►
From YouTube: HTM Hackers' Hangout - Feb 5, 2016
Description
Come chat with HTM committers and talk about current work and future plans.
A
We've
got
some
engineers
in
the
office
in
redwood
city.
Waiting
to
talk
to
us
also
got
David
Ray
the
programmer
of
HD
of
Java
and
Jeff
foal
was
going
to
give
us
an
update
on
some
visualization
stuff,
but
for
now,
if
you're,
following,
if
you're,
actually
following
along
and
I,
don't
think
you
could
see
this
in
chat.
A
A
This
is
really
more
of
a
30,000
foot
view
and
I'd
like
to
do
maybe
once
a
month
or
so
kind
of
keep
the
community
in
sync
with
what
us
developers
are
actually
doing
with
HTM
flick
across
the
board.
Applications
algorithm
enhancements
and
that
sort
of
thing
so
I
have
an
agenda
and
I
will
just
share
one
of
my
screens.
Here
we
go
so
here
was
the
agenda
of
going
to
to
run
through
and
then
so
for
some
people
probably
call
out
some
things.
I
want
them
to
talk
about
this.
A
So
let's
do
the
app
updates
later,
but
I
do
want
to
talk
about
some
of
these
other
things.
I,
don't
know
not
sure,
hey,
Richard,
just
joining
hi
Richard
I
was
going
to
ask
him
to
talk
to
be
involved
in
the
windows
discussion.
So
since
you're
here,
let's
let's
do
that.
First,
no
I
should
or
something
that
you
might
not
know
actually
exists
that
helped
us
out
with
issue
popped
up
on
the
mailing
list.
Recently,
hey
Marcus
I
heard
that
I
didn't
say
anything.
Can
you
hear?
A
Yes,
there
you
go
okay,
okay,
so
you
can
hear
some
people
and
you
can
see
other
people,
but
maybe
not
at
the
same
time
anyhow
so
on
the
new
decor
wiki.
By
the
way,
if
you
go
to
the
Duke
record,
github
repo
and
go
to
the
wiki
there's
a
Windows
development
page,
which
Richard
is
almost
entirely
filled
with
details
on
how
to
build
new
kurkure
on
windows.
So
you
probably
do
want
to
have
some
some
windows
and
development
chops
to
do
this.
A
But
we've
had
a
couple
of
people
have
some
success
getting
things
working
with
windows
following
this,
this
readme.
This
is
where
pointing
people
so
I
just
want
you
to
know
that
it
is
possible
to
build
and
run
your
pic
on
Windows.
If
you
know
what
you're
doing
it's
not
as
simple
at
this
point
as
just
you
know,
pip
install
new
york
on
windows
like
it
like
it
is
on
OSX
for
the
most
part,
but
we're
getting
there
and,
and
hopefully
we'll
get
closer
and
closer
to
that
state.
A
Oh
yeah,
back
to
the
the
wiki
page,
so
we
have
also
been
working
ahead.
Early
talk
too
much
about
this,
but
on
the
hackers
page
we've
been
talking
about
this
move
towards
bamboo,
which
is
a
continuous
integration
product
from
a
glass
again,
so
we
have
an
open
source
licence
for
this
product,
and
so
we're
running
these
bamboo
servers
and
Austin
and
I
well,
Austin
more
successfully
than
I
have
has
been
building
out
these
the
CI
systems,
these
pipelines
build
pipelines
from
depicted
core
we're
still
working
on
it.
A
B
Basically,
we're
trying
to
consolidate
all
of
our
resources
related
to
testing
to
those
locations.
We
have
currently
Travis
CI
and
a
fair
Phil's
and
her
longtime
Cappetta
Charlie
I'm
joining
for
the
growing
separate
version
of
the
pipeline,
independent,
of
course,
with
bamboo.
We'll
just
have
one
single
system
we
built
all
ages
and
the
next
person
at
the
windows
and
process
the
building
off
the
jobs
when
we
can
have
all
the
same
functionality
as
we
turn
around.
A
Yeah
I
do
PA
that
each
one
of
these
builds
can
have
a
linux,
build
an
OSX
build
eventually
a
windows
build
and
each
one
should
be
able
to
run
independently
on
the
same
platform.
