►
From YouTube: HTM Chat / AI / Neuroscience / Community chat
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
Okay,
so
this
was
from
this-
is
the
research
meeting
from
yesterday
no
Friday
excuse
me:
I
went
through
Lucas
who's
presenting
was
presenting
on
this
gave
me
a
bunch
of
more
information
notes
about
related
papers.
Specifically,
this
lottery
ticket
hypothesis
is
apparently
hot
topic
and
the
machine
learning
world
right
now.
A
Good
morning,
Falco
good
morning,
David
sorry
I'm,
your
hat,
David
you're
excited
were
doing
spacial,
pooling
what
the
woot-woot
was
for.
You
saying
I
I
have
to
apologize.
You
guys
ahead
of
time.
I
have
I
hurt
my
neck
over
the
weekend.
Moving
a
ping-pong
table
and
I
haven't
quite
recovered
I'm
having
a
hard
time.
Turning
to
the
left
and
looking
up
so
at
some
points
during
the
day,
I'm
gonna
be
taking
some
stretching
breaks,
I'm
gonna,
try
and
drink
a
lot
of
water.
Remind
me
to
strange
I'll,
probably
be
reminded
by
the
pain
anyway.
B
A
Meeting
there's
I
mean
there's.
This
seems
like
a
great
opportunity
for
HTM
inspired
ideas
to
work
in
a
machine
learning
system,
so
I
think
everywhere.
I'm
pretty
excited
about
this
I
mean
if
we
can
make
a
nice
impact
by
I,
don't
know
breaking
some
benchmark
performing
really
well
on
some
benchmark,
especially
with
noise
involved.
A
C
A
A
D
A
A
Anyways
go
to
my
screen.
Yeah
and
I'm
gonna
talk
about
some
forum,
stuff,
okay,
so
yeah
that
we
had
a
great
hackers,
hangout
I,
don't
for
you
guys
if
any
of
you
were
there,
some
of
we
do
our
hackers
hangouts
on
YouTube,
not
twitch,
so
in
case
you're,
just
watching
twitch.
Here's
a
link
to
this
thread
on
our
forum
and
we
had
a
good
hackers
hangout
with
several
people,
join
here's
of
the
link
to
that
video
or
sending
sending
the
elbow
links,
and
so
there's
some
discussion
about
this
hangout.
A
We
had
Marcus
who's,
one
of
our
researchers
on
a
research
team
or
primary
researcher.
Aside
from
Jeff
at
the
moment,
that's
researching,
like
sensory
motor
integration
type
of
stuff.
He
was
online
during
this
video
and
took
a
bunch
of
questions.
Serious,
that's
true
anyway.
Marcus
is
great.
It's
very
happy
to
have
him
on
so
he
took
a
bunch
of
questions.
A
A
Me
turn
discord
on.
Let
me
join
discord,
chat
I
mean.
If
anybody
wants
to
chat
with
me
here.
Let
me
send
the
discord
link,
I'm
gonna,
get
a
voice,
chat
and
I
may
want
to
turn
this
down
if
I'm
getting
on
voice
chat,
so
anybody's
free
to
jump
on
voice
chat
if
they
want,
because
I
wanted
to
have
an
understanding
of
this
idea
of
neurons
antennae
and
I.
Think
there's
I,
don't
think.
There's
I
just
pulled
some.
That's
not
what
I
wanted
to
do.
A
Neurons
antenna
is
I,
think
a
familiar
idea
for
me:
I
don't
but
I
just
wanted
to
understand.
I
didn't
leave
the
paper.
The
widespread
I
have
not
read
this
paper
yet,
but
the
summary
says
so
we're
talking
about
dendritic,
dendritic
synapses
and
one
of
the
key
parts
of
HTM
theory
is
in
sequence,
memory.
A
The
temporal
memory
part
of
it
is
this
idea
of
how
dendritic
distal
dendritic
segments
are
what
they're
there
for
and
I
think
that's
what
they're
relating
this
to
you
know
treating
these
distal
dendritic
segments
as
calling
them
antennae,
because
they're
I
think
things
are
far
away.
It
is
sort
of
like
an
antenna,
you
know
except
we're,
not
looking
for
light
or
signatures
were
or
pattern
doing.
D
A
D
A
That's
what
the
paper
wide
neurons
have
thousands
of
synapses
is
all
about,
so
they
are
monitoring
dendritic
and
somatic
apartments
in
the
awake
primary.
The
visual
cortex,
looking
at
layer,
five
pyramidal
neurons,
so
Mata
and
distal
apical
truck
dendrites.
So
surprisingly,
high
functional
correlation
yeah
so
like
that,
probably
because
the
distal
dendrites
are
a
predictive
of
the
neuron.
