►
From YouTube: AI & Neuroscience Chat: creating a software data flow diagram out of the layers in a cortical column
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
A
A
A
Where
did
my
oh
I?
Don't
know
why
these
text
things
keep
disappearing,
hello,
okay,
they're,
just
going
away;
I,
don't
get
I,
don't
get
it
anyway.
It
works
like
there.
They
are
they're
right
here,
but
they're
not
over
here,
okay,
whatever
anyway
so
Wednesday
research
meeting
and
then
the
caveat
on
that
is
the
we
haven't
had
great
research
meetings
that
we've
only
really
had
one,
and
it
was
only
about
20
minutes.
It's
just
bad
luck.
You
know,
I
started
I
want
it.
I
told
you.
A
If
I
wanted
to
do
this
and
then
we
just
had
some
people
out
of
the
office
and
then
and
this
kind
of
sucks,
but
all
next
week,
I'm
gonna
be
out
I'm
on
personal
time
off
spring
break.
My
kids
are
out
of
school
I've
got
family
visiting,
so
there
won't
be
any
research
meetings
next
week
because
I'm,
the
one
who
strains
up
but
the
week
after
I'll
be
back
and
I'll
be
streaming
research.
Media
I
mean
this
will
be
part
of
my
schedule.
A
I'm
planning
on
doing
this
throughout
the
rest
of
the
year
and
then
Thursdays
I
will
be
working
on
community
projects
and
this
if
you're
interested
in
HTM
and
how
to
build
HTM
systems
I'm
gonna
be
working
on.
This
is
on
my
twitch
page.
So
you
can
see
my
schedule
here
on
my
twitch
page
as
well
as
there
we
go
as
well
as
all
about
building
HTM
systems.
So
after
getting
some
feedback
from
the
HTM
community,
I
think
it's
a
good
idea
to
build
this
out.
A
So
what
this
is
is
a
website
that
that
I'm
doing
to
to
make
all
these
tutorials
and
visualizations
about
how
to
build
HTM
system.
So
like
here's,
an
encoder
and
then
it
shows
you
like
that,
encoder
running
and
visualizations
of
it,
etc.
I
want
I'm
gonna
build
this
out.
I've
only
made
it
to
like
encoding
time
so
I
have
some
visualizations
about
that.
But
I
haven't
gotten
through
I
want
to
do
words
and
multiple
encodings
and
then
get
to
the
spatial
pooling
and
then
I've
got
some
big
ideas
here.
A
For
for
how
this
is
going
to
look
there's
already
of
some
things
in
place,
but
I
think
this
is
going
to
be
really
interesting
and
I
want
to
start
working
on
on
this
as
a
part
of
my
weekly
schedule,
so
every
Thursday
I'm
going
to
be
working
on
filling
this
document
out.
It's
gonna
be
an
interactive
document
on
how
to
build
all
aspects
of
HTM
system
through
encoders,
spatial
pooling
and
temporal
memory,
and
probably
anomaly,
detection
and
prediction.
A
B
A
A
That
was
weird.
This
software
I'm
using
online
are
way
OBS.
It's
called
okay,
yours
I,
have
to
re,
specify
all
of
my
text
documents
and
I
never
used
to
do
that
before
before
a
little
pop
up.
So
just
give
me
one
second
should
be
easy
titles,
what
a
pain,
duration,
all
right,
okay,
so
at
least
on
my
the
two
screens
I'm
going
to
be
popping
back
and
forth,
except
for
this
one,
we
won't
see
that
hi
mark
I'm
struggling
with
something
stupid
there.
There's
one
seems
like
every
time.
A
I'm
doing
this
and
I
don't
know
what
I
need
to
do
to
fix
it.
Okay,
there's
two
as
you
guys
can
tell
I'm
still
sort
of
new,
it's
a
twitch
thing,
and
then
my
text
is
all
whack
out.
So
that's
fine
I'll
just
leave
it
I,
don't
I,
don't
really
care.
Okay,
so
everything
looks
good.
Now
that
looks
good.
That
was
good.
A
There's
one
of
our
older
members,
Gary
Gowen
created
this
overview
of
the
cortical
column
and
the
layers
inside
of
it,
based
on
this
video
from
last
week
on
Wednesday
that
I
recorded
with
Jeff
here's
the
link
to
that
video.
That's
still
on
twitch
for
now,
so
you
can
watch
it
and
see
all
the
chat
and
everything
if
you
want
to
so
we're
sort
of
gonna
go
over
his
diagram
or
his
sort
of
call
out
of
the
different
layers.
A
So
we've
got
because
I
like
I
like
the
way
he
built
this
out
and
then
I've
also
got
a
graphic
that
I
took
from
that
presentation.
So
I'm
gonna
draw
this
back
out
on
on
the
whiteboard
here
and
then
we're
gonna
try
and
turn
this
into
like
a
data
flow
diagram.
So
so
so,
let's
read
through
this
first
and
and
talk
about
these
layers.
So
so
we
know
one
where
we're
not
gonna
mess
with
or
we're
not
modeling
one
right
now,
I'm
just
going
to
take
it
out
just
to
simplify
it.