So
we
don't
have
to
have
dr
cio
a
beer
out
beer
over
here
for
this
one,
and
we
can
have
a
lot
more
control
over
the
versions
of
the
operating
systems
testing,
as
we
basically
built
those
agents
that
are
running
these.
These
builds,
so
we
won't
be
tied
to
SC
I
and
and
what
version
of
the
potato
they
want
us
to
use.
A
For
example,
we
always
use
the
latest
visit,
which
is
great
because
we'll
find
issues
faster.
A
lot
of
times.
Somebody
comes
out
and
says
you
know:
latest
ubuntu
is
failing
because
of
XYZ
and
we'll
get
a
leg
up
on
those
issues,
because
we'll
know
immediately
as
soon
as
it
happens-
and
we
can
also
still
you
know-
run
bills
from
an
older
version
of
ability
or
even.
A
You
move
towards
the
future
to
do
exactly
what
we
wanted
with
these
different
operating
systems
that
we
want
to
build
on,
and
this
is
a
platform
that
we
can
take
advantage
of
in
the
HTML
ones.
To
move
towards
this
I've
offered
help
getting
him
set
up
as
well,
but
in
the
end,
that's
up
to
you
and
how
you
want
to
run
your
project
did,
but
but
I'm
I'm
excited
about
this
I
think
this
is
going
to
be
great
when
it's
done
right
now.
A
If
the
plan
is
to
build
out
the
new
pic
or
a
new
pic
pipeline
sort
of
in
parallel
and
damn
do
with
what
we
currently
around
out
there
and
once
we've
got
something
working
will
kind
of
just
switch
over.
So
that's
the
plan,
we're
making
good
progress.
Snap
super
fast,
because
infrastructure
work
isn't
the
most
fun.
B
A
What
that
one
yeah,
the
idea,
be
a
you
know:
Senate
our
work
is
open
source
everything
we're
doing
out
of
it
too.
So
our
projects
a
couple
projects
that
hopefully
will
talk
about
it,
we're
building
those
as
well
on
on
bamboo,
so
we're
going
to
have
a
whole
suite
of
our
open
source
software
or
building
a
safe
system,
and
that's
it's
really
going
to
help
simplify
things
in
the
long
run.
D
So
yeah
you
can
use
the
gazebos
place,
several
memory
is
usable
and
there
is
I
think
we
are
in
review
right
now
that
allow
you
to
use
if
you're,
okay,
that's
a
good
work
on.
So
with
that,
when
that
is
merged,
we
can
use
version
yeah
and
it
works
on
different,
has
passed.
It's
just
not
producing
identical
results
to
a
Python
version,
with
a
certain
configuration
of
parameters
so
that
investigations
underway,
but
yeah.
A
D
Checking
yeah:
this
is
a
really
big
step
once
this
is
done
and
just
to
remind
everyone,
I
think
it
was
Richard
initial
version
of
of
the
temple
memories
of
CPD
and
then
taking
and
taking
a
be
working
on
all
of
the
last
discrepancies
intersections
intended
for
integrating
it
into
a
bailar
salsa.
That's
worth.
Oh.
A
Yeah
we've
been
waiting
for
this
a
long
time,
just
as
a
reminder
what
this
is
kind
of
milestone,
I
mean
once
this
is
complete
and
we
can
say
you
know,
there's
algorithm
parity,
basically,
between
Python
at
sea,
then
there
is
a
potential
that
someone
could
come
along
and
write
their
own
client
for
the
c++
back
for
Luca
core
and
expose
whatever
language
bindings.
They
want
with
fleet
algorithm
package
in
new
core
and
I.
Think
that's
a
big
deal,
I'm,
not
sure.
F
F
D
A
To
ask
questions
through
the
QA
interface,
someone's
already
told
me
that
I
need
to
share
the
links
one
talking
so,
if
you're
looking
at
a
Google+
page.
For
this
event,
the
there's
a
link
in
the
comments
that
goes
to
a
wiki
page.
That
would
be
pages
all
the
links
of
everything
that
I'm
showing
right
now
and
I
did
and
while
we're
on
the
subject
of
temporal
memory
of
Jana
Dorothy
had
a
question
of.
