►
From YouTube: HTM Hackers' Hangout - Mar 1, 2019
Description
Discussion at https://discourse.numenta.org/t/htm-hackers-hangout-mar-1-2019/5516.
HTM Hackers’ Hangout is a live monthly Google Hangout held for our online community. Anyone is free to join in the discussion either by connecting directly to the hangout or commenting on the YouTube video during the live stream.
Columns Paper: https://numenta.com/neuroscience-research/research-publications/papers/a-theory-of-how-columns-in-the-neocortex-enable-learning-the-structure-of-the-world/
Columns+ Paper: https://numenta.com/neuroscience-research/research-publications/papers/locations-in-the-neocortex-a-theory-of-sensorimotor-object-recognition-using-cortical-grid-cells/
More info on all these topics at http://numenta.org.
A
A
So
at
in
the
forum
page
for
this,
so
I've
listed
these
topics
out
and
I'm
going
I,
don't
even
have
the
controls
right
anyway.
Oh
there
we
go
okay,
okay,
first
thing,
I
thought
three
yeah
I'm
gonna
mute
everybody
also
for
now
so
Python
2
is
dead.
There
is
no
Python.
2
is
like
end
of
life
is
like
January.
After
that,
there's
no
guarantee
tip
install
new
pic
will
ever
work
on
any
environment
created.
After
that
point
in
time,
so
that's
a
big
problem.
We
want
like
dip.
Install
new
pic
was
a
big
deal.
A
It
got
a
lot
of
people
available
a
dit
available
to
a
lot
of
people,
so
we
have
a
big
some
big
decisions
to
make
the
Numenta
about
how
we're
going
to
handle
this.
You
know
that
we
have
options.
One
do
nothing,
that's
an
option.
Two
is
update
new
pic
to
use
Python
3,
which
would
be
new
pic
2,
which
would
be
a
bunch
of
cleanup
and
we
would
take,
and
then
we
have
a
bunch
more
decisions
to
make
right.
A
We're
gonna
make
all
these
decisions
in
March
I've
got
a
meeting
scheduled
when
everybody's
in
the
office
at
the
same
time
and
we're
gonna
get
it
figured
out
before
the
deadline
comes
so
by
mark
by
January
we're
either
going
to
have
a
new
pic.
That's
Python,
3,
compliant,
so
that
people
can
pip
install
new
pic.
It's
not
gonna,
be
any
enhancements.
It's
really
just
going
to
be
clean
up.
The
API
is
gonna,
be
as
similar
as
possible
and
we
might
remove
the
OPF.
A
That's
a
part
of
it.
Just
for
cleanup.
I,
don't
know,
there's
a
lot
of
stuff
on
the
table.
Anyway,
it's
a
big
deal
and
it's
a
lot
of
work
and
there's
a
lot
of
decisions
to
be
made.
We're
not
taking
it
lightly,
I,
just
wanna,
let
you
guys
know
and
I
am
paying
attention
to
what's
going
on
in
the
fork.
The
community
fork,
which
is
the
next
thing
I
wanted
to
talk
about,
and
David
you're
on
online
it'd
be
great.
If
you
could
give
a
quick
update
because
I
see
a
lot
of
activity
well,.
B
A
I,
just
nothing
in
particular,
but
anything
you
want
to
know.
That's
perfect
and
and,
like
I
said,
everything
is
on
the
table
right
now.
We
haven't
made
any
decisions
about
what
we're
going
to
do,
but
we
know
that
new,
pic
as
it
is
right
now
is
not
going
to
work.
It's
going
to
be
basically
a
legacy
project
after
January.
A
Unless
we
do
something
I,
don't
think
we
want
to
do
that
so
and
definitely
if
we
I
in
particular,
because
I'm
gonna
be
doing
a
lot
of
this
work,
I
mean
it's,
it's
gonna
be
Louis
and
myself
that
are
gonna,
be
doing
whatever
work
for
the
most
part
and
subitize
e
has
time
there.
That's
gonna
be
doing
this
transition.
This
Python
3
migration.
If
we
end
up
doing
it,
I'm
gonna
look
through
the
community
for
pretty
thoroughly
to
see
what
you
guys
have
done.
A
I
know
that
you've
pulled
swig
out,
I,
think
and
PI
bind
is
in
place
now.
That's
definitely
something
I
would
want
to
do.
