►
Description
In part one of the presentation, visiting scientist Jeremy Forest gives a brief overview of the plasticity mechanisms in the brain. He goes over how neurons interact and change over time with different plasticity events in dendritic spines, and covers topics such as molecular mechanisms, synaptic activations, and structural plasticity.
Throughout the presentation, Jeremy points out the biological mechanisms, such as time-scale interactions, that could potentially be modeled in AI systems and bring many benefits, such as continual learning and efficiency.
Jeremy talks about the dynamic neuronal activities that impact plasticity mechanisms in part two.
Part Two: https://youtu.be/rLpb3J4AHXE
- - - - -
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
A
But
what
I'm
going
to
talk
about,
and
hopefully
we
can
discuss,
don't
hesitate
to
like
interrupt
and
ask
questions
anything
like
that,
because
it
would
be
more
fun
that
way.
But
it's
just
it's
food
for
thought
right.
It's
not
it's
not
the
complete
set
of
mechanisms,
it's
not
the
full
overview
of
what's
happening
in
the
brain
at
all
time.
A
It's
it's
still
simplified,
and
but
hopefully
it's
enough
for
for
for
you
to
to
to
think
about
some
stuff
that
might
be
missing
from
from
ai
systems
and
some
stuff
to
dig
around
and
again
right,
so
learning
can
be
studied
at
different
levels.
You
have
behavior,
you
have
brain
networks,
you
have
specific
brain
regions,
you
have
the
neural
novel
and
then
you
can
do
protein
and
molecular
pathways.
So
I'm
not
going
to
talk
about
all
of
that.
I'm
going
to
talk
a
little
bit
about
some
of
that.
A
A
If
we
go
back
to
a
while
ago,
that's
from
the
seventies,
the
classical
experiments
on
on
long-term
dissociation
and
long-term
depression
and
spiked
spectrum
independent
plasticity,
you
can
see
on
the
left
here
that
after
tetanic
burst
stimulation,
which
is
basically
a
high
frequency
stimulation
of
a
neuron.
You
have
four,
while
an
increase
in
the
amplitude
of
how
the
neurons
respond
to
stimulation.
A
So
that's
derived
from
hebb
from
blitz
and
lomo's
work
and
on
the
right
here.
You
can
see
that
classical
figure
of
stdp,
where,
if
the
past
synaptics
response
comes
from
be
before
the
presynaptic
one,
you
have
a
depression
and
if
you
have
the
rebirth,
then
you
have
a
potentiation
of
the
past
synaptic
potential.
A
Okay,
sorry
about
that
yeah!
Please,
last
question:
if
you,
if
you
feel
like
I'm
familiar
with
some
of
the
dragon,
I
might
have
missed
some
stuff.
So
potentiation
and
depression
are
two
opposite
way.
A
synapse
can
respond
to
a
stimulation.
If
you
have
on
the
left
here
a
depression,
it
means
that
after
that
event
occurs,
the
synapse
will
respond
less
to
the
same
type
of
stimulation,
meaning
the
weight
is
going
to
be
less.
B
C
B
D
C
In
our
theories,
there's
a
very,
very
simple
explanation
for
this:
if
you
have
some
context
that
is
making
a
prediction:
if,
if
the
activity
arrives
before
the
context,
well,
the
context
couldn't
have
predicted
the
input
right.
If
the
context
is
before
the
thing
that
happens,
then
it
could
be
predicted.
So
you
can
look
at
this
very
simply
saying
some
event
occurred.
C
D
A
In
that
case,
you
you're
gonna,
it's
it's
been
done
in
the
hypocampus
and
so
you're
going
to
stimulate
the
afferent
nerve.
So
the
nerve
coming
to
the
hippocampus
you're,
going
to
stimulate
that
very
with
a
very
high
frequency
and
then
you're
going
to
look
at
how
the
cells,
the
power
synaptic
cell,
respond
to
that
stimulation.
So.
A
C
A
A
A
A
A
What
that
means
is,
if
you
have
some
stimulation,
it
can
result
in
what's
called
early
ltp.
In
that
case,
you
have
one
train
of
stimulation,
let's
think
of
it
as
a
simple
stimulation,
and
if
you
have
four
trains,
let's
think
of
it
as
multiple
stimulation,
then
you
it's
going
to
result
in
what's
called
late
ltp
and
the
major
difference
between
those
two
is
the
amount
of
time
the
modification
stays
early
rtp
here
after
two
and
a
half
hours
you
you,
you
go
back
to
the
basal
state
of
the
synapse.