A
A
It's
because
there's
like
these
two
circuit
to
sensory
motor
circuits
that
are
running
here
so
I'll
show
those
when
I
draw
the
picture,
and
why
is
my
phone
going
off?
I
get
so
many
spam
calls
lately.
So
many
spam
calls
okay.
So
this
is
a
lateral
output
layer
and
and
we're
also
labeling
it
as
that
we
think
in
this
first
loop
as
a
as
a
plate,
place
for
place
fields
or
place
cells,
but
it's
definitely
a
pooling
layer.
A
So
there's
it
has
some
stable
representation
of
a
place
in
there
and
it
doesn't
matter
this
is
so.
This
should
not
matter
where
this
should
be
a
stable
representation
over
changes
in
orientation.
Okay
and
the
orientation
changes
are
down
here.
So
layer
four
is
underneath
it.
Maybe
we
should
just
go.
Let's
go:
let's
go
sort
of
go
back
and
forth.
Hopefully
we
can
make
this
work
and
we'll
draw
a
layer.
Two
three
I'm
gonna
see
if
we
can
draw
this
all
in
one
column,
instead
of
the
the
two.
A
Let's
see,
if
we
can
do
that
because
that,
because
in
it
we
might
be
closer
to
actually
creating
a
data
flow
diagram
out
of
this,
so
so
here's,
let's,
let's
just
put
two
three
here
and
we'll
treat
it
just
like
one
I'm
gonna,
spread
these
out
because
we're
gonna
have
a
lot
of
arrows
going
back
and
forth
here.
So
this
is
four
so
things
we
know
about
this.
We
know
this
is
when
this
is
stable.
A
A
A
Let's
call
it
a
driver
signal,
I'm
I,
you
know
ideally
I'd
want
to
separate
this
out
between
distal
proximal,
but
sometimes
it
seems
sometimes
I'm,
not
I'm,
not
always
certain
that
digital
signals
aren't
driving
something
so
I'm
just
going
to
call
this
a
driver
which
is
like
a
primary
input
signal
and
then
we'll
have
another
color
and
we'll
call
this
like
a
modulatory
signal.
I
think
that
still
is
a
decent
representation
modulatory,
because
I
like
to
think
of,
if
you,
if
you
haven't
seen
this
I,
mean
I
I
like
to
think
of.
B
A
A
Output,
which
is
it's
it's
representation,
and
it
has
some
driver
input
and
a
modular
Tory
input,
so
it
can
and
I
drew
these
like
coming
from
the
side
and
typically
you'd
say
this
is
proximal.
Input
would
be
the
driver
input,
because
that's
what
causes
cell
activations
and
other
types
of
input
would
be
modules
for
input,
whether
it's
apical
or
distal,
but
just
to
simplify
it.
Let's
just
call
it
driver
and
modulatory.
A
Now
this
output
could
go
to
another
layer
over
here
and
and
maybe
it's
it's,
this
layers
driver
input,
maybe
it
splits
and-
and
maybe
it's
maybe
it's
also
input
to
some
other
layer.
I'm.
Just
saying
theoretically,
I
mean
if
we're
talking
about
data
flow
and
software
diagrams
I
want
to
I
want
to
break
this
out
as
generally
as
possible.
Perhaps
it
splits
and
its
modulus
or
input
to
some
other
layer,
I
mean
this.
All
these
things
are
possible,
I
mean
so
these
are
sort
of
the
tools
in
the
language.
I
think
we
should.
A
We
should
think
about
when
we're
talking
about
layers.
Is
the
output
has
an
opportunity
to
split
and
go
in
different
places.
The
input
could
be.
You
know
merged
from
from
potentially
two
different
places
representing
two
different
things.
What
we
really
want
to
call
out
is
this:
you
know
what
is
this
and
based
on
its
input
and
output:
hey
blue
neuron
dotnet.
Thank
you
for
for
the
the
bits
I
appreciate
that
so
shout
out
to
oh
you're,
a
computational
math,
a
neuroscientist
student
I,
assume
yeah
cool
I'll
check
it
out
so
anyway
yeah.
A
B
A
B
A
I
want
to
try
and
treat
these
layers
as
something
general
as
as
we're
trying
to
imagine
how
we
could
create.
We
can
create
architectures
based
upon
these.
These
frameworks
so
ignoring
this,
but
but
let's,
let's
talk
about
these
as
if
we
know
what
we're
talking
about
here.
So
we
know
that
there's
input
and
I'm
gonna
call.
This
driver
input,
I'm
gonna,
let's,
let's
change
this
I
always
make
sensory
input
blue,
so
I'm
gonna
change
these
just
to
confuse
everybody.
Now,
let's
go.
Let's
call
this:
let's
call
this
modulatory.
A
B
A
Not
pay
attention
necessarily
to
the
direction
the
arrows
are
coming,
but
what
colors?
They
are.
Okay,
so
so
we'll
draw
this
as
a
primary
input
from
the
left,
and
this
is
the
primary
input
I
would
say
to
the
whole
system.