Getting
rid
of
serialization
is
definitely
something
I
would
do
so.
There's
a
lot
of
alignments
there.
We're
gonna
do
what
makes
sense
and
we'll
be
in
communication.
You
know
I
will
definitely
as
a
community
manager,
so
I
don't
think
any
of
your
work
has
been
in
vain,
even
if
we
do
end
up
putting
out
a
new
pic.
That's
just
Python
3
compliant
just
to
put
that
stake
in
the
ground.
A
You
know
in
using
the
lessons
that
you
guys
have
learned
to
do
that,
even
if
it's
not
the
same
codebase
and
you
guys
continue
going
off
in
your
direction
because
I
do
not
know
what's
gonna
happen
with
new
pic
in
the
future.
We
one
of
the
options
we
always
have
is
to
start
fresh.
You
know
if
we've
got
if
the
theory
has
progressed
to
a
point
where
it
doesn't
make
sense
to
pull
all
hold
this
stuff
along
with
us.
It's
always
another
option
and
everything's
on
the
table.
So
I
just
want
to.
A
A
A
I
really
love
to
share
everything
in
it,
but
we
anonymously
submitted
it
to
a
conference
and
we
can't
really
share
it
until
the
conference
tells
us
whether
it
got
in
or
not
so,
I,
don't
want
to
say
any
details
about
it,
and
at
this
point
the
more
I
leak,
the
more
confusing
it
is,
and
the
more
confused
I
make
everybody
so
just
hold
tight.
The
paper
I
think
is
going
to
be
good,
I've
reviewed
it
a
bunch
of
several
times
and
I'm
gonna
I'm
working
on
a
video
that
will
explain
it
like
here's.
A
You
know
this
type
of
deep
learning
setup
and
here's
how
you
can
add
like
HTM
stuff
to
it.
You
know
what
I
mean
so
just
keep
that
I'll
get
that
done.
It's
like
the
main
thing
that
I'm
working
on
right
now,
but
we
want
to
make
sure
that
we
use
all
the
right
terms
and
everything
and
and
we're
communicating
properly
in
you
know
what
I'm
creating
versus
what
subitizing
Louis
are
writing.
A
Okay,
next
topic
was
swarming.
There's
a
Sam
hi,
sir
man,
who
is
usually
at
these
as
it
was
doing
something
with
swarming
who's
asked
me
questions
can
I
trust
you.
No,
the
swarm
gave
me
back
an
adaptive,
scalar
encoder
and
that's
an
encoder
that
we've
we
don't
even
use
anymore,
can
I
trust
that
the
parameters
are
good
and
I.
Think
short
answer.
No.
When
it
comes
to
swarming,
you
can't
trust
anything.
A
A
So
you
can't
trust
any
it'll
point
you
in
the
right
direction
sometimes-
and
you
almost
always
have
to
take
the
parameters
of
swarming-
gave
you
especially
the
encoding
parameters,
because
it's
decent
ant
is
at
least
finding
out
what
input
fields
contribute
to
prediction
accuracy,
so
it
can
at
least
get
you
in
la
pointed
in
the
right
direction.
But
for
the
most
part,
swarming
is
a
very
old
tool.
We
never
use
it
at
nomensa.
I,
don't
know
the
last
time
we
used
it.
Swarming
it's
been
years.
It's
it's
hard
to
use.
It's
we
don't
like
it.
A
It's
it's
buggy!
You
know
it's
barely
maintained,
so
I
mean
my
my
honest
advice
is:
don't
use
swarming
if
you,
if
you'd,
unless
you
know
exactly
what
you're
doing-
and
you
know
the
exact
parameters
that
you
want
to
permute
over.
In
that
case
it
can
be
really
useful,
but
that's
like
an
advanced
topic.
I
like
I,
couldn't
even
tell
you
how
to
do
that.
You'd
have
to
go
dig
through
examples
in
the
source
code
to
figure
out.
There's
this
permutation
stop
pie
file
that
you
have
to
define
the
different
parameters
you
want
to
permute.
C
A
So
I'm,
sorry
that
I'm
such
a
waste
of
space-
that's
true,
you
Jacob,
he
just
joined
it
heard
me
say
that
awesome,
okay
anyway.
The
last
thing,
unless
anybody
else
has
anything
is,
is
the
2d
object,
recognition
thing
and
I've
got
a
lot.
I
want
to
say
about
that.