It
doesn't
last
longer
than
that.
A
Late
ltp,
you
see
here
after
more
than
three
hours
is
still
there
right.
The
question
is
why
and
then
to
know
why
you
need
to
go
down
one
level
from
the
neuronal
level,
the
synapse
level,
to
the
molecular
level
and
that's
what's
on
the
right,
the
details
themselves,
don't
really
matter.
But
what
you
need
to
understand
is
that
for
early
ltp,
it's
going
to
recruit
molecular
pathways
and
changes
that
stay
mostly
at
the
level
of
the
synapse.
B
E
D
C
C
None
of
that's
really
important
what's
important
here
is,
I
think
so
you
said
it
earlier
since
I
said
it,
one
is
just
a
temporary
change
to
the
synapse.
The
other
invokes
what
it's
often
called
meta
metabotropic
retrain.
It's
basically
invokes
a
metabolism,
complex
chemical
reaction
where
it
produces
proteins
and
and,
as
jeremy
said,
one
of
the
things
that
can
happen.
It
can
generate
new
gates
in
the
cell
membrane.
So
imagine
you
have
these
modules.
These
molecules
have
to
come
across
the
gap
there
between
the
synapse
and
they
have
to
find
a
receptor.
If
you.
C
Then
more
can
get
in
so
one
is
like
a
growth
of
physical
change,
physical
change
to
the
cell
and
a
synapse,
and
the
other
is
just
a
chemical
change.
It
happens
briefly,
so
that's
the
difference
and
we've
often
you're
using
different
languages.
We've
used
here
in
the
past,
like
the
language
of
merge
sherman.
C
This
would
be
a
burst
versus
a
single
spike
right,
so
you
might
have
a
burst
of
activity
which
could
create
this
bit
of
a
tropic
effect,
and
then
I
think,
you're
talking
about
this,
germany
and
or
single
spike,
would
have
a
very
temporary
chemical
effect
in
some
sense.
Is
that
consistent?
What
you're
thinking
in
german
yeah.
C
Is
what
he
described
is
the
metabolic
effect,
which
is
you
invoke
a
metabol,
a
metabolic,
sweat,
word
pathway,
which
essentially
says
oh
turn
on
these
genes,
generate
some
proteins
build
some
new
receptors
change,
the
physical
structure,
and
it's
not
just
it's
not
just
new
receptors
at
the
synapse.
It
also
creates
new
receptors
adjacent
to
the
synapse.
B
A
B
E
B
B
D
A
So
if
I
go
a
little
bit
into
more
details,
if
you
have
an
incoming
action
potential
from
an
axon
that
synapse
onto
a
dendrite,
if
you
deploy
the
axon
but
doesn't
do
much
on
the
level
of
the
postsynaptic
dendrite,
then
you're
going
to
have
depression,
and
if
it
generates
your
other
neuron,
generate
an
action
potential.
Then
you're
going
to
have
fire
the
first
fire
of
the
second
and
that's
going
to
reinforce
the
synapse.
D
D
A
I
yeah
I'm
I'm
not
positive,
I'm
confident
to
say
with
100
confidence
that
it's
the
case.
The
thing
is
that
it's
one
of
the
mechanisms-
and
there
are
other
mechanisms
on
shorter
or
longer
time,
scales
that
will
interact,
and
so,
for
example,
if
this,
if
a
spine,
which
means
usually
about
spines
for
for
those
type
of
modification,
if
the
spine,
that's
just
on
the
side,
might
receive
stimulation.
C
A
A
Because
you
grow
a
new
spine,
you
grow
new
connections.
You
insert
new
receptors
so,
for
example,
the
empire
receptors.
Here
you
can
insert
more
more
of
than
at
the
synapse.
If
you
have
more
than
when
you
have
incoming
neurotransmitters,
they
interact
with
that
receptor.
More
and
you're
gonna
have
more
channel
open,
so
the
depolarization
is
gonna,
be
more
important.
B
E
A
B
C
I
don't
want
to
get
if
you're
going
to
get
to
it,
jeremy,
but
one
of
the
biggest
disconnects
between
neurons
and
artificial
neural
networks
is
the
fidelity
and
the
dynamic
range
of
synapse
right.
