►
From YouTube: HTM Hackers' Hangout - June 7, 2019
A
Hey
welcome
to
its
team
hackers
hangout,
it's
June
7th
I've
got
a
people
here,
waiting
and
so
I'll
get
to
everyone
and
we'll
take
questions.
I
notice,
people
are
watching
hi
cowntess
on
go
ice
on
you
on
chat
and
youtube
and
hello
breath.
Good
morning
everybody
we've
got
Marty
on
line
and
Paul
and
Marcus
and
David
who's
been
helping
me
with
my
twitch
stream
and
doing
the
react
stuff.
A
So
I
have
a
short
agenda.
First
I
want
to
show
off.
So
that's
the
first
thing
that
I
want
to
do
so.
Let
me
show
off
real
quick,
and
this
is
just
because
I
want
to
show
off
the
the
work
I've
been
doing
on
Twitch,
which
is
building
HTM
systems.
This
is,
if
you
look
down
below
that,
the
guy
on
the
left.
There
is
David.
Tremont
he's
been
helping
a
lot
with
this,
and
if
you
want
to
see
it
actually
working,
you
could
go
to
this
URL
I'll
put
it
in
chat
in
YouTube.
A
If
anybody's
watching-
and
you
can
actually
see
the
live
app
here-
and
this
is
the
latest
thing
that
we
created
just
a
couple
of
days
ago-
was
this
potential
fools
visualization,
so
we're
going
to
keep
working
on
this
and
eventually
we're
gonna,
add
pros
all
around
this
and
code
samples
and
and
all
that
and
really
fill
it
out.
So
it
looks
more
like
a
real
interactive
document
sort
of
like
this.
We
haven't
styled
any
of
this
at
all,
so
I
just
want
to
point
out
anybody
who's
missed
what
I've
been
doing
on
Twitch.
A
This
is
this
is
basically
it
and
the
code
base
is
wide
open.
I've
been
working
with
the
community
had
quite
a
few
contributions,
not
just
what
David
referred
couple
other
people
to
so
I've
shown
that
off.
The
other
thing
that
I
want
to
show
is
that
we
have
had
a
lot
of
new
projects
htm'
projects
over
the
past
year
or
two
that
I
mean
I,
haven't
really
thrown
up
much
fanfare
for
for
these.
A
But
if
you
look
at
this
projects,
page
which
again
I
just
threw
up
into
chat,
there's
a
ton
of
HDM
projects
in
here-
and
I
just
updated
this
this
morning
too.
So
some
of
these
are
really
new.
I
want
to
point
out
a
couple
of
them
this
this
one,
you
Taylor,
which
Marty
who's
the
pink
teacup
in
the
forum
and
right
now
currently
yeah.
A
He
has
been
working
on
this.
So
there's
a
lot
of
details.
It's
a
C++
invitation
very,
very
fast
I
haven't
tested
this,
but
I'm
gonna
take
his
word
on
it
because
he
used
me
other
HTM
said
were
also
extremely
fast.
Tiny
HTM,
for
example,
so
check
that
out.
If
you
want
to
run
parallel,
c++
HTM,
that's
a
lot
of
potential
there,
another
one
that
I
haven't
said
much
about
was
there's
one
in
scheme
and
I
think
this
has
recently
been
updated.
A
It
didn't
have
yeah
so
20,
so
this
is
so
there's
recent
activity
on
this,
so
someone's
programming
an
HTM
in
scheme.
So
that
could
be
interesting.
If
anybody
who
wants
to
get
into
that
I
think
scheme
is
a
type
of
list,
isn't
it
all
right?
Correct
me,
if
I'm
wrong
there
and
one
more
thing
I
wanted
to
point
out-
was
the
the
snowflake
medium
article
by
Jimmy,
which
I
really
liked.
A
We
just
had
a
Bay
Area
meetup
thanks
to
James
who's,
not
joined
us
right
now,
but
he
typically
will
pop
into
HTM
hackers
hangouts.
He
did
a
great
article
about
AI
and
and
what
what
is
it
is
what
it
is
today
and
and
how
HTM
is
different
and
why
he
put
it
into
a
plugin
for
his
database
called
snowflake,
so
he
took
Paul
lambs,
HTM,
Jay
s,
implementation
and
basically
injected
it
as
a
plug-in
for
snowflake
the
database.
Oh
one
more
thing:
sorry
I,
pre,
maturely
prematurely
turned
that
off.
