►
Description
In part two, visiting scientist Jeremy Forest continues his overview of the plasticity mechanisms in the brain and focuses on activity levels in neurons rather than dendritic spines.
He talks about how memory ensembles are dynamic and how neurons encode signals. He then explores different behavioral expressions of memory and context impacts plasticity mechanisms in neurons. He makes the case that all those mechanisms interact on widely different timescales and timescale should be considered when developing deep learning networks.
Part One: https://youtu.be/gbzMt4-3YhY
- - - - -
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://numenta.com/news-digest/
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
So,
okay,
last
week
I
talked
a
little
bit
about
how
neurons
can
interact
and
how
that
changes
over
time
with
different
plasticity
events.
So
we
talked
about
ltp
lcd,
some
of
the
mechanisms,
and
then
we
stopped
here
on
on
some
mechanism
for
local
signaling,
of
maintaining
or
increasing
or
decreasing
synaptic
strength,
which
is
an
as
an
impact,
basically
on
the
w
value
right.
A
So
there
can
be
a
lot
of
those
and-
and
one
of
the
message
was
that
there's
different
time
time
scale
that
needs
to
be
taken
into
account
and
all
of
these
different
time
scales
interact.
And
then
you
can
have
a
lot
of
different
results
at
the
level
of
spines,
which
are
the
basically
location
where
you
have
interaction
between
neurons
right.
A
So
ltp,
here
in
a
for
example,
on
this
specific
spine
is
gonna
cause
some
modification
in
that
case
an
increase
in
strength
and
which
is
associated
with
an
increase
in
some
specific
proteins
that
come
that
that
makes
that
spine.
Okay,
so
you
can
have
very
specific
input,
specific
increase
in
strength
right,
but
then
you
can
also
see
very
specific
decrease
in
strength.
So
in
the
body
value
of
of
neurons
right
and
like
last
time,
just
like
don't
hesitate
to
ask
question
or
interrupt
or
anything
right,
the
thing
is
you.
A
B
Yeah
I'll
I'll
try
out
something.
I
read
a
paper
once
I
read
a
paper
and
I
cannot
find
this
paper,
so
I
apologize.
I
can't
point
you
to
it,
but
someone
did
an
analysis
of
spines
and,
and
their
conclusion
was
that
spine
growth
had
less
to
do
with
the
strength
of
the
synapse
and
more
to
do
with
the
permanence
of
the
synapse.
B
B
From
my
understanding
all
right,
the
conclusion
was
that
if
you
looked
on
the
balance
this
this
author
and
I'm
not
making
this
claim,
this
author
made
the
claim
that
it's
really
more
about
permanence,
and
there
are
a
lot
of
things
related
to
why.
That
is
the
case.
But
it
was
a
convincing
argument
when
I
read
it,
it's
not
to
me
that
there
isn't
strength
and
changes,
but
I
think
it's
an
important
thing
to
think
about
to
realize-
and
this
is
where
our
our
concept
of
synapse
permanence
comes
from
came
from.
B
It
came
from
that
paper,
which
I
cannot
find
where
again,
it
was
more
like
hey.
You
know.
We
think
this
is
just
increasing
the
strength
of
it,
but
it
really
makes
it
harder
to
forget
and
a
synapse
which
has
a
very
small
spine
or
no
spine
like
a
filopedia.
They
call
it
can
almost
transmit
as
much
neurotransmitter
as
a
big
one,
but
it
goes
away
very
quickly.
That's
like
your
fourth
line
there,
so
I
just
want
to
throw
that
in
to
make
people
understand
that
it's
not
just
about
strength.
A
A
A
A
B
Hippocampal
cells-
remember
encode
also
in
code
time
and
episodic
time
of
events
and
so
part
of
what
they
do
is
record
things
at
different
points
in
time,
where
it's
not
clear
that
that's
happening
at
all
in
the
cortex
in
terms
of
like
forming
memories.
So
I
just
so
this.