The
whole
column
comes
in
through
layer,
four
I
think,
there's
I'm
sure,
there's
other
there's
some
other
input.
I
know
well,
obviously,
there's
lateral
input
coming
in
through
two
three
we'll
talk
about
that
in
a
minute.
A
A
Layer,
four,
and
so
the
output
of
this
connects
two
to
three
so
we're
gonna
have
a
primary
I,
think
primary
signal
heading
out
two
to
three
and
what
this
is
going
to
be
doing,
and
we
haven't
talked
about
this.
This
is
going
to
be
representing
basically
a
orientation
over
sensory
input
at
an
orientation.
Okay,
when
we
haven't
talked
about
the
orientation
layer
yet,
but
we're
gonna
get
there
if
you're
confused
about
it,
go
check
that
video
that
I
posted
in
and
chat
right,
I'm,
so
bad
at
this
they're
right
there.
A
So
so,
anyway,
in
through
here
up
through
to
three
and
I
think
now
we're
gonna
have
to
say
the
other
out.
There
is
an
other
output
from
this
and
here's
where
I'm
confused-
and
maybe
you
guys
can
help
me
if,
if
anybody's
knows
the
answer
to
this,
because
I
was
sort
of
wanted
to
work
this
out
with
you
without
running
to
Jeff,
to
explain
everything
for
me.
So
the
signal
from
4:00
to
5:00
and
let's
read
about
layer
five
now
so
this
set
well
output
from
layer.
B
A
A
B
A
That's
interesting
so
so
we
are
definitely
going
to
have
a
modular
Tory
signal
coming
laterally
out
of
here
and
I
believe
you
might
even
draw
this
as
a
modulatory
signal,
because
I
know
in
layer
two
three
there
are
lateral
connections
to
share
this
representation
across
cortical
columns,
as
well
as
connection
lateral
connections
to
cells
within
the
same
layer.
So
you're
going
to
get
sort
of
a
combination
of
sharing
and
a
temporal
pooling
here
so.
A
Okay.
So
so
that's
sort
of
that
loop.
There's
more
to
this
that
we'll
get
to
I
think
in
a
minute
when
we
talk
about
later
five,
but
let's
go
down.
I'm
gonna
draw
this
a
little
further
down
because
we
have
to
make
room
for
the
layer.
Five
I'll
draw
a
layer
five
here
or
not
not
put
anything
in
it.
Yet,
okay,
so
then
work
am
I
going
down
too
far.
Now,
okay!
So
then
we're
gonna
draw
a
layer,
6b
where's
that
6a
this.
A
A
It's
it's
not
pooling
and
what
it
represents
is
input.
We
yeah,
I'm
gonna
call
it
sensory
input,
sensory
input.
This
might
not
be
getting
direct
sensory
input,
but
it's
coming
from
the
direction
of
the
senses.
If
we're
higher
up
in
the
hierarchy,
we
can
still
sort
of
call
this
feed-forward
sensory
input
coming
up
up
the
hierarchy.
So
this
is
sensory
input
at
and
an
orientation
and
I'm
going
to
abbreviate
this
Ori.
So
it's
this
is
gonna
represent
some
sensory
input
at
some
orientation
and
this
will
have
many
columns.
A
A
We're
not
gonna
talk
about
the
thalamus
quite
yet,
but
there
there
there
that's
outside
of
this,
so
680
represents.
This
is
going
to
be
orientation
essentially
so
in
a
neuronal
cortex
there's
these
head
Direction
cells.
So
we
believe
this
is
the
equivalent
to
that
going
on
in
the
cortex.
This
is
going
to
be
an
orientation
layer.
It's
going
to
represent
orientation.
A
This
is
not
going
to
be
stable
like
this,
because,
as
you
move,
this
is
going
to
be
changing.
Your
orientation
will
be
changing,
it
will
always
change,
but
if
you
change
your
radial
orientation
and
respect
to
an
object
or
environment
that
you're
in
this
will
change,
and
so
these
are
like
head
Direction
modules.
We're.
A
I
think
that's
somewhat
are
similar
to
what
head
Direction
cells
work
in
the
intro,
Rhino
cortex
or
hippocampus
in
Toronto,
cortex,
okay,
so
we've
got
connections
here.
Major
connections
to
four.
This
is
a
big
connection
here
and
I
think
this
is
there's
a
connection
down
and
a
connection
back
up
so
now
I
don't
need
to
think
this
is
probably
we've
got
a
primary
driving
connection
up
to
here
and
I.
Is
this
a
module
Ettore
connection
or
a
primary
connection?
A
I'm,
not
certain
so
I'm
going
to
read
through
some
of
Gary's
notes
here,
6a
represents
looking
right
here:
rotational
degrees
of
freedom,
orientation,
head
direction,
major
connections
to
four
then
four
to
six,
a
six
eight
and
four
model
place
into
three
right
which
we
which
I
haven't
put
down
here
yet
so
we
could
also
label
this
place.
This
is
this
is
sort
of
what
this
is
representing
place:
sensory
input
at
orientation
and
orientation.
