►
From YouTube: How Active Dendrites Enable Prediction and Context Integration in Neurons - August 18, 2021
Description
Subutai Ahmad reviews the biology behind active dendrites and explains how Numenta models them. He first presents an overview of active dendrites in pyramidal neurons by describing various experimental findings. He describes the impact of dendrites on the computation performed by neurons, and some of the learning (plasticity) rules that have been discovered. He shows how all this forms the substrate for the HTM neuron, proposing that dendritic computation is the basis for prediction and very flexible context integration in neural networks.
Papers:
Bartlett Mel, Neural Computation 1992: https://direct.mit.edu/neco/article/4/4/502/5650/NMDA-Based-Pattern-Discrimination-in-a-Modeled
Poirazi, Brannon & Mel, Neuron, 2003: https://pubmed.ncbi.nlm.nih.gov/12670427/
Numenta Neurons Paper 2016: https://www.frontiersin.org/articles/10.3389/fncir.2016.00023/full
Numenta Columns Paper 2017: https://www.frontiersin.org/articles/10.3389/fncir.2017.00081/full
“Predictive Coding of Novel versus Familiar Stimuli in the Primary Visual Cortex”: https://www.biorxiv.org/content/10.1101/197608v1
“Continuous online sequence learning with an unsupervised neural network model”: https://direct.mit.edu/neco/article/28/11/2474/8502/Continuous-Online-Sequence-Learning-with-an#.WC4U8TKZMUE
‘Unsupervised real-time anomaly detection for streaming data”: https://www.sciencedirect.com/science/article/pii/S0925231217309864
“Active properties of neocortical pyramidal neuron dendrites”: https://pubmed.ncbi.nlm.nih.gov/23841837/
- - - - -
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
I
think
that's
like
another
meeting
and
karen
has
talked
about
in
previous
meetings
as
well,
but
this
is
sort
of
the
basis
for
some
of
it
and
I
I
should
apologize
in
advance.
I
just
slapped
this
together
this
morning
from
a
bunch
of
existing
slides,
so
I
may
have
missed
stuff
out
and
think
so,
but
anyway
you
got
a
paddle
slide.
B
C
A
This
is
from
also
from
another
talk
but
anyway,
and
the
title
kind
of
really
tries
to
say
what
what
we
think
is
going
on
with
dendrites
and
neurons.
So
I'm
calling
this
the
predictive
neuron,
the
neuron
is
predicting
stuff,
and
this
is
really
based
on
active
dendrites
and
I'm
going
to
try
to
show
how
active
dendrites
lead
to
prediction
and
context
integration
in
neurons
and
that
this
is
a
really
powerful
concept.
That
can
be
it's
a
very,
very
flexible
concept
that
can
be
used
in
many
many
different
ways.
A
So,
but
primarily
this
is
about
kind
of
the
biological
basis
for
this
and
a
little
bit
of
how
we
modeled
it
in
html,
okay,
so
neurons.
This
gets
to
the
question.
Charlene
was
just
talking
about,
so
this
is
the
typical
point,
neuron
model,
so
you
have
a
you:
have
a
bunch
of
inputs
coming
in.
You
take
a
weight
at
some
of
them
and
you
pass
it
through
some
non-linearity,
and
the
idea
is
that
this
happens
at
the
soma.
A
This
was,
I
think,
originally
proposed
back
in
1907,
and
it
is
still
the
current
by
far
the
current
model
of
how
neurons
works
in
in
computational
work
and
deep
learning,
so
in
the
perceptron
in
1962.
I
think
this
is
like
the
first
actual
computational
model
that
was
done
of
neurons.
They
used
this
and
then
in
deep
learning
from
1986
through
2021.
A
This
is
still
by
far
the
most
common
idea,
but
you
know,
as
we
know,
this
is
not
really
what
a
real
neuron
is
it's
much
more
complex
than
this.
This
is
not
a
neuron,
and
so
some
of
you
may
have
seen
this
video
I've
shown
this
before.
I
really
like
this
video.
This
shows
what
a
real
pyramidal
neuron
looks
like
and
you
can
show.
A
This
is
a
3d
reconstruction
and
you
can
show
that
there's
a
really
complex,
tree-like
structure
around
it
and
these
these
are
the
dendrites
of
the
neuron,
and
this
is
where
inputs
are
received
into
the
neuron
and
you
can.
You
should
be
able
to
tell
here
that
there's
just
a
lot
of
structural
complexity
and
there's
actually
a
lot
of
functional
complexity
in
here
as
well,
and
so
first
thing
is
that
the
entire
neuron
has
thousands
of
synapses,
and
if
you
look
at
the
the
point,
neuron
piece
of
it,
that
only
really
covers
a
few
hundred
synapses.
D
A
A
D
A
There
was,
as
far
as
I
know,
very
little
and
maybe
not
no
experimental
evidence
for
it,
but
he
did
a
detailed
compartmental
model
of
a
neuron
and
basically
suggested
that
what's
going
on,
is
that
these
these
neurons
are
detecting
these
small
patterns
all
over
the
dendritic
arbor-
and
this
was
this
paper-
had
a
huge
impact.
It
caused
a
lot
of
experimental
work
and
they
actually
verified
that
this
actually
does
happen
in
neurons.
A
B
Was
already
a
lot
known
about
modeling
of
dendrites
there's
these
they're
called
the
cable
models,
or
something
like
that,
and
so
there's
a
lot
of
people
trying
to
speculate
what's
going
on,
because
these
are
these
dendrites
are
non-linear
they're
not
just
like
collecting
everything
and
adding
it
together.
There's
a
lot
of
weird
dynamics
is
going
on,
and
so
he,
the
introduction
here,
was
like
saying:
hey:
it's
not
just
these.
These
cabling
properties
there's
a
it
wasn't
in
a
total
out
of
the
out
of
the
blue,
but
it
was
a
big
advance.
Yeah.
A
E
A
They
actually
verify
the
stuff
in
biology
and
so
from
the
biological
standpoint.
What's
going
on
with
this
neuron?
Is
you
have
this
complex?
You
know
dendritic
tree
in
the
very,
very
close
to
the
cell
body.
The
synapses
are
essentially
doing
the
point.
Neuron
thing
that
they're
taking
a
weighted
sum.
It's
a
linear
summation
of
the
inputs
and
if
you
get
enough
inputs,
the
cell
will
fire.
So
this
is
kind
of
the
classic
point
neuron
model,
because.
B
A
Yeah
and
most
of
this,
the
vast
majority
of
the
synapses
are
on
these.
You
know
more
distal
or
distal
means
further
away
from
the
cell
body,
so
more
distal
segments-
and
these
are
the
ones
that
are
pattern
detectors
out
there.