What's
the
biggest
resource
are
in
coral
memory,.
E
D
The
mix
is
from
a
cpu
standpoint
and
from
an
algorithm
standpoint,
the
biggest
loop
in
there
is
the
loop
that
figure
try
to
figure
out
which
cells
are
predicted,
because
that
has
to
look
at
all
of
the
cells
and
each
cell
figure
out
whether
it
has
getting
sufficient
input.
So
from
an
optimization
standpoint,
that's
the
biggest
thing.
This
version
does
include
some
smart
ways
of
doing
that.
So
I
don't
know
if
it's
a
true
resource
hog
in
this
implementation
that
it
is
exceptionally
well.
A
A
E
E
There's
one
issue
that
came
up
where
we
build
multiple
dynamic
libraries
that
are
loaded
by
python
and
if
multiple
of
those
compile
in
the
same
camera
schema,
then
if
your
script
happens
to
load
both
those
dynamic
libraries
and
get
an
error.
So
I
have
I
implemented
a
change
where
we
separate
the
kegerator
schemas
into
their
own
dynamic,
live
race.
That
way,
multiple
of
our
Python
extension
libraries
use
the
same
schema.
E
E
D
E
E
There's
also
one
of
the
disappointing
aspects
about
you
think
I'm
rota.
We
found
out
that
the
Microsoft
apology
doesn't
support
all
features
that
our
users,
so
when
we
implanted
it,
we
had
to
disable
the
Terrell
take
out
windows.
I
did
notice
that
some
patent
proto
community
member
had
backported
enough
of
the
functionality
to
work
on
visual
studio
2013.
So
it
also
were
definitely
to
that.
A
F
Sure
so
I
guess
last
November
I
think
it
was
one
of
the
one
of
the
community
members
who
had
been
a
contributor
to
HTM
dead
Java
brought
up
the
issue
of
vector
representation,
and
there
have
been
a
couple
community
members
who
talked
about
you
know.
Why
is
it
that
we're
using
32
we're
using
you
know
for
our
big
vectors?
We're
using
hints
java
hints
are
32
bits
and
isn't
that
a
waste
of
money
of
memory
not
really
a
waste
of
money
but
so
yeah.
F
F
F
F
Right
so
we
were
looking
at
that
and
I
do
some
of
that
too.
In
EHT,
nepO
I
wrote
up
a
sparse,
binary,
matrix
class
and
then
we've
had
some
contributions
done
by
his
call.
Name
is
chalu,
which
puts
his
name
he's
in
jail,
ooh,
okay,
he's
in
Spain,
but
anyway,
and
he
was
doing
some
visual
coding.
He
wanted
to
encode
I
guess
he
was
doing
an
in
nest,
experiment
or
something
like
that.
I
hate.
F
Anyway,
he
was
talking
about
how
his
input
vectors,
were
you
know
in
the
gigabyte
range
and,
and
so
he
wanted
to
minimize
memory.
So
he
worked
on
a
couple
of
different
implementations,
one
with
a
low
memory
profile
and
another
that
was
pretty
fast.
So
there's
been
some
work
around
this
already
and
I
started
looking
into
it,
because
I
was
also
looking
into
linear,
algebra
libraries
to
use
for
arm
some
of
the
more
advanced
aspects
of
the
K&N
classifier
and
then
you
know,
I
was
looking
for
a
library
that
could
handle
both.
F
You
know
the
some
of
the
algorithms
for
the
K&N
classifier
and
also
help
with
some
of
the
other
data
structure
representations.
So
I
had
I
got
this
idea
to
use
a
compressed
bitmap,
a
list
of
compressed
bitmaps
for
the
backing
data
structure
for
the
sparse
binary
matrix
and
so
I
talked
to
there's
a
guy.
F
His
name
is
Daniel
de
Meyer
and
he's
got
like
white
papers
on
he's,
like
the
authority
on
data
compression
or
at
least
bitmap
compression
he's
got,
he's
got
implementations
that
are
ported
over
c++
and
a
few
different
languages
and
I
wrote
him
about
this
idea
to
use.