That's
what
these
things
are
representing.
A
This
is
stable
running
temporal
pooling
this
is
unstable
and
it's
running
spatial
pooling
using
many
columns
and
temporal
memory
over
time-
and
this
is
this-
is
a
grid
cell,
like
grid
cell
module,
like
mechanism
that
selects
on
orientation
as
if
it
were
head
direction.
I
know
I'm
just
repeating
myself,
but
it
helps
me.
This
helps
me
to
understand
what
I'm
talking
about
so
so
I
think
something
has
to
drive
this
there's.
Nothing
else
driving
it
so
I
think
has
to
be
I
think
this
has
to
be
a
driver
signal.
A
Let's
just
do
that
a
driver
signal
from
four
to
six,
and
once
we
get
this
all
out,
we'll
try
and
clean
it
up.
You
know
turn
it
into
more
of
a
dataflow
type
of
diagram.
Perhaps-
and
let
me
have
a
couple
drawings
I
want
to
make
sure
I'm
not
getting
too
far
away
from
what
I
thought.
I
was
going
to
do
yeah
so
head
direction,
and
then
we've
got
another
link
back
because
we
need
to
get
the
orientation
back
to
make
a
prediction.
So
then,
so
this
is
definitely
a
modulatory
signal
back.
A
Let's
do
it,
let's
just
do
it
like
this,
and
this
helps
us
make
predictions,
because
we
once
we
feel
something
if
we
know
where
we
might
be
looking
on
it
or
how
we
might
be
oriented
towards
it.
We
can
make
a
prediction
about
what
the
next
sensation
is
going
to
be
given,
given
a
movement
as
a
part
of
the
the
input
here.
A
B
A
A
Okay-
and
this
is
actually
for
me,
the
more
the
more
familiar
loop,
because
that's
what
the
columns
paper
in
the
columns
plus
paper
really
sort
of
laid
out
so
so
we're.
We
still
have
two
three
and
now
we're
gonna
go
down
and
loop
to
level
layer,
five,
okay,
so
back
to
two
three
remember:
two
three
is
pooling
layer.
It's
pooling
on
place
in
this
model
based
on
orientation
at
is
there
sensation
at
orientation?
A
It
should
be
stable
as
a
rotonda,
knob
jekt,
an
independent
of
orientation.
This
this
should
be
independent
of
orientation.
So
this
is
like
the
sort
of
like
the
the
object
as
we're
moving
through
space.
Trying
to
you
know,
define
or
model
the
object
and
layer,
five
is
and
and
this
okay.
So,
as
Jeff
sort
of
stated,
there
are
two
layers
there.
There
is
an
A
and
a
B
and
there's
some
confusion
about
which
layers
a
and
which
is
B
and
so
I'm
we're
just
gonna,
be
real
hand-wavy
about
that
and
said.
A
We
don't
know
what
layer
we're
connecting
to
exactly
in
here,
because
it's
harder
because
a
lot
most
people
because
I,
don't
think
anybody
really
knows,
but
this
loop
is
going
to
be
basically
similar
to
this
loop,
okay.
So
so,
where
we
have
this
same
thing
happening
between
4
&,
6,
there's
another
layer
down
here:
that's
6,
B
2,
it's
pretty
close
to
the
same
thing,
so
layer,
5,
okay,
let's
forget
about
where
input
till
later
5
comes,
but
let's,
let's
first
show
that
the
same
thing
is
happening.
A
We've
got
a
driver
here
and
another
modulatory
signal
coming
coming
back.
Okay,
so
it's
the
same
sort
of
loop
happening
here
or
a
circuit
going
on
here,
where
there's
some
type
of
input
here
and
then
some
some
type
of
module,
that's
selecting
on
a
mapping
of
space
and
in
this
case
it's
two-dimensional
space
and
in
this
case
it's
radial
space,
okay,
okay,
so
well!
This
will
call.
This
is
a
lateral
output
layer.
So,
let's,
let's
and
I'm,
not
I'm,
not
certain
I'm,
really
not
certain.
A
If
there,
because
when
I,
say
lateral
output,
layer
I
mean
there's
their
signals
coming
out
of
it
and
going
elsewhere
outside
of
the
column,
I
am
NOT
certain
about
these
I
don't
know,
even
if
Jeff
is
I'm
going
to
write
them
both
and
just
let's
just
assume
that,
for
they
could
perhaps
be
primary
or
modulatory
signals
for
some
that
this
layer
could
be
primary
or
a
mileage
modulatory
signal
for
some
other
for
something
else,
I
don't
know,
I
mean
I
mean
eventually.
This
goes
through
the
thalamus
and
back
up
to
the
next
layer.
A
For
but
I
don't
know
about
I,
don't
know
much
about
this,
so
I'm,
just
gonna
wave
my
hands
and
say
you
know,
question
what,
but
there's
definitely
lateral
output
here
and
again
through
the
through
the
thalamus.
So
the
thalamus
is
certainly
involved
over
here
and
it's
gonna
make
a
big
general.