It
may
not
be
a
general
property,
mainly
yeah,.
A
A
A
A
A
Okay,
so
now
there's
also
a
lot
of
mechanisms
to
maintain
homeostasis.
So
what
that
means
is,
if
you
have
at
the
level
of
signups
or
at
the
level
of
the
neurons
that
is
in
its
entirety,
you
can
have
forever
scaling
up
or
down
of
of
connections
right.
You
have
to
have
ways
of
regulating
how
big
signups
can
grow
and
how
how
little
synapse
you
can
have
right,
and
so
there
are
some
mechanisms
to
constrain
the
the
whole
other
at
the
level
of
the
whole
neon,
how
to
constrain
and
regulate
synapses
as
a
whole.
A
Okay
and
now,
if
we
go,
I'm
not
going
to
spend
much
time
on
this
because
we
spend
a
lot
of
time
on
it
and
don't
write
so,
but
if
we
zoom
out
a
bit
more
from
the
synapse,
we
also
have
a
lot
of
linear
and
nonlinear
integration.
That's
going
on
the
level
of
the
full
neurons
and
there's
a
spatial
component
that
is
usually
not
included
in
in
deep
nets.
So
we
have
the
temporal
component
that
we
already
talked
about.
This
is
also
a
spatial
component.
A
So
what
that
means
is,
if
you
have
a
memory
event,
a
t0,
then
either
you
have
a
very
short
memory,
which
is
called
short-term
memory,
and
that
goes
away
very
quickly
or
you
have
long-term
memory,
because
you
had
the
plasticity
underlying
those
memory
that
allowed
it
to
form
and
be
stored
into
long-term
memory.
And
then,
when
you
want
to
recall
something,
you
have
a
retrieval
of
that
memory
and
either
you
are
going
to
completely
forget
about
it,
because
you're
going
to
rewrite
that
memory
or
you
have
a
consolidation
event.
B
B
B
A
A
I
think
it
shows
I
mean
I
might
have
not
understood
it
properly,
but
I'm
not
sure
it
exactly
show
the
firing
rate,
but
more
of
the
activity
in
the
sense
of
plastic
activity
of
that
neurons
in
relation
to
the
other
ones,
the
other
red
ones,
for
example.
So
the
encoding
of
specific
events
is
gonna
happen
and
you're
gonna
have
a
lot
of
modification
of
that
neurons.
Whether
we're
talking
about
if.
B
A
B
Okay,
so
again,
that
would
be
then
related
to
the
synaptic
state.
Not
activity
of
the
neuron,
so
a
state
of
the
synapse
can
doesn't
have
to
even
be
receiving
action
potentials.
It's
just
like.
Oh,
what's
the
size
of
the
spine
or
how
many
receptors
it
has,
and
so
that
to
me
would
be
like
the
fading
of
a
memory
but
you're
saying
it's
not
I
don't
get
it
and
it's
not
even
clearly.
What's
the
timeline
here
is
this.
A
B
A
B
B
A
Okay,
I
see
that
I
see
the
difference
in
what
you
mean
now.
Okay,
no,
I
think
the
first
way
you
framed
it
was
good
and
it
I
don't
think
it
means
that
those
neurons
are
not
active.
They
are
it's
just
the
subset
of
synapses
that
encode
or
the
subset
of
connection
that
encode.
That
memory
is
not
changing
much.
B
C
A
A
Okay,
so
if
you
another
question,
I'm
going
to
change
to
the
next,
which
basically
is
the
same
idea
that
those
memory
and
samples
are
dynamic
and
that
they
evolve
over
time
with
a
specific
example
of
some
memory
encoded
in
a
specific
brain
region.
In
that
case,
it's
going
to
be
the
hypocampus,
but
in
itself
it
does
not
matter,
but
these
specific
memory
and
sample
can
be
transferred
or
written
out
of
this
brain
region.
They
were
encoded
and
to
another
brand
region
that
is
going
to
be
its
final
storage,
location.
