►
From YouTube: HTM Meetup - San Francisco Bay Area | hierarchical temporal memory, numenta, neuroscience, artificia
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
Wanted
right
over
there,
so
there's
a
there's,
a
camera
rate
right
here.
You
know
camera
right
there
and
then
there's
that
one
sitting
on
the
table
and
that
was
sort
of
old
stick
to
move
it
around.
But
yeah
I've
been
doing
this
for
a
few
months
now,
just
just
for
the
past
few
months,
I'm
starting
to
have,
but
we
we
have
a
forum
right
and
inform
I.
Don't
know
if
you
guys
remember.
C
A
A
A
Engineers
that
they
could
just
like
the
gamers
we're
doing
live
coding
or
life
playing,
they
do
black
pony,
and
so
some
people
started
to
do
live
coding
and
using
the
Sam
gamer
setups
to
program
and
like
interact
with
people
at
a
chakra
and
it's
sort
of
catching
on.
So
now,
usually
when
you,
when
you
broadcast
on
switch
to
take
a
game
that
you're
playing
now,
there's
a
game
called
science
and
technology.
Oh.
A
A
Which
reminds
me
there's
people
meeting
us
here.
I'm,
assuming
is
your
my
excuted
I
think
everything
is
working.
I
always
got
the
double
check
because
you
never
know
sounds
fine,
so
somebody's
there
and
you
tell
them
I,
guess:
I
circuits
no
getting
to
know
people
after
a
while,
like
this
person
has
their
own
interesting
Channel.
Some
people
have
their
own
fish
channels.
You
know
this
is
the
chapter
I
saw.
A
A
A
There's
the
other,
so
there's
got
three
cameras
and
they're
just
sort
of
rotated,
all
right,
so
they'll
get
like
there's
a
wake
webcam
and
there's
this
cam,
which
is
mobile
and
I've,
decided
what
I
want
to
do
it
yet
and
then
there's
this
one
in
the
corner,
so
yeah
we're
streaming
Elvis
liked
by
people
who
are
watching
and
there's
a
there's.
An
info
beautiful
little
tend
to
show
up
and
ask
questions.
E
Talk
about
what
our
current
right
now
so
taking
our
biological
theories
and
how
we're
going
to
migrate
them
into
machine
learning,
and
we
have
a
pathway
to
do
that.
Slow
down
a
path.
So
I
wouldn't
say
it's
a
middle
ground,
it's
more
like
Heidi
Burge,
these
two
fields
together,
you
know
and
a
lot
of
the
work.
That's
going
a
lot
of
people,
the
debugging
we're
all
over
saying,
hey.
We
need
some
new
ideas,
inspiration
from
the
brain.
E
E
So
many
things
we
know
about
one
of
the
first
things
that
to
tidy
and
lowest
and
Lucas
are
working
on.
Right
now
is
to
take
some
of
the
fundamental
aspects
of
brain
science
like
the
boldest
activations
as
far
as
presentations
and
apply
those
to
neural
damage.
You
can
take
existing
publish
them,
no
matter
where
deep
network
and
have
less
sparsity
to
them
and
I'm
gonna,
try,
movies
and.
E
A
I
A
J
K
A
E
E
I
A
A
E
N
Because
of
the
time
difference,
I
watch
a
lot
a
place
rather
much
of
the
life
stuff,
but
he
actually
for
a
daughter,
right
now,
scene,
conference
and
so.
N
A
B
O
N
A
D
A
A
G
P
E
Q
B
D
K
Yeah
so
I'm
from
Europe
from
Prague,
so
I'm
hearing
there
or
one
more
because
I
work
for
singularity
University,
but
that's
just
some
kind
of
Eric
external
engagement
so
and
passionate
about
AI
and
neuroscience.
So
I
found
two
men
table
sometimes
super
spheres
years
ago
son
fully
cure
research
and
explain
to
mother
when
I.
First
read
this
all
to
your
stuff,
so
I
feel
that.
O
R
Cool
head:
my
name
is
Tommy,
while
I'm
originally
from
Finland,
where
studied
physics
and
neuroscience
here,
I
in
nine
years.
Working
currently
other
stuff
won't
start
so
hardware
stuff
I,
do
computer
wisdom,
data,
gurus
and
stuff
like
that,
I'm
interested
in
non-commercial,
a
I
approach.
These
different
concepts,
especially
local
learning,
distance
temporal,
including.
Q
Brian
I'm
back
in
and
software
engineer,
working
at
a
start-up
in
FinTech
and
2015
I
was
interning
here
in
Toronto
we
were
looking
at
the
temporal
polar
I
think
my
presentation
of
stability
before
that
I
was
very
interested
in
a
champ
dude
making
experiments
for
it.