A
This
strong
coupling
persists
across
their
activity
levels
because
it's
a
core
function
of
the
cortex,
that's
one!
It's
your
cut,
you're,
doing
a
temporal
prediction
constantly
as
you're
modeling
things
and
and
like
moving
through
spaces.
That's
why
that's
what
this
is?
Even
if
you
imagine
moving
through
space,
this
is
happening.
A
A
D
Found
that
the
activity
was
kind
of
alternating
between
how
the
cells
used
to
fire
and
how
they
are
so
so
imagine
rats
in
a
room,
their
reward
spread
around
the
room
and
in
these
consistent
places
within
an
hour
later,
those
places
are.
We
are
we've
rearranged
those
now
the
rewards
are
placed
in
different
locations
from.
A
For
more
info
on
this
lunch
hour,
Hector's
hangout
that
I
linked
to
earlier
so
he's
referring
to
that
alternation.
So,
essentially
you
learn
a
room.
You
learn
where
the
rewards
are
in
the
room.
You
know
you
come
back
to
the
same
room.
So
there's
some
identifying
feature
to
room.
You
come
back
to
the
same
room,
but
the
rewards
are
all
in
different
places.
So
over
time,
as
you
experience
these
two
different
situations
in
the
same
room,
you
create
I,
guess
different
grid
cell
activations.
What.
A
Others,
even
though
it's
the
same
room,
you've
created
two
different
mappings
of
grid
cells
on
to
that
room,
one
for
one
reward
arrangement,
one
for
another.
Sorry
so
you're,
like
tricking
your
brain
into
creating
two
maps
of
space
right
through
there
for
the
same
place,
and
this
idea
of
alt.
It's
not
really
tricking.
Your
brain
I
think
your
brain
just
does
this,
but
this
is
sort
of
the
trick
to
make
it
do
that.
You
know
as
this
experiment.
It's
really
interesting
experiment.
A
So,
okay,
so
the
grid
cell,
alternating
back
and
forth
in
time
when,
like
the
animal
needs
to
sort
of
refer
to
both
representations,
I
would
say
so
his
explained
as
the
same
as
in
the
ID
lab,
where
I
found
this,
as
in
a
union
merged
previous
blobby
experience
for
two
reference
frames
into
one
that
takes
into
account
directional
movement,
I,
haven't
seen
Gary's
work
in
a
while.
Gary
makes
these
really
cool
like
sim.
You
like
life,
simulations
they're,
really
neat
hit
like
this.
This
is
pretty
bad.
Let's.
C
B
B
A
I,
wouldn't
be
surprised
like
if
you
think
about
simultaneous
predictions,
it's
sort
of
similar
die
think
to
simultaneous
predict
shows.
You
can
simultaneously
keep
two
competing
concepts
in
mind
at
the
same
time,
until
you've
gathered
more
information
and
then
one
falls
away,
I
think
wonder
what
he
used
for
his
hex
representations.
He
did
all
this
in
an
old
old
version.
Of.Net
I
think
it's
like
vb.net
right
Gary,
if
you're
watching
on
YouTube
later
I'm,
pretty
sure
this
is
all
done
in
vb.net
check
the
post
date.
He
mentions
it
on
the
forum
you
can
do.
A
This
is
so
cool
though
I
love,
this
visualization
I,
don't
I,
don't
quite
get
the
hexagonal
thing
at
this
point,
but
but
I
think
what
he's
his
point
is.
He
sees
the
same
phenomena
in
the
the
system
he
created.
This
thing's
got
goals
and
everything
like
if
you
see
it's
going
to
get
food
and
it
stays
out
of
the
parts
to
hurt
it.
It's
got
all
these
sensor,
arrays,
all
these
little
things
around
it
or
sensors.
E
A
A
E
A
That
that's
what
I
started
doing
to
on
twitch
I
still
want
to
come
back
to
that,
because
I
think
that's
still
fun.
I,
there's
still
I
think
we
can
build
it
so,
where
we
left
off,
I
was
expecting
to
become
easier
to
just
include
the
alternating
between
timeframes,
part
than
watch.
It
go
on
its
mount.
Wait
alternating
between
time
frames
right,
alternating
the
like
representations
of
where
it's
at
between
time
frames.
A
Fixable
thanks
for
following
it's
much
appreciated
talking
about
brains,
they
have
any
precautions
about
brains,
we're
trying
to
well.
We
we
had
a
hackers
hangout
if
you've
been
watching
it
on.
The
discussion
is
like
all
about
the
representation
of
space
and
we're
in
an
orientation
of
objects.
