►
From YouTube: HTM Hackers' Hangout - Oct 4 2019
A
Welcome
to
HTM
hackers
hangout,
it
is
Friday
October
4th
I
want
to
just
go
over
a
couple
things:
I'm
I'm,
Matt,
Taylor
I'm,
the
mint
a
community
manager
I've
got
a
little
bit
of
an
agenda
cut
to
go
over
first,
and
this
is
alive.
Let
me
explain
the
meeting
first,
a
little
first
we'll
go
over
my
whiteboard,
so
I'm
gonna
talk
about
real
quick
for
the
Hangout.
What
is
this
meeting
and
then
we'll
do
a
quick
review
of
what's
going
going
on
with
memento
for
any
new
audiences
coming
in
here?
A
So,
first
of
all,
what
is
this
meeting?
So
we've
I've
done
this
as
a
monthly
way
for
years.
Actually,
we've
done
this
for
years
monthly
meetings
where
we
have
a
live
stream.
It's
usually
it's
usually
been
on
Google
Hangouts
for
anyone
in
the
community
and
when
I
say
the
community
I'm,
typically
being
our
mailing
lists
or
our
forums
or
you
step
mailing
lists
and
they
turned
into
forum.
So
it's
now
mostly
a
forum
but
I,
always
post
it
on
Twitter
too.
It's
open
to
anybody.
Sometimes
we'll
get.
A
You
know
eight
or
ten
people
join
most
of
the
time.
It's
you
know
four
or
five,
some
6
7,
something
like
that.
It
always
depends
on
people's
agendas
and
a
lot
of
them
are
across
the
ocean,
so
the
timing
doesn't
work
out,
but
anyway
we
do
this
meeting
really
to
sync
up
with
the
community
and
to
allow
people
in
the
community
to
bring
up
issues
they
want
to
talk
about
with
with
me
or
an
event.
Asipi
I
try
to
pull
in
folks
from
the
research
team.
A
A
A
I
was
thinking
about
this
and
there's
a
lot
of
people
that
probably
aren't
really
aware
of
the
history
of
what
we've
been
doing,
because
it's
we
tend
to
have
this
cycle
a
lament
at
the
company
that
I
work
for
researching
and
then
creating
applications
of
research
and
they're
going
back
to
research
and
then
creating
applications
of
research.
So
that's
sort
of
been
the
cycle
that
we've
been
going
through
for
years.
I
entered
the
company
like
right
around
this
transition
and
move
and
took
over
as
the
open
source.
A
A
A
We've
come
out
in
the
mid
to
hit
before
I
even
joined
that
they
were
just
coming
out
of
this
research
base,
where
they
had
figured
out,
like
the
sequence
memory
algorithm,
essentially
on
top
of
spatial
cooling,
and
that
depended
on
these
things
that
active
dendrites,
specifically
the
biology
biological
inspiration
for
this
like
phase
of
research
and
what
that
means
for
sequence,
memory
and
the
brain,
then
they
transitioned
to
attempt
applications
of
this.
In
you
know
the
application
space
most
of
mostly
focused
on
anomaly
detection,
because
that
was
with.
That
was
something
we
could
do
well.
A
We
don't
predict
very
well
right
now
without
a
more
realistic
cortical
structure,
I
think,
but
for
anomaly
detection
we
had.
There
are
actually
some
interesting
applications
there
and
I
still
think
there
are
there.
Just
a
lot
of
this
was
early
to
market
required,
really
fast
streams
of
data,
a
lot
and
a
lot
of
folks
around
this
timeframe,
weren't
ready
to
the
stream
status
anyway.
We
pulled
out
of
that
at
some
point
and
we
put
out
a
whole
bunch
of
open-source
application.
A
A
A
This
was
when
you
know,
Marcus
was
really
involved
here
and
turning
the
focus
on
grid
cells,
location
representation
and
the
hippocampus
and
around
cortex,
and
then
seeing
evidence
of
those
types
of
activity
patterns
in
neocortex
got
us
thinking
about
how
how
grid
cells
and
the
thousand
brain
sort
of
works
together
and
all
this
research
was
about.
That
was
about
object,
representation,
how
sensory
representations
are
used
through
movement
to
build
up,
object,
representations
and
we've
put
out
what
one
two
three
four
papers
or
something
from
the
during
this
period.
A
So
this
we
were
just
focused
on
writing
papers
honestly,
so
we
put
out
a
bunch
of
papers
about
this,
and
now
we
have
just
recently
been
transitioning
and
I've
been
talking
about
this
to
applications
again,
but
this
time
to
define
it.
An
interesting
thing
here
is
that
we're
not
taking
this
research
and
trying
to
apply
it
to
deep
learning,
we're
not
even
trying
to
apply
this
at
all.
I
think
this
still
has
some.
We
still
there's
still
some
legs
on
research
here,
I
think
there's.
No.
A
This
isn't
ready
to
be
turned
into
applications,
but
this
is
what
we're
still
using
we're
still
using
the
concepts
of
active,
dendrites,
essentially
sparsity
at
the
dendrites,
continuous
learning.