So
before
that,
before
we
go
into
that,
I
will
open
it
up
and
I
saw
Josh.
Disjoint
I've
got
everybody
muted.
A
So
if
anybody
wants
to
talk
unmute
yourself
and
we'll
chat
about
anything,
otherwise,
maybe
you're
just
waiting
for
me
to
go
into
this
okay,
so
I'm
gonna.
Do
that,
then
let
me
so
first
of
all,
I
want
to
emphasize
gotta
what
I
got
a
whiteboard
I'm
gonna
draw
some
stuff,
so
I
want
to
emphasize
that
we're
starting
this
whole
2d
object
recognition
project,
it's
really
stupid,
simple,
like
it
has
to
be
stupid,
simple,
the
more
complex
you
make,
the
environment
the
harder
it's
going
to
be
to
solve.
A
So
I
I
really
want
to
keep
it
so
simple
that
that
we're
focused
entirely
on
the
structure
of
the
network,
everything's
wired
together,
and
that,
for
example.
So
if
you
have
an
input
space,
that
is
just
really
small-
and
this
is
just
a
grid
or
a
2d
grid
or
whatever,
and
we
send
it
in
an
input
that
takes
up
this
part
of
it
and
that's
one
object,
and
then
we
send
an
input.
The
takes
of
this
part
of
it
and
that's
another
object.
A
Then
we
send
one
that
takes
up
the
middle
and
there's
three
objects,
and
we
want
to
classify
that.
That's
stupid,
simple!
That's
what
I
want
to
start
with
something
really
seized
or
something
as
simple
as
there's:
here's
a
box
over
here
and
then
here's
another
one
where
there's
a
box
over
there,
those
two
objects,
I,
would
say
inside
some
space
that
we're
going
to
classify
okay.
A
A
You
guys
hear
me:
okay,
cuz
the
mics
way
over
there,
so
you
can
yeah,
okay,
great
okay,
so
the
columns
paper
gives
us
the
idea
of
an
object
layer,
and
this
is
a
pooling
layer
and
it
also
gives
the
idea
of
a
of
a
they
call
it
a
sensory
layer-
and
this
is
where
spatial
pooling
has
happened
and
there's
many
columns
here
and
there's
you
know:
cells
acted
within
the
mini
columns,
except
instead
of
like
standard
temporal
memory,
spatial
pooling
as
we
do
in
hot
jam.
The
disco
context
is
not
coming
from
itself.
A
A
Okay,
this
is
proximal,
I
should
probably
use
color,
so
I
was
gonna
use
colors
anyway.
So
there
is
also
a
an
input
here
and
I
think
this
is
also
a
proximal
input
and
then
there's
an
apical
feedback.
This
is
all
in
the
columns
paper
and
there
is
sample
code
that
puts
this
together.
I
think
I
linked
to
it
in
on
the
forum
thread
in
the
third
object
recognition
thing,
so
this
one
is
going
to
have
a
sparse
representation.
So
what
happens
here
is
over
time
as
okay.
A
A
And
then
we
have
a
feature
which
is
a
and
it's
encoded
into
some
other
message
here,
and
these
are
coming
like
this,
so
the
proximal
input
or
the
excuse
me,
the
distal
input
happens
first,
if
we
know
the
location
that
we're
about
to
that
we're
at
or
that
we're
going
to
sense.
That
happens
first
and
that
creates
predictions
here
of
what
we
think.
A
Inputs
going
to
be
again,
this
is
all
just
columns
paper.
Okay,
this
whole
thing:
I
can't
I,
don't
want
you
to
explain
the
columns
paper,
because
we
have
lots
of
videos
about
this.
So
does
anyone
have
any
questions
about
this,
because
I
could
talk
about
the
proximal
and
put
here
the
pooling
here.
The
important
thing
is
this:
lateral
connectivity
and
I
never
really
talked
about
this,
because
that
this
is
the
part
we
really
want
to
test.
A
I
think
this
is
the
cool
thing
that
we
can
do,
because
in
our
world,
we're
not
just
sending
one
sensor.
We're
like
we're.
Gonna
have
a
little
collection
of
sensors,
they're
gonna
work
together
to
classify
objects
they're
going
to
inform
each
other.
That's
one
of
the
crucial
points
of
the
columns
paper.
So
this
would
be
one
sensor.