That's
where
you
know
here,
I'm
saying
eight
to
15
co-active
co-located
synapses
can
generate
dendritic
and
nba
spikes
and
I'll
show
that
in
a
second.
B
B
A
But
here's
an
example
of
the
type
of
experimental
evidence
you
see,
there's
tons
of
it
now,
since
the
since
part
of
this
paper
here
we're
looking
at
a
tiny
segment
of
the
dendrites
and
each
dot
here
represents
some
synapse
that
they're
actually
activating
experimentally
and
what
spines
is
just
you
can
just
think
of
it
as
a
synapse
here
and
if
you
can
see
this
non-linearity
that
if
you
activate
a
small
number
of
synapses
or
weights
simultaneously,
you
get
some
impact.
But
if
you
beyond
a
certain
point,
you
get
this
very
large
impact.
A
A
Yeah,
this
is
about
a
five
millivolt
difference,
so
it's
still
relatively
small.
I
think
you
need,
like
I
don't
know
thirty
forty
fifty
miles
to
actually
generate
an
action
potential.
So
there's
tons
of
evidence
for
this
now,
but
the
point
neuron
again
is
just
looking
at
you
know
this
less
than
ten
percent
of
the
synapses.
Let's
say
it's
like
two
three
percent
of
the
synapses
so
like,
like
all
of
the
computational
modeling,
that's
going
on
in
about
the
neocortex
that
basically
uses
point.
A
Neuron
models
only
looks
at
like
two
percent
of
the
synapses,
it's
ignoring
like
99
or
98
of
the
synapses
of
that
in
the
neocortex,
which
is
remarkable.
It's
a
huge.
It's
like
a
black
hole
like
you're
right
or
something
there's
on
our
sides.
That's
right!
There
are
very
few
people
that
really
modeled
this
stuff
there's
a
little
more
recently,
but
very
comfortable.
B
Basically,
it's
lost
it
leakages
along
the
dendrites.
You
know
leaky
pipes
if
you
will-
and
there
also
there's
a
lot
of
a
loss
at
the
junction
of
the
synapses
split.
So
there's
a
impedance
match
between
you
know
two
dendrite
branches
coming
together,
yeah
they're
like
leaking.
You
know
it's
like
leaky
pipes,
so
you
put
a
little
pressure
in
at
the
end
of
the
pipe.
That's
all
that
can
happen.
A
Yeah
and
if
you
were
to
activate
just
a
single
one,
you
would
like
not
even
notice
anything
it'd
be
just
like
this
little
thing
here.
Yeah
you'll
see
a
tiny
one,
but
if
you
get
a
cluster
of
these
things
activating
simultaneously,
which
is
what
you
get
with
sparse
codes
and
sparse
representations,
a
bunch
of
them
coming
together
like
eight
or
ten
of
them,
then
you
get
this
non-linear
event
that
happens.
That
generates
this
potential
increase
in
the
potential.
A
A
little
bit
more
in
a
second,
but
this
is
how
bartlett
and
yoda
koirazi
modeled
it.
They
basically
looked
at
this
and
said:
okay,
each
dendritic
segment
is
doing
a
linear
weighted
sum
followed
by
a
sigmoid
and
then
the
sum
the
cell
body
is
doing
summing,
all
these
up
and
doing
another
sigmoid,
and
so
basically
it's
just
like
a
two-layer
model.
So
they
said
it's.
You
can
look
at
a
neuron
and
say
it's
basically,
a
two
layer
receptor
on
the
two
layer,
deep
learning
system
and
it's
equivalent.
A
That
the
neuron
is
now
more
complex
than
you
might
have
thought,
but
from
a
machine
learning
standpoint,
there's
there's
not
much
use
for
this
there's.
You
can't
really
do
much
with
this.
It's
I
mean
not
not
that
you
can't
do
much
with
it.
It's
you
just
have
a
slightly
bigger
network
and
you
get
all
of
the
properties.
So
there's
no
reason
to
model.
E
A
So
yeah
and
the
computational
neuroscience
community
has
a
huge
impact
on
and
the
biological
the
neuroscience
community
in
general.
It
changes
your
notion
of
what
a
neuron
is,
but
from
a
machine
learning
standpoint
it's
just
a
slightly
bigger
network.
It
just
doesn't
really
change.
Why
go
through
the
complexity
of
modeling?
All
this
when
you
could
just
have
a
slightly
bigger
network,
it's,
and
so
that's
one
of
the
reasons.
I
think
this
is
not
really
caught
on
the
deep
learning
community
anyway,
but
they
had
a
huge
impact.
That's
not
true.
A
Yeah,
and
so
that
model
I
think,
misses
some
of
the
some
other
big
properties
of
it,
and
one
of
the
big
things
is
this
sort
of
plateau
potential
and
I've
taken
this
diagram
from
a
more
recent
paper.
But
this
was
one
of
the
things
that
were
behind
the
neuron
model
that
we
published
I'll
show
that
in
a
bit,
but
this
biological
result
is,
is
the
following
that
if
you
have
a
clustered
set
of
inputs
in
these
distal
dendritic
segments,
the
impact
on
the
cell
body.
A
Is
you
get
this
very,
very
long,
sustained,
depolarization
or
sustained
state
cell
body?
So
this
diagram
shows
that
you
know.
If
you
have
these
clustered
segments,
you
get
an
nda
spike
or-
and
this
you
get
this
big
long,
depolarization
at
the
cell
body,
where
it's
it's
sort
of
like
in
an
upstate,
it's
depolarized
by
you
know,
somewhere
five
here
they're
showing
the
20
millivolt
thing.
It's
could
be
somewhere
around
five
to
twenty
millivolts
and
this
lasts
for
hundreds
of
milliseconds.
A
So
as
long
as
like
half
a
second-
and
this
basically
allows
pushes
the
cell
body
closer
towards
having
an
action
potential.
So
now,
when
you
have
inputs
coming
in
they're,
going
to
respond
faster
and
they're
going
to
respond
more
strongly
than
you
would
otherwise
okay.
So
this
is
this
sustained
depolarization
plateau
potential?
Those
kind
of
terms
that
you
might
see
in
some
of
the
work
we've
done.
A
So
the
basic
idea
here
is
that
clustered
input
on
dendrites
leads
to
this
sustained
depolarization
exoma,
it
modulates
the
neuron.
It
puts
it
into
this
upstate,
which
makes
the
neuron
more
likely
to
fire.
So
if
you
have
a
set
of
inputs
that
normally
would
not
make
it
fire
and
you're
saying
that
cut
it
here,
not
exactly,
but
it's
similar.
The
same
same
set
of
inputs
now
will
make
the
neuron
fire
and
the
neurons
in
the
upstate
are
respond
faster
to
inputs
than
they
would
otherwise,
and
these
are
more
likely
to
win.