Space
around
objects
objects
how
they're
oriented
how
we
can.
A
A
C
D
A
Is
part
of
key
to
the
HTM
theory
that
we
propose
at
Numenta
is
that
they
each
have
their
own
reference
frame?
They
each
have
their
own
area
in
space,
where
we
take
sensory
input
that
we
perceive
or
in
that
frame
when
we
sort
of
like
attach
it
in
locations
relative
to
each
other,
and
we
think
group
we
use
grid
cells
to
do
that,
because
that's
sort
of
the
brains
mechanism
that
we've
found
that
happens
to
work
in
the
hippocampus
and
we've
seen
evidence
of
it
in
the
neocortex
as
well.
A
So
we
know
there's
this
type
of
mapping
happening,
or
at
least
this
representation
of
space
there's
a
there's
a
way.
Your
brain
represents
three-dimensional
space,
two-dimensional
space,
probably
any
dimensional
space,
and
it's
probably
using
the
spritzel
mechanism,
and
using
this
mechanism
we
can
theorize
about
how
objects
can
be
constructed
out
of
other
objects.
A
super
fascinating
stuff,
great
great
works,
thanks
Falco
for
the
explanation.
B
A
A
Okay,
so
I
was
just
reading
the
forum.
That's
another
topic.
If
you're
new,
we
have
a
forum
with
lots
of
friendly
people.
If
you
have
any
questions
and
you're
afraid
to
ask
me
on
twitch
or
you
or
whatever,
you
can
just
post
a
message
on
the
forum.
Somebody
hopefully
help
you
out.
Okay,
so
we're
back
to
this
2d
object,
recognition
challenge
that
I
was
working
on
when
I
first
started,
twitch
he's
referring
to
and.
A
Okay,
then,
have
to
write
code
to
manually
control,
all
that
extra
code
nation
Bagram
to
compensate
for
not
having
full
control
of
its
body.
In
that
case
it
would
also
be
a
force
sensor.
Antenna
dendrix,
starter
model
force
o
for
sensor,
so
this
is
a
little
more
detailed
than
I
thought.
Maybe
I
don't
quite
understand
that,
because
I
hadn't
actually
read
all
the
way
to
the
completion
of
this
thread.
B
A
But,
but
maybe
but
I'm
not
sure
what
more
to
say
about
that,
the
only
information
going
the
other
direction
is
whether
the
prediction
helped
right,
because
the
dendrite
has
to
know
if
it
becomes.
If
it's
if
a
dendritic
spike
occurs.
All
of
those
segments
that
contributed
to
the
dendritic
spike
need
to
know
whether
it's
that
correctly
indicated
the
firing
of
the
soma
so
that
they
can
increase
their
permanence
values,
increase
their
synaptic
weights,
but
I
think
that's,
that's.
A
A
A
A
Host
the
synaptic
PSP
postsynaptic,
something
from
the
distal
dendritic
regions
on
the
on
off
DCC
are
heavily
attenuated
by
propagation
to
the
soma,
heavily
okay,
so
heavily
attenuated
by
propagation
of
some
amines
affected.
If
but
wait,
but
that
spikes
initiated
within
local
dendritic
regions,
dendritic
spikes
right
or
are
we
talking
about
somatic
spikes
propagate
with
high
power
village?
The
student
must
be
dendritic,
spikes
to
the
soma
and
back
propagate
to
the
remainder
of
the
dendritic
tree.
Oh
so,
oh
I've
heard
about
this
before
I.
First
I've
heard
them
talk
about
this
at
research
meetings.
A
A
Segments
of
the
dendritic
spike
that
get
information
about
whether
the
cell
fired
the
entire
dendritic
tree
does
therefore
active
amplification
of
D
s,
which
I
think
means
appears
to
take
place
during
spike
initiation
in
the
difference.
So
when
the
soma
fires
there's
some
indication
back
to
the
dendrites
that
it
has
fired,
which
makes
sense,
this
seems
close
to
what
you're
describing
if
inhibition
is
optionally
included.
Then
back
propagation
can
be
stopped.
B
A
A
Would
be
reinforcing
the
spike
causing
loss?
There's
Marc
Brown
from
Jordan.
Weird
idea
keeps
coming
to
me
that
it
is
the
fire
together.
Wire
together,
thing
is
partly
or
wholly
a
local
thing,
based
at
the
synapse
level,
I
envision,
a
mechanism
where
the
products
of
the
metabolism
of
firing
interact
with
you
and
extracellular
chemicals
to
promote
growth.
I
extend
this
with
thought.
Some
residue
of
this
experience.
