►
From YouTube: HTM Hackers' Hangout - Apr 1, 2016
Description
Just hanging out and talking about HTM.
Links from the hangout:
https://github.com/numenta/nupic.core/issues/891
http://connected-car.devpost.com/
https://www.youtube.com/watch?v=XMB0ri4qgwc
https://github.com/rhyolight/paso-trucks
A
A
A
C
C
C
A
C
C
B
C
B
C
C
C
A
That
would
be
you
that
would
make
me
anger
too,
but
same
thing.
You
know:
don't
go
to
production
depending
on
other
people's
services.
The
fall
yeah
well,
unless
their
services
are
a
part
of
your
service,
I
mean
if
it's
a
production
service
of
me
being
dead
again
was
not
meant
to
be
lent
upon,
like
that.
Yeah.
A
Right:
yeah,
ok,
so
the
research
guy
says
it's
not
going
to
end
anytime
soon,
but
the
let's
assume,
let's
make
the
safe
assumption
that
they're,
probably
working
something
important
and
a
lot
of
work
on
it,
whether
they
can
join
in
the
next
time.
So
I
got
a
couple
things
I
want
to
talk
about,
which
is
basically
just
like
update
on
on
what
the
hell
Matt's
doing.
A
First,
with
the
children's
meant
doing
what
the
hell
is.
My
question
so
last
week
we
had
a
wasn't.
Last
week,
two
weeks
ago
we
had
a
meet
up
in
the
Bay
Area
a
hand
and
put
on,
and
we
had
a
good
turnout
with
hit
some
good
presentations
and
Aaron
right
and
I
had
a
little
brain
bailed.
We
put
presented
on
really
similar
topics,
which
was
you
know,
using
HTM
with
streaming
data
different
strategies
of
streaming
data
into
HTM.
He
used
Apache,
flink
and
I
was
using
in
flex
TBH.
A
I
think,
there's
a
good
amount,
at
least
of
ideological
overlap
in
the
way
that
we're
thinking
and
so
I
had
a
chat
with
Aaron
yesterday,
the
day
before
over
the
phone
for
an
hour.
So
just
talking
about
streaming
data
and
potential
future
landscape.
For
that-
and
you
know,
we've
also
talked
up
with
Austin
and
Sue
bataa
about
flank,
as
potentially
a
way
to
distribute
and
scale
HDM
processing
in
in
lieu
of
HTM
engine
or
taking
the
place
of
that
sort
of
in
the
applications.
A
Architecture,
which
I
think
is
a
lot
of
potential,
because
you
could
expose
data
sources
to
flink
pretty
easily.
So
if
you
can
create
an
H,
some
type
of
flint
architecture
with
the
distributed
HTM
processing
and
and
expose
a
bunch
of
different
streams
to
fling,
you
can
expose
them
all
automatically
to
HTM,
and
you
know
Aaron
and
I
are
brainstorming.
This
is
one
of
the
reasons
I
put
riverview
together
was
to
create
some
type
of
standard
for
consuming
screaming
data
and
I.
Think
the
more
I
think
about
it.
A
There's
these
particular
open
data
platforms
that
are
emerging
so
crowded
is
one
of
them.
There's
even
open
golf
is
getting
into
that
business
too,
and
we'll
probably
have
a
handful
of
them
at
some
point
that
are
providing
live
streaming.
Public
data
for
free,
without
an
authenticated
and
being
able
to
create
something
to
adapt
those
and
expose
them
to
the
flink,
our
infrastructure
or
whatever
else
good.
No,
no,
my
daughter
is
trying
to
bug
me
I
could
be
really
powerful
once
we
well
right
now.
A
I
think
that
if
I
had
a
few
weeks,
I
could
put
something
together
to
expose
all
the
temporal
streaming
data
and
so
Prada
to
in
a
river
G
like
interface,
so
that
anybody
could
just
have
a
little
Python
client
that
consumes
that
there
it's
it's
a
little
a
little
bit
tricky
though,
because
and
in
fact
with
soprano
guys
as
well,
because
they're
their
data
is
not
indexed
in
a
way
that,
like
calls
out
streaming
data,
you
can't
figure
out
which
of
their
data
streams,
is
really
a
data
stream.
