►
From YouTube: HTM 2D object recognition | community live-coding | artificial intelligence | hierarchical temporal
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
I
am
trying
to
set
up
something.
I
should
have
set
up
a
little
earlier,
but
I'm
doing
it
anyway.
I'm
getting
these
image
or
no
no
text
or
is
it
text?
We
go
duration,
all
right
here,
I'm
just
trying
to
get
these
little
things
in
the
right
spot,
so
I
got
them
all
set
up
now
and
it
took
me
a
little
longer
than
I
thought
stream
title
there
we
go
and.
A
A
A
It's
small!
It's
a
little
small,
but
it's
okay,
it'll
do
okay,
lock,
lock
chat
is
over
here
that
works.
Okay,
I!
Think
I
am
streaming
I'm
streaming
right,
yeah!
Okay,
sorry
for
you,
guys
on
YouTube
I
might
cut
the
stuff
out,
but
I'm
just
trying
to
set
my
twitch
stream
up
right
and
a
lot
of
the
things
you
cannot
really
test
until
you
go
live
so
I
think
this
is
somewhat
expected
based
duplicate
streaming,
eight
minutes.
Okay!
So
then
I
got
the
whiteboard
view
and
let's
paste,
oh
I
should
have
done
this
a
little
differently.
A
It
doesn't
matter
title
copy.
So
when
the
whiteboard
you
doing
with
the
same
as
the
hero
stream
title
copy,
whiteboard
paste
duplicate,
there
is,
and
then
the
number
of
viewers
I
will
take
on
the
hero
screen.
Hello,
one
viewer
I'm
just
sort
of
getting
started.
A
lot
of
stuff
I
want
to
do
today,
but.
A
A
They
let
people
know
that
I'm
streaming
streaming,
let's
see
live,
let's
just
say
streaming:
HTM
community
work
on
twitch
now,
yeah
I'll,
just
leave
it
at
that
and
I'll
copy.
The
URL
I
still
I,
don't
know
how
much
really
how
much
crossover
I'm
really
getting
on
Twitter
if
you're
coming
from
Twitter
and
you're
and
you're
coming
to
twitch.
Let
me
know
that
if
I
brought
you
to
twitch
that'd
be
really
interesting
to
know.
A
A
A
A
All
right,
Shaima
community
projects
on
this
squared
I'll
open
up
the
live
stream
and
see
how
this
works
I'll
open
up
voice.
So
if
anybody
wants
to
try
and
jump
on
voice
with
me
as
long
as
it's
on
topic,
I
want
to
I
do
want
to
work
on
community
stuff,
so
I
don't
want
to
just
like
do
general
neuroscience
chat
if
you
want
to
do
that,
we'll
we'll
do
that
on
Mondays.
A
A
Brand
new
I,
don't
okay
forum
and
there's
there's
still
more
PRS
now
one
of
them
is
from
from
will
I
think
on
on
the
2d
project.
The
other
was
is
on
HTM
school
visualizations
and,
and
that's
been
sitting
there
for
a
little
while,
but
I
want
to
look
at
it
because
I
think
I
might
want
to
try
and
get
if
I
can't
get
this
running.
I
should
get
it
running.
A
I'll
have
to
run
new
pic
for
this
because
I
think
it's
part.
It's
a
spatial
pool
or
it's
involved
with
the
spatial
pool
or
visualization.
So
I'll
probably
look
into
that
too,
and
then
we'll
we'll
do
we're
gonna.
Do
the
2d
object,
recognition,
tests,
we're
gonna,
look
at
tests
and
so
I'm
gonna,
look
at
that
gherkin
framework
and
see
if
we
can
write
some
language
and
environment
agnostic
tests
and
I'm
not
going
to
worry
about
visualizations
anymore
I'm,
just
going
to
define
tests
and
and
decide
what
is
the
challenge
going
to
be?
A
We
could
define
several
a
bunch
of
different
tests.
You
know
just
testing
like
setting
up
the
structure
of
the
system,
how
aftertouch
one?
What
do
you
expect
so
and
we're
gonna
have
to
define
some
terms?
You
know
what
do
we
mean
by
touch
one
and
move
and
all
that
stuff?
So
I
think
we'll
talk
about
that?
A
A
It
I
looked
into
twitch
to
see
if
maybe
you'd
replace
hackers
hangout
on
it
might
in
the
future,
but
I'm
not
ready
to
jump
yet
because
I
still
have
good
attendance
here
from
some
of
the
some
important
people
in
the
HTM
community,
so
I'm
gonna
do
that
tomorrow.