He
is
a
compressed
bitmap
scald,
roaring
bitmaps
they're
like
the
fastest
ones
in
at
least
in
the
Java
world,
I'm,
not
sure
about
the
c++
world
but
yeah,
and
to
use
those
compressed
bitmaps
in
Java.
F
You
know,
as
opposed
to
using
32
bits,
as
you
know,
to
represent
it
in,
but
in
Java
a
bit
gets
is
basically
translated
to
a
short
I
think
well,
when
you're
using
binary
arrays,
those
are
truly
actually
bitter
raids,
but
when
you're
using
a
binary,
primitive
type,
then
it's
a
short
anyway,
so
roaring
the
map.
So
daniel
meyer,
I
wrote
him
about.
He
got
intrigued
about
it.
F
Incredibly
huge
data
structures
mean
it's
going
to
obviously
save
memory,
but
he
also
said
it
should
improve
performance.
There's
been
some
work
on
I
subscene
on
Stack
Overflow
that
actually
using
the
32
bits,
because
that
is
the
register
size
in
the
JVM
that
that
actually
is
faster
than
using
partial
sizes,
so
I
I'm
kind
of
up
in
the
air.
So
I'm
going
to
write
a
benchmark
and
test
it.
F
Do
this
translation
I
think
I
finished
according
the
temporal
over
to
it,
but
I
just
the
special
pool
is
going
to
be
a
larger
thing
because
it
actually
uses
the
sparse
bitmap,
sparse,
binary,
matrix
arm
so
and
then
I
have
to
substitute
all
of
the
array.
Utils
the
Python
utilities
that
I
you
know.
Basically
ported
numpy
looks
like
and
try
to
get
that
in
using
these
roaring
bitmaps.
So
the
whole
idea
is
to
just
get
it.
Working
with
using
SD
are
using
bitmap
representations
for
sdrs
and
seeing
how
fast
it
is
so.
F
Right
right,
right
and
then
and
Daniel
Lamarre
also
suggested
using
boolean
arrays
to
so
we're
going
to
try
we'll
try.
Maybe
we'll
try
a
couple
of
things,
but
first
I
know
I'm
going
to
try
this
bitmap
thing,
because
that's
a
few
community
members
have
already
you
know
they've
inquired
about
it
and-
and
it
seems
like
you
know,
if
we
get
this
working,
that
it
could
be,
you
know
a
huge
performance
boost
for
HTML
java.
That's
great
thanks!.
F
A
C
C
A
C
C
G
C
Album
in
drinko,
if
they
are
cooperating
instructions
to
build
a
possible
by
contributions.
It's
interesting
in
the
sense
that
contain
this
illness
and
then
run
you
through
anywhere
on
the
machine
and
say
you
keep
that
Pokemon
by
some
distribution
on
your
business
suit.
Okay,
Brian
move!
Is
this
possible
by
contributions
from
where,
and
that
is
what
blooming
the
app.
C
A
So,
thank
you
very.
If
anyone
had
a
hard
time
hearing
that
the
audio
from
the
offices
isn't
very
good,
apparently
so.
The
whole
idea
behind
this
each
team
studio
for
the
unicorn
so
directory
in
the
midst
repository,
is
to
make
it
easy
for
anyone
cross-platform
to
to
run
from
dupid.
So
Marian
and
her
team
has
done
a
great
job
like
bundling
a
portable
Python
installation
to
be
installed.
With
this
electron
alpha
and
electron
beam,
a
JavaScript
environment,
they
can
run
of
Windows
or
Linux
or
whatever
just
comes
bundled
with
python.
A
Is
you
want
to
python
and
I'm
Vivek,
which
you
may
not
have
either
so
she's
saying
now?
You
could
try
and
write
yourself,
but
you
do
have
a
running
locally
for
it
to
work.
So
thanks
for
that
update
and
then
the
main
thing
you
know
it's
going
to
allow
you
to
take
data
files
that
that
you
have
and
watch
them
run,
HTM
and
see
anomalies
happening
on
the
sort
of
live
on
your
own
data,
which
can
be
kind
of
hard
if
you're
trying
to
do
it
through
compiling
and
writing
everything
locally
yourself.