The
thalamus
is
involved
in
this,
and
this
loop
and
I
haven't
talked
about
this
yet,
but
there's
also
a
fallible
cortical
that
will
loop
down
here
from
6a.
A
A
A
B
A
I
could
be
cleaner
about
this
and,
like
move
this,
keep
all
the
module
of
trace
signals
coming
from
one
place.
Let's
I
might
do
that.
So
so,
let's
do
that.
Let's
say
our
laterals
going
this
way
and
then
there
could
be
a
or
there
is
a
reciprocal
loop
there
and
then
we'll
and
we'll
just
keep
this
clean
and
say
we
don't
know
about
the
signal
laterally.
This
is
I,
don't
think
that
goes
through
the
thalamus
laterally,
but
I'm
pretty
sure
this.
A
If
we're
talking
about
a
driver
coming
coming
out,
laterally
I'm,
pretty
sure
that
goes
through
the
thalamus.
Okay,
just
make
sense
I
going
in
the
right
direction.
You
guys
is
this
clearing
anything
up
for
anybody,
it's
sort
of
making
a
little
bit
better
sense
for
me,
I
kind
of
like
the
keeping
the
dataflow
primary
versus
modulatory
separated.
Perhaps
I
don't
know
if
that's
going
to
be
useful
or
not
never
tell.
A
Okay,
so
where
are
we
at
2/3
the
place
circuit?
Let's
follow
the
let's
follow
this
through.
Let's
follow
the
data
through
now
that
we've
got
a
diagram
here,
let's,
let's
sort
of
just
put
all
the
caps
on
all
things
here
and
we're
gonna
start
here
right
here.
So
so
the
input
is
going
to
hit
layer.
Four.
A
A
I'm
going
that's
where
I'm
pushing
this.
Ok,
so
now
get
my
colors
right
here.
So
it's
a
driver
two
to
three
and
it's
a
driver
two
to
six,
eight.
Okay,
so
two
three
drives
five
I
forgot:
let's
just
call
it
five
two
three
drives
five:
maybe
I
should
I,
don't
know
where
this
is
gonna
end
up.
I'll
just
put
it
here.
Two
three
drives
five:
let's
just
follow
the
driver
signal,
so
this
is
five.
A
A
So
so
the
first
module
Ettore
signal
is
actually
coming
from
two
three.
So
we're
gonna
have
a
two
three
back
to
four
and
then
we'll
have
a
roundabout
here
because
it
informs
itself
and
then
there
is
six
a
up
to
four.
That's
that's
the
first.
This
is
the
first
sensorimotor
loop,
the
orientation
one
right
and
then
we've
got
six
be
back
to
five.
This
is
the
second
sensorimotor
loop.
So
if
you
can
imagine,
I
mean
this.
This
is
what
we
got
here.
A
A
Alright,
dang
it.
Okay,
you
got
that
as
I
break
out
that
I'm
gonna
move
along
I'm
gonna
erase
this
this
one
and
I'm
gonna
try
and
simplify
this.
Does
this
encompass.
Let
me
ask
you
this
first
minibus
pay
attention.
There's
this
diagram,
this
simplified
diagram,
it
compass
all
this
information.
As
far
as
the
arrows
and.
A
Covered
everything
I
might
not
have
covered
the
ladee.
I
got
the
lateral
out
here
the
in
here,
and
so
basically
this
is
in
this
is
out.
This
is
the
out
right
here
coming
out
of
layer,
five
and
that's
gonna,
be
like
our
output
signal
or
our
motor
representation.
Whatever,
however,
we're
interacting
with
the
reality,
we're
representing
it's
coming
out,
there
graph
layout,
yeah
all
right
so
so
I'm,
assuming
that
some
sort
of
on
the
right
track
here.
So,
let's,
let's
wipe
this
out.
A
Okay,
I
got
kind
of
a
messy
representation
here,
but
I
want
to
maybe
isolate
the
two
loops.
Let's
say:
let's
try
and
get
this
like
input
coming
in
from
one
direction:
output
going
a
different
direction,
maybe
it'll
just
go
up
down
all
right.
Let's
say
input
coming
up
because
that
just
seems
more
natural.
If
we're
talking
about
a
cortical
column,
this
is
actually
would
be
the
right
orientation.
Everything
about
bottom-up,
like
coming
up
screw
the
cortical
layer.
I
don't
know,
maybe
maybe
it's
not
the
right
way
to
think
about
it,
but.
A
A
Ok
and
2/3
is
going
to
be
a
part
of
both
loops.
So
let's
put
that
right
in
the
middle
2/3
and
then
we've
got
like
6,
6a
and
6b
on
either
side.
6A
is
gonna,
be
this
loop
6
beers
that
loop,
so
6a
6b,
all
right
and
now,
let's
start
drawing
the
arrows
and
see
if
we
could
see
if
this
makes
any
any
more
sense
at
all
got
my
ok
driver.
So
input
comes
in
right.
A
The
first
thing:
let's
do
a
boat:
let's
do
this
simultaneously
layer,
4
is
gonna
drive
to
3
and
excuse
me:
alright,
that's
the
first!