So
if
you
want
to
join,
live,
you'll
have
to
go
to
youtube.
So
look
at
the
youtube
channel
down.
A
In
my
twitch
description
and
my
twitch
page,
we're
going
to
talk
about
the
Python
3
movement
is,
is
still
I'm
not
going
to
be
stress
free
until
we
were
Python
3.
Do
we
have
an
HTML
5
on
three
and
I'm?
Going
to
talk
about
the
canoe
pic
and
the
community
fork
and
naming
thing
this
is
particularly
in
Python
3.
How
we're
going
to
name
things.
A
Usually
when
you
have
a
company-
and
you
want
to
open
source
some
code
that
you
developed
internally
or
you
have
to
go
through
some
review,
although
all
of
it
was
open
source
and
they
shouldn't
research
repo
anyway,
it's
I
still
want
to
look
at
it
because
it
was
constructed
in
a
private
repository
and
reviews
a
license
and
all
that
stuff
and
oh
yeah
falco
are
you
on.
Are
you
there?
Let's
see
I
can
see.
A
I
can
see,
you
know,
I,
don't
think
he's
on
yet,
but
so,
if
you're
watching
I,
don't
think
it
makes
I
think
the
object.
2D
object.
Recognition
project
is
perfectly
designed
for
what
we
need
to
well,
I,
wouldn't
say
perfect,
but
the
way
I
designed
it
was
specifically
to
test
without
orientation,
because
we
don't
know
how
it
really
works.
Yet
so
just
one
of
those
sensorimotor
loop
are
we
going
to
actually
be
testing
or
exercising
in
this,
so
I
think
it's
fine.
The
way
it
is
now
moving
forward.
I
think
we
want.
A
A
I
see
that
I
must
just
be
missing
some
of
the
chats
that
role
by
I'm
sorry
guys
tack
down
us
and
you're
optimistic.
You
can
get
this
kick
into
five
years.
You
mean
the
the
2d
object.
Recognition
thing
oh
come
on,
we
have
to
just
we
haven't
even
got
to
the
point
where
defining
test
its
will
define
the
test
and
I
think
then
we
have
to
start
building
solutions.
A
A
A
Let's
just
work:
let's
I'm
gonna
I'm
gonna
do
formed
later
I
think.
Well,
let's
just
work
on
object.
Recognition,
I'm,
gonna
have
to
come
back
and
I.
Don't
want
to
just
go
through
all
the
all
of
the
forum
posts
while
I'm
sitting
in
front
of
you
guys.
Let's
do
something
a
little
bit
more
meaningful.
So
let's
do
some
technical
work.
A
A
A
Back
to
where
I
was
so
this
this
paper,
let
me
get
it
and
I
haven't
read
this
paper
in
a
while
I
think
it
was
a
SDR
paper.
It's
yeah,
so
I
mean
I,
don't
even
have
it
downloaded
here
it
is
the
HTM
spatial
fool
or
neocortical
algorithm,
it's
in
frontiers,
let's
and
he
is
it
probably
in
supplementing
with
supplementary
material
yeah.
A
Hopefully,
using
new
pics
spatial
polar
to
do
these
tests
and
it's
testing
entropy,
which
measure
which
measures
the
efficient
use
of
columns
I've
talked
about
that
before
when
we
talked
about
boosting
which
sort
of
boosting
entropy,
when
we,
when
we
do
that,
you
want
the
entropy
so
that
so
that
the
semantic
meaning
can
spread
has
a
space
to
spread
through.
If
you
get,
if
you
will,
if
you
latch
on
to
to
something,
then
then
you
overload
it
anyway.
A
B
A
A
B
A
All
right
so
I'm
gonna
try
to
I'm
gonna,
try
not
to
get
too
distracted
here.
I'm
gonna
turn
discourse
off
I'm
gonna
leave
chat
on
so
you
guys,
certainly
ping
me
if
you
have
questions
or
anything,
but
forgive
me
if
I
don't
take
any
off-topic
stuff,
because
I'd
love
to
try
and
knock
this
out.
This
is
my.
A
This
is
my
little
live
streaming
server.
It
goes
in
pulls
twitch
every
minute
to
get
how
many
viewers
I
have
and
and
updates
text
files
on
my
computer
and
then
the
broadcasting
software
reads
the
text
files.
That
seems
to
be
the
best
way
to
do
it.