I
was
doing
my
doctorate
so
from
there
I
came
out
here
and
rest
is
history.
S
B
S
S
New
and
it
struck
me
except
I-
never
you
know
still
in
gaming
graphics
for
a
long
time
kind
of
kind
of
have
a
ball
now
and
trying
to
get
back
into
dip
a
hot
topic
at
noon,
ventas
doing
a
kind
of
approach
to
it
and
I
just
kind
of
want
to
see
what
everybody
else
is
doing
and
trying
to
take
that
in
and
let's
see
you
know
what
worse
the
truth,
everything
I
think
you
guys
have
a
lot
of
good
thanks.
M
A
F
U
V
R
W
I'm
I'm
a
research
visit
here
and
I'm
curious
about
things
of
of
us
providing
and
I'm
independently,
which
is
why
I'm
quite
interested
in
your
work
as
well
I,
read
on
intentions
five
years
ago,
also
also
looked
into
your
opponent
courses
on
their
shape
and
I.
Think
it's
it's
quite
interesting
because
many
high-level
intuitions
are
might
become
relevant,
also
mainstream
machine
learning
it
at
some
point.
So
it's
quite
interesting
that
you
kind
of
working
in
this
direction
pretty
much
aside
from
from
the
mainstream,
so
yeah.
A
B
D
E
B
E
E
F
E
J
M
J
E
E
E
The
perception
of
depth
from
the
world
perception
of
position
in
the
world-
it's
all
actually
every
single
thing
is
represented
by
popular
honest
truth
throughout
the
year
projects
that
represent
those
locations
kind
of
reference.
So
there's
a
reference
for
associated
this
room
and
you
know
actually
what
where
the
table
is
in
this
happen,
I
mean
relative
to
the
refugees
overrun.
F
E
E
December
this
year,
so
as
you
know,
and
so
your
intelligence
with
all
of
this
overall
model,
like
assigning
reference
frames,
that
everything
is
how
the
brain
works
and
that's
how
you
can
understand
and
then
the
question
about
general
purpose,
intelligence
or
flexibility
you
used
a
constrained
use
is
how
there
there
are
different
types
of
reference
for
interviews
and
it's
important.
What
type
of
use
you
want,
one
that
can
be
applied
to
many
different
types
of
problems,
so
the
competition
has
discovered.
E
F
F
Knowledge
of
any
sort
and
assigned
to
locations
in
a
reference
and
those
those
reference
frames
do
not
have
to
attach
to
physical
others.
They
be
processional
spaces
and,
just
like
you
can
move
between
a
physical
space
like
I
move
in
this
room
and
I
complain
my
emotions-
and
this
is
always
something
different.
When
you
move
mentally
through
these
phases,
which
is
everything
it's
mess
with
new
spaces,
you
can
move
through
the
space
language.
R
A
Been
a
couple
years,
what
I
used
to
run
to
a
year
and
did
that
for
three
years
and
six
months,
starting
in
2013,
but
we
went,
we
changed
our
focus
from
applications
of
each
tune
back
to
research,
okay,
duska
alone.
I'll
talk
a
little
bit
about
this
in
a
moment
was
a
very
angular
sort.
You
know
presentation
my
guys,
we're
rapping
on
yeah,
listen,
I!
Think.
A
B
A
I'm
not
doing
those
currently
I'm,
focusing
on
I'm
working
on
an
interactive
tutorial
on
how
to
build
HTM
systems
and
I'm
building
I'm
doing
that
on
Twitter
and
watch.
My
Twitter
stream
idea
once
a
week
at
work
and
people
are
helping
me
build
it,
which
is
really
fun
so
I,
don't
know
how
to
I'm,
not
very
good
at
react.
I'm
new,
like
so
there's
somebody
in
a
community
named
David
Sherman
who
locks
on
almost
every
time
and
helps
me.
People
react
applications.
I.
A
R
A
Programming
language
and
then
I
was
like
screw
this
Java
stuff
I
mean
I'm
a
functional,
dynamic
program
right
now,
I'm,
really
dangerous
and
then
and
then
remove.
The
JavaScript
APIs
have
been
awful
program,
represents.
B
A
A
N
A
A
To
the
point
where
I
can
create
I'll
just
create
a
rack
component
and
I
know
how
to
get
my
d3
in
the
right
place
and
I
do
all
those
interactions
in
d3
and
then
all
they
would
react,
interactions
would
react
and
the
only
complicated
part
is
window
when
I
have
to
go
between
them
and
I.
Try
and
do
that
just
with
state
just
with
react
state
if
I
can
so
there's.