A
It
could
be:
oh
just
a
bag
of
data
spreadsheet
or
whatever
anyway,
there's
some
complexities
involved,
but
I
think
that's
the
way.
The
future
I
would
love
to
be
able
to
have
some
type
of
distributed
architecture
where
you
could
just
drop
in
new
data
sources
and
automatically
turn
models
on
that
start.
Consuming
those
data
sources
providing
predictions
or
whatever,
so
that's
one
thing
and
thinking
about
I've
thought
about
that
for
a
long
time,
I
still
think,
there's
a
lot
of
cool
stuff
that
could
be
built.
Oh
I,.
B
B
A
A
So,
no
to
be
any
language
you
just
have
to.
You
know
that
the
tricky
thing
is
so
crowded
does
not
have
a
registry
that
keeps
track
of
what
what
of
a
date
of
its
data
stream.
It
has
a
huge
catalog
data
across
the
whole
world,
but
you
can't
tell
what's
to
portal:
what's
not
so
I
wrote
this
utility
called
soda
tap
that
registers
like
the
tests
out
every
every
possible
data
source
and
and
kind
of
just
samples
it
to
see
if
its
temporal
or
not,
but
there's
some
issues
with
it
like
data
cardinality.
A
A
lot
of
these
streams
have
lots
of
different
streams
in
them
and
that's
not
called
out
that's
something
that
you
have
to
investigate
manually
to
find
out
anyway.
There's
more
work
to
be
done,
but
thats
future
vision.
You
know.
So
there
is
a
public
data
marketplace
emerging
today,
right
now
on
the
internet
that
nobody's
like
architected,
it's
just
happening
and
I
want
to
be
aware
of
that.
A
B
A
Okay,
another
topic
connected
cars.
Let
me
give
you
a
link.
How
do
I
do
this?
So
there's
a
hackathon
coming
up
chat.
It's
called
the
connected
car
hackathon
and
this
is
in
oakland
california.
So
it's
a
local
hackathon
it's
next
weekend
april,
ninth
and
tenth
and
newman
side
is
sponsoring
the
sacrifice.
So
it's
a
small
hackathon.
It's
not
a
huge
deal,
but
they've
got
good
data,
so
they're
using
an
API
called
automatic
which
is
connected.
A
Car
API
gives
you
potentially
information
about
locations
of
cars
tracks
and
then
a
bunch
of
metrics
from
the
cars
themselves
like
being
recorded.
It
live
so
I.
Don't
know
how
they're
going
to
do
this
at
the
Agathon,
but
apparently
we
are
going
to
be
able
to
use
that
API
to
build
interesting,
connected
car
applications.
So
because
this
was
a
I
thought
could
be
an
awesome
data
set
for
use
with
HCM
because
it's
got
GPS.
It's
got
screaming
scalar
numbers
coming
for
cars.
A
We
approached
them
and
asked
if
we've
also
I'm
going
to
have
a
breakout
session
at
this
hackathon
where
explain,
HTM
and
how
it
could
apply
to
this.
This
topic
and
we'll
have
a
little
booth
there,
where
we
can
give
up
materials
I
can
help
you
get
started
and
we'll
probably
some
schwag
and
stuff
so
anybody's
watching
us,
and
you
want
to
go
to
a
cool
hackathon
next
weekend
april,
9th
intense
in
oakland.
A
It's
connected
the
car
Devlin's,
calm
posted
it
in
the
chat
here
come
on
out
and
I'll
get
started
with
with
new
pic,
at
least
so
that's
one
other
thing
which
is
good,
so
you
may
or
may
have
noticed
that
I
have
started
a
new
video
tutorial
series
called
HTM
school
I've
been
working
on
this
for
a
week
or
so.
I
have
two
episodes
recorded
and
I'll,
be
working
on
a
third
episode
next
week
on
sparsity
stupid
representations.
So
it's
on
our
YouTube
channel
there's
episode
0,
which
is
just
an
HTM
overview
and
I.
B
A
You
know
I'm,
basically
regurgitating
stuff
that
Jeff
has
said.
So
it's
not
really
original
content.
You've
probably
seen
this
before,
but
I
wanted
to
make
a
10
to
15
minute
video
that
I
could
give
to
anybody
and
explain
what
I
do
or
what
we
do.
What
we're
trying
to
do
and
I
think
it
sheaves
that
goal
at
least
so
in
15
minutes.