B
A
E
B
G
A
Those
things
weren't
known,
but
I
have
this
diagram
from
this-
is
some
we
had
a
visit
from
a
neuroscientist
called
wayne
and
son
who's
really
into
this
stuff,
and
he
got
very
excited
about
our
theory.
I
don't
know
if
you
remember
this,
but
after
visiting
with
us,
he
went
back
to
his
lab
and
actually
tested
this
out.
A
B
A
Yeah
he's
a
very
energetic,
really
fun
person
to
talk
to
he's
very
knowledgeable
about
this
stuff.
He
came
to
our
lab
and
then
talked
about.
It
then
went
back
to
his
lab
and
tested
this
idea
out
that,
and
basically
what
this
shows
here
is.
He
was
doing
this
on
real
neurons
from
rad
barrel
cortex,
and
but
this
was
the
slices
you
can
control
the
timings
of
these
pulses
very
precisely
so
the
red
this
blue.
A
I
think
now
this
black
line
here
shows
what
would
happen
normally
like
you
have
no
plateau
potential,
and
then
you
stimulate
the
cell
and
it-
and
you
know
the
stimulation
is
integrated
into
the
cell
body
and
then
eventually
you
get
an
action
potential
and
with
the
exact
same
now,
if
you,
if
you
depolarize
the
cell
body,
to
simulate
a
plateau
potential
and
do
the
exact
same
stimulation,
there's
a
five
millisecond
difference
in
these
experiments.
So
despite
the
neuron
spikes
five
milliseconds
earlier
than
it
would
otherwise.
So
this
was
very
reliable
in
this
experiment.
A
B
A
E
B
A
It
is
possible
that
if
you
get
a
bunch
of
these
clustered
things
and
set
different
parts
of
the
doing
your
garbage
simultaneously
again,
which
will
probably
be
very
unlikely-
you
get
multiple
nmda
spikes
neuroscientists,
that
I've
spoken
to
think
that
it's
possible
to
cause
an
action
potential,
but
they
don't
no
one's
kind
of
seen
that
it's
an
important
question
in
some
sense,
because
we've
talked
about
this
idea
of
these
active
predictions
where
maybe
context
alone
can
cause
things
to
fire
so
yeah.
But
we
haven't
really
directed
all
of
that.
A
A
A
A
Very
similar
to
kind
of
ideas
of
predictive
coding
and
other
things
like
that
is
that
they
know
cortex
is
constantly
predicting
its
input,
and
I
guess
you
found
this
quote
from
1951
lashley,
who
was
hebb's
supervisor.
I
think
his
thesis
supervisor
saying
prediction
is
the
most
important
and
the
most
neglected
problem
of
serious
physiology,
and
so
basically,
what
we
propose
in
that
paper
is
that
neurons
are
actually
constantly
predicting
their
inputs
based
on
these
contextual
ideas.
A
So
this
so
this
led
to
this
idea
of
the
htm
neuron
model,
and
this
diagram
on
the
left
shows
how
we've
modeled
the
complexity
of
a
pyramidal
neuron
here.
So
the
idea
is:
there's
this
proximal
segment,
which
is
the
typical
inputs
which
are
in
this
linear
weighted
sum.
Then,
if
you
have
a
bunch
of
independent
distal,
dendritic
segments
that
are
getting
contextual
input
and
each
one
of
those
you
know
you
get
a
sum
of
them
followed
by
a
threshold
or
a
non-linearity,
and
these
would
depolarize
the
cell
body
and
then
in
the
apical
areas.
A
Here
we
have
another
set
of
contextual
signals,
feedback
signals
or
other
signals
that
also
impact
the
cell
body.
So
I'm
not
actually
going
to
talk
too
much
about
this.
We
can
do
that
at
a
later
meeting.
There's
separate
kind
of
interesting
phenomena
that
happen
with
the
apocalyptics,
but
mostly
these
are
all
exotropic.
A
So
the
basic
idea
is
that
these
proximal
synapses
that
we
usually
show
in
green
on
this.
You
know
feed
forward
inputs,
they're
the
ones
that
actually
cause
the
spikes.
They
define
the
classic
receptive
field
of
the
neuron,
the
distal
synapses
cause
dendritic
spikes
and
they
put
cell
into
a
depolarized
state.
What
we
call
a
predictive
state
and
when
you
detect
context,
it's
actually
predicting
the
input.
That's
going
to
come
next,
so
that
kind
of
so.
B
Basically,
this
we're
punching
one
cell
fires
through
to
the
greenhouse.
It
tries
to
find
other
neurons
that
would
have
predicted
that
it
should
have
active,
so
it's
basically
forming
synapses
from
other
axons
other
neurons.
That
would
have
could
that
were
active
just
prior
to
that
and
say,
oh
well,
maybe
these
neurons
predict
my
they
were
predictive,
so
that's
what's
floating
in
the
context
block
there.
It's
just
trying
to
it's
very
simple.
This
is
I'm
gonna
try
to
find
a
pattern
of
activity
that
predicted
my
own
becoming
active.
G
A
Learning
rules,
but
that's
the
way.
That's
the
basic
idea,
as
a
con
cell
is
constantly
trying
to
predict
this
input
and
we'll
try
to
find
contextual
patterns
which
will
allow
it
to
win,
because
the
context
predicted
it's
a
little
bit
redundant,
but
basically,
you
know
clustered
inputs
on
dendrites,
detect,
sparse
contextual
patterns
that
causes
depolarization.
A
A
So
these
are
the
learning
rules
that
jeff
talked
about,
and
basically
we're
saying
is
that
whenever
cell
becomes
active
suppose
there
was
no
prediction.
This
is
this
number
two
here,
then
it
will
try
to
grow
connections
by
subsampling
cells
that
were
active
in
the
past
and
that
thereby
the
cell
will
be
more
likely
to
then
predict
it's
in
become
active.
If
you
get
this
contextual
activity
again,
if.
A
A
B
I
think
it's
worth
mentioning
this.
You
know
I'm
running
up
against
us
all
the
time
when
people
say
well.
How
do
you
know
these
things?
You
didn't
have
any
biological
evidence
for
this.
You
know
type
of
stuff
and
it
shows
the
power
of
good
theoretical
thinking.
This
theoretical
thinking,
which
is
constrained
in
some
sense,
by
biology
that
auctions
by
biology,
can
lead
to
the
predictions
and
confidence
that
these
things
are
actually
happening.
B
Even
though
the
time
we
came
up
with,
we
don't
actually
have
all
the
data
for
it
and
there's
a
lot
of
neuroscientists
who
are
very
uncomfortable
that
they
just
don't
like
the
idea
that
you're,
you
say
something
is
happening
even
though
you
didn't
measure
it,
but
we
can
say
pretty
confidently
that,
from
a
theory
point
of
view
you
can
strain
by
biology.