Honestly.
I
tried
it
with
with
a
web
server
and
it
was
just
Anna
Anna
browser
overlay
and
it
just
didn't
work
well.
I
had
to
get
WebSockets
involved.
This
is
so
much
simpler.
Sometimes
it's
simpler
just
to
write
files
in
read
them,
then
HTTP
for
sure.
A
A
B
A
A
A
Hate
doing
that
without
being
an
inertial
environment,
but
my
Python
environment,
so
jacked
up
anyway.
It
just
is
what
it
is.
Okay,
so
at
least
I
know,
new
pic
is
installed
properly.
So
let
me
go
get
I.
Think
I
have
HTM
papers,
I'm
gonna,
open
that
in
a
new
shell
I
think
it's
an
NCI
HTM
paper,
I,
don't
I,
don't
have
it
on
this
Mac.
This
Mac
is
a
little
bit
newer
than
my
other
ones.
So
let's
go
get
it
clone
it
copy.
Let
me
make
a
fork:
I
think
I
have
a
fork.
A
I
do
okay,
so
I
have
a
fork
already
so
I
just
need
to
clone
this
and
set
it
as
origin,
and
then
the
other
one
will
be
the
upstream.
So
this
is
typical.
Get
stuff.
I'm
gonna
go
into
my
in
TA
directory.
That's
my
note.
In
the
Mendte
projects
get
clone
HTM
papers,
oh
it
already
exists.
I
didn't
it
didn't
have
a
okay!
So
it's
here
it's
here
anyway,
I
thought
I
had
it
Howard
was
my
remotes
I've
got
my
origin.
Is
mine
and
upstream
is
good.
Okay,
get
a
full
upstream
master.
A
Make
sure
my
origin
is
this
clean
in
case
I
create
a
PR
or
something?
Okay.
All
is
good.
Please
all
should
be
good
all
right,
all
right,
so
here's
where
we
are
I
want
to
go
into.
Neuro
can
no
frontiers
right,
it
was
frontiers
and
then
the
HTM
spatial
poor
frontiers,
the
HTM,
spatial
polar
and
here's
where
we
are
I
think
I
can
get
rid
of
this,
and
we
can
just
worry
about
this
guy.
A
Wear
and
tears,
wear
and
tear,
as
you
filled
the
profile
out
yet
now,
don't
know
much
about
wear
and
tear
been
around
since
February,
so
a
relatively
new
user
asking
or
trying
to
run
the
code
for
this,
both
the
1d
and
2d
examples
give
me
errors.
Okay,
the
errors
being
I
can't
see
this
very
well.
Let's
see
if
I.
B
A
It's
the
1d
case,
lots
of
parameters.
Does
it
a
mean
number
of
active
columns
trace
back
and
then
what
is
all
this
Jess
I
think
you
I
think
you
just
put
this
sort
of
just
denote
yeah,
so
I'm,
just
gonna
reformat,
this
I.
Do
this
a
lot
people
people
don't
post
code,
very
well,
okay!
So
that's
this
better!
Now
I
can
read
it
better,
all
right!
So
here's
what
they're
trying
to
Train.
So,
let's
see
I'm
just
gonna
I-
haven't
tried
this
at
all.
A
B
A
A
Spatial
object
has
no
attribute
local
density.
So
that's
not
even
the
error
that
he's
getting
mean
number
of
active
columns.
So
what
happened
with
mine
so
ran
something
something
the
trace
back
came
up
in
plot.
This
there's
probably
a
way
to
prevent
the
plot.
Let's
look
at
the
code.
Train
SP,
probably
some
options
in
here:
hey
Richard.
How
are
you
doing?
I
need
like
a
I
need
like
a
noise
or
something
and
when
a
chat
pops
up,
maybe
I
can
do
that.
It'd
be
nice
because
sometimes
chat
pops
up.
Oh
I,
have
no
idea
what.
A
Also
play
sound,
it's
just
so
quiet
I,
don't
hear
it
I
guess,
anyhow,
hi
Richard
I
am
I'm.
Trying
to
somebody
had
a
question
about
one
of
our
older
papers
on
HTM
spatial
pool
or
trying
to
run
the
code
that
should
run
and
having
a
problem,
and
this
error,
special
puller
object,
has
no
attribute
local
area
density
makes
me
makes
it
seem,
like
perhaps
they're
running,
an
old
version
of
or
they're
running
a
new
version,
maybe
the
new
pic,
because
this
may
have
been
one
of
the
breaking
changes.