A
G
Yeah
sure
so
mark
was
asking
a
few
weeks
ago
about
getting
some
help
with
visualizing
CSV
files,
basically,
output
of
opf
and
I
thought
that
was
founded
interesting
to
me,
so
he
and
I
have
been
working
on
a
little
sort
of
a
little
mini
app
to
just
read
and
view
CSV
files.
A
second
I
guess
I
could
I
share.
My
screen
to
show
it
looks
like
see,
can
do
that.
B
C
G
So
there's
kind
of
the
thing:
a
work
in
progress:
that's
definitely
a
work
in
progress
and
the
it's
I've
just
been
working
with
mark
on
it,
and
so
actually,
if
anybody's
finds
us
useful,
I
would
love
to
get
some
feedback
on
what
people
find
useful
and
what
what
their
needs
are.
The
project
is
up
on
the
new
pic
community.
Well,
new
pic
visualizations
mess
until
we
get
a
better
name
for
it
and
it's
a
simple
web
app
that
uses
at
digraphs,
which
is
a
Oh
kind
of
a
little
tool,
open
source
tool
for
chori.
G
G
G
G
G
G
G
This
is
obviously
not
very
secure,
very
good
idea
to
pulse
this
publicly.
So
if
someone
were
to
post
it
publicly,
we
need
a
method
for
not
enabling
this
feature,
so
you
can't
just
type
in
a
path
to
some
file
up
and
have
the
public
read
it
so
in.
But
if
this
were
ever
posted
publicly,
then
I
suppose
we
would
have
a
different
kind
of
interface.
What
I'm
imagining
here
is
eventually,
as
you'd,
be
able
to
pan
back
and
forth.
G
G
G
I
don't
think
it
really
matters
enough.
You
know
the
things
we're
doing
are
using
any
like.
You
know,
latest
features
of
node.
Oh
all,
right.
If
people
interested
in
the
architecture
the
it's
a
it's
a
node
app
using
angular
as
a
framework
for
the
the
client
and
we're
using
socket
I/o
for
communication
between
the
server
and
the
client.
A
D
G
A
You
know
it
because
I'd
like
to
use
this
for
some
of
our
demos,
like
the
hot
gym
tutorial,
which
currently
uses
Matt
hot
lip
as
a
bunch
of
extra
coat
just
dedicated
to
Matt
pop
them,
I,
really
hate
matt,
lawton
so
having
having
something
that
can
just
like
be
watching
this.
The
CSV
file,
that's
really
easy
to
populate
with
dementia
or
results,
and
then
just
show
applause.
It's
probably
it
would
be
really
nice.
G
A
D
No,
not
here
but
I,
could
give
an
update,
yeah
great,
there's
two
sides
to
it:
there's
the
healthy
research
and
then
alex
is
work.
It'd,
be
kind
of
an
app
version
of
that
called
include,
with
working
with
Austin
and
on
the
NLP
research
side
would
be
working
a
lot
with
the
cortical
iOS.
Dr,
is
trying
to
see
exploring
different
applications
of
that
and
more
interesting
me
trying
to
take
that
as
input
and
they're
just
outputting
SDRs
weekend,
because
that
is
input
into
temporal
family.
D
A
couple,
cooling
and
odd
and
experimental
of
those
algorithms
and
try
to
see
clears
up
if
you're
saying
use
cases
in
natural
language,
we're
using
sequences
and
type
of
bullying
and
so
on.
That's
a
long
term.
Condo
recent
project
that
we're
working
on
it
go
is
an
app
that
Alex
and
Austin
are
doing
that.
Helps
you
visualize
some
of
the
results
and
interact
with
the
mom.
D
A
B
F
A
D
A
Okay,
all
right
thanks
for
the
update
tokbox,
it
seems
to
do
them
to
Newkirk
visualization
I'll
talk
briefly
about
these.
These
other
couple,
things
I'm
sure
my
screen,
like
I,
said
you
guys
need
to
run,
go
for
it,
but
I'll
take
advantage
of
having
people
here
to
talk
about
each
gym
in
the
cloud.
Okay.