N
A
Got
a
little
bit
of
a
taste
of
it,
because
just
this
past
week,
I
there's
I
needed
this
component.
That
I
want
data.
I
want
this
component
to
stream
data
and
I
have
all
the
diagrams
on
the
page
update
as
its
streaming
data.
So
it's
like
generating
data,
but
all
the
diagrams
just
updated,
and
so
this
is
I
hated
doing
this,
but
the
guy
helping
me
David
Truman,
had
me
create
something
called
up
high
order
component
and
react
that
you
wrapped
your
her
page
with
and
then
provide
it
like
ingest
the
data
into
it.
A
A
N
A
That's
the
type
of
payoff
you
get
from
using
react
when
you've
got
a
big
team
and
you've
got
all
these
standards
and
I'm,
just
like
a
team
of
one
but
I'm
also
I
want
it's.
What
everyone
is
using,
so
I
want
people
to
be
able
to
come
to
it
and
recognize
the
structure
of
it:
football,
hey,
there's
Mark
Brown,
you
said
my
hi
Mark
Brown,
Eddie
I,.
D
D
A
D
W
D
That
I'm,
like
as
someone
who's,
worked
on
a
web
browser
before
I'm
like
no
rely
on
the
native
control
to
put
the
text
there
and
you
read
it
don't
don't
write.
It
constantly
is
a
terrible
idea
from
like
just
the
the
responsiveness
of
that
a
technocracy
of
Botox
there's
just
like
there
are
some
things
that
reacts.
Those
people
tell
you
to
do
that.
It's
just
terrible
yeah
no
rely
on
the
browser
to
don't
use
JavaScript
to
set
the
text
of
a
box
like
the
browser
to
control,
take
the
input
and
use
of
that
way.
There's.
M
H
D
D
D
D
B
A
Styling
things
like
lots
of
big
big
company
students.
Now
those
style
pages
and
you'll
see
the
styles
of
last
names.
It'll
have
digits
in
them
and,
like
that's,
the
width
of
what
it's
supposed
to
be
like
some
and,
if
they're,
not
starting
it
based
on
what
it
is.
They're
styling,
if
they're,
adding
style
information.
B
A
B
O
A
A
B
A
J
N
A
B
A
A
A
Spices
I
move
from
Paso
Robles
and
then
that
didn't
work
out
an
hour
and
Knoxville
but
lacell
the
area
Watson
though,
which
is
sort
of
close
to
the
coast.
Really
nice
I
only
complain
to
the
house
is
super
small.
Everything
else
is
fine.
It
gives
are
happy
in
the
school
wife's,
happy
John,
I'm
happy
because
I'm
in
J's
yeah,
this
is
so
much
better.
So,
like
I
said,
I'll
complain
when
I
have
a
small
house
but
who
doesn't
even
California.
You
know.
X
A
A
Yeah
it
was,
you
start
out
promise
here.
Y
Y
Y
J
B
L
A
A
A
A
A
A
J
A
This
is
gonna
happen
and
what
what
usually
will
people
do
when
you're
interacting
directly
with,
is,
though,
you
could
push
panel
and
interact
on
discord,
I
think
I'm,
just
gonna
david
turn
it
and
I
repair
programming,
and
that
might
be
what
you
end
up
doing.
I've
never
been
on
that
side
of
it.
So
I
don't
really
know
what
to
do.
J
A
K
R
J
K
K
A
A
Have
research
implementations
of
yourselves,
so
there's
two
things
about
that
world
right.
You
can
learn
from
what
we
see
the
cells
doing,
and
so
you
can
see
like
that.
I've
used
that
as
a
tool.
So
we've
done
that.
There's
also
the
idea
of
how
do
these?
How
do
these
systems
emerge
or
learn?
How
do
they
already's?
How
does
this
learn
emergent?
A
Levels
of
murder
kind
of
these
networks
emerge.
That's
a
bit
of
a
deeper
topic.
There's
a
lot
of
ideas
out
there.
A
lot
of
people
figured
out
ways
they
could've
burned.
His
name
is
exactly
how
I
think
it
was
still
under
research.
You
know
we're
not
necessarily
focused
on
understanding
that,
knowing
what
we
know
how
they
work,
we
want
to
apply
how
they
work.
A
K
B
A
I'm
not
gonna,
that's
a
good
school,
but
what
I
usually
do
when
we
have
need
of
like
this
is
sort
of
go
over
the
state
of
things,
the
state
of
the
community
of
the
company,
where
we've
been
we're
going
and
all
that
jazz.
Let
me
flip
this
to
the
whiteboard
screen,
so
we're
not
going
in
this.
What
we
do
I
am
the
community
manager
for
dementia,
so
I've
been
managing
the
community
since
2013.