A
metabolic
residue,
makes
the
synapse
acceptable
to
the
soma
based
firing
wave
as
the
local
backprop
training
wave
as
a
vision,
divided
event,
so
I
think.
D
A
A
I'm
understanding
this
am
I
missing
something
because
this
type
of
back
propagation
that
we're
talking
to
it's,
not
the
I,
don't
think
it's
a
now,
and
now
it's
it's
not
the
same
type
of
back
propagation
that
we
talked
about
neural
networks,
because
that's
going
throughout
huge,
hugely
deep
hierarchies.
We're
talking
about
back
propagation
of
information
within
local
areas
and
I.
Think
that's
what
the
king
is
sort
of
pointing
out
here
now
this
extension
that
some
residue
of
an
experience
makes
a
synapse
susceptible
to
the
soma
based
firing
wave
yeah.
That's
how
that's
sort
of.
A
A
I,
don't
trust
this
right
away?
I
think
we
need
a
paper
right
brains,
activity
doesn't
influence
of
physiological
building
up
their
sex
cells.
What
you
think
can't
be
inherited,
but
now
it
looks
like
we
may
need
to
rethink
this
Oh
ourselves.
We
passed
information
down
several
generations.
Okay,
I
think
I'm,
not
services
related
at
least
we're.
Not
thinking
about
this.
All
so
Gary
says
don't
stop
there
keep
going
almost
everyone
reading.
This
is
probably
wondering
background
reading.
A
A
What
are
you
reading?
The
predicted
model
assumed
is
some
part
of
the
dendritic
Arbor
fires
on
a
sensed
pattern,
making
the
prediction
that
we
will
fire
again
soon.
The
prediction
is
a
partial
depolarization,
but
not
a
fire.
Another
wait
soma.
Its
friction
is
central
to
this
young
later,
but
not
much
later.
The
main
Sola
fires
with
big
events
on
the
solar
level
and
the
action
potential
fans
out
through
the
output
axon
and
back
up
the
dendritic
tree
muscle.
That's
how
the
dendrites
work
the
back
propagation
wave
that
we're
calling
it.
A
A
B
A
I,
don't
think
so,
the
the
basic,
the
sequence
memory
of
each
everything
is
built
on
that
that
is
not
tentative.
That
is
something
that
we're
pretty
certain
is
gonna
happen
or
is
happening
basically
this
this
paper
that
just
came
out
weird.
When
did
this
come
out
2019
days
ago,
we
wrote
about
this
eight
years
ago,
so
we
published
our
first
white
paper
basically
explaining
this
eight
years
ago,
and
that
is
like
core
to
HTM
theory.
That
is
not
tentative.
That's
I
think
that's
like.
If
that
doesn't
work
the
way
we
think
it
works.
We're.
A
A
It
really
needs
to
somewhere
in
the
cortical
mass,
be
found
a
whole
bunch
of
smaller
brains
with
the
moving
straw
fields
of
the
outside
world.
In
fact,
I
was
currently
hoping
that
the
theory
would
end
up
needing
to
include
at
least
some
of
the
individual
neurons,
and
in
that
case,
T
V
D
only
becomes
even
more.
A
E
A
A
E
A
A
A
Let's
talk
about
this,
this
is
so
cool,
okay,
so
Marty
get
it.
Maybe
you're
not
online,
but
I
know
you're,
probably
watching
later
I'm
very
I'm
super
excited
by
all
the
work
that
you're
doing,
and
this
particular
is
awesome
because
it
validates
cuz
here.
I
know:
you're,
writing
your
own
systems
and
you're,
finding
the
same
things
that
we're
finding
and
I
think
this
is.
A
So
he's
basically
taken
the
things
we
found
in
our
paper.
How
can
we
be
so
dense
which
is
attempting
to
take
aspects
of
spatial
pooling
and
apply
it
to
deep
learning
networks?
Can
you
can
see
all
this
and
just
thread
if
you
want
to
read
it,
but
essentially
this
is
the
direction
the
log
part
of
our
research
team
is
going
right.
Now
we're
trying
to
figure
out
how
can
we
apply
htm'
ideas
to
deep
learning
networks
and
and
what
Marty's
done
is
basically
validated
this?
A
Because
with
noise,
it's
particularly
it's
really
good
at
defeating
noise
and
that's
something
deep
learning.
Networks
are
not
very
good
at
so
like
we're
talking,
you
know
these
types
of
numbers,
recognition
and
pattern
recognition.
No,
this
is
noise,
accuracy,
I,
think
it's
a
pattern,
recognition,
accuracy,
classification,
accuracy
or
something
that
are
really
low
with
typical
cnn's.