B
A
A
This
is
not
a
good
warning
because
it
makes
it
seem
like
with
something
internal
broke,
because
let's
go
look
at
the
code
and
SP
metrics,
oh
I
think
there
is
something
broken
here:
let's
go
look
framework,
Oh
or
they're
using
an
old
version
of
HTM,
but
that
shouldn't
be
the
case
because
that
is
already
installed
that
should
have
been
so
so
I
can
see
some
of
the
problems
coming
up
here.
First
of
all,
what
version
of
new
pic
and
HTM
Research
are
expected
to
have
installed
here?
A
A
B
A
B
A
B
A
B
A
Being
passed
in
so
this
SP
should
have
a
property
private
property
that
this.
So
this
is
a
bad
I.
Hate
say
this:
this
is
bad
code.
This
function
is
has
this?
Is
a
code
smell
called
in
inappropriate,
intimate
wait?
What
is
it
intimate?
Something
like
something
knows
too
much
about
something
else,
so
this
function,
reconstruction
error
knows
that
this
SP,
that's
being
passed
into.
It
will
have
a
private
property
called
local
area
density.
Underscore
local
area
density
underscore
typically
means
it
shouldn't
be
used
outside
of
itself.
You
know
it's.
A
A
I'm
gonna
have
to
I
might
have
to
see.
I
can't
replicate
this,
because
why?
Because
matplotlib?
So
that's
that
kind
of
sucks.
So
let's
I
want
to
try
and
replicate
this
and
and
then
I
can
put
a
breakpoint
in
the
code.
So
I
need
to
move
past.
This
stupid,
Matt
flop
libero
just
by
this,
is
in
402
I'm,
just
going
to
prevent
matplotlib
from
trying
to
plot
anything.
A
B
B
A
It's
not
worth
looking
at
okay,
okay,
where
were
we
here,
I
was
in
the
HTM
spacial
polar
train
SP.
This
is
where
I
did
the
other
thing
and
then
I'm
going
to
SP
metrics
line
554,
so
SB,
no
open,
okay,
the
shortcuts
I'm
expecting
to
work.
Don't
work.
That's
us
new!
Oh
okay,
SP
metrics!
There
we
go
so.
B
A
B
B
A
A
A
A
B
A
C
C
C
A
A
A
A
B
A
B
A
Like
a
configuration
for
this
experiment
to
run
change,
data
set
at
don't
do
that
boosting
random
data
set
random
bar.
Oh,
it's
the
it's,
the
it's
the
tics,
it's
the
apostrophes
perfect
I
should
have
looked
a
little
closer
at
the
error
message:
shouldn't
I
have
okay,
so
I
don't
want
these
little
ticks
right.
Here,
barely
see
them,
but
there
they
are.
Okay.
Please
work
this
time
all
right.
Okay,
now
I
can
debug
it
and
that
breakpoint
should
catch
Python
to
development,
just
not
for
very
much
longer.
Okay,.
A
B
A
A
A
A
Does
it
even
have
it
so
it
does
not
have
the
what
was
it
called
local
area
density?
It
does
not
have
a
local
area
density.
That's
for
sure
this
might
be
a
legitimate
I.
Think
I'm
gonna
have
to
tell
Luis
about
this.
I
think
I'm
gonna
have
to
file
a
bug
about
this
because
I
don't
really
know
I,
don't
really
understand
what
the
problem
is.
It
seems
like
it
might
be
an
issue
with
swig,
because
this
is
a
swig
object
and
it
may
be
that
we
changed
yeah.
A
A
A
Not
just
kind
of
call
it
a
bug,
we're
gonna
label
it
a
bug
and
I'm
gonna
sign.
Not
might
not
me
I'm
sign
Louis,
because
the
one
that
will
end
up
working
on
it
certainly
needs
a
form
of
constant
accessory
up
honestly
yeah.
This
just
looks
bad.
You
can't
any
time
you're
doing
your
dot,
underscoring
anything
that
that
is
not
part
of
your
class
hierarchy.
A
A
B
B
A
Wrapper-
and
it
has
no
local
area
and
city
property,
where
did
you
say,
Richard
probably
needed
some
this
problem?
This
probably
needs
to
be
addressed
in
this
in
the
cease
to
by
adding
a
some
type
of
constant
accessor
I'm.
Not
gonna
worry.