B
B
Was
a
paper
that
we
read
a
while
back
this?
This
idea
of
this
transferring
I
mentioned
this
last
week
always
struck
me
as
odd
and,
and
there
was
a
paper
that
was
we
reviewed
here
at
momentum
some
number
of
years
ago,
which
fortunately
made
the
point
that
memories
are
not
transferred
from
the
hippocampus
to
the
cortex,
but
that
they
are
independently
formed
at
different
rates.
B
And
yet
somehow,
if
you
remove
the
hippocampus,
the
ones
in
the
cortex
do
not
form,
but
I
forget
the
details
of
the
paper,
but
they
it
was
pretty.
It
was
pretty
persuasive
that
it
just
the
idea
that
they're
being
transferred
it
may
not
be
right,
it's
more
like
they're
they're
occurring
simultaneously
in
both
places,
but
at
very
different
time
skills.
A
Yep
yeah
yeah,
I
did
not.
I
tried
not
to.
I
realized
it's,
maybe
a
little
bit
more
too
much
detail
already
so
yeah
yeah.
The
the
main
idea
was
to
to
to
make
the
point
that
memories
are
dynamic
and
don't
stay
in
the
same
state
all
the
time
and
viewing
in
the
context.
For
example,
continual
learning
during
the
single
purpose
of
remembering
everything
in
the
same
way
might
not
be
the
best
way
to
look
at
things.
A.
A
A
A
High
information
on
some
specific
inputs,
so,
for
example,
here
those
two
inputs
are
not
really
well
represented
by
any
of
the
neurons
that
are
encode,
that
series
of
input
and,
if
you
add
new
neurons,
you
can
specifically
tune
those
neurons
to
encode
those
specific
input
signals
and,
for
example,
at
least
in
mice.
These
neurons
have
been
shown
to
help
with
discriminating,
very
close
stimuli.
B
B
C
A
Yeah,
there's
definitely
a
consensus
for
for
mice
and
in
the
olfactory
bulb
and
hypercampus,
there's
less
of
a
consensus
for
human
olfactory
bulb
and
even
hypercampal
neurogenesis,
while
the
actual
the
hypothesis
is
that
it
occurs
a
lot
at
early
ages.
But
then
by
adolescence.
It's
almost
gone
but
because.
B
One
could
argue
that
you
know,
obviously,
when
you
have
new
neurons,
you're
you're
you're,
going
to
be
losing
neurons
too,
and-
and
one
can
argue
that
this
is
a
way
of
this,
helps.
You
learn,
like
you
just
shown
this
figure,
but
it
also
forces
you
to
forget
things
and,
and
so
again
in
the
hippocampus.
That's
okay,
these
episodic
memories,
we're
forgetting
of
all
the
time
performing
new
ones.
All
the
time.
That's
fine!
What
happened
to
me!
You
know
six
years
ago.
B
It
may
not
be
so
relevant,
but
in
the
cortex,
where
you're,
basically
just
you're
learning
the
structure
of
the
world,
the
more
permanent
structure
of
the
world,
then
it
wouldn't
make
as
much
sense.
It
would
make
it
easier
to
learn
new
things,
but
you'd
also
be
forgetting
fundamental
things.
You
knew
so
it
kind
of
makes
sense
that
you
wouldn't
see
this
much
in
the
cortex,
but
you
would
see
it
in
the
hippocampus.
A
Yeah,
it's,
it
seems
to
be
a
mechanism
to
adapt
to
environments
or
to
stimuli
that
you
can't
predict
in
the
sense
that,
for
example,
you
can
with
with
a
vision,
you
have
a
specific
light
spectrum
that
you
can
see,
but
with
context
or
others.
There
is
such
a
broad
input
space
that
you
cannot
predict
everything,
that's
going
to
be
an
input,
and
so
you
have
adaptive
mechanisms
to
be
able
to
adapt
according
to
the
environment,
you
are
in.