A
So
my
big
responsibility
was
a
project
called
new
pic
and
that
was
open
sourcing
in
2013.
This
was
my
baby
for
a
long
time.
We
didn't
have
any
releases,
we
didn't
have
any
compilation,
you
didn't
have
any
CI.
We
had
nothing
like
an
open
source.
We
just
made
it
public.
So,
oh
there
was
a
lot
of
work
that
went
into
this
and
just
this
last
year,
I
think
was
2018.
We
have.
We
have
put
new
pig
into
maintenance
mode,
May
10,
which
honestly
made
me
unhappy.
B
A
Is
Python
2,
7
I
found
2
7
and
we're
we're
not
going
to
update
it
to
2,
3.0
and
and
what
what
tends
to
happen
with
our
company
is
and
then
I
go
back
and
and
justify
this
by
talking
about
our
mission
because,
like
Jeff
said,
I
was
heartbroken
by
this,
but
it
was
also
my
idea,
and
the
reason
is
the
mission
that
we
have
momentum
is
more
important
than
a
software
project.
We
get
this
right
because
our.
I
A
To
reverse-engineer
the
neocortex
system
does
a
summation
of
it.
Yeah,
that's
just
reverse
engineer
the
neocortex
right
and
the
second
is
to
apply
what
we've
learned
apply:
our
knowledge
to
create
machine
intelligence.
However,
we
wanted
to
find
that
fit
your
create
machine
intelligence.
So
when
you
do
this,
you
have
to
inform
this
has
to
inform
how
we
do
that
and
then
they
test
it
out
and
then
we
come
back
and
say:
ok,
how
did
that
work
out?
A
A
T
J
E
A
D
E
B
E
Fact,
but
the
neocortex
is
also
built.
A
repetitive
element
called
the
court
of
a
cop
and
there's
150
thousand,
at
least
in
the
near
court,
they're
virtually
identical,
and
so
what
we
we
decided
to
do
is
focus
on
one
of
those
cortical
columns
dogs.
If
you
understand
what
that
one
Campos,
then
you
can
basically
understand
the
whole
thing,
and
so
we
just
started
doing
a
research
purely
on
that
hundred
thousand
neurons
and
a
square
millimeter
of
neocortex
and
when
we
focus
on
that
we're
not
thinking
about
the
hierarchy.
So
it's
like
the.
I
E
M
A
A
A
A
A
Are
still
really
new
you've
got,
you
know,
changes
like
database
bullion,
which
is
interesting.
Taking
of
depriving.
Oh
man
has
HDMI.
Yes,
because
the
concert
communication
he's
under
good
job
making
a
decent
implementation
animals
online
at
MIT
that
I
said
you
can
see
firsthand
every
game.
Just
like
okay,
there's,
a
few
more
JavaScript
wants.
A
A
W
E
Columns
exist,
the
brain,
there's
a
lot
of
stuff
happening
this
particulars
and
then
three
years
ago,
significant
progress
on
the
seminar
and
basic
trekker.
What's
going
on
here
and
that's
much
harder
problem
to
understand
and
resulted
now,
columns,
columns,
Boston
and
friendly
papers,
the
three
neuroscience
academic
papers:
no,
no,
no
sign
papers
about
this
topic
is
quite
fascinating.
Solution
understand
how
it
works
built
upon
this.
But
basically
you
take
this
in
your
spending.
E
E
The
big
idea
is
there,
and
so
now
we're
trying
to
fill
their
remaining
pieces
mostly
right
now,
this
market
myself
markets
are
in
full-time,
not
I'm,
working
for
a
lot
of
the
writing.
Another
one,
we're
sort
of
we're
sort
of
filling
in
the
details
here,
we're
documenting
it
we're
trying
to
and
and
that's
where,
pseudos
going
to
conquer,
because
once
once
we
can
still
see
the
end
of
the
tunnel
for
understanding
our
condos
and
how
the
whole
thing
works
in
the
big
pictures
in
place.
Having
the
index
does
what
it
does.
E
E
B
I
A
Hope
you
guys
recognize
if
you
follow
the
trends
and
sort
of
like
a
cycle.
I
mean
people
really
thought.
We
were
really
close
to
AGI
here,
really
Thunder
close
each
guy
here
and
there's
a
I
really
think
we're
close
to
AGI
wants
to
get
here
and
I.
Think
most
of
us
that
have
paid
attention
longer
sort
of
has
anticipating
what's
about
to
happen
here,
but
the
thing
I'm
gonna,
sorry.
A
All
right
thing,
I
want
to
point
out
here,
is
I,
wouldn't
put
what
we're
doing
on
the
same
continuum.
Is
this?