These
things
must
be
happening
and
just
in
neuroscience,
there
aren't
a
lot
of
theorists
that
do
this
kind
of
work,
but
yeah.
This
shows
that
it
really
does
work.
A
A
If
you
have
those
properties,
plus
the
learning
rule
that
we
had,
you
can
actually
get
a
very,
very
powerful
sequence
memory,
and
so
basically
it
works.
Something
like
this.
Here's
an
example
where
you
have
a
layer
of
cells
and
there's
no
predictions.
There's
no
depolarized
states
in
there.
When
you
get
some
feed
forward
input,
all
of
the
cells
in
the
mini
column
will
become
active
because
no
one's
anyone
else,
they'll
fire
at
the
same
time.
A
So
you've
got
the
sort
of
burst
of
activity,
but
if
you
did
have
predictive
input,
so
let's
say
these
little
red
cells
here
at
time,
t
equals
zero.
Were
depolarized
so
they're
predicting
their
input
when
you
do
get
the
input,
these
depolarized
cells
will
inhibit
the
other
cells
in
the
mini
column
and
you'll
get
this
very,
very
sparse
code
that
represents
this
prediction
in
the
context
of
the
prior
input.
A
Again,
I'm
not
going
to
go
into
too
much
detail
here
without
this
was
kind
of
the
basis
for
our
sequence,
memory
and
then
yeah,
and
then
these
cells
in
turn
will
cause
a
future
prediction,
so
they
are
going
to
depolarize
other
cells
within
that
layer.
That
will
perform
the
prediction
for
the
next
and
we
showed
in
a
bunch
of
work
that
this
could
lead
to
really
powerful
properties.
It,
oh
yeah,
it's
so
powerful
continuity,
learning
sequence
memory
in
when
you
give
it
streaming
data
you
can.
A
B
Large
motivation
for
this
think
about
the
picture
on
the
left.
You
got
it
that
shows
five
minute
columns
being
active.
That
represents
your
input.
Those
mini
columns
represent
your
input.
It
says
these
seven
mini
columns
represent
the
current
thing
I'm
sending,
but
then
you
want
to
regularly
represent
that
input
in
many
different
contents
right.
So
it's
the
same
input.
Those
five
mini
columns,
but
in
different
contexts,
so
picture
on
the
right
says:
yeah,
it's
the
same
input.
B
It
was
saying
by
mini
column
directive,
but
now
I'm
doing
this
picking
individual
cells
in
each
one
so
that
you
can
think
of
it.
As
that's
I
used
to
think
of
the
melody.
The
mini
columns
represents
the
note
and
the
melody,
but
you
have
to
have
a
unique
representation
for
that
note
in
all
possible
melodies.
Therefore,
the
the
one
on
the
right
says.
That's
that
note
at
this
location
in
this
particular
note,
and
if
you
think
about
that
you
have
to
have
a
kind
of
representation
everywhere
in
japan.
So
there's
like
a
fundamental
idea.
A
And
we
can
go
over
this
in
more
detail
in
a
future
research
meeting,
because
we
have
a
lot
of
new
people
here.
You
know
another
way
to
think
about
this
is
that
the
context
is
invoking
a
very,
very
specific
sub-network
within
this
population
of
cells.
So
this
you
know
this
set
of
black
dots
is
a
sub-network.
A
A
So
here,
as
a
picture
from
from
that
paper
or
from
the
top
of
the
paper
that
this
lower
layer
of
cells
is
exactly
the
same
thing
I
showed
before,
but
instead
of
getting
context
from
other
cells
near
in
the
same
layer,
it's
getting
a
context
signal,
that's
representing
location,
that's
analysis,
location
signal
and
when
you
have
that
again
with
the
exact
same
code,
you
are
now
representing
sensory
features
that
are
specific,
specific
to
particular
locations
on
the
reference
frame.
That's
on
that
object,
so
this
is
a
way
you
can
represent.
A
A
B
We
never
showed
this,
but
if
it
turns
out
that
you
actually
went
through
the
same
set
of
motions
each
time
like
if
you're
doing
the
movement,
but
if
I
learn
to
move
my
finger
in
a
certain
way
relative,
an
object,
then
you
learn
as
a
sequence
too
right.
Yeah
the
same
network
can
learn
both
like
a
motor,
sensory
context
and
a
sequence
context.
B
So
it's
not
like
you
have
to
have
one
or
the
other
you
can
have
both
it'll
just
say:
whatever
is
predictive.
That's
what
I'm
going
to
learn.
So
I
always
use
the
example
of
drawing
yourself
off
with
a
towel.
After
taking
a
shower,
you
tend
to
do
it
in
the
exact
same
sequence.
Every
time
you
don't
really
realize
that
if
you
observe
yourself,
so
it
didn't
start
that
way.
But
if
you
start
doing
it
that
way,
then
there's
a
sequence.
A
G
A
A
So
just
a
quick
summary,
real
neurons
are
not
simple
point
neurons
synapses
on
active
dendrites,
detect
dozens
of
sparse
contextual
patterns,
sustained
depolarization
adsomal
leads
to
faster
firing
and
there's
a
diverse
set
of
local
unsupervised
learning
rules
that
are
also
there
in
biology
and
then,
when
you
start
modeling,
this
basically
dendrites
enable
a
very
flexible
context.
Integration
capability
and
layers
of
neurons
so
talked
about
the
idea
of
a
predictive
neuron
with
active,
neurons,
active
dendrites
and
fast
infantry
networks.
You
can
learn
complex
temporal
sequences.
A
A
It's
also
the
these
ideas
are
the
basis
for
implementing
dendrites
in
deep
learning
that
we're
working.
This
picture
is
from
the
paper
that
I
just
released
with
calhoun.
That
just
shows
again
that
you
can
take
the
same
layer
of
cells.
You
can
there's
many
many
different
ways
of
feeding
in
context
and
you
can
get
all
of
these
capabilities,
so
top-down
feedback,
as
people
thought.
G
A
B
We
have
a
long
way
to
go
to
your
biologists,
except
this
too
yeah.
There
are
some
who
are
very
excited
about
it.
We've
mentioned,
but
there's
other
suggested
clues
about
any
of
this
stuff,
and
so
it
hasn't
become
sort
of
like
a
universal
idea.
Like
the
point
neuron,
I
think
it
will
be
at
some
point
in
time.
I
hope
it's
in
our
lifetimes,
but
there's
a
fundamentally
thinking
of
what
neurons
do
in
some
ways.
A
B
A
E
That
you
know
some
early
stages,
they
were
ignoring
certain
parts
of
the
neuron
and
to
me,
there's
you
know,
function
falls
form.