A
But
when
we
add
this,
instead
of
doing
a
rail
ooh
activation,
we
all
do
a
que
winners
and
also
enforce
potential
pools
with
initial
weight
settings
and
run
this
boosting,
which
means
that
you
have
to
run
this
at
each
time
step
if
you
want
to
get
really
good
numbers.
That's
the
problem
because
currently
deep
learning
doesn't
allow
you
to
do
that
very
easily
to
run
processes
after
each
data
point
you're,
usually
ghulam
together
thousands
of
data
points
and
run
this
huge
cumulative
mathematical
function
over
it
to
find
the
you
know:
what
do
they
call
it?
A
D
A
Do
this
by
themselves,
but
with
boosting
it's,
it's
that's
what
that's
like
the
secret
sauce,
because
you
have
to
take
this
and
will
and
I
we're
gonna
get
to
boosting
and
building
HTM
systems
at
some
point
too,
and
we'll
and
I
think
that
would
be
a
good
visualization
is
if
we
show
a
histogram
of
knowledge,
representation
or
active
duty
cycles.
Let's
say
a
histogram
of
the
active
duty
cycles,
which
is
basically
how
active
many
columns
are
over
time
and
without
boosting.
A
A
Can
can
represent
a
lot
so
they
do
when
you
let
them
just
run
the
first
ones
that
win
just
keep
winning
on
all
the
passes.
So
what
the
boosting
does?
Is
it
like
squashes
them
down?
It
pushes
them
down.
It's
like.
Okay,
you
guys
are
too
active,
we're
not
gonna.
Let
you
represent
these
new
paths,
this
new
patter
coming
in
we're
going
to
randomly
pick
something
another
mini
column,
and
this
is
happening
in
your
brain
to
I
mean
this.
Is
this
is
brain
inspired
this
homeostatic
idea
of
homeostatic
regulation?
This
is
completely
brain.
A
Inspired
who
stings
like
not
accelerating
this?
We
add
accelerating
the
strengthening
of
it's
like
artifice.
It's
not
artificial
I
mean
this
is
what's
going
on
in
your
brain.
It's
it's!
This
inhibit.
It's
part
of
the
inhibitory.
I.
Think
it's
in
the
I
did
a
video
on
boosting
and
I
and
I
pointed
to
a
paper
that
I
was
reading,
so
I
would
go
there.
I
could
probably
find
it
htm'
school
whose
stick
that's
the
boosting
one
now.
A
A
A
Scholarpedia
anyway,
when,
when
you
a
used
to
work
here,
he
was
he
explain
this
all
to
me,
you
a
is
a
real
hardcore.
Neuroscientist,
like
he's
he's
a
PhD
of
neuroscience.
He
worked
in
a
wet
lab
with
monkeys
for
before
he
came
and
worked
with
us,
so
he
seemed
to
nobody
was
talking
about
and
he's
you
know
he
knew
about.
He
knew
what
this
was
he's
like
yeah.
This
is
explain
to
me
how
it
worked
in.
D
A
A
If
everything
goes
well,
oh
man,
if
I
can
still
move
I
should
be
able
to
have
something
fun.
To
show
I
got
a
new
guitar
I
know
that's
super
exciting
to
all
of
you.
I
will
show
you
the
guitar,
hopefully
Friday
I'm,
hoping
it
shows
up
on
Friday
I've
had
my
saxophone
in
consignment
for
months
and
months,
and
nobody
would
buy
it.
And
yesterday
my
wife
comes
home
and
she's
like
you
got
this
check
from
village
music
and
they
had
signed
they
sold
it
and
just
sent
me
the
check
I
can
get
the
guitar.
A
A
A
B
A
D
A
A
B
A
A
A
D
A
A
It's
a
real
it's
a
pretty
hot
day
yesterday
and
today
we're
pretty
hot
during
Watsonville.
It's
supposed
to
be
well,
I
mean
hot
for
Watsonville,
all
right,
so
I'm
going
to
I
have
a
copy
of
this
locally
and
I'm.
Just
gonna
peruse
around
in
here.
I
really
wish
I
could
still
sneeze
I'm
in
building
a
team
system,
so
I
don't
want
to
be
in
their
gym
so
before
I
do
anything.
A
A
I'll
go
it
I,
don't
know
what
I'm
gonna
do
here
aside
from
just
perhaps
set
the
stage,
because
because
what
I
want
to
do
is,
if
you
guys
remember
our
last
pH
TMS
session,
we
were
working
on
potential
pools
and
initial
permanence
values.
At
this
point,
I
want
to
create
a
spatial
Pooler
class,
and
this
is
I,
am
gonna.
I.
Think
I've
been
thinking
about
this
and
I
think
it
should
be
pretty
functional.
