►
From YouTube: HTM Hackers' Hangout (Dec 2, 2016)
Description
Discussion about new boosting code and topology.,
A
To
the
hdmi
kurz
hang
out,
it's
December
second
2016.
Everyone
all
you're
going
to
hear
this
really
annoying
beep
coming
from
some
smoke
alarm
or
something
in
the
office
that
needs
to
be
replaced.
The
battery
needs
to
replace
so
I'll
just
apologize
ahead
of
time.
I
know,
none
of
us
engineers
can
figure
out
a
fix
it.
So
so
we're
going
to
talk
today
about
topology
and
boosting
and
I
think
we
we
might
go
straight
to
you
a
who's
sitting
on
my
right
here.
You
may
want.
A
Let's
go
stand
over
here
very
good,
so
one
of
the
questions
I
know
I,
know
David
you're
you're
on
listening
in
and
one
of
the
one
of
the
questions
is:
what's
the
deal
with
the
new
boosting,
wiser
and
new
boosting
code
and
what's
the
difference
between
the
old
boosting
way
in
the
new
boosting
way?
Should
we
try
this
thing
now,
if
you're
using
a
new
pic,
should
you
try
boosting
out
again
that's
a
good
question,
so
maybe
you
can
explain
why
you
did
it
and
give
a
couple
references
because
there's
the
whole
paper.
B
So
the
extended
version
of
the
description
can
be
found
in
the
spatial
pool
of
paper,
which
is
on
bio
arc.
Half
now,
and
the
whole
idea
of
posting
is
to
encourage
columns
to
be
used
efficiently.
By
that
I
mean
every
column,
is
a
spatial
cooler
should
participate
in
representing
the
input
and
in
the
code
we
implemented
by
applying
booster
factor
to
each
column.
So
in
biology
we
call
that
a
homeowner
homeostasis
homeostatic.
B
A
A
A
A
C
B
B
C
A
I'm
not
muted,
say
you're.
Can
you
hear
me
David
sure,
okay,
okay,
sorry
so
I
just
pasted
the
video
into
chat
and
what
it
is
is
it's
a
comparison
of
the
way
boosting
worked
and
for
us
change
and
after
and
then
the
first
part
of
the
video
shows
it
before,
and
you
can
see
that
it's
not
very
effective.
Like
there's
this
phase
where
boosting
starts
and
at
a
typical
apply
either
no
boost
factor
or
you
know,
one
boost
factor,
there's
no
gradient
at
all
and.
B
D
D
C
D
C
C
A
A
Before
the
change,
this
is
before
the
change
and
so
boosting
kicks
in
and
it
starts
affecting
sort
of
pseudo
random.
Randomly
the
green
ones
are
the
ones
that
are
affected
by
the
boost,
and
it's
always
max
boost,
so
the
green
ones
are
boost,
factor
is
max
boost
and
the
red
Windsor's
want
and
that's
pretty
much
the
way
the
old
boosting
worked,
and
it
didn't
seem
to
have
anything
like
that.
We
wanted.
D
C
C
A
A
And
the
active
duty
cycles
in
the
middle,
it's
a
red
to
green
gradient,
so
the
redder
it
is
the
more
active
the
columns
are
the
greener,
the
less
active
and,
on
the
right
hand,
side
where
it
says:
learning
SP,
active
columns
ignore
that
these
are
the
boost
factors
and
so
there's
a
much
more
dynamic
gradient
of
boost
factors.
It's
not
just
max
boost
or
nothing.
Yes,.
B
A
C
I,
don't
know
what
happened,
but
when
I
clicked
on
that
link,
I
have
a
hidden
screen
now
with
your
and
I'm
listening
to
the
video.
Well,
there
it
is
okay,
so.
E
A
A
Is
the
old
boosting
applied?
You
know
you
give
it
a
max
boost
right,
that's
how
you
turn
boosting
on.
If
boosting
is
one
that
means
the
boomers
are
always
one.
That
means
it's
effectively
doing
nothing
to
the
column
overlaps
course
and
before.
If
you
put
max
boost
10,
those
boost
backers
would
either
be
for
each
column,
1
or
10,
and
so
there
was
no
gradient
there.