So
if
there's
a
part
of
a
structure
that
you
don't
understand,
you
don't
ignore
it.
It's
you
know.
For
a
long
time,
we
ignored
junk
dna
yeah
until
oh,
okay,
this
oh
micro,
oh
god,
all
the
stuff,
that's
that's
in
there.
So
you
know
something
that
actually
says.
Okay,
here's
the
reason
for
the
morphology
and
maybe
there's
a
reason.
Why
there's
so
many
morphological
permutations
of
neurons
yeah?
You
know
as
as
well
why?
B
Why
are
there
these
thousands
it's
interesting?
I
went
back
and
read
this
this
like
paper
about
one
of
the
graduate
students.
It
was
just
a
it's
just
like
a
quick
paper.
Well,
it
took
a
while
right,
but
it
was
like
describing
what
I
wanted
for
my
things
that
you
can
read
and
you
know
I'm
kind
of
naughty
back
then,
but
I,
this
very
question
was
already
in
my
head
back.
Then
it
was
like
dude.
We
have
to
explain
why
all
these
synapses
are
there,
how.
D
B
B
But
it
just
struck
me
that
same
thing,
kevin
just
tried,
because,
oh
my
god,
like
ninety
percent
of
the
synapses,
are
just
people
that
don't
have
a
theory
for
them.
I
can't
be
just
you
know
what
the
way
you
find
neuroscience
was
to
describe
it.
They
would
say
like
well,
it
modifies
the
action
of
the
neuron.
Well,
that's
true,
but
what
does
that
mean?
You
know
modulatory
effects.
The
hell
is
that
you
know
what
why
yeah?
What
is
the
modular
difference
and
how
is
it
working?
We
have
to
understand
in
the
context
of
a
circuit.
A
A
H
Okay,
we
make
that
for
the
specific
prediction
you
could
test
it
using
this
and
this
type
of
experiments
would
that
be
interesting
to
you,
yeah
or
diffuse,
that
broadly
and
to
whoever
wants
to
try
or
to
to
test
it.
I
think
that
was
my
first
thought
and
the
second
thought
is
the
allen
brain
actually,
but.
A
A
H
Sure
I
mean
it's,
I
think
it's
one
thing
like
theoretical
driven
neuroscience
is
one
thing:
that's
really
powerful.
The
the
loop
between
experimental
and
theoretical
neuroscience
is
crazy,
powerful,
so
yeah.
I
think
this.
If
you
have
even
one
of
the
experiments
of
the
experiment,
to
propose
and
right
now
I
see
a
multiple.
I
think
that.
A
H
A
A
H
Once
you
set
up
like
the
whole
experimental
scheme,
I
mean
it
would
be
interesting
for
people
to
just
do
it
and
on
that
front
that's.
My
second
point
is
the
allen
institute
actually
just
opened
up
what
they
call
open
scope?
It's
basically,
you
suggest
to
them
a
bunch
of
experiments
to
run
from
a
theoretically
driven
hypothesis,
they're
going
to
run
them,
and
then
you
get
the
data.
A
A
C
C
A
Yeah
at
the
end
of
it
all,
and
so
that's
really
not
something
we're
able
to
do
right
now,
but
I
agree
it
would
be
great
to
this
is
the
kind
of
thing
you
know
we
could
collaborate
with
another
lab
if
they
want
to
do
this
yeah.
So
if
anyone
is
listening
to
this
research
meeting
and
wants
to
help
out
with
that
we're
more
than
open
to
that
there's
a
lot
of
stuff.
That
could
be
that
I
mean
if
people
want.
I
could
talk
about
this
experiment.
It's
it's
kind
of
cool.
G
H
A
I
don't
think
I
have
the
movie
easily
available
in
first.
I
think
it's
in
another
presentation.
I
would
have
to
find
it,
but
basically
there's
a
bunch
of
predictions
that
come
out
of
the
temporal
memory
and
spencer
and
yiyu
had
run
some
experiments
already
where
they
had
data.
That
would
where
the
design
was
very
close
to
what
we
might
want
to
test
some
of
these,
and
so
we
actually
work.
A
A
I
could
show
this
in
a
little
bit,
but
they
had
a
movie
where
they
he
basically
took
a
camera
into
the
rat's
cage
and
moved
it
around
as
if
it
was
a
mouse
running
around
in
their
little
nest,
and
then
they
basically
showed
that
movie
over
and
over
again
to
a
mouse
while
they
were
recording
from
v1
and
v2
in
here
using
optical
recording.
So
they
got
hundreds
of
neurons
that
we're
recording
and
the
basic
idea
is
that
this
is
a
30
second
sequence,
a
natural
sequence
that
you're
seeing
over
and
over
and
over.
A
Again.
If
you
remember
correctly
the
if
you
remember
from
previously
what
I
was
showing
when
you
don't
have
a
prediction:
what
you're
going
what
you
expect
to
see
as
a
burst
of
activity,
so
more
dense
activity
and
then,
as
you
learn
to
make
predictions,
it
should
become
really
sparse
the
activity,
that's
the
basic
kind
of
prediction
and
there's
another
prediction
which
I'll
talk
about
in
a
second
and
so
basically
that's
what
they
saw
in
their
in
their
trials.
So
here
this
is
the
first
trial
the
first
time
the
mouse
sees
the
sequence.
A
This
is
a
30-second
long
sequence,
so
the
x-axis
goes
from
0
to
30,
and
these
are
kind
of
neurons
firing.
So
the
first
time
it
sees
it.
You
get
this
much
activity
and
the
20th
time
you
get
this
really
sparse
activity
here
and
the
same
thing
in
al,
which
is
kind
of
like
v2
mouse
and
here's
a
plot
of
the
density
of
the
activations
as
it
sees
more
and
more
stuff
and
it
goes
becomes
very
sparse-
we're
not
quite
sure
what
these
spikes
are,
but
overall
it
becomes
very
sparse
over
time.
A
D
A
These
clusters
of
cells
that
occur
in
so
these
are
like
clusters
of
three
cells
at
a
time,
four
cells
at
a
time,
five
cells
at
a
time
and
so
on.
So
the
likelihood
of
these
clusters
occurring
at
the
same
point
in
time,
is
much
higher
than
you
would
expect
randomly
and,
more
importantly,
when
you
look
at
the
sparse
code
at
a
particular
point
in
the
sequence,
when
you
look
at
a
particular
point
in
the
c
sequence,
you
get
the
same,
sparse
code
recurring,
whereas.
A
The
probability
of
it
firing
is
pretty
uniform
throughout
the
sequence,
but
when
you
look
at
a
particular
point
in
the
movie
you
get
the
same,
you
basically
got
the
same
subset
of
cells
that
are
repeatedly
occurring,
so
this
sparse
code
is
very
predictive
of
that
frame
in
that
movie.