This
whole
thing
would
be
the
column
associated
with
one
sensor,
and
we
have
another
sensor
over
here
and
it's
got
its
associated
structure.
A
You
know
it's
cortical
column
and
these
object
layers
are
going
to
be
communicating
laterally
to
share
representations.
So
not
only
is
this
going
to
get,
it's
is
going
to
activate
its
cells
here,
based
on
its
own
sensory
space.
You
know,
calculations,
it's
going
to
be
informed
by
all
the
other
sensors
around
it
and
they're
going
and
project
it's
its
reality.
So
this
is
sort
of
this
is
internal
reality.
Okay,
I
love.
Using
these
terms
like
this
is
internal
reality.
Nobody
sees
this
except
this
cortical
column.
This
is
what
the
current
the
column
projects.
A
This
is
its
representation
of
the
object
for
the
most
part.
It's
not
even
really
labeled
at
this
point,
I,
don't
think
it's
just
a
representation,
but
what
we
can
do
is
we
can
say
at
a
higher
level.
We
can
look
at
all
of
these
columns
and
see
the
different
representations
that
each
one
of
the
paths
and
we
can
write
classifier
on
them,
so
we
can
tell
which
ones
are
thinking
what
object
it.
So
we
can
have
a
probability.
A
Okay,
it's
everybody
follow
me
yep,
okay,
all
right,
so
the
hard
part
comes
in
when
you
get
to
the
comms
plus
paper,
so
we're
still
going
to
have
yes
I'm
going
to
talk
about
columns
plus
this
is
a
much
more
complicated
paper.
I
haven't
done
a
video
on
this
there's
a
reason.
I've
got
like
all
these
drawings,
honestly,
so
many
drawings
about
columns,
columns,
plus
and
and
but
you
really
got
to
read
the
paper
and
figure
this
out,
because
so
not
only
do
we
have
the
object
later.
A
Okay,
so
we're
going
to
keep
this
from
the
columns
paper
and
now
and
we
have
the
sensory
layer
so
we'll
call
this
object
and
call
this
sensory.
So
we
still
have
this
interplay
that's
going
on
here
and
we
still
have
a
location
coming
in,
but
it's
not
coming
in
from
now
we're
defining
where
it's
coming
from
right.
A
So
now
we
always
draw.
We
always
draw
this
as
collection
of
grid
cell
modules,
so
this
location-
and
this
is
I-
think
most
of
the
confusion
and
especially
with
Marty
who's,
a
I
hope
you're
watching
at
some
point
Marty.
This
is
basically
the
grid
cell
mat
mechanism,
the
grid
cell
module
mechanism
occurring.
It
was
really
tempting
to
think
of
it
as
an
encoder
right,
because
it's
because
that's
the
way
it
looks
like
in
like
the
grid,
solid
GM,
school
episode
and
and
it's
I
even
showed
you
know
an
encoding
window.
A
That
says
you
could
easily
that's
it's.
You
could
totally
make
an
encoder
that
way
and
Marty's
done
that,
but
this
is
not
an
encoder.
This
is
a
layer
of
cells
and
the-
and
the
big
difference
here
is
an
encoder
is
just
one
way
it.
Basically,
it
takes
input
and
it
transforms
it
into
something
else.
This
is
getting
feedback.
It's
it's
projecting
location
here
and
it's
getting
this
feedback
back
here.
So
they
can
predict
what
it's
gonna
feel
the
next
time
it
goes
somewhere
because
that's
very
different
than
just
being
an
encoder.
A
This
has
to
be
a
layer
that
a
layer
of
cells
like
a
simulated
layer
of
cells,
I
think
that's
the
big
confusion
here.
You
can't
just
encode
it.
We
have
to
simulate
this
layer.
This
is
going
to
be
a
layer
in
the
network
that
is
running
red
cell
module
logic
over
the
2d
space,
so
so
I'm
gonna
figure
out
how
to
make
that
work.
So,
oh
quick
overview,
there
is
no
orientation
in
this
experiment
space.
A
There
is
no
displacement
in
this
experiments
plate
in
this
experiment,
space,
which
is
different
than
what
I
started
with
so
I
confusingly,
gave
a
description
that
included
displacement
cells.
Sorry
about
that
I
was
confused.
We
I
think
we
would
need
displacement
cells.