A
A
A
If
there's
going
to
be
something
like
an
OPA,
if
I
expect
it
to
be
in
another
repo
like
new
pic
that
PI
perhaps
but
there
is
a
network
API
I,
believe
in
in
new
pic
CPP
that
has
that
should
eventually
I
don't
know
if
it
has
PI
mind
interface
to
it
yet,
but
I
hope
that
it
will
eventually
have
a
five-iron
interface
to
the
network,
API,
so
open
question.
If
anybody
knows
the
answer
to
that,
let
me
know
alright
right.
A
So
next,
up
on
our
agenda,
I
talked
about
updates
and
some
of
the
new
HTM
projects.
Let
me
before
we
go
to
to
Marcus
and
do
a
little
QA
on
research.
Anyone
in
the
community
at
the
moment
have
what
to
highlight
any
of
their
work.
That
I
didn't
touch
upon
or
have
any
comments
on
what
I
just
went
over.
A
A
So
if
I
understand
correctly,
the
reason
why
systems
like
grid
cells
and
displacement
cells
are
necessary
in
the
neocortex
is
movement
while
receiving
sensorial
aside
sensor
information,
the
fluctuating
data
due
to
movements
needs
to
be
compensated
for
otherwise
would
seem
like
a
stream
of
completely
new
and
unrelated
information
when
a
cloud
passes
or
light
switched
on
colors
changed
brain
still
recognizes
everything.
Wouldn't
a
similar
system
like
crystals,
be
required
to
compensate
for
changes
in
color.
A
So
there's
this
idea
of
color
and
another
question
regarding
another
interesting
comment
that
Marcus
made
according
to
a
paper
that
was
discussed,
a
probe
in
a
subject
like
in
an
environment
Marcus
described
as
a
jungle
gym,
which
is
like
a
three
dimensional
sort
of
something
that
a
rat
can
go
in
all
three
dimensions:
easily
produced
a
deformed
a
grid
pattern.
Is
it
possible
that
this
has
the
effect
of
a
wraparound
grid
as
if
the
pattern
was
printed
on
a
piece
of
paper
and
they
wrapped
around
a
3d
environment?
A
Is
it
possible
they
absolute
distances
between
vertices
remain
constant,
but
that
seen
from
the
top
of
the
production
of
the
pattern
is
what
causes
the
distortion
so
Marcus.
Take
it
away.
Okay,
yes,
sir,
there
were
a
few
things
there.
I
can
talk
for
a
little
while
based
on
those,
because
there
are
kind
of
three
things
there.
So,
okay,
so
grid
cells
and
displacing
themselves.
A
So
to
predict
the
input
to
predict
your
sensory
input,
you
need
some
context.
You
need
something
that
represents
context
of
some
sort
and,
and
sometimes
the
right
context
you
need
is
a
location
and,
for
example,
if
you
know
the
location
of
your
retina
and
the
reference
frame
of
the
object,
you're
sensing
and
you
know
its
orientation,
you
know
which
way
it's
pointed
that
is
enough
to
to
to
figure
out
what
the
eye
is
going
to
sense.
A
So,
in
general,
like
in
some
ways,
I'm
stating
something
obvious
on
purpose
to
build
up
from
it
in
general
to
predict
what
you're
gonna
sense.
Next,
you
need
some
sort
of
context.
Sometimes
that
is
going
to
be
a
location.
Sometimes
it's
going
to
be
like
what
just
happened.
What's
what
few
things
just
happened?
That's
gonna
cause
me
to
predict
what's
next,
so
you
use
the
example
of
the
passing
cloud.
A
For
example,
well,
that's
like
a
movement
sequence
of
the
cloud
moving
I
could
make
a
connection
there
to
grid
cells
in
the
sense
that,
as
a
cloud
moves
across
your
visual
stream
you're
the
location
of
your
retina
relative
to
the
cloud,
their
location,
your
retina
and
the
space
of
the
cloud
is
being
moved.
It
is
moving
because
of
because
of
the
clouds
movement
which
is
which
is
like.
A
So
if
you
do
have
grid
cells
tracking,
that
there's
kept
there's
something
kind
of
nice
there
that
it
will
that
you
can
use
that
as
a
clue
to
figure
out
what
movement
is
occurring
or
the
the
nicer
extreme.
The
nicer
version
is:
if
all
the
clouds
are
moving,
you
can
use
that
as
a
clue
to
figure
out
your
own
movement.