At
that
point
in
time,
which
is
a
pretty
remarkable
doctor,
I
was
thinking
about
that
yeah.
It's
just
like.
B
B
A
A
So
the
blue
bar
is
what
actually
occurs
so
the
number
of
unique
cell
assemblies
increased
the
yeah,
the
number
that
we
were
able
to
count.
This
is
sort
of
six
cells
at
it,
firing
at
a
time
versus
what
would
be
predicted
by
just
a
probabilistic
model
where,
as
the
number
of
right
as
the
order
increases
you
it's
less
and
less
likely
that
it's
going
to
occur
by
one's
poisson
and
the
other
one.
The.
E
A
E
E
E
A
A
A
E
B
Used
you
know
well
in
this
way,
but
the
data
was
already.
There
was
no
new
experiment,
yeah
yeah,
but
just
just
finding
like
you
know
how
many
people
are
observing
and
seeing
how
who
might
have
tested
these
ideas.
They
have
the
data.
Well,
they
don't
have
the
data.
It's
been
the
paper
that
this
paper
came
out
in
2016,
even
though
the
white
papers
earlier
so
at
this
point,
there's
a
couple
hundred
citations
or
so
and
and
so
it
might
be-
I
don't
know
I've
never
done
it,
but
it
seems
a
lot
of
work.
B
Experimental
techniques
keep
cropping
on
up,
and
you
know
some
of
them.
Some
of
the
posed
experiments
would
be
very
difficult
to
do
with
today's
technology,
but
you
know
add
one
more
technique
and
also
they
become
accessible
yeah
most
most
experimentalists
do
not
want
to
be
told
or
what
they
suggested.
B
Idea
you
want
to
test
it
most
of
them
like
screw
you,
you
know
it's
like
things
see
some
sense.
The
theories
have
to
be
well
accepted.
They
have
to
be
sort
of
in
the
in
our
kind
of
work
we
do
have
to
be.
You
know,
has
to
be
sort
of
prevalent
already
before
they'll
spend
any
time
thinking
about
testing.
You
know
just
just
something
to
be
aware
of
they're,
not
sitting
around
like
hey,
you
know
you
go
and
say.
Would
you
like
to
test
this
idea
a
little
bit.
A
So
this
is
a
good
example
of
of
one
of
those
michael
berry
visited
us
and
he
also
read
our
paper.
I
was
pretty
excited
about
it
and
he
did.
He
did
sign
us
as
well
in
here,
but
they
did
a
similar
thing
where
they
showed
sequences
of,
in
this
case
gabor
patches,
and
they
found
a
bunch
of
findings
which
are
completely
consistent
with
the
temple
memory
in
here.
So
it's
so
this
could
this
is
one
of
those
papers.
A
G
One
thing
that
keep
my
interest
in
this
presentation
was
one
of
your
last
slides.
We
talked
about
past
states
acting
as
contextual
input,
and
I
know
you've
studied
that
type
of
thing
before
through
the
lens
of
hdm
and
I'm
kind
of
curious
in
the
current
dendrites
project.
Together,
you
go
to
the
estate
as
contextual
input.
D
G
A
B
Mean
there's
a
difficulty
in
going
from
the
biology
to
applying
it
to
machine
learning
right,
and
I
think
that's.
The
kind
of
thing
that
super
tight
came
up
with
a
couple
of
years
ago
is
the
idea
that
you
can
do
this
sort
of
incrementally,
but
you
don't
get
all
of
it
right,
you're,
just
getting
pieces
of
it
yeah,
because
we're
trying
to
not
create
these
biological
networks,
we're
just
trying
to
improve
machine
learning
networks
initially.
Is
that
related
to
that
question?.
A
Yeah
yeah,
and
so
the
recurrent
stuff
will
could
come
very
easily
onto
this
right
now.
I
think
we're
focused
on
showing
the
impact
of
this
for
continuing
learning
in
a
pure
feed
forward
network,
but
it's
much
more
natural
to
do
this
in
the
in
within
the
context
of
our
requirement
network-
and
you
know
temporal
inputs,
sequence
of
inputs.
E
F
Yeah,
do
you
have
any
like
data
or
determinants
on
how
the
generalization
capabilities
are
of
this,
so
does
the
prediction
was
to
rely
on
memory
and
sequences
that
were
observed
in
the
past?
Or
can
you
also
extrapolate
to
new
sequences?
That
kind
of
follow
the
same
pattern
or
like,
for
example,
every
sentence
ends
with
a
period
and
I
can
see
a
completely
new
sentence
that
I
never
read
before,
but
I'm
still
not
surprised
by
a
period
coming
in
the
end.
For
instance,
yeah.
A
So
we've
talked
about
it
in
the
context
of
the
location
input.
So
if
you
have,
if
you're
sensing
stuff-
and
you
know
the
location
of
this
on
the
object
that
you're
you're
sensing,
you
can
make
a
prediction
of
the
inputs
and
with
grid
cells
and
with
this
memory,
the
exact
sequence
of
your
movements
doesn't
matter
at
all.
A
As
long
as
you
can
predict
the
right
context,
you
can
predict
the
sensory
input,
and
so
you
could.
You
could
do
that
with
with
grid
cells,
and
this
is
something
that
you
can't
really
do
easily
in
a
recurrent
neural
network
in
an
artificial
network
because
they
tend
to
fixate
on
the
memorize
specific
sequences,
as
opposed
to
doing
this
kind
of
path,
integration,
and
so
this.
So
that's
one
example
where
you
know
the
sequence,
specific
sequence
of
movements
doesn't
matter.
You
can
still
make
a
prediction
on.
A
B
The
way
the
sequence
memory
was
presented,
it
doesn't
do
any
generalization
at
all
and
we
kind
of
recognize
that
as
the
limit
of
it
and
and
what
we
have.
But
what
we
do
know
is
that
we've
come
out
on
this
and
we
haven't
really
published
this
or
really
worked
on
one,
but
in
the
context
of
sensory
motor
in
learning
with
the
location
signal.
B
What
we
believe
is
going
on
is,
if
you
see
something
novel,
that
what
what
you
you
clue
in
on
subsets
of
the
of
the
sequence,
whether
it's
the
proper
direct
sequence
or
if
it's
a
sensory
motor
sequence
or
it's
sort
of
like
when
I
look
at
a
new
object,
and
I
don't
know
what
it
is.
I
look
for
subsets
of
it
that
have
a
familiar
arrangement
relative
to
each
other,
and
then
I
generalize
from
the
subset
it's
like
if
I
hear
a
new
melody,
but
but
it
has
some
similarities
to
previous
melodies.
