►
Description
This is just a recap of the Deep Learning portion of the event.
A recap by Lucas Souza, Numenta Research Engineer.
Numenta Research Meeting - Aug 7 2019
Discuss at https://discourse.numenta.org/t/deep-learning-reinforcement-learning-summer-school-2019-recap/6434/2
A
B
B
D
B
Had
three
hundred
people
out
of
like
five
thousand
applicants,
so
it's
it's
a
lot
bigger
numbers,
Wow
hardened
to
so
just
oh
I,
think
this
chart
disclaimer
it
mainly
because
you
it's
gonna
go
in
the
web.
This
is
just
render
notes.
It
doesn't
do
any
justice
to
the
lectures.
It's
not
an
access
to
review,
I
just
focus
on
some
areas
that
are
interested
to
our
research
here
and
I
tailor.
The
presentation
took
55
my
hands
just
like
one
arm
and
I
have
an
expected
powdered.
B
E
B
Andy's,
goodness
we
had
researchers
from
other
fields
as
well.
A
lot
of
neuroscience
is
some
people
from
biology
from
physics,
etc
and
yeah,
or
even
very
short,
actually
I'm
short
and
I
looked
all
in
the
speaker's
we
just
and
we
had
like
from
undergrad
to
professor's
and
we
even
had
likes
girl.
Just
came
out
of
high
school
publish
a
paper
yoshua,
so
he
had
like
whole
spectrum
and
the
main
organizers
were
Yahshua
and
reach
certain
who's
here,
beating
the
beer
at
the
bar.
We
talked
about
booking
people
that
was
fun
yeah,
so
I.
B
B
But
a
lot
about
networking
as
problem,
so
this
was
like
every
we
had
this
kind
of
different
mixtures
and
was
organized
by
safar.
So
it's
the
same
Sammy
school,
that's
going
on
for
30
years,
I
mean
Mila
and
vector.
These
are
institutes
that
Albert
Aman
from
Toronto
and
they
are
organized
by
Amy's,
which
observe
Miller's,
yoshua,
bengio
and
vector
is
just
him.
So.
B
B
The
size
of
these
Institute's,
like
specially
Nealon,
we
look
like
huge.
It
has
I
think
about
500
people,
including
students,
professors
and
all,
and
they
are
forget
it
too,
for
universities
in
ultra
and
basically,
all
students,
income
which
are
studying
machine
learning
at
these
universities.
They
go
through
a
selection
process
before
and
really
it's
kind
of
sitting
on
top
of
the
universities
they're
on
trial
right
now,
especially
in
this
machine
learning
field,
so
they
are
quite
big
and
I
think
you
ought
Shaw.
Has
this
go?
B
B
Of
these
people,
they
were
not
looking
for
a
job
like
most
of
them
were
very
happy
into
their
PhDs
or
professorship
or
something,
but
all
these
companies
they
wanted
to
hire
and
the
Dominos
interest,
like
everyone
is
just
focused
on
like
research.
These
are
like
hard
core
researchers.
They
spend
their
summer
going
over,
like
tons
of
slides
of
math
and
they're,
not
interested
in
going
to
the.
B
B
Of
weird,
so
the
only
company
that
drew
everyone's
attention
was
like
deep
mine
and
mainly
because
the
mine
works
with
research,
and
that
was
like
an
opportunity
to
continue
working
with
research,
but
in
the
industry
and
I.
Think
that's
why
the
mental
speech
was
also
a
bit
feeling
for
a
lot
of
people,
because
it
was
appealing
like.
B
B
That
that's
interesting,
no,
like
Facebook,
ki
and
Microsoft.
Would
you
like
big
ones
who
are
not
their
boss?
Just
the
only
one
which
was
pure
research
focus
was
mine.
All
the
other
companies,
like
small
startups,
they're,
trying
to
solve
a
problem.
They
mean
there's
Jesse,
but
most
people
didn't
care.
No,
they
didn't
work
for
a
Khan
credit
card.
I,
don't
know
something
like
that.
They
only
care
about
yeah.
C
B
Don't
think
they
got
anything
for
me
and
they're
looking
for
machine
learning
research,
but
maybe
that's
not
what
they
want.