The
Matt
live
is
my
problem
that
this
codes
problem-
and
this
is
old,
so
I'm,
gonna,
I'm,
gonna,
I'm
gonna,
keep
my
complaining
to
new
code,
be
instances
where
ever
it
has
an
ogre
property
may
need
to
add
a
constant
accessor
in
the
C++
bindings
I'm
going
to
CCU
Richard.
A
A
Down
how
long
did
that
take
me?
You
guys
timing.
A
half
an
hour
had
Dredd
stuff
like
that,
because
I
saw
that
yesterday
and
I'm
like.
Oh,
that's,
that's
gonna,
take
a
while
mark,
see
the
last
post
and
how
to
generate
place.
Fields
and
grid
cells,
see
abstract
of
linked
posts.
We
here
we
show
that
a
large
fraction
of
medial
in
Toronto
cortex,
neurons
fire,
specifically
when
mice
or
at
given
distances
and
directions
from
spatially
confined
objects.
A
These
object,
vector
cells,
are
tuned
equally
to
a
spectrum
of
discrete
objects,
irrespective
of
their
locations
in
the
test.
Serena
yeah,
so
I
am
familiar
with
object,
vector
cells
to
a
point
at
least
I've
heard
them
being
discussed
a
bunch
in
research
meetings,
especially
between
Marcus
and
Mirko.
So
they've
got
a
bunch
of
ideas
about
that.
I'm
I
haven't
been
keeping
up,
that's
pretty
deep
into
research.
If
you
want,
if
you
want
to
as
a
window
into
that,
just
keep
watching
research
meetings
because
they'll
eventually
come
up
again.
A
This
algorithm,
decodes
rat
squeaks,
been
repeated
question
of
a
possible
research
problems.
Here
is
an
interesting
one.
I
want
to
skip
this
I
might
take
a
look
at
it
later:
okay,
okay,
oh
yeah!
This
was
the
question
from
Freeman
hi
Freeman.
If
you're
watching
thanks
for
your
interest
in
for
starting
this
topic,
I
think
it's
interesting.
I
didn't
quite
understand
what
you
meant
at
first
I
was
just
thinking
along
the
lines
of
anomaly
detection.
A
Sorry,
so
it
matters
whether
it's
going
up
or
down.
So
your
arch
sort
of
trying
to
do
a
sequential
classification
and
so
I
think
that
you
would
need
to
do
run
temporal
pooling
with
it,
and
some
profiling
is
hard,
we're
still
trying
to
figure
out
how
that
works.
So
he's
just
saying,
after
discussing
with
Paul
thanks
Paul
for
helping
him
out,
I
understand
it's
possible
to
prevent
nupic
from
getting
accustomed
to
the
spikes
in
the
presence
of
seasons.
A
A
A
A
Discussions
on
the
idea
of
pooling
input
in
the
context
of
orientation,
so
just
for
with
the
post
I'm
at
right
now
wait
a
minute.
I
thought
you
were
I
thought
you
were
still
referring
to
border
or
object,
vector
cells,
so
you're
talking
about
different
attacks.
Slope
is
part
of
the
issue
with
this
time.
Step
is
critical,
yeah
and
he's
saying
that
maybe
if
he
can
get
the
prediction
or
a
history
of
predictions
or
even
maybe
you
could
do
step
ahead
prediction
like
you'd
predict.
Oh,
you
could
get
the
prediction
and
your
other
point
was.
A
It
is
oh,
oh,
the
attack
slope
thing:
yeah
I'm,
not
missing
anything
on
this
because
I
think
I'm,
not
sure
I.
Certainly,
if
he's
trying
to
do
sequence-
classification,
you
just
you
can't
do
it
with
new
pick
today.
So
I
want
to
be
clear
about
that,
but
he
might
be
able
to
do
step
up
step
down
identification
with
a
combination
of
anomaly
detection.
A
So
how
big,
at
this
time
step
like
how
big
a
change
is
happening
so
so
the
only
thing
you
can
do
that
I
can
think
of
with
new
pick
is
have
some
step
ahead.
Classification
of
what
you
think
the
prediction
is
going
to
be
I,
don't
know
if
it's
valuable
to
have
it
warned
versus
many
steps
ahead.
Maybe
one
would
be
just
as
good
as
as
many
and
won't
be
as
intense
to
do.
A
Just
like
the
anomaly
likelihood
window
does
a
simple
does,
a
pretty
simple:
you
know
mean
type
of
thing:
I,
don't
know
exactly
what
it
does.