This
is
sort
of
dependent
upon
all
of
these.
The
neural
networks
of
yesterday,
you
know
and
and
and
the
hype
cycle
this
come
along
with.
Is
we
don't
think
at
least
I
still,
don't
think
that
the
weak
or
narrow
AI
that
exists
today
is
going
is
exist
on
the
same
continuum
as
but
will
eventually
bring
us
AGI
or
really
general
intelligence.
A
But
that
being
said,
there's
an
opportunity
here,
because
there
is
so
much
hype
around
this
to
try
and
ride
this
wave
and
to
get
some
attention
or
the
biological
work
that
we
have
done
for
for
years
and
years
now.
So
that's
one
of
the
reasons
why
we've
been
focusing
on
the
machine
learning
ecosystem
and
deep
learning
systems
and
they're
trying
to
see
how
can
we
bring
the
ideas
that
we've
discovered
into
that
world?
That
was
my
introduction
to
soon,
as
I
have
to
say
a
few
of
the
things
that.
J
J
E
But
yeah
it's
great,
but
we
totally
told
Sam,
what's
going
on
in
routine
learning
and
neural
networks,
I
just
think
they're
going
to
need
to
they're
gonna
hit
a
dead
end
and
they're
gonna
need
these
biological
principles
and
a
lot
of
people.
The
machine
learning
were
coming
to
that
conclusion
as
well,
so
so
we're
not
adversarial
at
all.
It's
just
I.
T
T
So
the
method
we
have
two
missions,
one
is
figuring
out
the
operating
principles
of
the
neocortex,
and
so
that's
what
we've
been
focused
on
for
many
years
now,
just
purely
focused
on
the
neuroscience,
and
the
second
part
of
the
mission
is
to
see
how
those
principle
how
we
can
use
those
principle
to
build
intelligent
systems
could
be
think
those
principles.
Should
you
have
a
core
of
intelligent
systems
and
we've
kind
of
ignored
that
mission
for
quite
a
while
now,
and
it's
there's
always
being
this
question
of:
when
do
we
get
back
to
it?
T
What
is
the
right
timing
for
it,
and
we
think
the
timing
is
the
right
timing
is
actually
now
and
the
reason
for
that
is,
as
Jeff
mentioned,
there's
a
bunch
of
work.
We've
done
and
we
can
kind
of
see
the
framework
of
how
the
neocortex
works.
How
does
what
does
a
cortical
column
do?
What
is
its
principal
operating
functions?
What
are
the
properties
or
the
kind
of
algorithms
and
whether
the
layers
doing
we?
Never?
T
We
don't
have
it
all
figured
out,
but
we
see
the
structure
of
what
it
is
and
now,
instead
of
filling
a
lot
of
the
details,
and
the
son
felt
that
now
that
we
have
this
structure,
we
can
there's
dozens
of
principles
that
are
embedded
in
that
big
theory.
We
can
now
start
taking
those
principles
and
applying
them
to
practical
systems,
and
we
should
be
able
to
see
dramatic
improvements
in
certain
areas.
So
in
this
room
you
know
we
right
look
at
Jeff
room
if
you
think
about
theoretical.
T
So
you
think
about
the
sequence,
memory
of
the
original
white
paper
and
the
CLA
algorithms.
So
we
introduced
the
notion
of
STRs
in
their
sparse
distributive
are
sophisticated
representation,
there's
sparsity
in
general,
we
had
a
completely
different
neuron
model,
so
this
is
much
closer
to
what
real
neurons
have
look
like
it's
nothing
like
what
big
learning
neurons
look
like
and
within
that
there's.
You
know,
there's
concepts
like
active
dendrites
and
a
bunch
of
other
stuff.
T
V
T
The
core
of
all
of
this
was
to
build
a
predictive
model
of
sequences,
so
we
had
neurons
esta
as
the
core
and
we
had
mentioned.
You
know
many
columns
and
a
way
of
and
along
with
these
learning
algorithms
there's
something
called
structural
plasticity,
so
we're
changing
the
network
structure
as
we
go.
T
These
are
all
coming
from
the
neuroscience
hardcore
neurons
and
in
fact,
if
you
talk
to
nurse
and
if
they're
studies
that
shows
something
like
30%
of
those
synapses
in
your
brain,
operating
the
neocortex
turnover
or
every
few
days,
so
there's
amazing
rewiring
going
on
in
the
brain.
There's
nothing
like
that
in
deep
learning
today.
So
this
was
all
of
the
core
of
their
original
set
of
algorithms
that
we
released
and
with
the
newer
stuff,
there's
more
principles.
So
we
have
a
location-based
representations
and
specifically,
using
you,
know,
grid
cells
with
this
notion
of
composite
objects.