They
just
want
to
apply
machinery
and
I.
Think
they'll
have
better
luck,
just
getting
someone
smart
from
the
industry
and
trained
in
on
machine
learning
or
just
using
those
rather
than
trying
to
hire
someone
to
just
enter
some
kind
of
paper
published
movies.
B
B
B
Apply
as
a
student
from
Brazil,
then
they
gave
me
feel
funding
since
I'm.
Not
a
student
anymore,
I
didn't
accept
the
funding,
because
my
condition
change
so
I
paid
for
part
of
it,
and
it
meant
this
thing
for
another
benefit,
so
I
paid
for
subscription
and
everything
and
at
the
paid
I
think
what
kind
of
ticket
yeah
so
yeah.
They
hadn't
liked
a
lot
of
money.
It
was
really
well
organized
and
to
pay
for
all
these
might
be
expensive.
B
C
C
B
B
Up
until
last
year,
it
was
separate
they
had
a
separate
application
process
now
and
now
it's
just
joined,
but
so
we
have
been
learning
for
four
days
and
then,
when
14
learning
for
five
days,
so
in
deep
learning
day,
one
we
started
with
yokoo.
We
just
gave
like
an
introduction
to
neural
networks
and
just
a
comment.
These
are
mainly
my
notes,
I
transcribe,
to
this
light
when
I
wrote
it
and
most
of
them
make
sense.
But
now,
like
cuing
Slayer,
we've
taught
many
contexts
a
lot
of
them
don't
make
any
sense,
but.
B
Are
eventually
going
to
be
online
one
bad
thing
that
a
lot
of
people
complain
is
that
they
didn't
produce
the
slides
even
not
even
before,
or
after
usually
what
I
like
to
do
when
I
go
to
these
things.
Is
that
I
look
at
this
life
before
so
I?
Don't
have
that
effect
of
every
slides
of
surprise
things
so
I
know
what's
coming
and
I
can
prepare
for
it,
but
since
they
didn't
believe
we
had
this
and
why
you
went
to
this
presentation.
B
B
So
you
got
gave
an
introduction
to
neural
networks
that
was
really
good
class.
He
has
this
his
videos
on
why
he's
working
at
cuckoo
AI
right
now
didn't
have
like
any.
He
talked
about
over
permit
parameterization
as
an
approach
escape
saddle
points
and
that
we
might
have
a
winning
ticket
in
the
neural
network.
So
that's
exactly
the
problem.
We
were
broken.
What.
D
B
B
It's
a
benefit
because
when
you
start
a
network,
it's
like
you
have
all
this
ticket
and
one
of
them
is
gonna,
be
a
winning
one.
So
that's
that's
way
of
escape
desirable,
so
it
is
a
benefit,
but
you
have
a
different
way
of
skating
saddle
points.
Maybe
you
don't
need
the
over
permit
relation
and
that's
kind
of
the
problem
we
also
have
been
working
and
later
we
had
Brandon
Taylor
from
back
to
talk
about
CNN's
2000
product
product
Ori
talk,
he
talked
a
little
bit
about
state-of-the-art
networks
with
a
squeeze
and
excitation
networks.
B
That's
actually
the
first
time
I
heard
about,
and
the
only
difference
on
that
is
that
you
have
this
new
layer
with
weights
each
channel
adaptively.
So
not
all
channels
have
the
same
way
when
creating
output
feature
map,
that's
how
CNN's
work
and
I
thought
it
relates
a
lot
to
attention.
You
basically
assign
different
ways
to
different
futures
and
tension.
Does
that,
but
in
been
paroled
in
a
pain
prevention
shared.
B
That's
from
that
2017,
working
and
I
think
it's
the
latest
winner
of
the
original
Jack.
The
frontier
on
CNN's,
is
extending
to
non
including
data.
A
lot
of
people
talk
about
that,
and
that
includes
a
grass
convolutional
neural
networks
or
neural
networks
that
work
on
a
spherical
data
or
any
kind
of
fun
all
including
data
and.
C
B
C
B
He
also
one
thing
interesting
to
mention
is
that
begs
Norma
fact
robustness
and
he
showed
like
a
few
slides
showing
that
and
that's,
except
maybe
a
solution,
so
fix
up
is
just
better
way
to
initialize
the
network
so
that
the
network
is
stable
without
requiring
background.