D
C
D
D
B
In
some
sense,
but
you
remember,
there's
never
well
in
any
kind
of
distributed
representation,
there's
never
an
essential
neuron
there's.
You
know
at
some
point
you
wipe
out
enough
and
then
the
thing
collapses,
but
you
know
it
slowly
degrades
anyway.
This
is
an
interesting
property
supervised
says:
it's
not
clear!
You
have
to
do
this.
It's
only
occurring
parts
of
the
brain.
There
are
other
ways
you
can
achieve
it.
This
may
be
just
one
of
those
biological
things
that
helps
brains,
adapt
quickly.
B
E
B
B
B
Connect
to
with
with
some
value
if
they
don't
find
anything
to
connect
to
the
receipt,
so
even
existing
neurons,
the
dendrites
and
the
axon
tips
are
growing
and
receding
and
growing
and
recepting
trying
to
find
useful
connections.
Creating
a
new
neuron
doesn't
really
help
help
you
any
further.
You
still
have
to
send
your
accent,
someplace
and
figure
out
where
it
goes.
So
it's
from
a
theory
point
of
view.
B
It
really
wouldn't
matter
you're
trying
to
form
connections
to
places
you
haven't
had
before
whether
you
grow
the
existing
neurons
or
you
put
a
new
neuron
in
it's
kind
of
the
same
thing
I
mean.
How
would
where
would
the
new
neuron
know
where
to
go
it?
Doesn't
it
doesn't
know
where
to
send
its
axon
there's?
No
like?
Oh,
I
need
to
send
it
to
this
other
section
over
here,
there's
no
way
of
knowing
that.
So
it's
based
you
saw
when
you're
born.
You
have
these
very
confused
connections.
B
B
D
B
Development
you've
got
this,
it
tells
you
where
to
basically
send
send
your
axon
to
dendrites
most
of
your
axons
and
and
then
actually
start
experiencing
the
world.
Some
of
those
things
are
useful
and
some
of
them
aren't
and
the
ones
that
don't
seem
to
be
useful,
get
pruned
back,
yes
agreed,
but
that
doesn't
seem,
but
I
don't
think
that's
happening.
B
C
A
B
A
Different
type
of
glyosol
are
going
to
play
different
types
of
role
in
regulating
interaction.
So
glass
are
going
to
play
a
role,
but
microglia
are
going
to
play
a
role
so,
for
example,
they
can
remove
spines
on
specific
neurons
and
with
I
didn't
go
into
that
here,
but
instead
of
speaking
of
of
a
spine,
that
is
a
pre
and
a
past.
Synaptic
element
we're
starting
to
speak
more
about
a
tripod
tight
spine
where
you
have
an
glia
element
associated
with
that
usually,
and
the
glial
element
is
also
able
to
regulate
the
synaptic
strength.
B
A
That's
not
true
anymore,
I
think
because,
for
example,
there's
it's
been
shown
that
glass
cell
can
have
calcium
waves
and
those
calcium
waves
can
go
pretty
far
and
actually
two
neurons
that
are
are
not
connected
can
be
and
very
well.
Relatively
distant
can
be
activated
via
a
glass
cells
that
propagates.
A
B
A
A
B
But
just
imagine
imagine
this:
what
you're
ultimately
generally
trying
to
do
is
to
find
connect
an
axon
to
a
dendrite
that
have
a
a
coincidence
in
firing
or
that
are
you
know
that
they
can
productively
form
a
synapse
which
is
really
like
could
be
like
like
heavy
and
type
of
learning.
There
is
no
other.
B
B
There's
no
way
in
the
distance
there's
somebody
who's
might
be
useful
that
you
form
a
synapse,
it's
hard
to
imagine
what
kind
of
you
know
gradient.
That
would
be
someone
who
could
detect
that
over
long
distances
we're
talking
about
specific
axons
and
specific
dendrites.
It's
not
a
particular.
It's
not
like
a
general
thing.