Your
own
self
motion,
that's
kind
of
like
a
form
of
optic
flow,
that's
kind
of
the
least
interesting
thing,
I'm
gonna
say
so.
The
the
I
think
you
made
a
good.
A
He
pointed
out
something
good
when
you
talk
about
color
being
similar
to
orientation
or
end
and
and
you
bring
up
displacement
cells
I
think
the
right
way
to
think
of
this
I
might
draw
a
little
bit
on
the
board
there.
The
right
way
to
think
about
what
we
often
call
displacement
cells
is
I'll,
actually
draw
a
diagram
from
I've,
been
reading,
some
just
content
recently
and
a
diagram
from
1981
on
it,
which,
which
is
kind
of
similar
to
things
that
drop
before
it,
similar
to
some
stuff
in
our
papers.
A
So,
let's
see
oh,
can
you
hear
me
write
down?
It
sounds
good
cool,
so
this
is
notation
from
multiple
papers
from
Jeff
Henson
in
1981.
I
could
point
to
a
specific
one,
but
the
idea
that
you're
going
to
transform
between
like
what
I,
what
I'm
sensing
I
guess,
I
would
say
like
what's
what's
out
that?
What's
actually
there.
A
Okay,
it
could
be
interpreted
it
as
that,
but
like
this,
for
example,
can
be
that
this
can
be
as
simple
as
like
here's
what's
landing
on
my
retina
like
here's,
an
oriented
edge
or
something
like
that
here,
is
that
edge
in
the
reference
frame
of
the
thing
I'm
sensing
or
so
you
know,
I
said
what's
actually
there,
but
with
at
the
core.
What
we're
doing
is
applying
some
kind
of
transformation
so,
like.
A
I'm
trying
to
write
this
in
very
general
terms,
but
this
is
it's
some
sort
of
transformed
version
of
this
like
it's
like
a
normalized
I'll
say
it's
like
an
normalized
version
of
this
sort
of
where,
like
the
normalized
input,
substant
like
if
you
were
looking
at
a
coffee
cup,
you
were
looking
at
the
top
of
it.
You
can
see
it
from
lots
of
different
viewing
angles.
A
A
What
this
is
what
what
this
can
be.
So
when
we
talk
about
displacement
cells,
we're
talking
we're
usually
referring
to
we
usually
referring
to.
Where
is
one
part
of
an
object
in
the
reference
frame
of
the
whole,
like,
for
example,
where
is
the
handle
and
the
reference
frame
of
the
cup
or
in
the
in
the
framework
paper?
A
The
new
mental
logo,
where
is
it
in
the
space
of
the
cup
and
this
idea
of
having
yeah
some
sort
of,
for
example,
Louise
that
specific
example
in
the
column,
the
frameworks
paper,
we
have
a
logo
and
a
cup
the
transformation
being
applied.
There
is
location
in
the
space
of
the
cup.
Imagine
if
you're
not
familiar.
Imagine
a
coffee
mug
with
something
mighty
logo
on
it,
give
it
a
location
in
the
reference
frame
of
the
cup.
A
Then
you
get
this
sense
of
like
what
I'm
sensing
I'm,
something
like
I,
don't
know
my
retinas
getting
hit
by
some
mix
of
photons
of
following
within
some
range
and
then
like,
what's
actually
out
there,
what
color
do
I
actually
perceive
that's
out
there
and
I
and
I
think
it's
correct
to
say
that,
like
this,
the
similar
diagram
you're
trying
you're
trying
to
infer
what
is
the?
What
is
the
the
mapping
from
this
to
this
and
you're?
Trying
to
you
want
to
track
this.
A
You
want
to
you
want
to
keep
track
of
this
at
the
current
moment.
You
want
to
keep
track
of
keep
track
of,
like
you
know
this
greenish
thing
actually
maps
to
blue
or
that
are
like
I'm
and
I'm
inferring
that
the
what
the
lighting
in
this
room
is
such-and-such.
Therefore,
the
dress
is
black
and
blue
or
whatever
movie.
So
let
me
say,
let
me
try
and
infer
what
you're
trying
to
say
so.
A
A
Yes,
I.
Think
I.
Yes,
it
worked
for
like
a
second
there
you
cut
out
all
right,
it
might
have
been
me.