So
fix
up
is
also
recent
were
giving
2016.
Not
a
lot
of
people
have
been
using,
but
as
he
presented,
it
seems
that
it
works
fine
and
it
might
be
a
good
replacement
for
background.
So
we
can.
We
can
try
and.
B
F
F
B
B
Was
talking
about
moving
beyond
humanism
and
how
we
should
look
to
a
diverse
society
in
the
new
food
to
where
we
have
humans,
hybrid,
artificial
agents
which
can
coexist
interesting
part
of
you
we
had
this.
A
lot
of
professor
was
I
really
like
to
start
everything
in
his
name.
Let's
talk
about
AI
researchers
being
world
builders,
and
he
discussed
the
paperclip
analogy,
we're
very
interested
some
Nick
Bostrom
knowledge
that
he
says
if
you
optimize
an
AI
is
a
single
objective
of
creating
paper
clips.
B
B
Just
do
one
thing
and
that's
like
every
startup
dream
is:
you
know
like
to
convert
the
whole
world
into
its
objective,
and
so
you
already
had
that
scenario
and
what
stopped
them
from
doing
that
that
we
have
several
startups
Frank
maximize
their
single
objective,
so
they
compete
in
this
space
and
that
is
probably
gonna
happen
AI
as
well,
so
you
might
have
like
money.
I
would
sing
objective.
B
Law,
professors
focus
a
lot
of
ethics.
She
was
talking
about
that.
Every
decision
we
do
comes
with
an
application
and
they
were
working
on
defining
like
an
ethical,
scarf
or
machine
learning.
Project
I
shall
talk
a
lot
about
how
machine
learning
researchers
should
worry
about
what
they
are
doing,
their
ethical
implications
in
every
small
decision.
B
Even
small
decisions
like
you
just
make
network,
which
is
ten
times
bigger
and
performs
better,
then
you
were
certainly
consuming
ten
times
more
energy,
that's
and
that
comes
economy,
or
we
can
insert
a
lot
of
kind
of
bias
into
your
products
and
that
shouldn't
be.
That
should
be
the
concern
of
every
machine
learning
research
and
you
shouldn't
just
say:
oh
my
boss
asked
me
to
do
it
so
just
really.
B
Had
Derek
Coffee
is
a
natural
language,
processing
researcher.
He
also
talked
about
that
so
that
every
tech
conceived
has
a
has
a
possible
rule.
Use
and
developers
should
take
into
account
so
when
they're
making
something
they
should
think
of
all
the
other
ways
that
people
are
using
that
and
if
they're
any
bad
ways,
and
he
also
talked
about
privacy
and
on
the
early
Industrial
Revolution.
B
We
had
many
social
issues
and
they
seemed
at
the
time
inevitable,
but
we
kind
of
fix
them
and
we
don't
have
kids
working
in
factories
twelve
hours
per
day,
and
if
it's
privacy
that
the
similarities
day,
it
seems
inevitable
that
privacy
is
gonna,
be
gone,
but
we
might
fix
it
and
society's
gotta
fight
back
and
it
might
fix
privacy
in
your
future
optimistic
fiance
and
the
only
general
consensus
I
find
is
that
at
consideration
should
be
made
by
every
developer.
I
think
that's
solid
argument.
B
B
F
F
B
C
B
B
To
teach
like
one
semester,
credit
course
in
one
week
and
then
it's
just
that
one
thing
and
you
end
up
learning
that
thing.
Why
are
you
just
covering
a
lot
of
advanced
topics
and
works
more
like
a
conference,
this
one
was
kind
of
a
mix
kind
of
trying
to
convey
two
or
three
years
of
credit
cards
in
two
weeks,
so
it.
B
Where
you
go
learn
and
you
go
home
and
do
it's
more
the
kind
of
summer
school
that
people
try
to
attend
every
year
to
get
like
some
more
exposure
which
what's
going
on
in
the
field,
so
these
do
I
think
was
a
mixed
stock,
so
by
Ranger
Cheng
was
really
interesting.