B
I
won't
be
able
to
predict
every
note,
but
in
some
subset
of
that
melody
I'll,
be
able
to
predict
the
next
note,
because
it's
in
a
certain
genre,
or
certainly
it's
familiar
to
other
motifs.
I've
heard
in
the
same
way
that
I'm
looking
at
some
physical
object
and
manipulating
or
touching
it
if
some
subset
of
the
components
that
I
in
some
subsequence
of
it
are
similar.
Then
I
make
predictions
based
on
that.
B
F
Okay,
so
let
me
just
rephrase
it
so
I
sure
that
I
understood
it
right.
So
I
guess
there
are
different
types
of
generalization.
So,
for
instance,
if
I
see
a
new
object
that
is
kind
of
similar
to
an
object
I've
seen
before
I
can
sort
it
into
that
same
category,
because
similar
mini
columns
will
be
activated
but
not
exactly
the
same
ones
but
similar
ones.
And
if
it's
a
similar
context,
it
would
be
the
same
with
the
context
vector
that
it's
not
an
exact
niche.
But
it's
kind
of
a
good
enough
man.
B
It's
more
that
there's
a
subset
of
the
object,
I'm
observing,
whether
it's
a
melody
or
a
physical
object,
there's
a
subset
of
it.
That
has
a
consistent
relationship
that
I've
observed
before,
and
so
I
can
make
predictions
from
that.
It
would
be
the
same
mini
columns
and,
and
initially
it
would
be
like
I
would
even
be
sort
of
invoking
the
same,
maybe
even
some
of
the
same
cells
in
those
meeting
columns.
It's
like
oh
yeah.
This
is
now
I'm
starting
to
recognize
this
thing.
B
B
F
Okay,
and
do
you
have
any
I
don't
know,
do
you
have
an
option
of
having
like
a
hierarchy
in
there?
So,
for
instance,
if
we
think
of
this
sentence
example,
every
sentence
ends
the
period,
and
maybe
I
know
this
because
I
know
the
general
sentence,
structure
of
noun
verb
and
so
on,
and
I
recognize
this
sentence
structure,
even
though
I
haven't
seen
these
exact
words
in
this
order
that
this
is
like
a
higher
order.
Type
of
sequence.
B
A
G
B
So
I'm
not
focusing
on
it
right
now.
It's
it's
kind
of
it's
just
there's
too
many
things
to
think
about
at
once.
So,
but
as
we
get
a
better
idea
of
the
actual
representations
that
are
occurring
in
cortical
columns
and
the
processes
are
going
on
there,
then
then
we
can
say:
oh
what
do
these
outputs
represent
and
how
can
I,
how
can
I
hook
them
together
in
different
ways
and
different
columns
and
different
ways?
Some
hierarchical,
so
I'm
not
and
at
the
moment
we're
not
doing
that.
D
B
D
The
biology
I
mean
essentially
right,
there's
a
classic.
You
know
if
you're
thinking
about
sequence
learning,
if
you
think
about
sequence,
prediction
and
stuff,
there
are
classic
non-biologically
inspired
new
ways
of
doing
this
right,
and
so
I
was
just
wondering
how
those
two
contrasted
right.
You
know
htms
versus
this,
you
know.
Is
it
robustness?
Is
it
a
neuron
saving?
You
know
those
those
kind
of.
A
B
B
C
B
B
B
This
non-biological
system,
or
called
biological
inspired
system
works
better.
So
why
should
we
think
about
the
biology
yeah?
That
is
not
what
we
want
to
do
here
in
momentum.
You
have
to
be
cognizant
of
that.
But
if
you
take
that
approach,
then
you
would
have
never
discovered
the
sensory
motor
learning.
You
would
never
discover
this
stuff
and
talk
about
the
thalamus.
You
never
discovered
the
stuff
marcus
is
doing
with
the
you
know
the
details
of
how
we're
going
to
do
this
this
this
graph
structure.
So
it's
it's.
E
A
B
D
A
B
A
So
that's
typically
what's
measured
in
deep
learning,
but
we
showed
you
as
with
complex
temporal
sequences.
We
can
get
the
same
accuracy
as
lsdns,
but
the
htm
model
was
much
better
at
responding
to
changes
in
the
input.
So
it's
because
yeah,
correct
and
robustness
yeah
as
well.
So
we
when
the
when
the.
If
there's
some
change
in
the
statistics
of
the
input,
then
the
hdm
model
responds
very
very
quickly
to
it.
Whereas
the
lstm
takes
a
long
time,
because
it's
essentially
a
batch
based.
A
B
A
E
A
It
so
yeah,
but
we
showed
that
from
anomaly.
Detection
prediction
is
just
you
know
if
you're
making
predictions
you
can
detect
when
you
have
anomalies,
because
that
the
predictions
are
incorrect,
that
burst
of
activity.
So
we
use
that
in
all
these
real
time,
but
a
bunch
of
real-time
data
from
various
companies
and
we
labeled
it
with
anomalies,
and
this
technique
did
much
much
better
at
anomaly.
Detection
than
other
techniques
there's
a
whole
benchmark
that
we
have
around.
This
called
that.
E
A
Have
an
example
of
that
somewhere
here
yeah:
this
is
this
is
like
our
first
big
example
like
this
little
slight
drop
in
here
was
when
an
expert
looked
at
it.
They
could
say:
oh
yeah,
that's
the
problem
with
the
system.
This
is
like
this
big
industrial
machine.
This
is
the
temperature
sensor
and
then
here
is
where
the
system
actually
failed.
The
temperature
went
way
down
so,
but
this
this
occurrence,
like
weeks
before
the
or
several
days
before
the
actual
failure,
was
predictive
of
this
failure.
A
A
B
B
E
A
A
Way,
you
have
to
adapt
to
dynamic
environments,
you
want
to
detect
anomalies
very
early,
so
that
graph
above
was
an
example
of
forcing
to
minimize
false
positives
and
false
negatives.
But
I
mean,
if
you
think
about
it,
the
grain
works.
This
is
how
the
frame
works,
there's
no
batch
system
in
there,
but
we
don't
store
the
entire
data
set
and
make
multiple
passes
through
it.
We
just
see
things
once
and
that's
it.
C
C
B
They
do
this,
I
mean
they
do
this,
not
because
there
isn't
a
graded
response.
The
further
out
you
get
it
because
what's
what's
happening
is
the
way
we
model
it
is
that
you
get
a
dendritic
spike
and
that
could
occur
anywhere
along
the
dendrites.
So
but
I
didn't
talk
about
dna
spikes
too
much,
not
too
much
yeah.
So
it's
it's
not
like.
There's
this
there's
this
graded
response
to
further.
We
get
to
the
to
the
sum
up.
It's
like
we
kind
of
modeled
it.