That's
the
first
thing.
That's
all
that's
the
first
thing
that
happens,
but
it's
also
gonna
drive
6a.
So
it's
we're
gonna
get
2
signals
right
coming
out,
therefore,
based
off
of
this.
So
here's
a
question:
is
this
happening
at
the
same
time?
I?
Don't
think
so,
I
think
what's
happening?
Is
we
go
here?
We
go
2
2,
3
first,
and
we
and
we
get
feedback
from
it
right.
A
I
think
that's
what
should
this,
because
this
is
a
well.
This
is
the
prediction.
This
will
be
the
this
will
be
the
prediction
for
the
next,
so
actually
this
is
we're.
Gonna
go
both
places
simultaneously.
This
is
all
this
is
gonna
be
happen
at
the
same
time,
because
we're
going
to
our
sensory
input
here
is
going
to.
A
If
we
already
know
where
we're
at
this
is
gonna
have
a
representation
in
it
already.
So
that's
what's
going
to
get
updated
by
this.
If
we
don't
know
where
we're
at,
we
have
no
current
orientation.
Oh
I,
just
I
miss
chat
first
diagram,
this
post
thalamus
excitation
for
a
level
4.
Where
does
that
fit
I,
don't
know
I'm
the
thalamus
is
I,
don't
know
I'm
trying
to
get
through.
A
A
Controlling
scale,
it's
particularly
being
able
to
control
the
grid
cell
module
and
orientation
module
representations
in
layer
6
to
basically
expand
or
contract.
So
that's
what
we
that's.
What
Jeff
thinks
the
thalamic
signal
coming
into
layer
4
could
be
could
be
used
before
it's.
It's
certainly
related
to
oscillations
and
wave
behavior
I,
don't
think
it's
more
than
forming
representations,
I,
think
it's
warping
representations
or
it's
supplying
a
scale
to
the
model
like
it's,
it's
whatever
you're
attending
to
it's
fitting
it
somehow
I
don't
know
exactly.
A
But
let
me
let
me
move
on
and
try
and
get
this
this
straight
in
my
head,
so
from
layer
4
we're
driving
these
two.
This
is
the
orientation
loop.
This
is
orientation.
This
is
place
and
we
get
feedback
to
decide
to
make
predictions
about
what
we're
going
to
feel
next,
which
also
makes
predictions
about
where
we're
going
to
feel
it
next,
based
on
the
movement
movements
making.
This
is
this.
This
is
what
I
haven't
drawn.
A
Okay,
so
we've
got
two
six,
a
and
two
three
and
now
we
have
to
acknowledge
that,
at
the
same
time
that
this
is
occurring,
this
driver
input
gets
to
two
three
and
it
is
immediately
driving
layer.
Five,
as
this
loop
is
resolving,
this
loop
is
also
resolving.
This
is
that
we,
you
have
to
think
about
this
happening
at
the
same
time.
So
so
again,
we've
got
layer
five
in
the
same
way
that
layer
four
is
primary
input
for
layer
six
and
gets
modular
gets
modulus.
We
input
back
from
layer,
six,
a
layer,
five
and
layer.
A
B
A
Modulus
or
modulus
or
E,
and
then
get
rid
of
this
one
or
this
one
just
so,
we've
got
a
cleaner
sort
of
diagram
and
everything's
nice
and
isolated.
So,
but
now
we
have
to
think
about
thatÃs,
so
yeah
I
was
thinking
about
adding
a
time
ticker
to
this.
That
would
what
one
of
the
things
we
do
in
the
columns
plus
paper
is
we'll
put
like
a
let's
take
a
look
at
that.
Actually
cuz,
that's
a
good.
That
is
a
good.
A
A
A
Articles
about
weird
okay,
there
used
to
be
a
link
to
download
here.
That's
weird
I
have
this
somewhere.
Let
me
just
search
my
we
just
search
my
hard
drive.
I've
probably
got
six
copies
of
it.
Honestly,
okay
hold
on
a
moment,
but
there's
some
good
diagrams
in
here
here
it
is
no
that's
not
it.
Where
is
it
using
grid
cells
for
coordinate
transformations
there.
B
A
A
So
this
location
would
be
the
location,
layer
and
the
sensory
layer,
so
the
the
sensory
layer.
We
would
call
that
layer,
four
and
the
location
layer
would
be
6a
in
this
case
or
this
or
five
and
six
B.
It's
the
same,
the
similar
loop
right.
So
so
it's
basically
motor
input
to
layer
four
as
one
to
his
driver
input
to
the
sensory
layer,
Oh.
B
A
Of
the
location
there,
maybe
this
is
not
gonna.
Maybe
this
doesn't
work.
Maybe
this
is
outdated
move
because
we
have
a
movement
signal
coming
in
here
and
that
would
be
coming
into
it
have
to
be
coming
into
either
6b
or
6a.
So
I
bet
that
so,
if
the
input
sixty
or
six
Bay
is
six
a
is
the
one
coming
from
the
thalamus
loop
right.