She
talked
about
state
of
the
art,
being
instance,
segmentation
and
I.
Think
some
my
slides
are
okay,
but
they're
not
showing
they're
like
what's
the
bottom
thing
said
yeah.
The
bottom
thing
is
like
covered.
B
These
are
the
difference
between
so
this
is
regular
classification
tell
if
it's
a
cat
or
not
a
cat.
This
is
semantic
segmentation,
that's
basically
pixel
classification,
so
you're,
seeing
all
these
at
grass
all
these
are
cat.
This
is
an
object
detection
when
you
have
like
multiple
objects
named
it,
so
semantic
segmentation
can
do
that
if
we
have
like
two
cat
you're,
just
gonna
like
mix
up
the
pixels
and
they
state-of-the-art
now
is
doing
instant
segmentation
that
you
you
can
capture
like.
D
B
2017
state
of
the
art,
so
the
system's
you
have
to
be
a
very
advanced
and
you
can
actually
like
I
can
take
a
picture
of
this
room.
I
can
remove
three
people,
nobody
know
the
difference.
I
just
know
it's
kind
of
scary,
so
we
evoke
we
had
our
sin
in
and
then
we
worked
fast
arson
and
fast
arson
and
all
the
mascots
in
there
and
it
just
keeps
getting
faster
and
faster.
B
D
B
D
B
B
B
There
is
a
volumetric
data,
it's
just
boxes,
it's
4d
tensor
its
pixels,
but
with
added
channels-
and
there
is
point
cloud
most
of
the
3d
data-
is
in
form
of
point
clouds,
just
bag
of
points,
but
the
issue
bag
of
points
is
ordering
variant
and
there
is
no
locality,
and
so
you
have
some
waters
which
work
on
that.
Some
newer
models
are
working
mainly
on
boxes.
You
have
techniques
to
convert
a
point
cloud
to
a
box
of
you'd
like.
B
And
she
show
like
recent
work
on
since
interests
and
since
interest
decided
to
create
the
same,
for
example,
you
can
build
a
room
and
you
can
train
this
by
getting
several
rooms
and
you
mask
a
few
object
and
you
put
the
network
to
build
that
object
which
is
gone
so
you
can
show
different
rooms
and
you
can
mask
the
bed
and
confine
the
network
to
figure
out
where
the
bed
is
going
to
be,
and
then
you
can
use
that
network
to
generate
like
new
room.
So
you
have
like
a
new
space.
B
B
B
The
whole
idea
is
that
you
can
connect.
You
can
have
a
memory
from
network,
it's
not
limited
to
a
small
small
space
if
you
have
a
way
of
knowing
where
you
saved
that
specific
memory.
So
it's
like
you're
using
the
attention
to
look
to
know
where
the
memory
you
want
to
look
for
the
data
you
want.
So
you
can
have
like
this
large
memory
associated
with
small
networks.
It's.
C
C
C
B
E
B
B
C
B
B
How
that
this
is
actually
quite
interesting
about
the
concern
is
prior
and
it
comes
up
in
another
talk
as
well.
What
does
that
mean
the
consciousness
prior?
So
it's
this
idea
that
you
have
a
high
dimension,
abstract
representation
space,
which
you
have
like
a
it's
a
very
high
dimensional,
and
you
know
just
concepts
and
factors,
and
you
reason
based
on
that
and
that's
gonna
direct
your
attention
towards
low
dimensional.
It's
other
right,
so
you.
B
It's
the
other
way,
I
have
this
low,
dimensional
conscious
spot,
so
you
can
represent
spins
in
low
dimension
and
you
can
reason
faster
and
little
dimensional
things
and
then,
when
you
need
to
attend
to
something
high
dimensional,
you
just
attend
to
that
specific
things
you
want.
So
you
can
like
just
to
generalize
like
we
can
reason
about
the
room,
thinking
about
chairs,
table
people,
and
then
we
need
the
specifics
of
the
table.
Then
we
can
access
a
higher
dimensional
representation
of
the
table.
B
G
B
You
need
at
least
a
mapping
right
that
tells
you
that
these
representations
the
same
as
that
representation,
so
that
that's
good
done
by
this
thing
to
Magnus.