If
we
talk
about
agency
of
movement,
if
this
model
were
to
start
contributing
to
the
movement,
that's
being
generated
and
fed
back
in
I
think
you
need
displacements,
then,
because
movement
means
something
different.
You
know
it's
not
just
an
input
to
the
system
anymore.
A
It's
an
input
to
the
system
that
has
effect
that
a
direct
effect
on
where
you
are
and
where
you're
going.
You
know
it
means
more
than
than
just
an
input.
It's
like
core
to
your
definition
of
space.
You
don't
I
mean
okay,
I
wanted
to
make
sure
I
didn't
miss
anything
and
then
I'll
take
like
questions
and
stuff,
but
the.
But
the
big
thing
is
all
this
is
in
the
columns
paper
in
the
columns
plus
paper
and
there's
a
good
amount
of
code.
A
There's
a
lot
of
charts
in
those
papers
to
get
those
numbers
in
the
charts.
We
had
to
run
these
simulations,
so
there's
sort
of
the
small
implementations
of
these
Network.
These
networks
I'm
describing
not
the
whole
thing,
but
one
of
this
one
and
one
of
this
one
in
two
different
places
and
supporting
material
for
these
papers.
A
What
I
want
to
try
and
do
is
put
them
together
and
that's
really
the
the
crutch
of
the
of
the
project
and
and
we
should
be
able
to
test
it
like
we
could
write
tests
I
could
we
could
define
an
input
space
and
I?
Could
I
could
fill
in
part
of
it
and
say
this
is
an
object.
This
is
an
object.
This
is
an
object
and
then
train
train
of
system
that
maybe
just
has
random
movements
or
maybe
has
a
scripted
movements.
A
D
One
thing
I
wanted
to
add
on
the
the
reason
that
you
need
the
feedback
also
to
the
and
the
grid
so
modules,
and
they
you
can't
consider
it
an
encoder
is
because
the
the
representations
of
space
and
the
grid
cell
modules
are
specific
to
the
object
and
that's
actually
important
in
the
Collins
plus
paper,
because
you
could
be
predicting
to
three
objects
and
you
want
to
have
their
representations
of
the
space
all
at
the
same
time
until
you've
narrowed
down
what
specific
object.
You're
sensing
that's.
A
A
great
point,
and
as
this
thing
moves
like,
if
we
have
a
sensor
moving
through
space,
you
might
have
a
hit
here
and
a
locational
light
up
in
the
center
future
will
light
up,
and
you
have
a
really
dense
representation
of
in
this
object
later,
because
what
it
sees
is:
okay,
everything
I've
felt
in
that
location
and
any
object
before.
Is
it
because
it's
gonna
light
up
here?
So
if
you've
had
six
objects
that
feel
similar
in
that
place,
they'll
all
light
up
and
as
you
go
to
another
place,.
A
A
Right
and
if
you
think
about
it
well,
every
this
space
and
every
column
is,
and
the
representations
in
the
spacer
are
gonna,
be
unique
right
because,
because
the
sensors
all
have
different
experiences
with
features
in
space
over
time,
so
this
really
can't
be
shared
between
objects,
and
we
know
that
I
think.
That's
our
excuse
me
between
cortical
columns.
This
is
what's
sure
right.
There's
some
computation
down
here
in
the
guts
of
the
system.
There's
a
representation
of
space.
It's
not
shared
because
it's
it's
impossible
to
share
I!
A
Think
if
they're
all
initialized,
you
know
locally
right
there
and
there's
no
I,
don't
think,
there's
any
reason
why
they
would
share.
But
this
is
important
to
share-
and
this
bugged
me
at
first
a
lot
because
I
thought:
why
do
these
all
have
to
be
unique?
Why
does
every
cortical
column
have
to
have
a
unique
representation
of
space
it
feels
like
it
would
be
more
efficient
if
they
all
use
the
same
representation
of
space
and
then
they
could
do
direct
comparisons.
A
C
A
One
of
the
tricky
things
orientation,
I
think
is
a
big
deal
like
we
seem
to
have
a
whole
collection
of
cells
to
represent
orientation
and
then
around
cortex
called
head
Direction
cells,
and
it
makes
sense
that
there's
probably
something
similar
in
the
cortex
that
have
to
do
with
orientation.
It's
a
hard
problem,
but
the
research
team
has
been
like
thinking
about
this
for
a
long
time.