A
A
Out
of
the
diagram
state
and
we're
going
to
keep
it
somewhere
else
and
we're
going
to
use
a
spatial
pool
or
class
for
the
logic
there
now
there's
a
lot
of
ways
we
could
go
about
doing
this.
I
am
in
favor
of
keeping
it
object-oriented.
I
am
not
necessarily
in
favor
of
mutable
functions,
so
I
think
we
will
try
and
do
a
hybrid
object-oriented
functional
approach
here,
which
is
sort
of
what
nupoc
does
as
well
and
I,
never
liked
how
it
did
it
until
I
thought
about
doing
it
myself
and
now.
A
A
Potentially
other
things,
but
I
want
to
make
all
of
the
functions
as
much
as
possible,
functional
and
immutable.
So
you
pass
it
in
arrays
of
data
representing
the
input
and
an
empty
array,
essentially
of
what
the
active
mini-com
activations
are,
which
are
empty
and
it
passes
it
basically
populates
your
array,
that's
what
new
pick
does
we
don't
have
to
do
it
just
like
that?
But
the
point
is
that
we're
not
gonna
maintain
all
of
that
state.
A
Initially,
at
least
in
the
spatial
cooler
we're
going,
we're
gonna
we're
gonna,
keep
it
out
and
the
spatial
cooler
is
going
to
be
like
a
track.
It's
gonna
keep
track
of
how
much
data
it
seen.
It
might
keep
a
cache
of
some
things,
but
this
cells
state,
the
cellular
state,
is
just
simply
in
our
like
an
array
of
it's
just
a
data
structure
of
weights,
it'll
be
for
each
mini
column,
there'll,
be
an
array
of
indices
of
the
potential
pools
and
a
matching
array
of
permanence
values
that
each
one
matching
one
of
those
cells.
A
A
Connection,
we
need
those
two
values,
so
we
I
think
we'll.
We
might
not
keep
track
of
that
in
the
spatial
polar
it'll
just
return
that
that
way,
we
can
keep
the
state
arrays,
primitive
and
that's
sort
of
the
idea
is
because
I
want
this.
Data
I
want
these
visualizations
to
be
super
close
to
the
data,
so
I
want
to
keep
the
data
primitive
operate
upon
it.
A
Using
a
spatial
cooler
class
potentially
track,
what's
been
happening
as
we
interact
with
that
class
in
the
class
state,
but
the
actual
state
of
the
cells
is
going
to
be
primitive
and
it's
going
to
be
kept
outside
and
well
useful.
We'll
have
that
as
a
page
level
variable
that
all
of
these
diagrams
can
then
have
an
opportunity
to
display
and
publish
the
sources
in
the
page
to
be
public,
will
I
publish
the
source
in
the
page
I,
don't.
A
Mean
you
mean
the
code:
the
source
code
is
all
going
to
be
open.
Source
I
mean
simple,
HTM
is
right
here,
I
mean
so
we're
going
to
have
n
source.
We're
going
to
have
spatial
Pooler
will
probably
create
a
algorithms
folder.
That's
special
cooler-
and
you
know
our
building
HTM
system
simply
depends
upon
this
project.
So
all
of
the
clean,
clean
HTM
code
is
going
to
be
in
here.
It's
going
to
be
heavily
tested
and
it'll
be
no
visualizations,
and
then
the
building
HTM
systems
is
all
the
diagrams.
C
A
Put
the
snippets
in
I'm
just
sort
of
driving
down
the
whole
visualization
pathway
first
and
then
we'll
write
the
prose
and
put
in
the
text.
You
know
just
like
I
did
with
this.
So
it's
going
to
be
like
this
for
encoding
numbers
except
the
spatial
Pooler
and
lots
of
descriptions
about
what's
occurring
and
all
that
jazz.
A
A
So
just
thinking
about
where
I
might
like
to
get
Marty's
on
dis
board,
where
I
might
like
to
put
tests.
So
let's
I
always
want
to
start
with
tests.
So
let's
create
an
algorithms
folder
here
and
it
might
be
fun
to
just
write
a
few
tests
so
that
when
we
start
the
stream
after
the
research
meeting,
we
just
have
a
place
to
start
so
I.
A
A
Sort
of
a
behavior
driven-
this
is
not
a
three:
it's
a
behavior
driven
testing
platform
for
node
and
I.
Don't
know
I
just
I
like
the
I,
like
the
syntax
I
like
describing
tests
this
way.
So
it's
it
makes
sense
to
me.
So,
let's
create
in
here
a
new
file,
I
call
it
spatial
cooling
will
call
it
spatial
cooler
tests
and
we'll
call
the
class
the
spatial
cooler.