So
first
of
all,
this
is
not
like
HTM
engine
HD
image
is
much
more
production
rate
than
this
HTM
in
the
cloud
thing.
A
So
basically,
this
is
a
mostly
Jonathan
Mackenzie's
work,
because
I
asked
on
the
mailing
list
for
somebody
to
build
this
I
needed
like
a
REST
API
that
has
you
know,
HTM
running
mind.
I
want
to
create
some
models
through
it.
I
want
past
data
and
I
want
to
get
predictions
back.
That's
that's
basically
it!
So
it's
a
very
poor
man's
version
of
that
I
have
touched
me
like
it's
about
two
months
old.
Now.
I
have
enough
time
to
really
work
on
this,
but
but
I
have
plans
for
this
thing.
A
It's
it's
a
very
little
application
right
outfit.
It
was
potential
to
maybe
tear
out
of
the
HTM
guts
and
put
HTM
engine
in
its
place
at
some
point
or
something
more
resilient
than
what
we
have
right
now
right
now,
you
can
start
it
up.
You
can
run
maybe
three
or
four
models
at
the
same
time
before
you
run
out
of
resources
and
then,
if
you
set
it
up
data
to
those
models,
eventually
you'll
run
out
of
resources.
E
A
And
I
did
and
so
did
Jonathan.
So
we
look.
I
worked
with
Jonathan
to
kind
of
outline
all
the
different
potential
ways
we
could
do
this
and
that
using
the
H
game,
simple
server
was
was
one
he
would.
It
had
put
a
little
bit
more
functionality
than
that,
so
he
just
runs
on
which
was
okay
with
me.
I
was
just
looking
for
results,
so
so,
basically
it's
just
a
REST
API.
A
A
The
data
is
coming
in
a
totally
different
rates
from
all
these
different
sensors,
so
the
aggregation
parts
com
there's
lots
of
things
that
need
to
get
done
to
fix
this
up
and
they're
all
defined
in
all
these
issues
here
and
as
well
as
in
HTC
hitc
HTM,
the
cloud
there's
a
bunch
of
relevant
issues
that
need
work.
So
these
are
things
that,
if
somebody's
interested
in
working
on
the
HTML
occasions,
you
don't
want
to
get
down
under.
A
A
E
A
C
F
Has
there
been
any
observation
of
new
pics
slow
down
like
running
after
running
like
ours
like
what
happens
when
segments
you
know
when
you
have
like
a
ton
of
segments
that
get
created
and
stuff?
Is
there
going
to
be
any
work
on
I?
Don't
know
if
you
guys
have
already
implemented
or
not
because
I
know
you
get
rid
of
the
synapses,
but
you
still
have
a
segment
so
which
means
that
any
lists
that
have
to
go
through
segments
and
find
segments
that
aren't
used.
F
F
F
F
E
There's
been
some
recent
changes
around
how
exactly
thing
to
work
when
you
hit
the
limits
and
at
some
point
there
was
some
cases
where
there
was
some
exceptions,
I'd
be
from
person,
apses
and
that's
been.
That's
has
been
resolved,
yeah,
Jayne
and
saying
should
be
the
ones
talking
right
now
that
they've
been
working
on
some
of
the
stuff
recently
but
yeah.
F
G
D
A
Ok,
everybody:
it
sounds
like
a
wrap
it
up
now.
So
I
just
want
to
say
thanks
to
everybody
who
joined
everybody,
who's
watching
I'm
gonna,
do
a
typical
YouTube
thing
down
sit
you've
enjoyed
this
YouTube
broadcast.
Please
hit
the
like
and
subscribe
button
at
the
bottom
of
your
screen.
Yeah,
oh,
and
if
you
like
this
new
format,.
A
Hangout
thing:
if
thats
this
is
better
than
the
old
one,
which
is
new
development
progress
review.
Please
let
me
know
I'm
going
on
with
some
feedback
and
see
what
you
guys
say:
I'm
trying
to
do
this
for
the
community.
This
is
useful
for
the
community,
that's
better
than
I'm
happy.
So
let
me
know
please
thanks
guys
thanks
everybody
take
care
and
have
a
wonderful
weekend
and
I'll
see
you
soon.