T
You've
been
following
along
in
our
research,
this
and
transformations,
and
probably
one
of
the
biggest
things
that
we've
done,
which
we
published
in
the
columns
paper.
Is
this
idea
of
column
voting
or
voting
across
cortical
columns,
and
in
here
this
allows
cortical
columns
to
very
rapidly
infer
what
objects
that
seeing
in
what
percepts
it
seen
based
on
sensory
data.
It
can
do
it
over
time
and
it
could
do
it
over
space,
and
you
know
so.
If
you
put
all
of
this
together,
you
get
a
kind
of
what
we
call
the
thousand
brains.
Theory.
T
T
Of
the
most
amazing
things
about
cortical
columns
is
that
cortical
column
appears
everywhere
throughout
the
throughout
the
neocortex,
but
every
part
of
the
column
receives
motor
information
and
an
Aquos
letter
information,
whether
it's
visual
areas
of
put
motor
information
and
motor
areas,
receipt,
motor
information
and
sensory
information,
so
cortical
columns
are
inherently
sensory
motor
and
your
entire
brain
is,
since
we've
looked
at
a
very
deep
level,
there's
really
no
such
thing
as
a
motor
area
and
sensory
area.
Everything
is
a
sensory
motor
area,
so
this
is
really
key
to
to
everything
that
we're
doing
so.
H
T
Grains
theory,
which
kind
of
redefines
how
the
neocortex
works
and
that
redefines
how
we
think
about
hierarchy
and
how
multiple
cortical
columns
aren't
connected
together.
So
I'm
not
going
to
go
through
the
details
of
this,
but
you
can
read
about
it
in
our
papers,
but
if
you
kind
of
step
back
and
take
a
look
at
this,
these
there's
a
huge
rich
road
map
of
things
that
are
not
talked
about
at
all.
In
deep
learning
that
are
built
out
of
solid
neuroscience
principles
built
out
of
studying
the
neocortex
in
neocortex
is
the
only
limit.
T
The
only
intelligent
thing
that
we
can
all
agree
is
intelligence
and
it
behooves
us
to
so
see
how
these
principles
can
be
then
apply
to
practical
systems,
and
so
the
the
kind
of
omission.
So
the
second
part
of
the
mission
really
involves
looking
at
all
of
these
stuff
and
starting
to
apply
them
to
practical
machine
learning
systems
and
practicability
learning
systems.
What's
interesting
is
that
very
recently,
the
nature
of
what
people
are
talking
about
in
deep
learning
has
changed
a
little
bit
over
the
last
year.
People
have
realized
that
people
learning
has
really
fundamental
limitations.
T
So
deep
learning
systems
are
not
very
robust.
You
can
make
tiny
changes
to
inputs
and
tiny
changes
to
it,
I
think
in
a
completely
full
system,
so
they're
not
robust
at
all.
We
could
really
cool
that
they
don't
generalize
very
well.
If
you
try
to
train
it
on
something
a
new
thing,
you
have
to
come
up
with
thousands
and
thousands
and
examples
of
that.
You
think
before
we
really
learn
it.
This
deep
learning
systems
are
not
continuously
learning
they're,
not
flexible.
T
You
have
to
have
a
lot
of
manual
tweaking
to
get
any
system
working
well,
so
in
order
to
really
get
a
machine
learning
system.
Well,
you
have
this
ship
of
machine
learning
engineer
with
the
system
together.
Working
the
brain
is
nothing
like
that.
Deep
learning
systems
have
batch
training,
they're,
not
continuous
learning.
You
need
get
tons
of
label
data
sets
and
batch
training
scenarios
that
really
train
them.
You
can
think
about
all
of
those
properties.
It's
nothing
like
the
brain.
You
know
we
are
very
flexible.
T
We're
continuously
learning,
there's
no
batch
painting,
there's
no
parameter
to
be
gained.
Nothing
like
that
is
necessary,
so
I
think
these
principles
are
at
the
heart
of
getting
deep
learning
systems
and
practical
systems
in
general
to
work
in
an
intelligent
to
be
truly
intelligent
and
flexible
and
embedded
systems
the
world.
So
that's,
basically
the
mission
that
we're
on
in
that's
already
in
the
second
half
of
the
mission
is
starting
to
apply
these
things
to
commissioner.
T
L
T
A
T
T
T
T
T
H
T
Order
to
do
that,
you
need
to
figure
out
the
location
of
your
sensor
in
the
reference
frame
of
the
object,
so
you
can
figure
out
how
many
represent
objects
in
the
Cimmerian
weight.
You
can
actually
learn
very
very
quickly.