There's
videos
about
it,
but
yeah
you
could
do
another
a
whole
other
anomaly.
Detection
on
the
anomaly
score.
I,
don't
think!
We've
tried
that.
But
it's
that's
interesting.
That
could
certainly
be
done.
A
A
Okay,
so
Jeff
introduced
the
big
idea
that
the
initial
function
performed
on
input
that
comes
into
the
cortical
column
is
to
pool
those
inputs
with
an
orientation
signal
and
the
discussion
you
concluded
that
the
output
of
this
function
would
be
equivalent
to
place
cells.
Styria
makes
a
lot
of
sense
to
me
to
fill
some
gaps
in
earlier
models
like
sensory
orientation
scaling.
A
However,
where
I
became
a
little
stuck
and
understanding
of
theory
is
with
what
happens
after
the
initial
function
and
the
white
board
diagram
Jeff
depicts
the
output
of
place
cells
being
used
as
the
input
to
the
layer
which
gets
its
context
from
grid
cells
representing
location.
One
of
the
things
I
want
to
do
for
because
I'm
gonna,
I'm
gonna
get
rid
of
this
I
know
what
to
do
here.
A
One
of
the
things
I'm
going
to
do.
We
have
we're
supposed
to
have
an
AI
chat.
Ai
chat,
Monday
at
1:00,
is
I'm
going
to
I'm
going
to
take
that
the
cortical
column
diagram
that
Jeff
true
and
I'm,
going
to
summarize
it
and
I'm
going
to
in
a
way
I'm
going
to
try
and
summarize
it
and
I
think
the
way
I'm
going
to
summarize.
It
is
attempt
to
do
a
data
flow
diagram.
A
So
this
might,
this
would
be
if
this
will
be
a
fun
thing
to
do,
because
I
I
don't
quite
get
the
whole
the
whole
picture
yet
I'm
still
trying
to
put
it
together.
So
this
will
be
a
learning
experience
for
me
too.
So,
if
you
haven't
watched
this
live
stream
that
Paul's
talking
about
and
you're
going
to
attend
a
I
chat
on
Monday.
You
might
want
to
check
it
out.
It's
got
a
link
to
it
here
anyway.
A
A
It
seems
like
the
outfit
of
the
first
function
is
the
location.
The
output
of
the
first
function
is
in
a
location,
though
it's
it's
really
all
about
orientation.
I
thought
the
air
it's
a
place.
It's
not
a
location
is
not
a
place
right,
so
the
place
that
is,
that
gets,
triggered,
sort
of
or
emerges
from
an
orientation
and
a
sensation
or
those
as
you
as
you
move,
the
orientation
I.
Don't
know.
I
don't
have
this
in
my
head
all
the
way
yet
I
might
try
and
pitch
this
to
Jeff.
A
Sometimes
so
he'll
hear
he's
actually
looking
it
out.
Let
me
let
me
burn
through
this
before
I,
try
and
figure
it
out.
Sometimes
it's
best
to
read
the
whole
thread,
not
understand
and
come
back
and
see
what
everybody
filled
in
okay.
Additionally,
I
wasn't
really
able
to
identify
on
the
whiteboard
which
of
the
layers
would
represent
a
stable
representative
representation
of
the
object
just
from
the
lines
on
the
whiteboard.
A
It
would
make
sense
to
me
if
l5
or
performing
a
pooling
function
and
its
outputs
and
its
output
depicted
stable
object
representations.
This
would
explain
the
lateral
connections
with
l5
and
other
columns
for
voting.
That's
the
one
that
Jeff
didn't
seem
to
have
an
idea
about,
or
he
didn't
say
anything
about
them,
but
he
did
say
they
were
voting,
but
not
really
necessarily
about
what
I
was
I.
A
Keep
thinking
that
has
to
do
with
motion
has
to
be
like
voting
about
what
most,
what
what
motion,
or
it's,
either
motion
or
time,
I
always
think
of
motion
in
time
as
being
very
similar
things
in
the
brain
anyway,
because
because
the
brain
experiences
time
through
motion,
so
I
think
it's
I,
think
it's
technically
for
the
brain,
the
same
thing
time
and
motion:
okay,
just
for
the.
Where
was
I.
This
would
explain
the
lateral
connections
with
l5
and
other
columns
and
the
projections
through
the
thalamus
to
another
region
supporting
composite
objects
in
hierarchy.
A
A
A
B
A
A
This
point:
I
I'm,
gonna,
defer
to
Jeff
and
what
I
do
is
I
just
sent
him
an
email,
I
say:
Jeff,
here's
a
good
question.