Here's
what
HTM
is
and
why
we're
working
on
why
it's
important!
That
was
the
main
goal
of
that.
First
episode:
oh
I'll,
give
you
a
link.
A
I
don't
know
if
this
is
going
to
work
on
youtube
or
not,
but
I
think
I'll
throw
these
links
into
the
youtube
description
if
you're
watching
this
on
youtube
just
link
to
the
video.
So
that's
the
first
episode
the
next
one
is
is
really
I
mean
these
are
really
basic
we're
starting
from
like
nothing
here.
A
And
so
there's
there's
that
and
might
I
have
this
overall
plan
20
to
30
episodes.
I,
don't
know
how
far
along
I'm
going
to
get.
If
there's
a
lot
of
views
and
the
popular
or
keep
going
if
it,
if
I'm
not
getting
a
lot
of
feedback
and
people,
not
many
people
are
watching
them,
I'm,
probably
not
going
to
keep
doing
it,
but
we'll
just
we'll
just
see
what
happens.
A
A
They
are
oh
yeah,
so
I
I
go
over
like
what
an
oar
is,
what
an
ex
or
is
what
didn't
and
is
and
then
talk
about
more
about
sparsity
in
general
and
the
properties
of
a
sparse
array
versus
a
tensor
a
and
you
know
very
obviously
leading
into
STRs
SDRs
is
a
pretty
big
topic.
Those
I'm
pretty
sure
I'll
do
three
or
four
episodes
on
that's
on
SD
are
so
we'll
see.
A
Yeah,
so
the
groundwork
is
laid
in
that
first
bit
array
video,
which
makes
a
little
bit
easier
to
do
so.
The
last
thing
I
wanted
to
talk
about
is
something
I'm
very
excited
about
talked
about
this
in
the
mailing
listed
on
chat
is
I've,
got
all
this
garbage
truck
data.
So
it's
funny
stories,
I
moved
from
cupertino
to
Paso
Robles
California,
which
is
where
I'm
at
now
it's
about
three
hours,
south
and
hooked
up
on
next
door.
Next
door
is
a
social
web
site
for
neighborhoods
and
met
a
guy.
A
It
might
aber
Hood
named
Jed
Dawson,
who
is
like
the
I
for
the
local
waste
management
company,
so
the
Paso
Robles
waste
and
recycle
company.
They
run
all
the
garbage
trucks,
all
the
solid
waste
trucks,
garbage
recycling
yard
waste
all
that
stuff,
and
apparently
his
family
is
involved
in
the
owners,
and
you
know
he
decided
he
wanted
to
do
computers
instead
of
be
a
garbage
man
at
some
point,
so
he's
like
the
computer
guy
for
their
garbage
company
and
he
provides
software
for
like
neighboring
garbage
company.
A
General
manager
has
been
nice
enough
to
approve
the
public
release
of
this
data
if
he
calls
it
there,
the
breadcrumb
data
from
their
trucks
and
it's
just
latitude
longitude
and
speed
and
a
timestamp,
but
more
for
all
their
trucks
over
the
past
year.
So
I've
got
all
of
that
data
online
and
I've
been
working
with
it.
It's
been
very
fun,
I
can
I
was
your
birthday,
so.
B
A
You
know
I
was
a
little
bit
worried,
but
so
they've
assured
me
that
the
data
is
fine
being
public
and
it's
going
to
be
stamped
with
an
open
data
license
soon,
where
I'm
just
trying
to
get
the
official
approval
from
their
general
manager
or
email
that
you
can
license
this
as
open
data,
which
is
means
essentially
means
anybody
can
do
whatever
they
want
with
it
that
period
but
yeah,
I'm
hoping
we'll
find
some
interesting
thing.
So
the
nice
thing
is
I've
got
a
contact
there.
So
I
can
ask
him
hey.
A
What's
what's
what
are
any
interesting
anomalies
happen
over
the
past
year
and
you
can
tell
me
yeah
look
at
truck
60,
so
right
now,
I'm,
processing
truck
60
because
he
says
something:
anomalous
happened
that
they
noted
in
truck
60
over
the
past
year.