A
A
A
A
Let's
just
say
it
creates
the
tent
a
potential
pool
for
each
mini
column.
All
right,
that's
clear
enough
right
if
it
creates
a
potential
pool
for
each
mini
column,
all
right
that
we
are
also
going
to
test
that
it
just
while
I
have
this
in
my
head
creates
an
initial
permanence
value
for
each
cell
in
each
potential
pool
of
each
mini
column,
pretty
explicit
right.
But
that's
that's
what
we
exactly
what
we're
going
to
test
here.
A
A
A
A
A
I'm
gonna,
I'm,
gonna,
I'm,
gonna
I
always
do
this,
so
just
bear
with
me:
I'm,
going
to
read
ascribe
some
of
this
stuff
and
and
narrow
down
these
tests
as
I
go
along.
Let's,
let's,
let's
make
this
a
little
bit
more
granular
when
you're
writing
tests
right
and
unit
tests,
write
them
as
granular
as
possible.
Let's
say
spatial
polar
instantiation
of
Wynn,
creating
the
potential
pools.
Okay,
let's
so,
let's,
let's
write
a
function
for
that
and
then
this
is
going
to
be
in
here.
A
When
creating
potential
pools
first,
we'll
say,
creates
a
potential
pool
for
each
mini
column.
Let's
move
this
here,
so
we'll
so
we'll,
so
this
will
be
the
spatial
pool
or
that
we're
using
to
test
when
creating
potential
pools.
So
so
first
I'm
gonna
just
write
as
many
tests
as
I
can
with
100%
connectivity.
A
Things
are
gonna
happen
here,
and
this
is
going
to
be
just
seriously
overly
descriptive,
but
there
we
go
the
connected
percent.
So
so
so
now
we've
set
up
our
our
tests
to
have
test
something.
So
we
know
we're
going
to
create
a
special
ER
and
we're
going
to
we're
checking
the
potential
pools
when
we
get
tell
it
it's
a
hundred
percent
before
you
say
it.
C
A
I'm
doing
test-driven
development,
it's
pretty
common!
That
I
will
write
like
flesh
out
a
bunch
of
tests
and
to
sort
of
define
an
API
before
I
write
any
code.
So
that's
sort
of
what
I'm
gonna
do
here.
Okay,
so
I
know
that
this
is
sort
of
a
way
that
I
want
to
create
a
special
puller.
I
know:
I,
want
to
test
that
you
know
go
through
each
mini
column
to
make
sure
there
is
a
potential
pool
here.
It
doesn't
matter
what
type
of
connectivity
it
has
for
this
test,
but.
C
A
A
Alright,
let
me
run
this
and
see
I'd
love
to
just
keep
running
it.
Make
sure
this
fails.
We
should
have
one
failing
tests,
I
probably
broke.
Something
cannot
find
module.
Oh
yeah,
there's
no
spacial
Pooler.
So
let's
do
that
and
the
source
I
need
a
new
folder
called
algorithms
and
then
I'm
going
to
take
the
structure
of
one
of
these
create
a
new
file.
Spatial
cooler.
A
C
A
B
A
That's
all
I'm
testing,
which
is
accessible.
So
if
I
create
an
SP
here,
it
doesn't
matter
what
the
connected
percent
is
in
fact
I'm,
not
even
going
to
give
it
a
connected
percent.
It
should
just
default
to
something
I
should
be
able
to,
and
now
here's
where
I
make
the
decision.
How
do
I
get
to
how
do
I
get
the
many
columns
or
how
do
I
get
the
the
initial.
A
A
So
so
one
thing
is
the
potential
pools,
don't
change
for
the
life
of
the
spatial
polar
so
perhaps
we'll
attach
that
to
the
state
of
the
spatial
cooler
and
the
spatial
pool
are
when
instantiated
is,
must
be
directly
connected
to
an
input
space
like
in
order
to
create
a
spatial
cooler.
You
have
to
know
the
dimensions
of
the
input
space
that
it's
going
to
pull
over.
A
So
in
this
spatial
cooler,
we'll
start
we're
going
to
start
really
simple,
okay,
so
this
might
this
isn't
going
to
be
the
way
it
ends
up,
but
we're
going
to
start
simple.
We
need
to
tell
it
the
dimensions
of
the
input
and
we're
typically
do
this
with
an
array
and
say
a
hundred
things
and
let's
assume
at
least
initially,
that
we're
only
going
to
handle
one
dimensional
input,
we're
not
doing
topology
and
we're
not
doing
local
inhibition,
because
that's
going
to
simplify
everything
everything
drastically
so.