That's
what
we're
applying
when
we're
talking
about
building
out
these
deep
learning,
apps
that
were
working
on
now,
I
have
no
doubt
Jeff's
working
on
a
book
at
some
point,
he's
books,
gonna
come
out
within
the
next
year.
Probably
and
I
have
no
doubt
at
least
I
fully
expect
or
another
transition
back
into
research
from
the
Mendte
at
some
point
in
the
future.
A
A
A
A
A
Illusion
will
have
that
effect.
How
the
circuit
works
right.
It's
going
to
optimize
on
whatever
type
of
patterns
it's
getting,
so
it
doesn't
surprise
me
that
there's
some
variations
in
this
circuit,
but
I
think
there's
a
core
circuit
that
we're
trying
to
define
and
we
have
to
understand
how
motion
works
in
that
circuit.
Our
object
identity
works
in
that
circuit.
How
prefrontal
cortex
works
in
that
circuit,
how
our
goals
and
plans
and
all
that
stuff?
How
does
that
service
support
all
these
things?
A
A
A
Don't
know
the
way
I
read
it
is
the
way
they're
talking
about.
Affordance
is
almost
like
the
representation
of
something
not
including
how
it's
learned
at
all,
just
the
representation
of
some
action
in
response
to
a
stimulus.
Okay,
that's,
like
you,
have
a
representation
of
you
throwing
a
ball
in
your
brain
or
the
motions
it
takes
to
do
certain
things
in
your
brain
I.
Think
that's
the
level
of
an
affordance
like
imagine
catching
a
ball.
Affordances
would
be
the
quick
motions
that
you
take
to
grab
the
ball.
A
You
know
as
it
comes
towards
you,
where
there's
going
this
direction
or
that
direction
or
vector
you
know
you
do
that
you're
you're
you're
matching
like
the
movement
you're
selecting
this
is
what
I
like
to
think
about
it.
The
motion
that
gets
selected
you
can
call
it
an
importance
similar
to
object
selection
when
you're
moving
through
an
object
space.
If
you
reach
into
a
dark
the
dark
and
you
make
a
motion
cross
the
object,
the
the
union
that's
happening
to
identify
that
object.
The
sensory
input
is
matched
against
all
of
your
brain
space
of
objects.
A
A
ball
like
that,
you
know,
yeah,
that's
the
union
selection,
that's
happening
in
real
time.
Well,
anything
is
its
object
identification.
So
it's
most
affordance
identification.
Whatever
that
object
is
it's
being
stored,
its
associated
with
your
senses,
in
space
in
both
ways
right
all
and
you've
got
a
motion.
A
Ball
like
at
all
times,
you're
representing
an
object
and
ventral
screens,
are
representing
objects
through
you're,
interacting
with
the
dorsal
streams
you're
representing
motions
that
are
body
centric
and
space
right.
So
this
would
be
different
body.
Centric
reference
frames
versus
object,
centric
reference,
earrings
right,
okay,
but
it's
the
same
commutation
and
that's
what
I'm
trying
to
put
all
this
that
make
sense.
Mark
okay,.
B
I
believe
that
you're
also
seeing
not
just
a
binding
of
different
representations,
but
a
connection
between
the
sensation
and
the
action.
So
what
happens?
This
is
it.
He
uses
the
example
of
going
around
a
tree
in
front
of
you,
so
you're
trying
to
select
an
action,
emotion
and
there's
these
two
paths
affordances-
and
you
sense
this
opening,
and
it
connects
to
the
motion
to
drive
towards
that
opening.
A
B
A
A
A
B
B
It's
part
of
what's
happening
is:
is
that
I've
been
doing
this
for
a
long
time
and
I've
I?
Have
this
this
giant
vision
of
how
all
the
different
pieces
fit
together?
Some
of
it
makes
sense.
Some
of
it
doesn't
but
I've
been
trying
to
take
pieces
of
it.
That
seemed
like
the
best
cohesive
whole
and
put
them
on
the
forum
mm-hm
and
mostly
it
works,
but
people
are
asking
questions
I
try
to
fill
in
this
or
that
piece
of
it
and
it
scattered
all
over
the
forum.
B
I've
collected
it
together,
mostly
in
the
stream,
that's
called
the
brain
diagrams
or
something
like
that.
I
forget
the
title
of
it
and
I've
been
working
offline
here
to
try
to
do
the
key
thing
which
I
think
is
appropriate
to
HTM,
which
is
the
this.
This
concept
of
of
hex
grids.
It
takes
to
work
at
Kelvin
and
ties
it
to
the
work
of
momentum.
Nice.
A
B
Okay
in
one
of
the
papers,
I've
shared
the
the
paper
is,
is
how
brains
make
meaning
abstract
symbol.
Does
this
whole
thing
here?
Well
we're
talking
about
affordances.
This
has
a
diagram
that
shows
how,
in
this
part
of
the
cortex,
how
the
the
highest
levels
of
the
processing
stream
the
what
where
streams
where
they
come
together
in
the
temporal
lobe,
are
connected
through
fibers
into
the
planning
and
action
load.