So
I'm
currently
got
my
cpu
processing
truck
60
yeah,
and
it's
going
to
take
a
while
because-
and
you
know,
there's
a
bit
of
experimentation
that
I
need
to
do
here,
because
I
have
to
figure
out
at
what
the
the
geo
spatial
coordinate.
A
Encoder
has
I
think
just
one
parameter,
that's
scale
and
so
I'm
doing
a
scale
of
5
meters
and
the
other
thing
is:
should
I
subsample
or
not,
because
I'm
getting
data
every
15
seconds,
I'm
I'm
running
some
with
some
sampling,
some
without
subsampling.
It
just
takes
a
long
time.
So
I've
got
a
lot
of
different
experiments
to
run,
to
figure
out
the
best
set
of
the
best
setup
for
this,
but
it's,
but
it
looks
promising
already
so.
I
can
already
find
some
things.
A
I'm
like
yes,
I
see
that
that
track
is
not
more
anomalous
on
this
Thursday
than
the
last
Thursday
I
can
tell
why
so
at
some
point,
I'll
have
to
come
up
with
a
way
that
I
can
scour
all
these
tracks
and
you
know
automatically
find
the
anomalous
periods
and
violate
them
somehow,
so
that
I
can
manually
review
them
and
see
if
I
can
find
a
cause.
So
that's
what
I'm
working
on
and
I
pasted
a
link
that
will
be
in
the
description.
A
C
I'm
I'm
super
excited
to
get
in
on
that
data.
I,
don't
know
when
I'll
find
the
time,
but
I
find
it
really
controlling.
B
A
Yeah
yeah,
it's
a
great
data
set
really
and
the
cool
thing
is
so
there's
a
couple
things
you
could
do:
I'm
taking
the
truck
centric
route,
so
I'm
going
to
create
model
for
each
truck
and
I'm
breaking
up
I.
Think
when
I
process
that
the
routes
I'll
take
any
for
our
break
or
a
break
of
four
hours
or
longer
I'm
going
to
soon
that
they
parked
the
truck
for
the
night
and
then
a
new,
a
new
sequence
starts
on
the
next
morning.
That's
that's
work
really!
Well,
it's
like
430am.
A
The
truck
starts
rolls
out
and
it
rolls
back
in
about
4,
30
p.m.
and
that's
their
day,
but
there's
other
ways
you
could
do
it
if
you
that
I
think
we
could
probably
figure
out
how
to
make
this
work
is.
Another
way
is
like
there's
certain
areas
of
town
that
get
a
lot
of
activity
like
there's
a
lot
of
garbage
trucks
in
the
more
popular
populous
areas.
A
I
would
like
to
be
able
to
draw
a
polygon
of
some
sort
and
and
analyze
an
area
and
have
a
model
for
different
areas
in
space
and
and
somehow
you
know,
process
over
time
that
ignoring
the
fact
that
there
are
individual
trucks-
just
something's
moving
from
here
to
here
from
here
to
here
this
area
over
time
and
I'm
wondering
if
we
could
train
a
model
to
recognize
the
activity
of
things
moving
in
space
in
a
area
instead
of
folk
start
object.
So
that
would
be
like
sort
of
the
next
experiment.
I
was
thinking
doing
yeah.
C
I
think
something
that
might
be
particularly
interesting
is
a
you
know.
You
just
moved
its
paso
robles,
I,
don't
know
how
long
or
if
at
all
your
house
was
occupied,
but
let's
say:
there's,
there's,
there's
a
change
in
residency.
What
or
a
new
house
or
new
property
is
built
or
you
know,
there's
a
new
road
or
you
know
our
roads
shut
down
or
something
the
interesting
to
see.
If
you
can
pick
up
on
that
kind
of
data
automatically
yeah
it.
A
B
A
Think
that
we'll
be
able
to
catch
some
of
that
stuff
that
I
don't
know
about
a
new
stop,
but
I
can't
figure
out
how
time-sensitive
it
is
yet
I
haven't
processed
enough
data
yet
like
if,
if
a
truck
that
usually
drives
by
30
miles
an
hour
suddenly
stops
for
30
seconds
at
that
spot
and
then
it
continues
on.
You
should
be
able
to
see
that
in
the
gps
track,
but
but
how
sensitive
is
HD
I'm
going
to
be
to
that
I'm,
not
sure?
Yet,
when.