So
this
guy
decides
where,
when
you
need
to
go
down
in
details,
so
when
you
need
to
go
down
details,
then
you
access
the
high
dimension
enough
interesting.
C
B
C
B
C
D
C
B
E
B
Be
doing
like
good
research,
so
he
talked
about
state
of
they
are
being
activity,
recognition,
activity
detection,
so
you
only
you
recognize
it
David,
you
also
localized
in
space
and
time.
So
you
know
when
directive
it
starts
with
enter
phineas
and
where
it's
happening,
that's
what
you
want
in
readers
like.
There
is
an
example
here.
So
you
you
know
all
these
people
they're
what
they're
doing
what
they
are
about
to
do,
and
you
can
also
have
like
a
time
frame
when
they
started
when
they
ended.
B
There
is
a
shop
early
recognition,
so
you
won't
ring
for
the
next
section,
so
you're
trying
to
recognize
activity
in
the
video
as
early
as
possible,
so
for
I
can
perform
to
stop
someone
from
shooting
people.
I
can
public.
You
want
to
recognize
activity
even
before
it
takes
place
somewhere
else.
We
can
predict
gun
and
there.
B
B
B
Of
group
activity
recognition
that
you
want
to
identify
what
a
groups
going
to
do
in
the
viewer
like
where
people
moving,
who
is
gonna
interact
with
who
and
there's
new
ideas
here
with
the
PO
social
bullying.
Where
do
Alice
name
of
spatially,
proximal
sequence,
proximal
sequence,
share
their
hidden
state
so
like
the
whole
idea,
is
that
you
can
share
knowledge
between
Alice
terms
of
things
that
are
happening
closer
in
the
video.
B
Would
be
like
an
added
stem
for
each
person
in
the
group.
Pods
and
I
have
like
a
thousand
person
in
the
video.
Then
they
would
share
their
hidden
state
socially
and
he
talked
about
generative
models
of
video
which
are
still
in
their
early
stage,
but
it's
likely
the
most
groundbreaking
application
of
machine
learning.
That's
going
to
come
in
the
few
years.
We
already
have
like
some
examples,
which
are
a
bit
scary
of
videos
being
generated
I
think
when
we
have
very
good,
maybe
even
like
change
the
movie
industry
or
I.
B
B
B
Doesn't
make
any
sense
and
but
when
you
go
to
videos
now
you
need
temporal
consistence,
which
is
whole
new
issue.
It's
not
that
easy.
So
you
need
to
find
ways
of
keeping
high
level
control
of
the
contact
you
can
buy.
This
kind
of
fishes
and
current
models
are
reducing
operational
headquarters
with
other
stairs
and
they
find
fun
way
of
inserting
Pryor's
and
incorporating
physical
physical
relations
between
objects.
So
you
don't
break
the
law
of
physics
in
your
video
which
written
partner
to
try
to
generate
a
realistic
video.
C
B
B
He
started
to
talk,
saying
that
Newton's
method
performs
well,
and
why
should
we
use
Newton's
method
and
then
that
I
think
it's
a
big
question.
It's
an
optimization
approach
to
try
to
find
the
roots
of
the
function
from
the
iterative
method.
I,
don't
know
exactly
the
details
of
the
matter.
That's
a
different,
optimization
method.
B
But
you
can
take
like
steps
towards
it
right.
You
look
at
the
secondary
and
take
steps
towards
it,
and
he
said
the
reason
we
don't
use
it,
because
it's
computationally,
expensive
and
because
of
the
behavior
they
are
getting
locally
in
each
region
of
the
space,
but
there
might
be
ways
of
moving
from
grade
in
the
same
thing
to
Newton's
method,
so
what'll
be
pretty
giving
to
that.
B
B
B
B
B
B
F
B
D
B
C
B
F
B
Optimization
that
we
had
Bayesian
deep
learning
project
girls
who
back
there
was
really
good
time.
So
the
difference
between
Bayesian
and
regular,
deep
learning
is
that
instead
of
doing
point,
estimation,
you're,
learning
the
distribution,
sorry,
assuming
all
the
non
Gaussian
and
you're
learning
mean
and
standard
deviation,
so
I
think
this
picture
kind
of
represents.