A
So,
instead
of
this,
let's
say:
input
count.
I
know
this.
Eventually
we're
going
to
change
this
and
we're
going
to
have
a
multi-dimensional
input.
It's
not
just
going
to
be
a
count.
It's
going
to
be
an
array
of
counts
for
the
different
dimension
sizes,
but
for
right
now
assume
1d
input.
That's
that's
the
assumption
that
we're
going
to
make.
So
all
we
really
need
to
tell
it
is
how
big
is
the
input
space?
It's
assuming
one
dimensional
input,
no
topology,
no
local
1d
input,
global
inhibition,
no
topology.
So.
A
A
A
A
A
A
So,
let's,
let's,
let's
put
these
as
constants
here,
input
count
actually
yeah
yeah
input
count
100.
A
A
A
D
C
A
A
A
A
Just
to
differentiate
input
index
versus
pool
index,
okay
because
of
the
pool
index
is
just
the
index
of
the
P
array,
and
the
input
index
is
the
actual
value
of
that
index,
which
should
point
back
to
the
input
space.
So
in
this
case,
cert
are
same
or
are
equal.
Equal
equals
34
and
put
index
index.
So.
A
As
soon
as
we
don't
move
this
down
to
less
than
that,
this
will
fail
because
we'll
be
missing
an
index
and
this
working
okay.
So
let's
I'm
gonna
start
with
that.
Obviously
this
is
gonna
fail
and
we
are
not
going
to
write
the
code
for
this.
We'll
do
that
this
afternoon,
but
at
least
I
have
a
test
and
we
have
sort
of
a
starting
point
and
I
won't
be
like
starting
a
Pleakley
blind.
A
C
A
B
A
B
A
I'll
have
to
make
some
frequent
trips
to
the
restroom,
because
I'm
trying
to
drink
a
lot
of
water
today,
it's
supposed
to
be
good
for
some
more
muscles
got
a
really
really
stiff
neck
today,
because
I
moved
a
ping-pong
table
over
the
weekend.
That's
what
happens
when
you
reach
40!
You
come
incapacitated
simply
by
moving
a
ping-pong
table.
A
So
I
was
just
sitting
here,
thinking
about
spacial
cooler
class
and
whether
it's
how
much
state
it
should
contain
and
I
know
I
was
just
saying:
maybe
it
will
keep
the
permanence
values
outside
of
the
spatial
cooler
just
so
react
and
get
to
it,
but
I
don't
think
we
need
to
do
that.
All
we
need,
we
can
keep
it
all
in
the
spatial
Pooler
and
just
create
local
excuse
me,
like
page
level
Global's
that
are
primitive
for
race,
that's
what
will
trigger
redraws
and
then
on
each
step.
A
We
just
replace
those
with
whatever
we
get
from
the
spatial
floor.
So
that
should
be
fine.
Permanence
is
an
active
column
indices.
We're
not.
There
will
be
active
column
indices
soon,
but
we're
not
quite
there
yet
right.
Now,
all
all
word
we're
going
to
the
only
state
we're
going
to
make
accessible
for
the
diagrams
that
are
the
potential
pools
and
the
permanence
is.
The
potential
pools
won't
change.
The
permanence
is
we
will
change
as
the
spatial
Poehler
learns
and
I
think
I.
A
A
No,
so
the
so
I
think
all
we
need
is
one
array.
It
would
contain
two
things
in
index,
two
in
the
input
and
a
permanence
all
the
time.
That
way,
we
don't
have
to
keep
track
of
any
input
cells
outside
of
the
potential
pool.
We
that
way,
that
array
would
only
contain
connections,
nothing
that
could
never
be
connected.
I
think.
A
A
We
could
also
have
an
array,
that's
the
same
size
as
the
input
and
have
you
know,
a
an
undefined
weight
and
then
not
keep
track
of
the
indices.
You
know:
what's
it
saying
that
the
index
is
the
same,
that's
the
other
way
to
do
it,
but
I,
don't
necessarily
like
having
undefined
values
and
arrays
it's
hard
and
it
messes
with
diagrams
I'm
going
to
do
that.
You
have
to
make
special
conditions.
D
C
A
C
A
Simple
HTM
is
really
just
meant
to
be
a
reference
implementation
of
HTM
that
goes
along
with
this
document
that
other
people
could
use
it.
You
know
and
then,
like
I
said,
I'll
probably
go
look
at
all
lambs
JavaScript
HTML
plantation,
which
is
most
likely
a
lot
better
and
more
efficient.
But
it's
that's
not
my
goal.
Efficiency
is
definitely
not
the
goal.
Just
ease
of
display
and
understanding
I.