You
know
once
you
have
a
you,
just
need
enough
information
familiar
and
very
modeled,
and
now
you
only
need
to
see
them
all
with
all
the
different
positions:
rotations
and
translations
masasa
community
dramatically
with
these
and
other
training
example,
teen.
N
E
T
Many
of
these
pieces
will
be
part
of
that.
If
you
have
a
fingertip
and
your
pen
recognize
an
object
by
moving
around
it.
That's
a
sensory
motor
inference
problem,
that's
very
analogous
to
having
your
eyes
popping
around
an
object
and
recognizing
it
and
that's
over
attention
mechanism
and
then
your
eternal
neck
imperative.
H
G
T
T
M
A
If
you
think
about
your
neuron,
you
got
all
your
senses,
you
know,
that's
your
STR,
all
those
taxes
and
whatever,
whatever
quest
on
that's
the
signal,
you're
getting
and
you
decide
that
they're
gonna
fire
what
you're
doing
one
way
to
look
at
it.
The
other
way
is
your
stamps
and
lots
of
other
STRs
of
other
cells
that
are
observing
you.
You.
J
A
Q
Doing
here,
so,
are
you
trying
to
just
demonstrate
that
these
biologically
derived
features
are
doing
augment
existing
techniques
and
then
at
some
point,
you're
gonna,
say:
okay,
we've
got
some
evidence
and
now
we're
gonna
build
like
HTM.
You
know
3.0
or
whatever
you're
going
to
call
it,
or
do
you
think
you're
just
going
to
keep
adding
on
these
features
to
like
Alice,
TM
or
something
and
eventually
it
becomes
like
a
completely
different
algorithm.
Yes,.
T
I
think
these
principle
is
being
essential
for
building
intelligent
systems
or,
and
you
can
think
of
this
being
as
a
set
of
design
principles
instead
of
design
patterns
for
building
intelligent
system
soon,
we
know
how
they
work
in
the
neuroscience
that
would
figure
out
how
to
get
best
ever
working
in
the
culture.
My
sister's
said
son
I,
just
kind
of
demonstrated
that
this
is
possible.
We
think
these
are
actually
core
properties
that
you
need
to
have
this
kind
of
blueprint
for
building
without
these
systems
are.
E
R
E
E
To
me,
you
know
their
own
network,
pretty
damn
simple,
they're,
really
simple
things
and
there's
something
attractive
about
that.
The
brain
really
complex
a
mess.
You
look
at
you
know
a
column.
Those
are
hundred
thousand
arms
are
divided
into
maybe
two
or
three
thousand
different
types
and
and
they're
all
these
different
properties.
It's
like
it's
not
thing
and
what?
What
maybe
what's
been
really
surprising
for
me,
the
last
three
years
realize
how
complex
a
function
it
actually
import.
E
E
T
E
P
T
E
Really
really
really:
the
hippocampus
is
a
very
old
system
evolved
a
long
time
ago
and
what
it
allows
an
animal
to
do
is
know
where
it
is
in
the
world.
Nobody
did
episodic
memories,
navigation
and
spaces.
It's
it's
like
a
complete
little
memory
world
of
the
animal's
life
and
that
thing's
under
a
lot
of
evolution
and
pressure
from
anyone
in
a
long
time.
So
it's
very
detailed
how
it
builds
a
model
of
the
spaces
that
you
walk
through
in
your
episodic
memory.
E
The
basic
theory
outlined-
and
we
describe
this
in
the
january/february
paper-
is
that
the
mechanisms
that
the
hippocampus
used
to
map
out
rooms
and
spaces
now
been
copied
and
make
many
many
copies
of
the
engineering.
Architects
and
the
oak
reduces
upon
the
same
basic
principle
to
every
tip
of
your
finger
and
part
your
scans.
Every
party
dilemma.
So
every
part
of
your
body
is
like
a
little
animal
moving
around
the
world.
I
think
part
of
your
century
to
raise
it's
like
it's
a
correct,
hippocampus
woulda
took
they
protracted
rooms
they
watch.
E
What's
going
on
is
like
every
part
of
your
century
raised
or
a
pretty
scary
part
of
your
eyes,
it's
like
a
little
rat,
exploring
the
world,
that's
the
basic
analogy,
and
so
it's
not
fundamentally
different
than
the
neocortex.
It's
just
an
older
system
that
was
evolved
under
a
certain
set
of
principles,
and
then
your
crutches
that
abstract
and
it
made
a
clear
version
that
can
be
applied
to
everything
and
so
they're,
really
not
fundamentally
to
a
super-tight
point.
Now.
People
talking
about
the
hippocampus
is
being
so.