The
dynamic
range
of
a
synapse
is
really
low,
and-
and
it's
and
it's
very
you
know
it's
just
very
low,
and
so
you
couldn't
represent
one
or
two
or
three
digits
of
accuracy
in
a
synapse.
It's
just
not
even
possible
they're,
just
noisy
things
that
kind
of
work
and
don't
work.
You
know
yes,
there's
a
chain
but
not
much.
A
C
C
A
C
No,
I
understand
that
too,
but
in
the
end,
what
are
the
synapses?
I'm
not
saying
deep
learning
can't
be
israeli,
I'm
just
saying
in
the
end,
if
the
system
requires
super
tight,
pointing
out
like
oh
yeah,
they
don't
model
these
two
types
of
the
short
term,
long
term
plasticity.
But
another
big
shortcoming.
Is
they
don't
model?
You
know
the
precision
that's
required.
Even
if
the
times
you
add
noise,
the
precision
that
you
end
up
using
is
this
non-existent
right
anywhere.
E
E
E
A
In
the
case
of
one
synapse,
you
rarely
gonna
have
one
synapse
doing
all
the
work.
That's
going
to
be
the
exception.
Usually
you
need
multiple
sign-ups,
so
the
difference
is
going
to
be
if
let's
take
free
sign-ups,
if
you
need
free
synapse
to
activate
around
the
same
time
to
to
create
deploration
sufficient
to
create
an
action
potential,
then
if
you
have
a
potentiation
of
one
of
those
signups,
you
might
need
only
two
synapse
activated
at
the
same
time.
A
A
A
It's
the
level
of
the
synapse
is
going
to
be
not
much
different
in
terms
of
timing,
but
in
terms
of
strengths,
and
then
you
go
through
all
the
summation
from
the
different
dendrites
at
the
level
of
soma,
and
then
you
fire
an
action
potential.
The
action
potential
might
be
a
little
bit
earlier
or
later,
depending
on
the
the
stimulation
from
the
dendrite.
B
Go
ahead
again,
sorry
relating
back
to
deep
learning
and
artwork
in
particular.
The
timing
stuff
is
something
we
sort
of
implicitly
assume
is
happening
and
that's
why
kind
of
this
stronger
neurons
will
win
out
in
a
if
there's,
a
local
inhibitory
system,
the
stronger
neurons
will
win
out
and
the
weaker
ones
will
not,
and
this
is
one
way
and
of
course
we
do
it
much
more
so
with
active
dendrites.
C
Of
the
neuron
in
the
temple
memory
and
the
context
where
we
we
model
all
the
time
that
relies
on
the
timing
of
the
output
right,
but
that's
unrelated
to
the
timing
of
the
synapses
it's
much
more
complicated
than
that
right.
That's
the
set
of
synapses
on
the
dendritic
branch,
which
causes
the
dendritic
spike,
which
then
equalizes
the
same
cell
body,
which
then,
when
you
have
you
know
it's
much
far
removed.
You
can
think
of
the
synapse,
see
those
little
channels
on
the
synapses
there's
an
ample
one,
there's
an
nmda
one!
C
Those
are
like
little
hoses,
letting
letting
you
know
current
come
in.
If
you
think
of
it.
That
way,
and
if
you
have
more
of
them,
then
the
current
can
come
in
faster,
but
it
doesn't
arrive
any
faster,
the
synapse
it
could
be
polar.
It
could
be
a
ramping
up
with
the
polarization
faster,
but
there's
no
timing
difference
in
that
in
that
synapse.
It's
it's!
It's
almost
instantaneous.
C
C
B
E
C
B
A
A
lot
of
complications
like
I
said
it,
I'm
I
was
trying
to
keep
a
high
level
overview
without
skipping
too
much.
But
if
there's
specific
points
you
want
to
talk
about
later
on,
like
tell
me-
and
I
can
do
like
specific
detail
point
on
on
very
specific
stuff-
okay,
so
ltp
is,
is
one
thing
it.
It
lasts
for
more
or
less
time,
but
it's
still
quite
a
bit
here.
We're
talking
minutes
right,
but
then
you
also
have
mechanisms
that
last
milliseconds
up
to
a
couple
of
seconds
top
top.
A
A
Here
we
we're
talking
about
stimulation
and
measures
over
100,
200
milliseconds,
so
on
a
very
short
time
scale,
you
can
have
very
short
adaptation
of
synapse
that
have
decreasing,
increasing
or
more
variable
responses
right,
so
repeated
stimulation
can
either
increase
or
weaken
the
synaptic
connection
on
a
very,
very
short
time
scale,
and
it's
gonna,
for
example.