A
It
was
very
All
or
Nothing
pretty
much
or
how
it
would
affect
the
columns
overlap,
scores
and
whether
it
would
be
involved
in
as
the
wedding
columns
or
not
so
as
a
UA
was
explaining.
He,
he
applied
this
much
more
gradient
way
of
calculating
the
booze
factors
and
works
a
lot
better.
And
if
you
look
at
the
boosting,
HTM
school
video,
you
can
see
it
working.
You
know
really
well.
B
Typically,
there
will
be
only
a
small
fraction
of
the
column.
You
just
spatial
quarter,
that's
being
used
to
represent
the
input.
So
if
you
are
not
using
all
the
resources
you
how
and
the
result,
you
can
see
that
reflect
the
motor
performance.
Like
prediction
accuracy,
in
fact,
after
the
change
of
the
boost
factor
of
the
examples
like
how
Jim
gets
other
performance.
A
So
it's
not
quite
merged.
It's
okay,
I
think
the
new
pic
or
one
is
merged.
New
pic
is
not
quite
merged,
will
have
to
cover
lily's
first
yeah,
so
we'll
probably
kind
of
release
today
about
will
get
it
rich
today,
so
that
all
that
coat
new
code
will
be
in
probably
today,
and
I
already
that's
what
I
ran
in.
D
A
So
that's
the
boosting
change
we
see
if
there
are
some
people
watching.
So
let
me
see
if
anybody's
chatting
anyone
has
any
questions
just
jump
into
chat.
I
don't
know
if
I
turn
the
qaf
on
I'll
also
check
the
YouTube
page,
real,
quick,
because
there's
a
way
to
add
comments.
There
yeah
nobody's
there,
not
a
good
otherwise
I
want
to
talk
about
topology.
So
I'll
move
on
to
that.
If
there's
no
questions.
A
A
A
new
pic
and
most
of
our
examples,
so
what
I
think
I'll
do
is,
I
have
a
whiteboard
here,
I'm
going
to
sort
of
do
a
whiteboard
thing
and
I'm
going
to
present
you
a
visualization
that
I'm
working
on
that
I
think
explains
topology
pretty
well,
and
then
we
can
discuss
it.
And
if
anyone
has
questions,
we
could
ask
I
think
this
might
be
good
good
video
to
have
to
show
to
continue
the
topology
discussions.
A
We've
been
having
on
the
forums
and
you
guys
can
please
correct
if
I,
if
I
do
anything
or
write
anything,
that's
wrong,
so
I'm
going
to
try
and
do
this
sort
of
three-dimensional
visualization
where
this,
so
this
would
be
like
the
spatial
cooler
cortical
sheet.
I'm,
assuming
you
know,
there's
a
bunch
of
cells
in
this
sheet
and
we
would
have
another
sheet
of
this
would
be
like
the
input
space
you
can
put
that
assigns
a
letter
providing
feed-forward
input
to
the
spatial
cooler.
A
A
So
there's
space
there's
two
ways:
Inuvik
and
in
current
HTM
three
that
pathologies
apply.
One
is
the
calculation
of
the
neighboring
columns
around
this
column
and
that's
involved
in
the
boosting
calculation
and
also
in
the
inhibition
calculation
right
and
so
getting
the
neighboring
columns
as
a
part
of
the
topology
that
makes
associations
between
this
column
and
columns
around
it.
The
way
that
this
these
are
calculated
is
that
is
it
simple
distance
calculation?
A
As
you
know,
you
might
know,
when
you
define
an
SP,
you
can
give
it
dimensions,
X
lies
or
just
in
the
any
number
of
dimensions,
so
assuming
you've
defined
a
two-dimensional
input
space
like
I've
sort
of
drawn
here
this.
This
would
be.
You
know
some
very
simple,
geometric
square
around
that
column
and
that's
the
columns
neighborhood.
So
that's
one
way,
topology
is
applied.
A
Another
is
you
know
this
this
column
or
this
input
space
sort
of
projects
to
that
column
it
has
a
receptive
field
again,
which
may
be
you
know,
bigger
or
smaller
than
its
neighborhoods
not
quite
same
way,
but
it
has
a
receptive
field
down
here
of
potential
connections.