A
For
us,
if
we
were
to
do
orientation
here,
it
would
be
like
giving
this
agent
and
orientation
right
and
then
and
then
it
could
turn
and
and
then
turn
like
that
and
turn
like
that,
adding
that
is
almost
as
complicated
a
ding,
a
third
dimension
in
some
ways
you
know
to
this
problem
from
the
HTM
standpoint
like
this
is
a
hard
problem
in
HTM.
The
orientation
problems
that
we're
still
working
on
and.
A
I
think
so
I
think
so
and
if
you
think
about
it
from
the
standpoint
of
your
senses,
every
one
of
your
senses
when
you
go
towards
something
has
a
different
orientation
right
towards
that
object
and
that's
all
being
calculated
somehow
and
we
look
in
relation
to
whatever
object,
you're
interacting
with
every
every
point
of
contact.
You
know
you're
modeling,
some
type
of
orientation
there,
somehow
very
efficiently,
yeah.
D
And
it's
necessary
because
you
know
I,
look
I,
look
at
a
you
know,
look
the
coffee
cup
and
it's
up
straight
up
and
down
the
edges
and
whatnot
that
I'm,
detecting
in
my
the
sensors
of
my
eyes,
are
completely
different
than
if
it's
tilted
to
the
right.
So
you
know
that
the
knowing
what
orientation
your
sensor
is
in
relation
to
the
object
just
is
just
as
important
as
knowing
what
location
your
app
in
relation
to
the
object.
Well,.
E
C
D
Is
turned
to
the
right
and
I,
you
know
move
my
you
know
close
the
joints
on
my
finger,
I'm
going
at
I'm
going
toward
the
object
at
a
different
orientation
that,
if
I
turn
my
hand,
the
you
know
the
other
direction.
I
close
my
finger
and
I'm
coming
out
from
a
different
direction
and
I
can
I
can
predict
what
I'm
going
to
feel
on
that
object
before
I
touch
it.
So
I
have
to
be
knowing
what
direction
I'm
coming
toward
the
object
from
in
order
to
be
able
to
make
that
prediction.
D
A
And
it's
more,
it's
not
just
object
orientation!
That's
really!
Now
what
I'm,
what
I'm
referring
to
at
the
core
of
the
orientation
problem?
It's
it's
not
about
that!
It's
about
sensor
orientation
and
it's
maybe
it's
the
same
problem
just
looked
at
from
a
different
perspective,
but
that
I
think
the
brain
solves
is
from
a
sensor
orientation
standpoint,
not
an
object
orientation.
Standpoint,
yeah.
D
Yeah,
it's
it's
similar
to
location
bottom,
you
got
egocentric
and
now
eccentric
is
the
same
for
the
orientation
right.
Somehow
I've
got
changed.
Orientate,
transform
the
orientation
of
my
sensors
in
my
entire
body
into
an
orientation
of
the
sensor
and
in
object.
Space
just
like
I
need
to
translate
a
location
in
relation
to
my
body
to
a
location
in
object.
Centric
space,
yeah.
A
That's
right
at
the
core
of
this
I
mean
this:
isn't
some
of
the
Jeff
talks,
but
all
the
time,
but
it
seems
pretty
obvious
to
me
that
there's
there's
there's
a
big
link
between
the
the
computations
that
are
happening
in
these
two
places
or
the
allocentric
the
egocentric
space.
As
far
as
sensor
in
relation
to
object
or
agent
in
relation
to
environment,
it's
the
same
problem
right
being
solved
in
both
of
those
situations.
Yeah
I
feel
probably
being
solved
in
very
similar
ways.
C
A
E
E
C
E
C
Within
the
filters,
so
I
was
thinking
of
just
doing
a
simple
thing:
first,
like
input
image,
one
convolutional,
filter
and
then
another
whatever
they
use
for
the
empty
convolution,
and
then
the
output
image.
That's
a
little
D
noise
from
the
D
noise
input,
image
I
believe
that
would
be
a
Gann
in
that
case.
So.
A
This
sounds
interesting
because
you
know
Siraj
revolved
a
ai
YouTube
guy.
The
machine
learning
guy
does
a
bunch
of
videos.
Anyway,
he
did
one
on
a
mento
once
and
he
concluded
that
HTM
was
pretty
much
an
auto
encoder
I.
Don't
think
he
got
the
whole
concept,
but
that
always
intrigued
me
so
maybe
you're
onto
something
well.