You
should
answer
this
because
it's
pretty
detailed
and
he's
the
guy.
That
knows
more
of
the
answers
than
me.
So
give
me
a
sec
I'm,
just
gonna
compose
an
email
offline
and
ask
Jeff
to
answer
this
and.
C
A
A
Because
it
seems
like
pretty
good
questions
and
I,
don't
know
the
answers,
and
maybe
there's
some
good
ideas
in
here.
This
still
doesn't
explain
what's
happening,
l6b
as
we
were
saying
why.
Why
would
l5
need
input
from
l23,
representing
specific
relation
to
the
object
and
from
l6b
also
representation
representing
a
specific
location
on
a
specific
object?
I,
don't
know,
I,
don't
know.
A
A
A
A
Anyway,
when
you
turn
in
a
second
system
changes
the
second
system
changes
when
you
move
and
turn
the
same
time.
Both
systems
are
impacted,
they
feed
off
each
other.
We
necessary
I've
been
thinking
about
this.
The
whole
morning,
I'd
like
to
speculate.
Please
take
this
with
a
grain
of
salt
when
you
move
in
a
straight
line,
essentially
you're
considering
features
in
two
dimensions:
features
come
closer
or
go
together.
A
Only
when
you
add
a
second
system
that
somehow
works
with
an
angular
input,
your
brain
can
build
a
model
in
three
dimensions.
So
this
is
somewhat
similar
to
we're
talking
about
this
object.
Recognition
project
we're
keeping
it
in
two
dimensions,
so
we
don't
have
to
deal
with
orientation
orientation
is
almost
like
another
dimension.
You
know
you're
adding
to
the
problem.
A
I
guess
it
is
it's
sort
of
like
adding
that
third
dimension,
then,
because
you're
no
longer
restricted,
but
even
in
two
dimensions,
you
can
have
a
an
agent
that
has
an
orientation
in
the
environment
and
as
a
part
of
its
movements,
is
turning
and
changing
its
orientation.
So
when
Jeff
saw
the
idea
of
our
object
recognition
project,
he
liked
it
and
I
thought
one
of
the
reasons
he
liked.
It
was
because
you
start
without
orientation,
and
you
may
be
able
to
add
it
at
some
point
in
the
future
as
a
part
of
the
recognition
tasks.
A
A
Okay,
there
is
something
at
work
to
form
the
cluster
of
features
that
we
can
identify
as
a
Moser
grid
cell
being
active
when
certain
inputs
are
present.
I
think
that
time
Demento
will
show
how
the
thousand
brains
theory
builds
up
to
doing
this
I
think
the
Calvin
tiles
method,
executes
coding
shows
great
promise,
but
until
I
get
demo
code
running
just
show
this
in
action,
it's
just
another,
pretty
theory
just
to
clarify
relationship.
A
A
This
is
this:
is
the
mini
columns
operating
within
a
cortical
column,
if
I
can
think
of
it
that
way,
I
have
to
think
of
it
within
a
cortical
column,
because,
honestly,
because
Jeff
restricts
the
research
thinking
for
Numenta
them
to
really
trying
to
understand
the
column
as
its
own
entity.
We're
it's
it's
very
hard
for
me
to
branch
off
and
think
too
much
about
columns.
A
Ok,
so
you
don't
you're,
not
restricted.
You
don't
have
that
restriction.
Are
you
talking
to
me
mark
or
tachyon
anyway?
I
just
want
to
make
sure
we
can
think
about
this
as
one
column
called
the
black
spots
or
mini
columns.
The
colored
lines
are
competition,
winning
lateral
connections,
ok,
the
red
one,
the
red
in
one
map
and
the
blue,
another
okay.
B
A
This
is
these:
are
these
seem
like
grid
cell
bumps
to
me
grid
cell
module?
Bumps
is
what
I
hear
them
called
whenever
I
see
them
in
the
math.
Even
the
mathematicians
call
on
this.
They
they're
sort
of
emergent,
Mexican,
hat
style,
bumps
that
emerge
in
in
these
grid
cell
patterns,
and
you
can
make
them
move
mathematically,
but
by
changing
by
changing
asila
Tori
features
at
least
I've
seen
that
so
I've
seen
you're
just
getting
sort
of
a
taste
of
the
research
that
I've
seen
as
I
continue.
Recording
research
meetings.