I
lost
you,
but
yeah
like
yes,
the
idea
that
multiple
inputs
will
be
kind
of
normalized
to
the
same
thing
based
on
some
form
of
transform,
so
I
argued
against
Falco,
and
this
idea
of
color
being
somewhat
represented
in
this
way.
So
it
sounds
like
you
that
he
might
be
one
I.
B
A
Say
that
at
the
high
level,
concepts
of
like
of
keeping
track
of
the
transformation
between
input
and
what's
actually
there
I
can
draw
the
analogy
there.
However
I
don't
know
representing
the
color
or
something
using
grid
cells,
I,
don't
see
any
anything
there
at
the
moment.
It
doesn't
stand
out
to
me,
but
I
can
I
can
step
back
and
thinking
in
terms
with
the
higher
concepts.
That's
what's
here
and
and
and
to
show
that
this
connection
exists.
A
I
think
I'm
following
you,
I
just
never
thought
about
it.
That
way,
so
there's
good
good
explanation
Oh
see
so
that
was
that
was
part.
Two
I
guess
the
the
third
was
totally
disconnected.
Let's
see
have
I
addressed
this
first
part
very
fully.
Well,
you've
certainly
provided
some
food
for
thought
for
me,
one
hopefully
for
him
too.
Well,
he'll,
probably
discuss
it
with
me
on
my
stream
later,
I
might
think
of
something
later,
but
yeah
we'll
see
what
happens.
Okay.
A
A
None
of
these
results
are
published.
They
both
they've.
Basically,
there
have
been
a
couple
posters
from
all
enough
skis
crew
and
so
Felco
asked
about
what
I
was.
What
I
in
a
research
meeting
called
the
jungle
gym.
The
this
rat
is
moving
through
this.
You
yeah,
we'll
just
keep
calling
it
that
this
little
this
little,
this
3d
triangular
Trant
triangular
lattice.
It
might
not
have
been
trying
Euler.
But
the
point
is
this:
lattice,
where
the
where
the
rat
can
climb
around
arbitrary,
almost
sarbat
arbitrary
directions
in
3d,
some.
B
A
A
Some
background,
so
it's
really
hard
I
would
I
would
draw
something
on
the
board,
but
it's
really
hard
to
draw
something:
that's
3d
and
messy
and
make
it
worth
trying,
but
so
in
in
that
Falco
asked
if
well,
first
I'll
throw
in.
He
asked:
if
are
the
fields
kind
of
uniformly
separated
or
are
they
are
the
distances
between
them
similar?
A
Basically
Kate
Jeffrey's
conclusion
or
that
their
groups
conclusion
I,
think
she's
working
with
Roddy
grieves
on
that
was,
and
it
still
it
hasn't
been
published.
Yet
so
I
mean
they're
still
working
on
this
is
that
the
fields
are
blobby
and
non-uniform,
and
that
was
basically
where
they
left
it,
whereas
all
enough
C's
lab
with
the
bats
they
also
confirmed
like,
or
they
also
reached
the
conclusion.
The
fields
are
like
these
3d
blobs
they're
not
uniformly
they're,
not
uniform.
Like
a
lattice,
however
they're
the
this
space
space
between
them
is
roughly
the
same.
A
The
the
spacing
of
them
is
I
mean
uniform
in
the
sense
that,
like
they're
just
the
distances
between
the
fields,
although
the
directions
between
them
might
be
kind
of
messy,
the
distances
between
them
are
roughly
uniform.
So
it
does
seem,
like
that's
plausible,
that
I
think
that
the
fields
themselves
are
are
kind
of
uniformly
spaced.
A
So
I
do
think
that
if
I'll
go
ask
if
what
could
be
going
on
here
is
that
this
2d
grid
is
kind
of
being
folded
into
3d
space
in
some
funky
way,
totally
plausible
yeah
that
could
that
could
be
exactly
what's
happening.
A
It
might
be
the
case
that,
like
as
the
rat
moves
around,
it's
always
kind
of
it's
almost
always
attending
to
a
2d
plane,
so
it
almost
perceives
itself
like
alright,
now
I'm
in
this
2d
plane
now
I'm
in
this
one
and
and
it
does
that
in
sort
of
a
way
that
that
that
I
don't
know.
It
follows
a
certain
set
of
pattern
like
it
always
comes
back
to
the
same
few
planes.