So
this
is
what
I'm
saying
when
I
say
the
affordances
are
tied
directly
together
to
the
performance
part
of
the
brain.
B
Part
of
the
Rinne.
Well,
if
you
think
about
the
forebrain,
what
happens
is
the
down
in
this
section
the
the
the
lower
frontal
lobe?
You
see
the
richest
connections
between
the
hippocampus,
the
hypothalamus,
the
amygdala,
the
the
lower
brain
structures
heavily
are
connected
to
the
planning
parts
of
the
brain,
so
I
see
that
what
happens
is
is
from
the
central
sulcus.
Here
to
this
there
you
see
a
instead
of
a
stream
that
is
elaborated
into
some
higher-level
detail.
You
see
some
higher-level
detail
unfolded
until
you
reach
the
sulcus
and
you
actually
see
the
motor
drives.
B
A
B
Okay,
so
this
unfolding
bit
going
from
the
lower
to
lower
frontal
lobe,
it's
it
is
a
value,
is
unfolded
into
a
proto
plan
and
if
you
have
enough
of
a
support
from
the
lower
from
the
rest
of
your
brain,
that's
gated
in
into
an
action,
this
expands.
What
and
where
right?
Yes,
yes,
it
does,
because
what
you're
seeing
up
here
at
the
at
the
inferior
temporal
is
the
the
highest
level
of
evaluation
of
the
what
and
where
streams
for
the
objects
and
spatial
coordinate
orientation.
Everything
is
brought
together
at
that
point
and
that's
your
experience.
B
D
B
You're,
seeing
in
the
upper
part
of
the
temporal
lobe,
is
your
combined
experience
to
here
and
now
so
you're
going
to
have
whatever
you're,
seeing
around
you
you're
going
to
have
the
objects
that
you've
selected
as
being
part
of
your
conscious
awareness
they're.
All
in
this,
this
blackboard
of
your
current
experience,
all
right
all.
B
That's
being
brought
down
to
whatever
how
I
still
can't
quite
imagine
what
the
the
subcortical
structures
make
of
all
this.
This
is
the
same
thing
that
would
make
a
lizard
work,
so
it's
getting
some
digested
version
of
this
to
make
decisions
and
decide
on
actions.
That's
what's
projected
to
the
forebrain
and
then
you
those
are
unfolded
using
you're,
very
clever,
cortex
and
wrapped
around
through
the
sensory
stream,
and
then
the
lizard
gets
some
sort
of
feedback
of
what
your.
B
B
B
B
A
Can't
explain
that
I
can't
explain
that.
So
let
me
ask
you
a
question
then,
on
the
diagram
that
you
showed
mark
there
were
all
these
arrows
sort
of
coming
out
from
from
the
center
and
a
moving
outward,
but
there's
the
is
that
consistent
with
the
yeah
yeah
there
we
go.
Is
that
good
I
mean
the
way
I
read
the
the
dorsal
stream
is
that
it
originates
in
striate,
cortex
and
move
sort
of
forward.
As
this
that's
a.
How
does
that
play
with
this?
These.
B
Are
roughly
what
you're
seeing
for
these
arrows
is
the
the
major
fiber
bundles
that
are
connecting
parts
of
the
brain
from
one
distant
area
to
another
distant
area?
There's
a
lot
more
short
loops
that
connect
every
every
gyrus
and
joining
gyrus.
As
you
can
see
in
this
this,
this
very
simplified
diagram
here
shows
that
there's
short
loops
and
longer
loops
and
long.
These
are
the
very,
very
long,
the
longest
distant
fibers
in
the
brain.
Okay,.
B
B
D
B
A
B
Here's
the
app
what
it
is
is
a
front-end
for
blender
you,
you
feed
this
thing
with
connections,
and
it
makes
neurons
model
neurons,
oh
and
you
can
make
lots
of
them
a
matter
of
fact.
This
is.
This
is
a
demonstration
where
they
were
going
to
show
how
their
this
probe
they're
fabricating
this
thing
in
blue,
it's
a
probe
for
sampling,
the
activity
of
neurons
in
a
column,
and
while
they
were
at
it,
they
put
together
a
model.
But
this
thing
is
loading
and
it's
it's
sensitive
to
your
mouse.
Give
it
a
moment.
A
B
B
A
B
A
hyper
column,
yeah,
yeah
I,
believe
it
is,
but
I
they
were
more
considering
about
the
job,
the
geography
of
a
column.
They
were
more
concerned
about
how
it
would
fit
in
with
the
the
psalmist,
so
they
model,
as
so
missus
simple
cylinder,
just
to
get
the
spatial
orientation
and
they're
looking
to
see
what
what
it
is.
They'll
be
sampling
when
they
insert
this
in
cortex.
A
A
You
could
interact
with
it
and
you
could
pick
out
the
pick,
the
cell
types
that
would
show
all
the
different
papers
that
reference
evidence
that
you
know
that
the
cell
type
was
there
and
how
it
interacted
with
other
things,
and
you
would
make
connections
from
this
cell
to
that
cell
and
each
link
would
have
references
to
papers.
That
would
be.