It
could
make
to
the
input
space.
A
So
how
this
receptive
field
is
defined
is
also
another
aspect
of
the
topology,
so
we've
got
the
distal
connections
that
that
element
of
the
topology-
and
these
are
the
lateral
connections
right,
proximal,
proximate,
proximal
connections
here
and
the
terms
that
were
used
for
this
is
is
column
neighborhoods,
typically
in
the
code,
and
that's
that's
what
this
boxes
column,
neighborhood-
and
this
would
be
weird
they're-
called
a
potential
fool
or
a
receptive
field.
In
the
code
it's
called
potential
pool
and
every
column
has
these
two
attributes.
A
So
one
of
the
mistakes
that
I
made
when
trying
to
understand
topology
was
making
too
big
of
an
association
with
the
dimensionality,
that's
defined
for
the
input,
space
and
spatial
cooler
with
the
topology,
for
example.
This
is
like
2d,
it
could
just
be.
1D
could
just
be
one
huge
array:
it's
used
for
the
calculation
of
the
receptive
field
and
the
column
neighborhoods
for
getting
that
distance.
How
what
columns
are
close
to
me?
A
A
One
of
the
things
that
I
didn't
quite
understand
is
why
don't
we
have
this
on
all
the
time
it
seems
like
it
would
be
beneficial
to
have
topology
on
all
the
time,
because
and
I
think
one
of
the
one
of
the
answers
that
I've
come
up
with
out
there
talking
to
everybody.
Is
we
don't
necessarily
need
it
at
the
at
the
scale
of
the
structures
that
were
simulating
at
this
point?
For
example,
we
typically
use
2048
columns
in
the
spatial
cooler
and
our
input
space
is
typically
on
four
to
six
hundred.
A
It's
sometimes
larger
it,
but
but
not
huge
and
and
the
interested
interesting
thing
is.
Let's,
let's
say
this
is
our
our
input
bits
just
all
in
an
array
a
lot
of
times,
we've
gotten
separated
up.
You
know:
we've
got
one
scalar
value
X
here,
another
scalar
value
may
be
in
inside
this
portion
of
the
input
space-
that's
being
encoded
into
this
and
then
like
date
over
here.
So
one
thing
that
that
really
makes
sense
is
why
we
thought
these
topology
is.
A
We
want
every
column
to
have
the
whole
picture,
because
then
it
can
represent
more
at,
like
smaller
portions,
of
more
of
the
data
than
like
just
one
little
piece
of
the
data
I
think
it's
better
to
distribute
the
representation
across
the
columns
than
it
is
to
have
a
bunch
of
localized
columns
that
fairly
strongly
associated
with
with
local
spatial
aspects.
Of
the
token
that
people
are
nodding.
At
least
that's
good.
A
C
C
A
A
A
A
E
A
C
So
say
those
three
input
in
in
the
bottom
array.
You
have
XY
and
say
the
last
one
was
called
okay,
XY
and
date
whatever
and
say
the
left
bit
in
the
spatial
pooler
up
on
top
the
very
left
one
had
a
connection
to
all
three
sections:
okay,.
C
E
E
Is
really
low
on
a
Sunday
during
the
day,
but
on
weekdays
it's
actually
high
during
the
day
you
think
about
like
a
hot
gym
or
energy
example,
you
might
want
I
think
one
columns
that
represent
the
combination
of
Sundays
and
during
the
day
and
low
values,
and
maybe
weekdays
and
high
values,
and
you
can
only
do
that
if
the
columns
is
getting
yeah
a
common
input
from
all
multiple
fields.
That
can
be
one
of
the
main
reasons
to
use
a
spatial.
E
E
A
A
A
We,
if
you
you
know,
look
at
all
the
encoding-
is
that
we've
done
that
I've
shown
on
hot
Jam,
an
HTM
school,
that
it's
very
important,
that
you
see
not
only
the
scalar
value
up
in
the
very
top
left,
but
you
know
the
date
and
the
weekend
encoding
is
happening
on
a
complete
opposite
spectrum
of
the
encoding.