A
I
think
you
guys
will
get
a
better
understanding
of
what
we
what
we
have
thought
about
and
what
have
not
thought
about,
because
I'm,
not
the
best
one
to
inform
you
about
that.
Like
like
I,
said
I'm,
not
on
the
research
team,
I
sort
of
have
a
window
into
research
and
I
mean
we.
The
the
math
is
way
over
my
head,
but
there's
some
super
detailed
models
of
grid
cell
behavior.
A
A
This
is
total
speculation
on
my
part,
but
there
could
be
that
many
column
activations
in
one
layer
can
trigger
grid
cell
module
activations
in
another
layer
and
there's
they
sync
together.
Somehow
I
mean
I
think
that
that
is
probably
how
some
of
this
is
working
and
and
a
sensation
can
spark
a
grid
cell
module
to
sort
of
sync
up
with
an
input
or
something
like
I
said.
This
is
speculation
on
my
part,
but
I
have
to
think
about
it
within
a
corporate
column.
A
So
I'm
glad
that's
closed.
So
I'm
glad
that's
closed.
So
the
yellow
spots
are
in
a
third
map
where
the
projections
from
the
red
and
blue
Maps
coincide.
So
all
this
hex
grid
stuff
is
that's.
It's
really
reminiscent
of
from
some
of
the
high-level
mathematical
stuff
I've
seen
people
do
with
grid
cell
modules,
so
before
Merkel
left,
maybe
I
can
get
this.
A
That's
that
separates
it's
anyway.
He's
got
a
model
that
emerged
where
he
he
would
have
bumps
like
this
over
here
and
they'd
have
some
pattern
and,
and
they
would
converge
in
the
middle
very
smoothly,
and
then
this
one
would
have
a
completely
different
pattern
of
grid
cell
bumps,
maybe
at
a
different
scale
or
a
different
orientation
and
but
the
the
line
between
them.
He
worked
the
math
out.
So
there's
this
nice
sort
of
convergence
of
you
ever
seen.
A
One
of
those
trippy
screensavers,
you
know
or
you've
got
two
patterns
that
could
have
converged
in
the
middle
and
and
I
think
that
looked
really
promising
to
be
like
that
idea,
because
you've
got
to
have
these
different
sort
of
grid
grid
cell
grid
fields
or
something
acting
together
across
your
whole.
Cortex
they're,
not
all
going
to
be
the
same
same
logic
is
okay,
good,
yeah,
but-
and
the
thing
is
I
mean
that
gives
me
a
confidence
that
we're
on
the
right
track,
because
I
know
Mirko
doesn't
read
the
okay
he
just
like
he
doesn't.
A
He
doesn't
have
much
of
a
public
persona.
I
guess
you
know,
he's
a
a
German
mathematician
and
super
smart
guy,
but
it's
hard
to
follow
his
work.
You
know,
and
before
he's
gone
on,
to
work
with
ill
Afiya,
the
paper
that
he
just
wrote
with
Marcus
he's
going
to
work
for
illa
who's,
one
of
the
authors
of
the
paper
so
he's
going
to
be
doing
similar
stuff.
He
came
to
new
Mendte
because
he
wanted
to
understand
the
mathematical
properties
that
he
that
he
could
he
here's
what
he
said.
He
said
I'm
paraphrasing
him.
A
He
said
he
saw
some
of
the
beautiful
concepts
of
mathematics
in
the
brain
exposed
in
some
of
the
our
some
of
our
theories
and
he
wanted
to
explore
it.
He
wanted
especially
grid
sails.
He
was
super
excited
about
grid
cells,
so
all
of
us
all
of
his
work,
I
think
in
one
way
or
another
had
something
to
do
with
grid
cells
and
how
they
could
how
they
can
emerge
and
how
they
can
interact
with
each
other
and
what
properties
they
have
very
interesting
stuff.
A
Okay,
all
right!
Moving
on
I
got
through
everything.
Okay,
let's
do
some
work.
Let's
do
some
work.
I've
been
doing
this
for
an
hour.
It
takes
me
an
hour
to
get
through
form
stuff
yeah.
So
that
could
be
so
that's
interesting.
You
can.
You
can
certainly
put
a
bunch
of
these
grid
cell
modules
together
to
form
other
representations
that
you
might
use
to
identify,
bigger
things
or
a
collection
of
things.
You
know
all
right,
so
I'm,
happy
I
got
that
done
here.
I'll
have
to
worry
about
this.
Not
to
worry
about
this.