E
What
a
lot
more
region
of
the
neocortex,
even
though
it's
structured
differently,
it's
sort
of
like
a
continuum
and
so
I
understand
out
of
the
neocortex,
relates
to
the
rest
of
the.
How
the
hippocampus
relates
to
the
next
to
the
network
is
not
too
much
different
than
asking
as
a
visual
region
would
relate
to
it's
sort
of
like
a
old
version.
It
was
like
making
out
a
fleet
of
modern
cars.
You
know
electric
Tesla's
whatever,
then
you
got
this
19.
E
B
E
Still
use
gas
talk
to
each
other,
it's
kind
of
like
a
capital
just
transferring
information
to
the
new
context,
they're
going
both
ways
and
they
both
doing
the
same
basic
thing.
It
deserves
out
that
there's
far
more
experimental
data
on
the
hippocampus,
far
more
and
so
people
talk
about
this
thing.
We
know
about
and
we're
trying
to
say
hey,
you
know
what
the
new
part
is
really
different.
It's
the
same
thing:
it's
a
cleaned
up,
modern
version
of
it.
E
Y
W
Y
E
Y
O
Y
Y
E
M
B
Q
Thinking
about
goals,
do
you
have
in
your
research
like
agenda
plans
to
study
like
how
you
would
come
up
with
a
a
conceptual
model
of
way
to
add
biases
or
preferences
Jeff
was
talking
earlier
about
nature?
Is
evolution
is
building
some
preferences
in
humans,
and
so,
if
you
wanted
to
say
we're
about
to
pick
up
trash
off
the
street,
well,
you'd
have
to.
Q
E
E
Y
E
Y
Y
X
Y
E
Y
F
F
E
Decide
we
could
say
a
car
we're
gonna,
divert
in
time
the
people
detector
and
when
the
people's
detective
defense
people
we're
gonna
play
the
patron
completely
ignore
everything
else
that
you
trying
to
do
so.
My
car
will
ignore
my
commands
with
the
with
the
accelerator.
If
it
see
the
person
just
gonna
I've,
never
stopped
I,
don't
care
what
you
tell
me
man
so
yeah,
it's
a
basic
things
to
do.
I!
Think
it's
not
like
this.
C
E
F
C
E
E
Well,
we
can
have
a
whole
separate
conversation.
Cause
I,
think
that
today
today,
before
you
guys
graduate
there's
a
new
book
by
hammock
on
Howard
Stern
Harris's
life
support,
it's
very
short,
but
it's
an
interesting
topic.
But
we
don't
talk
about
that
here.
I
have
some
very
definite
opinions
about
consciousness,
but
I
actually.
E
Y
Y
A
Unless
you
go
so
the
people
that
are
watching
that
fifth,
which
you
guys
are
already
like
fish
bin,
so
I
got
to
you
fishing,
but
every
I'm
doing
a
bunch
of
stuff
on
Twitch
and
it's
our
hyo
l,
IG
HT
underscore,
which
is
awful,
I
know,
but
I'm
doing
two
things
live
research
meetings
at
least
once
a
week,
usually
two
times
a
week
and
that's
just
in
this
room
or
in
my
office
in
in
watsonville
world,
is
room.
The
reason.
X
A
Are
research
new
zones
in
the
student,
sometimes
I'm
on
the
double
I'm
sewing
like
stringing
research
meetings
and
that's
going
to
interesting.
We've
got
some
good
interactions
join
the
forum
if
you're
interested
not
on
this,
because
that's
where
all
the
conversation
happens,
that's
right,
click
all
of
this
media
that
I've
done
creating
an
August
on
the
forum.
If
you
want
to
keep
up
with
what
we're
doing,
join
the
forum
and.
H
A
A
I
A
A
A
X
A
A
A
B
A
A
R
A
X
X
A
C
B
A
A
Interesting
because
I've
never
used
a
chat
yeah,
but
they've
just
I've
always
been
like
unloading.
A
couple
idea
like
this
is
how
it
works,
but
they're
doing
some
cool
stuff
out
of
uniform.
You
already
said
this
is
a
guerrilla
and
they
can
do
that
sort
of
thing
with
words
think
you'd
be
able
to
identify
words
in
projects.
G
A
A
G
A
A
You
have
trouble
seeing
the
many
polynomials
lawyers,
but
it
looks
as
if
they
stretch
so
we're
simulating
like
this
big
and
we're
stimulating
one
of
these
layers.
So
these
many
problem,
activations
very
likely
have
affection
in
other
layers,
have
some
way
of
taking
this
activation
or
anything
else
down
here.
I
think
this
has
something
to
do
with
some
modules
or
other
types
of
orientation,
location
of
slash
modules
that
are
some
kind
of
sink.
It's
sensory
network
or,
but
that's
still.