That's
important
that
every
column
has
has
reference
to
all
of
those
bits,
so
it's
sort
of
an
application,
specific
thing
that
these
are
the
types
of
applications
that
we've
been
building
for
HTM.
A
Now,
all
of
this
being
said,
topology
is
still
really
important.
It
imagine
in
the
future,
where
we
have
these
cortical
structures
that
go
on
and
on
and
on
for,
for
who
knows
how?
Long
when
we're
able
to
maintain
that
there's
no
way
to
do
global,
inhibition
anymore
you're
going
to
have
to
do
vocalized
inhibition,
so
we're
going
to
have
to
apply
topology
at
some
point.
So
it's
a
really
important
part
of
the
theory.
It's
just
something
we're
not
forced
to
be
to
use
right
now
for
thee.
The
applications
that
are
trying
to
build.
F
Come
with
just
so
I
think
it's
important
to
make
the
distinction
between
like
Justin
into
a
terminology
when
you're
talking
about
the
surface.
So
there's
one
you
can
talk
about
the
apology
of
the
all
the
input
bits
are
distributed
so
in
the
visual
sense.
So
this
is
like
the
two-dimensional
grid
kind
of
lean
and
as.
F
Presents
now
we
have
your
image
review,
which
you
represent
by
some
state
of
this
two-dimensional
grid,
and
then
these
data
points
can
itself
like
have
a
like
intrinsic
to
apology,
yeah,
and
so
there
are
two
questions
here.
So
one
is:
should
our
weather
spatial
cooling,
some
our
relatives
to
the
topology
of
how
the
input
bits
are
organized
and
in
the
like
individual
from
the
visual
view,
point
of
view,
encode
images
or
something
like
that?
Yes,
I
mean
there
seems
to
be
an
advantage
of
that.
But
then
you
can
also
ask
shoot.
F
A
F
If
you
have
a
say,
I
give
you
an
eight,
a
stream
of
like
your
collection
of
bits.
We
have
some
honoring
on
it
and
I
give
you
images
represented
in
that
so
I,
just
cook,
sorry,
/,
mutated
array
and
then
I
asked
you.
Can
you
tell
me
whether
this,
whether
these
inputs,
whether
this
I
actually
organized
in
a
grid
like?
Can
you
tell
me
that
by
just
looking
at
which
bits
are
one
for
an
image
or
not?
F
If
we
give
you
a
million
images,
so
that's
kind
of
the
these
questions
I've
evolved
years
old
for
for
visual,
it
seems
very
natural
that
if
I
give
you
a
bit
array
which
represents
just
like
black
or
white
black
or
white
image,
the
positions
and
if
I
co
mutilated
and
then
I'll
give
you
like
a
bunch
of
it.
Can
you
tell
me
that
it
really
comes
from
a
telephony
grid
for
itself,
so.
F
E
Yeah
good
good
example
in
the
brain
that
I
used
for
this
is
like
audio
audio
is
just
like
a
one-dimensional
stream
of
inputs,
but
in
the
brain
it's
organized
two-dimensionally
with
you
know
like
a
kind
of
like
a
spectrogram.
That's
the
easiest
way
to
think
about
it
and
you
can
almost
process
audio
using
images.
So
there
you
have
alumni
and
extreme
that
it
might
make
sense
to
organize
in
the
to
be
one.
A
E
C
So
question:
yes,
what
so?
What
makes
what
makes
what
makes
the
inhibition
scheme
a
local
as
opposed
to
global?
So
if
I
had
two
of
those
things
like
in
your
continuation,
arrows
man,
if
you
had
another
say
sheet
to
the
left
of
that
by
virtue
of
you
having
to
you,
would
then
have
local
inhibition,
no
because
the
global,
so
isn't
there
a
change
in
the
in
in
the
algorithm
to
for
that.
It's
not
just
a
simple
yeah
beyond
a
consideration
of
neighborhood.
It's.
A
A
C
A
D
A
C
A
A
Well,
that
is
pretty
much
what
I
want
to
talk
about
today,
and
this
is
all
been
really
informative
to
me.
I'm
figuring
out
how
I'm
going
to
present
this.
So
all
this
make
sense
and
community
can
ask
questions.