►
Description
An explanation of HTM sequence memory. -- Watch live at https://www.twitch.tv/rhyolight_
A
A
Let
me
do
it,
let
me
just
do
it
on
the
whiteboard,
so
the
question,
so
you
hear
a
song
right,
so
you
say
you
know
already
a
hundred
songs,
so
I
mean
you
know
more
songs
than
that.
Probably
probably
there
were
thousands,
maybe
tens
of
thousands
of
songs,
your
musician.
Certainly
so
all
those
songs
have
like
a
beginning
and
they
go
on
for
a
while.
In
the
end,
just
this
one,
maybe
shorter
this
one's
longer,
but
essentially
they're
sequences
of
patterns
right
and
they
all
have
a
beginning.
A
Some
point
and-
and
this
is
like
any
type
of
sequence-
has
a
beginning
and
an
end
and
a
bunch
of
stuff
in
between
it.
That
changes
over
time
and
what's
changing,
is
the
state
of
things
right.
The
state
of,
in
this
case
it's
the
state
of
a
song
melody,
but
you
can
think
of
spatial
state
changing
over
time
and
what
HTM
does
is
it
is
it
creates
this
structure
within
the
mini
column
in
the
cells
it's
in
the
mini
column.
This
is
not
a
neural
network
unit.
A
By
the
way,
this
is
a
mini
they're,
all
a
neural
network
layer.
If
you're
coming
from
the
machine
learning
world,
this
is
a
mini
column
and
these
are
pyramidal
cells,
not
point
neurons.
These
are
a
mill
cells
that
have
distal
connections
and
proximal
connections.
Okay,
so
that's
a
big
difference.
If
you're
coming
from
the
ocean
learning
world,
it's
a
toad,
it's
a
totally
different
structure.
It
allows
this
temporal
linking
of
sequences
right.
So
if
you
were
trying
to
learn
these
songs
spatially,
you
could
give
them
each
sort
of
a
fingerprint
and
a
spatial
fingerprint.
A
This
is
I
think
how
deep
learning
currently
tries
to
memorize
songs
they
might
batch
it
cut
it
up
into
pieces
and
then
store
each
of
those
as
sort
of
a
fingerprint
of
that
courted
song.
So
when
you
start
listening
to
soggy
start
recording
it,
you
process
that
and
then
compare
in
the
same
way.
You
process
your
other
songs
that
you
train
in
the
network
guy
and
then
you
compare
it
essentially
I'm,
not
the
expert,
so
I'm,
assuming
this
is
how
they
do
it.
A
A
So
if
you
think
about
like
the
note
C,
let's
say
C
in
the
ABCD
efg
scale,
there
might
be
seas
all
over.
The
song
might
have
a
ton
of
these.
This
might
be
in
the
key
of
C,
so
there's
C's
all
over
the
place.
There
may
be
a
few
C's
here,
but
some
might
not
have
any
seeds
right
so
that
note
that
C
could
be
all
you
had
was
a
seat.
Then
you
just
fared
Bing
I
can't
sing.
A
I,
see
you,
okay,
I,
don't
have
perfect
pitch,
but
if
you
just
heard
a
seat
whatever
you
could
not
think
of
anything
any
song
that
was
in
because,
unless
it
like,
unless
it
was
like
Beethoven
like
bump
bump
bump
I,
don't
know.
Maybe
that
would
trigger
that
because
it's
such
a
hard
beginning
of
a
song,
but
if
it
were
just
NEC
in
the
middle
of
song,
you
never
come
up
with
one.
A
If
you
had
to
a
mark
I'm
talking
about
how
sequence
pout
how
you
can
have
like
start
in
the
middle
of
the
song
and
match
a
sequence
that
a
song
that
you
know
like
technically,
if
you
have
to
like
an
interval
and
music,
they
call
it
interval.
If
you
have
two
notes
and
C
to
an
F,
for
example,
that
transition
now
has
meaning
right.
A
So
a
C
to
an
F
might
only
happen
in
this
song
here
and
maybe
once
there
so
now,
you're
narrowed
it
down
so
that
transition
C
to
F
the
these
many
columns
will
retain
so
so
if
this
is
a
see
that
this
might
represent
C
after
that
another
another
selling
here,
if
it's
active,
it
might
mean
C
after
E,
okay.
So
so
as
soon
as
you
identify
like
two
bits,
two
bits
in
the
seat
in
the
in
the
sequence,
you
can
start
narrowing
it
down
in
this
fashion.
A
By
sort
of
doing
this
broad,
it's
it's
it's
sort
of
like
a
quick
I,
don't
know
it.