In
that
context,
change
depending
on
the
brain
area.
You
study.
C
D
B
C
Because
the
nervous
system
is
very
simple
and
big
right,
in
fact,
its
nervous
system
is
completely
determined
by
genetics.
I
believe
even
the
synapses
like
many
of
them.
Many
of
them
are
determined
genetically
meaning,
all
the
brains
look
the
same,
and
and
so
it's
not
a
brain
that
rewires
much.
It
is
just
like
this
thing
that
it
is
and
early
nervous
systems
like
this
all
seem
to
be
that
way:
they're
very
genetically
determined
and
learning.
C
C
So
it's
not
to
say
that
this
stuff
doesn't
happen
in
the
cortex,
but
when
you
look
at
a
policy,
you've
got
to
realize
this
is
the
only
way
it
looks
by
modifying
synapses
where
in
the
human
brain,
we
believe
in
the
cortex.
Most
of
learning
occurs
by
the
formation
and
removal
of
synapses,
and
so
we
just
got
to
keep
that
in
mind.
A
Yeah,
I
think
I'm
gonna
talk
to
that
to
the
next
slide.
I
was
talking
about
more
short-lived
mechanisms,
so
you
have
the,
I
would
say
intermediate
mechanisms
with
ltp
ltd.
I
was
going
to
talk
here
about
the
the
very
much
shorter
duration
mechanism
that
happens
on
extremely
short
time
scale
and
then
there's
the
structural
changes
that
happen
on
on
longer
time.
Scale.
A
C
I've
talked
to
I've
talked
to
what's
his
name
like
I
do
a
visit
like
kendall,
and
he
wrote
a
letter
from
my
first
book
and
it
was
amazing
how
he
just
dismissed
completely
all
the
sort
of
synaptic
formation
in
the
cortex.
His
world
existed
completely
implicit
and
he
just
didn't
even
think
about.
There
might
be
other
ways
of
doing
his
eyes.
Well,
the
police
just
showed
you
everything.
C
At
least
you
can
acknowledge
the
fact
that,
oh
great,
you
know
it's
just
I
just
you
know,
this
is
a
sore
point
where
you
can
tell,
because
people
just
don't
think
holistically
about
this
stuff
enough.
You
know
they
don't
question
these
things,
though
so
anyway,
so
we
just
have
to
you,
have
to
sort
through
all
this
stuff
that
what
people
believe
and
you
have
to
just
look
at
it.
Clearly
I
don't
know
okay.
A
Okay,
the
main
takeaway
from
this
slide
I
didn't
want
to
get
too
much
into
detail-
is
that
some
of
the
mechanisms
plastic
mechanism
can
happen
on
different
time
scales.
So
you
can
have
very,
very
quick
adaptation,
whether
it's
increase
or
weaken,
synaptic
transmission
or
connection,
and
the
slide
I
showed
before
is
you
can
have
mechanisms
that
last
longer
in
time
right
and
that's
something
that
I
don't
think
I've
seen
really
in
in
deep
networks.
B
A
Okay,
so
so
next
we
we
get,
we
get
into
structural
plasticity,
right
and
functional
and
structural
plasticity.
So
when
you
have
modification
at
the
level
of
the
synapse,
one
problem
is
you're
going
to
have
dna
modification
that
are
going
to
be
be
read
by
the
cell
machinery.
That's
going
to
create
new
protein
right,
but
the
proteins
they're
going
to
be
created
at
the
level
of
the
soma
and
then
they
need
to
be
transported
to
the
spine.
That
needs
to
be
changed
right.
C
The
spine
is,
it
is
the
synapse,
it
is
the
neck
of
the
synapse,
so
synapse
is
not
excitatory.
Synapses
are
not
on
the
surface
of
the
dendrite
they're
on
these
little
protuberances
that
come
out
called
the
spine
and
the
size,
and
the
structure
of
the
spine
has
a
lot
to
do
with
the
permanence
and
set
efficacy
of
the
synapse.
A
Okay,
so
one
of
the
idea
is
that
there
is
a
way,
a
biological
way
of
tagging.
The
signups
that
are
have
been
active
at
a
certain
point
in
time,
and
then
what
that
means
is
those
new
protein
that
are
being
generated,
those
new
channels,
those
new
neurotransmitters.