A
A
Thing
here
was
that
the
studies,
because
they're
especially
Katie
Jeffrey's
lab,
has
done
a
fair
number
of
studies
of
various
types
of
3d,
whether
it's
like
the
rat
is
climbing
pegs
on
a
wall.
The
rat
is
calling
a
climbing
chicken
wire
running
through
the
spiral
staircase
and
one
thing
I
think
that
they've
found
through
talking
to
them
a
little
bit,
as
is
that
it
does
make
a
difference
whether
the
rat
kind
of
grew
up
climbing
walls.
If
they
didn't,
then
then
you
expect,
then
they
I
think
they
found.
None
of
this
is
conclusive.
A
All
this
is
just
like
individual
experiments.
You
easy
to
over-interpret
it,
but
early
results
with
grid
cells
showed
kind
of
these
pillar-like
fields
like
when
the
rat
is
the
climbing
these
pegboard
walls.
Then
a
single
grid
cell
would
often
have
a
field
that
just
kind
of
stretches
up
into
space,
but
I
think
those
were
rats
that
were
less
experienced
with
a
3d
world
where
early
during
development,
if
they
just
kind
of
live
in
3d,
if
they
navigate
3d
all
the
time,
it
almost
seems,
like
maybe
they're,
good
cells
learn
3d
space
differently.
C
A
And
we
there's
not
really
a
need
for
me
to
give
that
much
of
a
of
like
a
update
or
research
summary,
because,
like
a
couple
nights
ago,
we
had
a
meet-up
here
so
I
mean
going
Matt's,
YouTube
feed
to
see
talks
from
mini
talks
from
Matt
and
Jeff
and
Suba
ty,
where
they
kind
of
summarized.
What's
going
on
yeah,
it's
sort
of
unnecessary
to
give
research
updates
here
on
the
in
this
meeting
anymore,
since
we're
just
live-streaming
research
yeah.
A
At
the
same
time
like
if
you
just
jump
into
a
research
meeting,
you're
kind
of
being
thrown
into
the
middle
of
things,
yeah,
the
the
meetup,
would
it
give
some
good
stuff
yeah
I
mean
a
couple
of
us
are
working
on
the
high
level
question
of
how
does
the
neocortex
learn
models
of
objects
or
represent
models
of
objects?
An
object
here
can
be
like
a
big
I.
It
can
is
kind
of
a
loose
term.
It's
if
anything,
it's.
How
does
the
brain?
A
Then
the
others
of
us
are
working
on
applying
what
we've
learned
to
machine
learning
to
deep
learning.
So
there's
a
there's,
a
question
from
Youngstown
beau
on
YouTube.
He
says:
how
are
they
certain
that
grid
cells
didn't
react
or
during
the
observation
talking
about
here
at
jungle,
gyms
burnin,
I
would
say
they're,
not
certain
of
that.
That's
that's
very
plausible
that
that
that
they
are
that
that
a
lot
of
reentering
is
occurring
as
the
as
the
as
the
rat
is
like
it.
A
You
might
get
tired
stops
for
a
second
and
then
we
kind
of
reorient
and
then
starts
moving
around
again
and
in
those
times
maybe
the
grid
cells
firing
fields
changed
completely
because
they're
just
doing
these,
these
heat
maps
of
grid
cells
over
time
as
long
as
they're
as
long
as
their
sample
size,
isn't
too
large
as
long
as
they're
not
recording
the
rat
for
like
days
and
days
and
days,
you're
still
gonna
find
these
grid
fields.
So
you
know,
that's
very
possible.
I!
A
A
Very
biased
toward
detecting
things
that
can
be
discovered
in
a
heat
map.
If
you
have,
if
you
have
this
heat
map
of
you,
have
this
Trump
I'll
you
have
the
five
minute
trial
rat
running
around
a
box.
You
make
a
heat
map
of
that
of
all
the
places
where
a
Cell
fired,
and
you
look
at
that
and
anything
you
can
conclude
from
that
picture.
They're
gonna
discover,
but,
but
anything
that
gets
lost
by
that
picture,
there's
a
good
chance.
Those
things
are
harder
to
find
it's
just
kind
of
the
experimental
paradigm.
A
What
are
the
types
of
things
that
are
hard
to
find
that
you
wouldn't
find
in
a
heat
map
like
that,
like
one
thing
is
blows,
the
paper
is
here:
I
can
point
to
exceptions
where
they
did
succeed
in
finding
it
and
the
recent
who
was
it.
Who
was
it?