A
B
A
B
But
also
I'm
thinking
like
just
logically
walking
through
what
it
would
mean,
and
why
would
it
seem
very,
very
hard
to
imagine
that
could
happen
in
our
simulations.
We
can
make
anything
happen.
We
can
say
this
neuron
these.
These
two
neurons
should
form
a
synapse
and
therefore
we
don't
really
care
how
far
apart
they
are,
but
but
in
cortical
in
neural
tissue,
it's
very
hard
to
imagine
a
mechanism
that
would
be
that
specific
directed
over
long
distances.
A
At
the
behavioral
level,
yes,
well,
what
else
is
I'm
saying
that?
Because,
because
you
don't
see
a
behavior,
the
behavioral
level
doesn't
doesn't
mean,
there's
not
remnants
of
that
behavior
encoded
at
the
cellular
level
and,
for
example,
if
I'm
doing
the
second
training
on
the
right,
the
animal
is
going
to
forget.
But
then,
if
I'm
retraining
him
he's
gonna
be
able
to
relearn
that
training
much
quicker
than
the
first
time,
meaning
there's
still
something
in
the
cellular
level.
That
is
there
and
that
allows
for
that
training
to
happen.
Much
faster.
A
And
that's
also
the
case
for
neuromodulatory
system,
which
are
basically
broad
diffuse
system
in
the
brain.
So
you
have
cholinergic
neurogeneratic,
ceramic
and
dopaminergic,
the
specific
name
don't
really
matter,
but
what
you
need
to
take
away
from
that
is.
Those
systems
are,
depending
on
the
context
and
the
environment,
going
to
be
more
or
less
activated,
and
besides
those
plasticity
events
that
we
talked
about
previously,
those
additional
events
can
regulate
and
modify
the
plasticity
events
that
we
talked
about.
A
D
A
D
A
Not
as
closely
related,
I
think,
or
at
least
not
that
I'm
aware
of
because
historically
I
think
dopamine
in
our
sense
has
been
linked
with
reward
prediction
error
and
talked
a
lot
with
the
ml
crowd.
That
directly
saw
a
relevance
of
it.
I
don't
think
there's
been
a
specific
cross
link,
that's
been
made
for
the
other
systems.
D
B
B
Yeah,
I
I
know
I
mentioned
this
because
you
know
one
of
the
questions
I've
gotten
when
we
talk
about
modeling,
cortex
people,
they
ask
about
these
neuromodulatory
systems
and
I've
always
said
well,
they're
they're,
clearly
not
the
way.
We
remember
they're,
not
part
of
the
circus,
remembering
the
model
of
the
world
and
so
on.
B
They're
they're,
modulatory,
right
and
so
some
of
them
might
be
associated
with
you
know,
fear
or
someone
like
associated
with
you
know,
with
the
sex
or
romance,
and
things
like
that,
and
I
just
pointed
out
that,
from
my
perspective,
you
can
always
you
can
you
can
remove
these
systems
from
the
cortex's
information
processing
role
and
just
say?
Well,
we
have
switches,
we
can
speed
up
memory,
we
can
slow
down
the
formation
of
memory,
we
can
there's
various
sort
of
system,
level
controls
and
you
don't
have
to
emulate
these
systems
directly.
B
That's
been
my
argument
at
least,
and
some
of
them
might
not
have
any
role
at
all
in
terms
of
building
a
a
system
that
models
the
world
where
clearly
the
dopaminergic
system
is
one
that
involves,
rewards
and
effort
that
you
could
argue.
Well,
that's
a
good
learning
rule,
but
even
that
could
be
replaced
with
a
you
know,
a
switch.
This
thing
you
want
to
learn
now
and
not
learn
now
type
of
thing.
C
E
How
far
are
these
four
systems,
how
how
are
they
preserved
down
through
the
animal
kingdom
I
mean,
are
these?
Is
this
something
of
just
arbitrarily?