They
are
going
to
know
where
to
go
because
there's
a
tag
on
the
specific
sign
up
that
they
need
to
go
to
right.
That's
basically,
I'm
not
gonna
go
into
too
much
details
on
this
slide,
but
that's
basically
what
it
means.
A
A
Another
possibility
is
also
that
neural
activity
release
neurotransmitters
right
to
communicate
between
two
neurons,
but
you
also
release
other
factors,
all
the
molecules
that
will
have
other
effects
and
don't
are
not
necessarily
involved
in
communication
per
se.
So
one
of
those
is,
for
example,
bdnf,
doesn't
really
matter
what
it
means.
Doesn't
matter
really
matter
what
it
is.
The
point
is:
when
you
see
here,
b
on
the
bottom
and
active
synapse
besides,
releasing
neurotransmitters
that
will
communicate.
C
A
A
C
C
There
are,
there
would
be
a
lot
of
inactive
or
silent
synapses.
There's
some
like
synapses
are
there,
but
they
haven't
been
connected
yet
and
they're
expensive,
because
you've
got
you've,
got
this
stuff
they're
sitting
around
consuming
space
and
energy,
but
they're
not
doing
anything,
but
but
you
can
turn
a
silent
synapse
to
an
after
synapse
very
rapidly.
C
B
C
D
A
I'm
actually
not
100
sure,
but
my
guess
is
that
the
molecules
on
the
silent
synapses
are
not
all
pre
all
there.
So,
for
example,
the
channels
that
needs
to
open
for
signal
transmission
are
not
necessarily
there
or
if
they're
there,
there's
not
enough
to
cause
any
any
changes
in
the
synaptic
potential,
but.
D
B
A
A
So
so,
let's
go
back
to
the
second
slide
here.
These
empire
receptors
here
that
that
you
see
at
the
membrane
in
reality,
there's
like
a
bunch
of
them
right.
But
the
thing
that
here
is
not
shown
is
there's
also
a
bunch
of
them
that
are
not
in
the
membrane,
but
that
are
at
the
synapse
and
those
are
reserve
channels
and
those
can
very
quickly
be
inserted
into
the
migraine
and
be
usable.
A
C
You
can
still
think
of
that.
The
overall
learning
mechanism
can
still
be
heavy
and
learning
it's
still
the
same
thing
right,
it's
still
like
hey.
I
got
an
incrimination
and
I
have
a
post-synaptic
event
and
I
want
to
not
form
a
connection
between
them,
and
so
this
just
happens
quickly.
It's
it's
as
opposed
to
elsewhere
in
the
brain.
D
C
From
our
theory's
point
of
view,
we
didn't
really
care
right.
This
doesn't
make
any
difference
to
our
theory,
because
they
would
just
say
you
need
to
be
able
to
go
from
a
zero
synapse
to
a
one
set
up
and
also
in
our
software
simulations
or
whatever.
Of
course
you
can,
you
can
create
new
synapses
really
nearly
as
fast
as
you
want
right.
It's
just
a
another
point
in
the
matrix
right.
So
but
biology
is
constrained
by
this
a
bunch
of
other
constraints,
so
we
modeled
our
sequence
memory.
We
didn't.
C
C
A
A
C
So
so
we,
so
we
basically,
but
we
are
our
mechanism,
for
we
knew
that
we
wanted
to
learn
something
and
learn
it
quickly
and
the
mechanism
we
came
up
with
learning
involved.
Creating
the
stamps
is
so
there
was
a
disconnect
and
like
hey.
How
could
this
be
biologically
possible?
So
so
then
I
had
to
conclude
that
there
must
be.
There
must
be
something
like
assignments
apps,
because
there
must
be
a
way
of
creating
synapses
very,
very
effectively,
creating
symptoms
very,
very
fast
and
and
that
led
to
the
sort
of
idea.
C
Well,
there
must
be
something
like
assignment
synapses.
That's
there
there's
a
connection
but
not
effective,
and
then
I
found
that
that's
true
and
then,
of
course,
then
we
speculated
I
said.
Well,
there
must
be
a
lot
of
them
in
the
hippocampus,
because
that's
where
we
know
best,
and
then
I
talked
to
some
researchers
about
that,
they
said
yes,
so
these
these
primal
cells
and
hippocampus
have
tons
of
them
so
that
all
fit
together
nicely
from
the
medical
point
of
view.