c,
it's
below
95
percent.
So
I
guess
it's
not
it's
not
as
impressive
as
it
as
you'd.
Originally
think.
G
D
B
Yeah,
so
what
they
did
in
the
gating,
they
gave
80
percent
of
the
hidden
units.
So
you
know
so
one-fifth
of
the
hidden
units
are
allowed
through,
whereas
here
in
this
in
this
in
the
split
network
case,
they
also
have
exact.
You
know
they're
still
only
allowing
20
of
the
units
to
get
through,
but
there's
no
redundancy,
there's
less
redundancy.
G
G
You're
training,
each
sub
network
on
each
new
task,
but
now,
even
within
that
subnetwork
you're
constraining
some
weights
and
saying
weights
that
have
been
that
are
sort
of
shared
with
another
task:
you're
constraining
them
a
little
more
and
so
those
are.
Those
are.
Those
are
not
getting
updated
so
much,
but
they
already
have
so
much
capacity
in
their
sub
network
because
they
have
even
when
they're
at
the
87
percent
sparsity
level.
They
still
have
260
units
that
they
can.
They
can
choose
to
to
train
and.
D
G
D
G
B
Yeah
I
mean
even
in
the
even
in
the
in
all
of
these
you
cases
you
have
individual
synapses
that
have
to
be
repurposed
to
multiple
tasks,
and
so
it
makes
sense
that
si
or
ewc
would
help
in
those
cases-
and
I
guess
once
you
have
the
context,
dependent
gating,
you
have
much
less
interference
between
the
different
tasks
and
so
each
synapse,
with
both
xdg
and
synaptic
intelligence.
A
given
synapse
is
only
used
for
a
small
number
of
tasks,
a
much
smaller
number
of
tasks.
H
Yeah,
I
feel
like
it.
That
would
be
super
interesting
to
know
the
effect
of
the
overlap,
because
it
could
be
an
interesting
explanation
to
say:
okay,
because
there's
some
overlap,
the
overlap
destroys
some
of
the
connections,
so
it
actually
has
to
use
a
very
few
neurons
to
encode
this
class
because
sometimes
through
the
overlap,
there's
like
mix
up
between
the
weights-
and
that's
could
be.
Why,
like
the
synaptic
import,
the
importance
distribution
is
changed
like
that
yeah.
D
D
H
D
E
A
D
E
I
was
just
going
to
say
the
hypothesis.
The
hypothesis
that
I
had
was
that
with
imagenet,
the
different
tasks
are
somewhat
similar
because
they're
all
dealing
with
natural
sort
of
objects
and
they're
all
sort
of
the
same
general
distribution
of
natural
images,
but
with
bermuda
mnist.
It's
like
a
totally
different
task
and
even
though
it's
still
coming
from
you
know
the
base
mnist
data
set
when
you
permute
it
all
of
a
sudden.
All
that
structure
is
lost.
So
then,
yes,
you
can't.