This
isn't
100
active,
but
the
way
we
model
it
is.
B
You
have
the
soma.
You
have
the
practical
synapses!
That's
one
thing!
Then
we
model
all
all
the
other
synaptic,
all
the
other
dendrites.
If
we,
if
you
saw
that
picture
of
the
hdm
neuron
we
are,
we
are
we
model.
Just
as
a
set
of
coincidence.
Detectors
we
don't
have.
We
don't
pay
any
attention
to
how
far
out
it
is
because
and
that's
not
100,
biologically
correct.
B
B
Dendrites
are
leaky,
but
all
biological
membranes
are
leaking
right
that,
but
once
you
get
a
spike,
the
whole
point
of
a
spike
is
spikes.
Don't
leak
spikes
propagate
despite
the
leaking
right
right,
but
if
you
just
do,
if
you
do
a
a
membrane,
a
voltage
potential
like
I
just
drop
the
voltage
or
raise
the
voltage
that
will
leak
out
right,
but
a
spike
is
a
self-replicating
process.
B
B
B
We're
going
to
have
as
part
of
the
phallic
thing
I
was
talking
yesterday,
jibs
go.
That
model
has
a
makes,
an
assumption
about
where
things
are
in
a
dendrite.
So
that's
not
our
theory
isn't
complete
in
that
regard,
but
that's
how
we
model
it
today.
So
if
you,
if
you
see
that
htm
neuron,
we
take
all
those
we
take
all
of
those
dendritic
segments,
we
just
consider
it
as
an.
B
A
The
one
big
difference
to
everything
said
is
the
stuff:
that's
at
the
very
top
here
this
typical,
what's
called
the
apical
tuft
or
apical
dendrites
those
have
another
really.
They
also
have
these
nmda
spikes.
But
then
there's
another
thing
called
the
calcium
integration
zone
here,
and
that
has
a
very
different
impact,
which
is
actually
quite
important.
But
we
didn't
talk
about
that.
B
A
E
Is
is
there
a
is
a
well-known
description
of
what
that
that
someone
does
in
some
way
I
mean
what
the
calcium
integration's
on
right,
because
the
nda
on
everywhere
else.
It's
basically
it
raises
the
action
potential,
so
the
thing
can
fire
earlier.
Just
is
there?
Is
there
an
easy
way
to
say
what
the
purpose
of
the
calcium,
what
the
purpose
is
or
what
it
does?
Well,
okay,
I'll
settle
for
the
latter.
A
Very
easy
way
to
say
it:
you
know
it's
kind
of
like
if,
if
these
nmda
spikes
happen,
we
think
of
it
as
a
prediction
and
this
neuron
fires
earlier
right
in
this
case,
if
one
of
these
nmda
spikes
fire,
this
area
gets
depolarized
and
then,
if
you
get
input,
you
get
a
very
strong
burst.
Output
from
the
cell.
These.
E
A
Type
of
prediction,
but
it
has
a
very
you
know-
can
have
and
there
are
five
cells.
It's
burst
very
strong
response
also
triggers
like
really
strong,
very
quick
learning
throughout
the
cell,
which
is
another
okay.
So
it's
it's
there's,
there's
a
lot
of
layers
to
to
what
it's
doing
yeah
yeah.
So
I
think
it's
worth
maybe
another
meeting,
but
I'm
not
sure
if
we
even
got
to
your
question
yet.
Oh
no,
I
was
just.
C
B
So
the
simple
answer
is
that
when
a
cell
generates
an
action
potential,
the
real
spike
right,
it
creates
a
back
action
potential
which
travels
back
along
the
dendrites.
Okay,
that's
the
training
signal
that
says.
Okay,
I
actually
fired
now.
So
you
guys
somebody
needs
to
know
that.
So
you
have
a
training
signal
to
say
your
prediction
was
good.
Something
like
that
that
baxter
potential
is
very
complicated
because
in
general
it
will
not
reach
the
end
of
all
the
dendrites.
B
So
it
gets
it
peters
out,
but
the
evidence
suggests
that
the
dendrite
path
that
was
recently
depolarized
because
of
a
back
because
of
an
nba
spike,
the
back
section
potential
preferentially,
will
travel
along
that
path
and
can
reach
the
end.
So
in
some
sense,
maybe
that's
not
important
from
a
theory
from
a
information
point
of
view,
but
what
we
want
to
do
is
we
want
one
segment
of
the
dendritic
branch
to
go
undergo
training.
B
A
B
Theoretically,
it
has
to
exist
and
so
sure
enough
that
they
now
have
shown
that
the
back
entry
potential
will
reach
the
end
of
the
dendrites
under
under
the
right
conditions.
So
there's
just
a
bunch
of
free
eyes
kind
of
hanging
around.
After
that.
I
don't
remember
the
actual
chemical
processes
kevin.
I
don't.
I
just
remember
that
it
happens.
The.
A
B
There's
obviously
some
something
hanging
around
whether
you
say
ions
or
if
there's
some
channels
and
having
clothes.
I
don't
remember
what
it
is
yet.
Maybe
look
at
them,
but
you
know
there's
a
tremendous
amount
of
details
about
this
stuff
that
you
know
that
make
your
head
spin.
If
you're
trying
to
remember.
C
B
It's
going
to
know
that
that's
where
the
ending
spike
occurred
and
it's
going
to
preferentially
train
that
one.
So
there's
got
to
be
not
only
to
have
to
get
back
an
extra
potential
get
to
that
gendered
excitement,
but
that
tender
children
have
to
know
that
I'm
ready
to
be
learned,
I'm
ready
to
be
trained
because
I
just
turn
around
so
plasticity
is
not
directly
affected
by
distance.
B
B
B
To
any
sequence,
some
things
are
longer
and
shorter
notes,
along
with
shorter
melodies,
different
physical
motions.
So
we
have
a
theory
about
that
which
you've
never
published.
Where
that
timing
signal
is
coming
in
on
the
apical
dendrites,
it's
coming
from
someplace
else,
so
there's
a
way
that
the
neurons
can
encode.
So
you
might
say,
the
apical
dendrites
in
that
case
are
detecting
based
on
the
time.
B
B
Like
you
know,
not
just
what
the
next
note
is,
but
when
it's
supposed
to
occur
right,
the
band
and
dendrites,
don't
encode
the
we
didn't
know
that
it
wouldn't
encode
that
in
time,
but
the
but
the
april
dendrites
would
likely
be
the
place
and
might
talk
about
the
whole
theory
about
how
that
occurs.
But
so
there's
there
you
can
see
like
oh
we're,
differentiating
one
set
of
dendrites
from
another
set
of
dendrites
based
on
the
function.
What
they're,
representing
a
coincidence
of
coincidences,.