Is
it
something
the
mammalian
brain
has
all
four
of
these
things,
or
it
goes
down
to
reptilian,
or
does
anyone
know
whether
these
these
systems,
all
four
systems,
are
conserved
down
to
some
level
in
the
in
the
in
the
genetic
tree.
B
But
kevin
you
can
look
here.
The
source
of
these,
the
cells
that
actually
distribute
these
neurology
leaders
are
all
like
in
the
brain
stem
in
the
basal
ganglia.
You
can
see
that
in
the
entrance
right,
so
they're
these
and
they're
also
spreading
to
parts
of
the
old
brain
they're
going
to
the
cerebellum
and
various.
You
know
the
middle.
B
A
I
I
just
wanted
to
relate
this
kind
of
context-dependent
modulation
with
the
dendrite
artwork
we
recently
did
and
how
maybe
there
is
something
to
dig
here
about
how
context
or
different
context
can
modulate
the
same
input
on
on
the
same
neuron.
We
talked
a
little
bit
about
that
already,
but
I
think
it
was
and
it's
an
interesting
point
to
to
think
about.
D
A
Okay,
I'm
almost
done.
I
just
wanted
to
mention
this
briefly,
which
is
that
those
mechanisms
that
we
talked
about
are
not
set
in
tone
along
the
lifetime
of
an
animal
or
an
organism,
and,
for
example,
when
you
go
from
development
to
postnatal
development
is
gaba.
You
have
a
gaba
switch
where
gaba
comes
being
excitatory
in
the
umbrella
development
becomes
inhibitory
in
personnel
development
and
those
those
needs
to
be
taken
to
encounter.
B
Unless
you
wanted
to
grow
your
model
in
an
embryo.
I
I'm
not
sure
why
you
would
think
you'd
want
to
emulate
that
it
seems
like
that.
Would
you
could
just
eliminate
that,
so
you
know
we're
not
really
interested
in
how
brains
know
how
to
grow
in
the
in
the
womb,
but
assuming
they
have
this
structure
at
birth,
how
do
they
learn?
So
I
guess
I
push
back
on
that
idea
that
we
might
want
to
do
both.
A
B
Right,
but
why
pre-wire
it
the
way?
You
know
that
you
have
to
create
a
growing
brain
in
a
womb.
I
mean
it's
or
something
equivalent
to
it.
You
could
just
start
with
you
know
at
birth
and
say
here's
the
kind
of
pre-wiring
you
have
and
let's
go
from
there.
I
see
no
purpose
at
all
in
sort
of
emulating
or
modeling
or
simulating
or
anything.
B
I'll,
just
let's
leave
it
at
that.
I
I
guess
you
know,
you
know.
The
point
is
not
how
the
brain
you
know.
Look,
we
don't
have
to
figure
it.
We
have
to
grow
neurons
from
the
neural
plate
and
have
them
go
up
in
their
little
baby
columns
and
then
grow.
I
mean
we
don't
have
to
do
any
of
that
stuff.
That's
like
forget
it.
It's
you
come
out
alone.
You
got
a
brain,
it's
got
some
wiring
and.
A
A
A
B
A
D
Yeah
yeah,
this
stuff
is
definitely
not
taken
into
account.
I
would
say
the
only
thing
that
comes
close
is
the
reinforcement,
learning
things
where
you
do
control
we,
you
have
a
sense
of
kind
of
how
fast
things
should
learn,
depending
on
how
far
away
it
is
from
reward,
and
so
there
is
a
temporal
difference
or
time
time
component
there,
but
nothing
like
the
flexibility
of
the
stuff
that
you
talked
about.
D
Another
yeah,
I
mean
you
talked
quite
a
bit
about
homeostasis
as
well
in
different
places
and
that's
in
in
machine
learning,
there's
a
lot
of
batch
normalization
and
homeostasis
at
the
level
of
activity,
but
in
terms
of
weights,
some
people
have
done
work
on
weight
normalization.
But
you
know,
I
don't
think
that's
been
a
big.