Has
it
eight
to
call
it
a
search,
but
it
really
sort
of
is
a
search.
You
can
search
your
whole
memory
space
just
by
performing
a
union
right,
so
you
can
perform
a
union
of
the
representations
between
the
one
note
and
the
next
or
the
tread
that
transition
itself
and
then
compare
it
to
the
Union
of
things
that
you've
seen
the
narrow
things
down
so
I
think,
that's
generally
how
you
do
it
and
temporal
memory
works
that
way.
A
So
if
you
learn
a
bunch
of
sequences
and
then
you
play
C
e,
it
will
get.
It
will
return
all
the
bits
that
represent
all
the
songs.
I
mean
if
you
were
doing
object.
If
you
were
doing
object,
pooling
right,
you
did
it,
but
it
would.
It
would
give
basically
give
you
all
the
locations.
This
is
what
temporal
memory
would
do
would
give
you
this
this
this
and
this,
and
it
would
give
you
probabilities
for
each
one,
so
you
could
predict
where
you're
at
what
its
gonna
be.
A
Next,
you
know
the
problem
with
temporal
memory
is
because
we
don't
have
a
way
to
pool
over
this
and
identify
the
sequence,
like
the
object,
identification
of
the
song
that
we're
in
the
sequence
that
we're
in
we
might
know
where
we're
at
and
what's
coming.
Next
excuse
me,
we
might
know
what
the
system
itself
might
make
a
good
prediction
of
what's
coming
next
or
make
interesting
anomaly
indications,
but
it
we
can't
tell
which
song
we're
in
alright.
So
that's
that's
one
thing,
that's
tricky
that,
maybe
that's
a
maybe
that
answers
your
question
up
here.
A
A
That's
why
we
always
draw-
and
the
current
theory
will
draw
two
layers-
will
draw
one
like
this
and
then
we
have
this
or
this
circuit
we
think,
is
repeated
twice
in
the
in
the
common
cortical
column
column
down,
and
this
one
will
have
SP
like
spatial
pooling,
so
we'll
have
mini
columns
and
cells
per
call.
You
know
some
temporal
memory
essentially
running
and
I'm
gonna
call
it
TN,
but
it's,
but
it's
the
mechanism
of
temporal
memory.
A
It's
only
temporal
memory
when
we're
when
we're
distantly
connected
to
ourselves
right
sorry,
I
was
ignoring
chat
for
a
bit,
but
I'll
get
I'll
get
back
to
that
and
this
one
we
want.
We
want
to
pool
so
and
it's
we
want
to
pool
and
form
an
object
representation
based
upon
the
activity
of
this
over
time.
A
So
assuming
we
have
attention
one
thing
like
we're
within
the
reference
frame
of
one
object:
that's
not
let's
say
it
like
that,
because
those
are
the
terms
that
were
we're
moving
towards
it,
dementia
talking
about
reference
frames,
so
assuming
that
we're
defining
one
reference
frame,
an
object,
sort
of
an
object's
reference
frame
as
we
move
through
that
reference
frame-
that's
what's
happening
here,
moving
through
the
reference
frame.
This
is
narrowing
down,
narrowing
down
narrowing
down
to
form
a
representation
that
we
can
match
against.
Things
of
you
know,
okay,.
A
But
doesn't
it
learn
those
transitions
in
the
context
of
all
the
previous
notes
yeah,
but
we
only
had
two
notes.
We
didn't
have
I
mean
the.
What
we're
trying
to
match
against
was
two
notes.
So
those
two
notes
have
all
these
different
possible
beginnings
right.
We
don't
know
the
beginnings
yet
as
we
narrow
as
we
listen
to
more
well,
that's
when
we
can
narrow
it
down
completely
and
lock
it
in
on
a
sequence
and
know
exactly
what
sequence
grab,
but
we
have
to
have
more
input
over
time
and
that
whole
model
doesn't
involve
movement.
A
You
know
you,
don't
you
can
think
of
it
in
two
ways,
with
or
without
movement,
because
there's
two
ways
that
your
brain,
your
brains,
input,
changes,
I,
won't
write
that
down.
There's
two
ways,
your
brains
into
it,
changes
one
is
that
reality
changes.
You
know
the
world
changes,
the
state
of
the
things
has
changed
or
two
you
move
through
reality,
and
so
your
brain
has
to
decide.
What's
where
the
line
is
where
that
line
is
drawn
between
reality
and
your
effects
on
reality,
that's
one
of
the
things
it
learns.