A
A
A
A
A
I'm
gonna
have
to
ask
I'm
gonna
have
to
ask
that
question
because
I,
don't
think
I
understand
that
so
I'm
gonna
I'm
just
trying
to
relate
the
idea
of
these
time
steps
here
in
the
columns
plus
paper,
starting
with
a
movement
coming
into
what
would
only
be
either
orientation
or
location
layers
which
hold
grid
cell,
like
modules
or
orientation,
or
head
Direction,
like
modules,
which
would
then
I
think
modulate
the
sensory
layer
which
has
already
gotten
a
sensory
input
right,
I
think
I
think.
Ideally
we're
gonna
get
sensory
and
movement
information
at
the
same
time.
A
If
we've
got
motion
information
being
fed
somewhere
into
layer,
six
which,
where
is
it
going
and
is-
and
is
that
simultaneous
with
one
or
is
it
going
to
be
two
I,
don't
know
so
I'm
gonna,
I'm
gonna
dig
that
out
and
figure
that
out,
but
I'm,
but
I
know
so
based
on
this
on
that
Sonoma.
So
this
is
outside
of
columns
Plus.
This
is
the
columns
paper,
so
I'm.
A
Not
certain
if
this,
if
if
this
is
2,
these
are
both
2
at
the
same
time
or
if
this
is
2,
and
then
this
is
2,
you
know
I'm
gonna,
that's
a
good
question.
That's
something
to
figure
out
so
I'm
I'm,
gonna,
I'm
gonna,
investigate
that
I'm
gonna.
Ask
particularly
I'm
gonna,
ask
Jeff
about
that,
one
in
the
office
on
Wednesday
and
and
see.
If
we
can
take
this
diagram
and
add
these
times
these
time
ticks
to
it.
A
A
So
he
drew
the
the
Flajnik
loop
down
here
and
he
said
there
was
a
ton
of
data
coming
from
layer,
6
a
to
the
thalamus
and
then
there's
a
loop
back.
The
thalamus
then
loops
loops
back
into
to
be
a
part
of
the
input.
I
think
this
is
in
the
same
layer,
it's
this
is
not
going
off
to
another
level
of
the
hierarchy
or
anything.
A
A
We
know
there
is
and
I
don't
know
if
this
is
modulatory
input,
but
we
know
there's
at
least
Jeff
is
drawn
here
that
there
is
let's,
let's
draw
the
thalamus
in
red,
because
why
not?
I
are
headed
head
it
and
read
earlier
and
we'll
just
say:
this
is
the
corresponding
layer
of
cortical
cells
that
that
operate
with
this
cortical
column
and
the
thalamus.
A
A
A
A
For
those
of
you
don't
know,
Jeff
Jeff
Hawkins
is
my
boss.
He
founded
the
company
Numenta
questions,
so
one
is
time
timing,
especially
between
the
two
loops.
The
two
SMI
loops
would
be
great.
If
we
could
add
those
time
ticks
to
all
these
and
then
because
then
we
could
sort
of
animate
it
out
and
play
it
out
like
what's
happening
at
the
same
time.
A
A
What
you've
got
here,
I
mean
if
you're
talking
about
layer
6
in
our
model
layer
6
is
going
to
be
either
location
or
orientation
information.
So
it's
like
in
in
the
context
within
while
attending
and
a
thing
and
object.
It
would
be
somehow
the
current
location
and
orientation
information,
not
necessarily
the
sensory,
well,
that
it
would
have.
It
would
have
some
sensory
information.
You
know
coming
because
it's
getting
feedback
over
time.
A
A
So
it's
like
a
new
input
space
on
itself
when
it
gets
rerouted
I
think
maybe
it's
just
it
needs.
Perhaps
it
just
needs
to
know
the
current
aspect
or
point
of
view
to
the
object
under
inspection
before
it
can
apply
a
scale.
I
mean
it
kind
of
it
sort
of
needs
to
know
where
am
I
at
now
before
I
have
to
know
the
current
situation
before
I
can
imagine
it
bigger
or
smaller
or
scale.
A
My
expectations
I
got
to
know
what
I've
got
right
now:
I'm,
assuming
because
I
like
assume
that
let's
say
you're,
running
right,
you're,
running
and
you're
running
as
things
pass
you
and
you
and
you
attend
to
them.
They're
gonna
change
scale
constantly,
so
this
could
be
like
the
current
state
of
it.
It's
its
current
location
orientation
relative
to
you.
So
the
thalamus
knows,
as
you
flow
by
it,
that
to
make
to
predict
what
the
next
state
will
be
at
there
to
inject
back
in
there.
A
How
big
that
things
should
be
as
it
passes
you
the
whole
time.
Does
that
make
sense?
I
think
that's
could
be
what
the
downest
is
doing.
At
least
that's
the
way.
Jeff
explains
it
so
I
think
I'm
explaining
it
the
right
way
that
he
would
explain
it,
but
we'll
see
so
so
what
were
the
other
questions?
I
got
one
on
timing,
so
we're
gonna,
try
and
resolve
the
timing,
the
the
thalamus
loop
like
what
is
it?