There
were
recently
two
papers
that
came
out:
one
was
from
Stanford
people,
Kia
Hardcastle,
the
butler
and
the
other
was
from
can't
think
of
their
name
right
now.
A
That
involves
grid
cells
and
rewards
and
learning
reward
locations
rapidly.
I'll
post
this
link
somewhere
Oh
and
one
interesting.
They
found
an
interesting
thing
they
found
was
that
the
activity
and
the
internal
cortex
and
hip
and
her
hippocampus
over
time,
if
they
were,
if
they
like
you
know
if
they
break
it
up
into
time,
slices
of
say,
like
50,
milliseconds
I,
don't
remember
the
exact
time
slice
size.
A
They
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.
But
then,
an
hour
later,
those
those
places
are
are
rearranged.
Those
now
the
rewards
are
placed
in
different
locations.
A
They
found
that,
as
the
rat
is
running
around
this
room
now
it
is
sort
of
jumping
back
and
forth
between
representing
the
room
as
if
as
if
the
rewards
were
in
their
previous
locations
and
as
if
they're
in
the
new
ones.
Basically,
if
you,
if
you
try
to
decode
where
the
rat
thinks
the
rewards
are,
this
is
I'm
oversimplifying
a
little
bit,
but
if
you
try
to
decode
from
the
neural
activity
where
the
rat
thinks
the
rewards
are
for
it,
the
activity
is
like
it
kind
of
goes
back
and
forth.
A
First,
it
thinks
is
in
the
old
book.
Old
old
places
now
thinks
this
in
the
new
ones
now
than
the
old.
Now
in
the
new
there's,
this
temporal
juggling
of
of
like
representing
different
possibilities
that,
where
the
so
the
word
this
comes
with,
what
this
comes
down
to
is
the
grid
cells
are
firing
where
they
used
to
fire.
Another
fire
in
the
new
location
now
they're
firing
where
they
used
to
fire
now
they're
firing
in
their
new
location.
Is
this?
Could
you
compare
this
like
an
intersection
of
two
concepts?
A
Right,
yeah,
you
know
it's
both
yeah,
it's
both
of
them,
yes,
and
so
it
seemed
there
was
some
sort
of
the
heat
map
would.
If
you
were
looking
at
this
as
a
heat
map
you
see
like
oh,
the
cell
fires
and
its
old
location
and
then
in
its
new
location.
But
if
you
look
at
it
spread
out
over
time,
you
would
say
like,
but
it
never
does
both
it's
like
the
it's
like
the
it's
like.
The
cortex
is
jumping
between
modes
old
mode
and
new
mode.
A
That's
interesting
well
this!
Well!
This
feat
is
going
on
I'm
gonna
find
that
that
that
paper
I
feel
bad
for
forgetting
the
author's
names
yeah.
Well
he's
finding
it
still
open
for
questions.
You
guys
trying
to
read
some
of
the
chat
for
anything
Mark
Brown
dimensions
like
there's
some
optical
illusions
that
seem
uncomfortable
and
he's
saying
like
a
Necker
cube,
is
one
of
those
extra
looking
cubes
because
he
said
because
it
triggers
repeated
grid
formation.
You
know
you're
trying
to
navigate
or
through
an
object
in
it
and
you're.
A
Every
time
you
go
from
one
point
to
another,
it
doesn't
make
sense
and
you,
like,
maybe
you're,
jumping
to
another
grid
formation
or
something
just
just
to
throw
in
that
other
paper
I
mentioned
is
called.
The
internal
cognitive
map
is
attracted
to
goals
with
the
first
author
being
Charlotte
Charlotte
bukhara.
But
anyway
that
was
a
cool
study.
A
D
A
D
A
A
The
temporal
memory
has
this
strange
thing
with
any
sort
of
shared
subsequence
where,
if
like,
if
suppose
you
like
suppose,
ABCD
efg,
appears
and
multiple
sequences
like
sometimes
it
occurs
in
multiple
contexts,
the
temporal
memory
will
basically
have
to
relearn
it
and
all
of
it,
this
context
and
that
learning
process
takes
a
while
because
it
has
to
kind
of
first
it
learns
like.
First,
it
has
to
learn.
Xa
then
XA
be
the
next
ABC
you
know,
and
then
it
has
to
learn.