B
I'm
concerned
about
for
the
area's,
it
stops
and
starts
so
like
if
you,
if
you
have
an
area
and
you
have
objects
which
are
present
in
there
at
time,
zero
and
they
move
and
they
move.
If
they
stay
in
the
area
for
an
extended
track,
and
then
they
tend
to
repeat
that
on
a
daily
basis,
then
maybe
you
could
get
some
traction
there,
but
what?
If
they
just
start
in
that
area
and
like
five
minutes
later,
they
leave
the
area.
And
you
get
a
lot
of
that.
B
A
Don't
neither
for
the
area
centric,
oh
I,
don't
know
yet
somebody
on
our
mailing.
This
was
trying
to
do
this
a
month
or
two
ago
and
I
don't
didn't,
have
any
good
advice
for
them.
I,
don't
know
how
far
we
got.
But
you
know,
all
of
our
current
stuff
is
set
up
sort
of
an
object
centric
on
that's
not
right.
A
And
it's
a
great
data
set
to
try
something
like
that,
because
it's
because
it's
free
anybody
can
check
it
out
and
use
it.
So
if
anybody
wants
to
try
something,
that's
never
been
try
with
HTM
go
to
this
paso
trucks,
github
repo
and
try
and
figure
out
how
to
do
an
area
center.
Gps
analysis
of
the
truck
truck
tracks.
C
It
may
not
be
I,
wouldn't
say
it's
not
a
good
idea,
but
I
would
consider
trying
to
organize
it
by
route
rather
than
my
truck
particular
truck.
I'm
guessing
you
know
when
you
have
the
cycle
of
you
know
trash
days,
particular
truck
is
going
to
have
a
different
route
each
day
and
that
route
is
is
going
to
be
cyclical.
C
So
it's
always
going
to
be
no
same
day
of
the
week
every
week,
but
on
a
different
day
of
the
week,
is
going
to
be
a
different
realm
because
they
have
to
be
able
to
share
the
trucks
that
way
and
it
I'm
suspecting
that
not
every
truck
always
takes
the
same
realm.
So
a
particular
route
might
be
serviced
by
any
number
of
trucks.
Yeah
I've
seen
that
so.
A
It
looks
like
there
are
primary
trucks
and
I've
only
I've
only
looked
at
four
different
tracks
out
of
these,
like
20
or
30
of
them,
and
it
looks
like
there
are
some
primaries
that
typically
will
do
the
monday
through
friday,
routine
every
day
is
a
different
route,
but
then,
like
that,
they'll
be
just
a
wednesday.
That's
missing!
Well,
they
didn't
go
out
at
all.
So
I
you
have
to
assume
some
other
truck
did
that
route,
because
there
was
proper
truck
or
I
took
a
day
off,
I,
don't
know
so
yeah
we're
gonna
have
that
problem.
A
That's
going
to
be
anomalous
when
this
other
truck
that
usually
does
another
route
is
suddenly
doing
a
completely
different
route,
but
is
that
something
that
the
end
user
would
want
to
know
about?
I,
don't
know,
maybe
that's
valuable
for
them
to
know.
Maybe
they
let
their
drivers
make
those
decisions
on
their
own
and
it
would
be
useful
for
them
to
have
that
theory.
Well,.
A
C
A
Cuz
they're
all
going
to
be
a
little
different
and
I've
already
noticed
little
in
discrepancies
between
routes.
You
know
I'll
go
flat,
5
Thursdays
and
look
at
the
routes
over
top
of
each
other
and
this
one
route.
For
some
reason
they
did
counter
clockwise
itself
off
once
and
I.
Don't
know
why
but
new
pic
notices,
but
it's
a
great
data
set
and
I
think
this
is.
A
This
could
potentially
turn
into
some
really
interesting
demo,
so
what
we
can
do
with
GPS
data,
so
I'm
super
excited
if
Jed
Dawson's,
watching
thanks
dude,
it
was
really
awesome
of
you
to
help
me
get
this
data
out.
Oh,
that's!
All
I
got
guys,
that's
the
that's
the
Hangout!
Thanks
for
watching
everybody
and
happy
HTM
hacking,
and
hopefully
you
will
see
one
of
you
guys
at
disconnected
car
hackathon
or
whatever
anybody
has
any
feedback
on
HTM
school.
Let
me
know.