What's
it?
What's
it
injecting
and.
B
A
Yeah
yeah,
that
could
be
it
okay,
so
this
is
where
we
got
after
an
hour
or
so
talking
about
more
about
cortical
columns
is
some
direction,
I
think
towards
simplifying.
You
know
this
this
diagram
over
here
to
something
perhaps
a
little
a
little
simpler,
that
we
can
relate
more
towards
software
architecture
will
says.
Another
thing
that
might
be
useful
to
visualize
is
the
amount
of
information
that
flows
between
the
layers.
Yeah.
That's
true
like,
for
example,
I
know
this
is
super
thick.
A
A
Okay-
and
he
did
this
whole
thing-
okay,
cortical
layers,
author
Casey,
I-
think
this
works
no
cortical
layers
there.
It
is
layers
layer,
2,
3,
notes,
summary,
so
this
guy
has
done
some
really.
Nice
got
some
really
nice
notes.
So
there's
he
did
one
that
was
like
the
whole
cortical
column.
I
swear,
I,
thought
I
thought
there
was
one:
let's
just
do
from
wait:
where
is
it
posted
by
Casey
Casey,
Walsh,
Walter.
A
So
here's
like
he
took
a
ton
of
notes
on
on
layer,
six
and
I
perused
through
this,
but
I
didn't.
If
I'm
sure
we
could
find
a
lot
of
information
if
anybody
wants
to
do
a
little
research
and
help
me
answer
these,
these
questions,
the
timing,
the
size
of
the
inputs.
Perhaps
you
really
might
be
able
to
find
it
everything
you
need
in
an
HTM
form
because
there's
a
ton
of
neuroscience
stuff
posted
on
this
forum,
especially
in
HTM
Theory
neuroscience.
That's
the
now
posted
over
here.
There's
there's
a
lot
of
resources
in
there.
A
If
you
can
sort
them
out,
and
this
stuff
from
Casey
is
is
particularly
I
think
high
high
quality.
It's
got
a
lot
of
references
and
stuff.
So
here's
what
happens
in
layers,
5,
&
6,
a
lots
of
conversation
about
this
I
mean
we
don't
have
time
to
weigh
in
it.
I,
certainly
don't
have
enough
knowledge
to
weigh
in
a
lot
a
lot
of
these
neuroscience
conversations,
but
there's
a
quite
a
few
people
on
the
forum,
including
you
know,
Mark,
Brown
and
Casey,
and
there's.
A
A
Okay,
information
density,
we'll
call
it
that
okay,
all
right
so
I-
think
we're
gonna
wrap
up
this.
This
chat
this
little
session
and
I
will,
when
I
go
back
to
the
office.
I'll
I'll
probably
show
some
of
this
to
Jeff
or
subbu
tie
and
asking
these
questions
and
to
see
if
they
have
any
anything
that
can
help
inform
us
and
I'll
post
about
it
on
HTM
forum,
sound
good
appreciate
you
guys
taken
taken
time
to
watch
the
show
this
week
next
week
on
AI
neuroscience
chat.
A
We're
going
to
talk
about
an
AI
license
called
the
responsible
AI
license
it's
something
that
I've
seen
floating
around
since
we
have
an
open-source
artificial
intelligence,
codebase
I
thought
it
would
be
a
good
idea
to
review
that
license
and
we're
just
going
to
talk
about
it.
I'm
going
to
talk
about
alive
and
what
I
think
about
the
license.
A
I
haven't
even
looked
at
it
at
all,
barely
I
mean
I
could
sort
of
get
the
gist
of
it,
but
I'm
going
to
talk
about
whether
we
think
we
should
maybe
adopt
that
license
and
we'll
see
how
it
goes.
So,
we'll
talk
to
Sartre
about
responsible
AI
a
little
bit
in
the
next
in
the
next
chat.
If
you
want
to
cover
so
it
won't
be
that
technical.
If
you
want
to
come
talk
about
the
philosophy
of
AI
or
what
have
you,
you
know
the
the
effect
it
could
have
on
humanity
and
safety
issues.
A
A
Let's
go
rate.
Somebody
let's
go
read
somebody
it's
twitch.
So
let's
raid
somebody
sometimes
I
forget
to
do
this.
So
let's
see
Who
am
I.
Who
do
I
know
that's
currently
line.
We
could
always
go
back
to
rhyme,
ooh
I,
don't
know
this
guy,
Dino,
D
and
nugget
I.
Don't
know
this
guy
either.
Let's
go
back
to
write
new
doing
his
C++
stuff.
A
Let's
rate
him
he's
a
good
guy
rhyme
ooh,
what's
his
name,
eight
three,
five:
four
thanks
for
sticking
around
those
of
you
guys
are
sticking
around
for
the
raid
whoops
I,
gotta,
I
gotta,
say
his
name
right.
First
of
all,
okay,
this
I
think
I.
Did
it
yeah
all
right,
so
we're
gonna
be
ready
to
raid.
Let's
go
raid
them
thanks.
Everybody.