A
Why
a
then,
why
a
B
and
then
why
ABC
in
the
repeating
the
repeating
element
problem
is,
is
the
special
case
of
this,
where
the
little
sequence
subsequence
length
is
1
or
2
or
whatever
you
want
to
call
it.
So
I
would
say
that
the
temporal
memory
is
this:
it's
the
the
class
that
is
called
the
temporal
memory,
it's
sort
of
in
its
academic
form
of
just
being
this
pure
algorithm.
A
Came
up
with
these,
like
one
class
of
solutions
to
these
problems
back
then,
and
we
can
keep
just
making
up
answers
to
how
this
could
be
solved,
how
how
this
system
could
be
kind
of
biased
toward
trying
method
bias
toward
trying
to
reuse
previous
sequences
rather
than
thinking
a
then
a
a
then
a
a
then
a
a
are
all
separate
sequences
that
have
varying
lengths.
But
at
this
point
we're
just
kind
of
we
just
be
making
stuff
up
at
this
point,
so
so
yeah
they're,
the
repeating
inputs
problem
is
something
that
we
think
will
settle.
A
A
Improving
the
neocortex's
ability
to
to
predict
its
input,
if
something,
if
something,
like
suppose
suppose,
suppose
the
neocortex
juggles
a
few
hypotheses,
one
of
them
is
going
to
be
like.
Okay,
this
sequence
is
something
familiar:
it's
not
new,
it's
not
a
novel
sequence.
It's
it's!
It's
the
ABC
that
I'm
used
to,
don't
don't
don't
say
it's
a
different
ABC.
A
C
C
Yes,
campus
kind
of
Drive
an
attention
span
and
stuff
like
that
could
also
play
out
a
deciding
factor
does
make
the
decision
on
which
one
weighs
more
mmm-hmm,
like
you
can't
tell
the
difference
between
0.333
to
contribute
to
be
up
to
fifty
digits
versus
up
to
sixty
that
your
brains,
not
gonna
yeah,
there's
gonna
be
the
same
number,
even
though
they're
not
the
same
number
and
you're
in
or
you
look
at
a
random
pad,
and
you
can't
figure
out
any
patterns
from
it,
because
it's
random,
it's
something.
That's
kind
of
same
thing:
yeah.
A
Cool,
hopefully,
that
was
useful,
there's
just
some
speculation
in
there
and
trying
to
give
you
your
little
context
done
that
yeah
there's
this
problem
in
the
temporal
memory,
because
it's
sort
of
been
pulled
that
back
to
it's
like
academic
form.
It's
not
in
this
practical
form
ahead
like
if
you
want
to
apply
the
temporal
memory
in
production
systems.
Right
now,
you
need
to
add
a
couple
things
to
it,
like
better
cleanup
of
connections
that
don't
prove
valuable
or
these
kinds
of
issues
and
those
issues
are
there.
B
Anything
else,
last
minute
comments
or
questions.
I
had
a
quick
comment
on
on
the
the
repeating
endpoints
problem,
because
it's
kind
of
an
area
I'm
interested
in
a
couple
of
things.
One
is
so,
if
you
think
of
a
sequence
as
an
object,
an
object
should
have
semantics,
so
there
should
be
semantically
similar
sequences.
So
if
you
think
about-
and
you
know
a
cortical
column-
sort
of
learning-
a
more
structured
thing-
a
more
complex
thing
as
it
sees
it
more
often-
then
you
could
see
the
subsequence
becoming
feature.
B
There's
a
you
know
themselves
a
feature
of
multiple
sequences
and
the
other
one
that
the
other
idea
is
that
timing
could
be
involved.
So
you
know
if
instead
of
saying
I'm
doing
lol
lol
lol.
Instead,
you
say
I'm
doing
L
for
X
amount
of
time.
Then
you
know
that
that's
a
different
way
of
a
sort
of
encoding
from
him.
I
shouldn't
have
him
to
worry
about
L
versus
l,
L
versus
LLL.
A
Totally
agree
with
both
of
those
and
somehow
having
this
I,
don't
know,
I,
don't
know
the
right
terminology,
but
but
before
for
this,
but
some
notion
of
grouping
the
input
stream
into
into
these
pieces
like
ABCD
efg,
is
like
we
don't
think
of
that,
as
this
strange
sequence
of
an
a
followed
by
a
B,
we
think
of
it
as
like.
No
it's
it's
a
it's
as
part
of
the
alphabet,
whereas
other
ones,
others
sub
sequences
are
too
totally
random
to
us.