So
this
is
the
point
estimate,
and
this
is
the
sorry
distribution
you
were
learning.
So
this
is
another
way
of
seeing
that.
B
We
very
low,
very
high
and
you
end
up
having
probabilities
which
are
not
calibrated
to
the
reality,
and
you
can
kind
of
fix
that
use
temporary
scaling.
But
there
are
other
ways
of
fixing
if
you're
doing
Bayesian
deep
learning,
you
kind
of
get
the
probability.
The
right
probabilities
right
away,
because
you're
modeling
uncertainty
as
well.
They're,
not
modeling
as
participations.
B
So
it's
a
it's
an
intercept
vacation
for
exploration
as
well
since
they're,
naturally
stochastic
or
modeling
a
certainty
so
yeah,
but
the
good
thing
about
it
that
when
you
get
to
result,
you
also
know
how
certain
how
uncertain
you
are
on
that
prism.
So
that's
a
good,
buy,
very
deep
learning,
I
think
there's
a
lot
of
research
on
it
recently.
B
Courses
in
the
afternoon,
so
autumn
closures-
maybe
I'm
just
explaining
pictures.
So
it's
basically
there's
idea
of
compressing
data
into
latent
representation
services
upon
that
and
this
latent
representation
it's
on
smaller
size
than
the
original
idea
of
any
kind
of
compress
any
image
and
we
create
so
an
Internet
application.
For
example,
is
you
want
to
send
image
to
your
phone
and
then
you
can
send
in
the.
E
B
B
What's
interesting
here
is
that
you
have
these
new
applications
of
atom
cores
like
glow,
which
can
generate
image
very
similar
to
games,
so
I
think
Ganz
right
now
are
the
state
of
the
art
for
generative
models,
but
our
encoders
about
are
closing
that
gap.
There
Gary
chickens
and
if
that's
basically
unsupervised
the
difference
again.
Is
that
again
you
have
this
other.
You
have
this
two
networks
which
you're
doing
one
is
trying
to
get
generate
better
image.
The
others
try
to
classify.
But
you
are.
B
E
E
B
B
B
This
is
how
game
for
Michael-
and
you
just
respect
you
got
Rio
de-
do
get
generate
the
network,
and
then
you
get
the
discriminator
networks
of
trying
to
classify.
What's
we
were
not
and
they're
like
competing,
so
the
generate
is
getting
better
because
it's
trying
to
pull
the
discriminator-
and
here
usually
you
can
use
you
just
insert
Brenda
noise
and
then
engineer
it,
but
you
can
also
use
priors,
so
you
can
set
up
your
problem
in
the
way
that
you
can
generate
something
which
specific.
What
do
you
want
to
generate?
B
F
F
B
That
was
interesting.
Maybe
you
should
talk
more
about
this,
so
that
was
bleak.
Frigid
start
and
the
talk
is
it's
called
deep
learning
in
the
brain
and
the
whole
talk
was
about
how
deep
learning
is
feasible
in
the
brain.
So
he
started
the
talk
saying
long
ago,
a
drink,
Jah,
victims
kool-aid
that
back
propagation
is
the
way
we
were
on
that
and
I
want
to
show
you
that
the
brain
came
back
properly
and
it
was
very.
B
It's
not
good,
and
he
defined
like
that
three
issues
that
why
people
usually
say
backpropagation
can't
work
in
the
brain
once
you
don't
have
an
error
term,
and
then
he
showed
this
solution
of
a
clearing
propagation
and
says
a
brain
alternates
between
free
phase,
which
has
no
external
feedback
and
weakly
claimed
faced
with
here,
external
environment
notice,
the
network
and
the
difference
between
the
correlations
is
the
gradient
of
they're
not
entirely
sure
right
now
how
this
works.
I
wish
I
had
access
to
this
life.
I
do
have
some
a
lot
of
pictures
yeah.
E
B
E
F
B
E
B
B
C
C
E
C
You
can
go
to
great
extremes
to
do
so.
It'll
come
up
with
tremendous
arguments
for
the
existence
of
something
I.
Just
you
know,
and
so
I
have
to
ask
myself
all
the
time
like
well
am
I
wrong
about
this.
True,
and
the
other
point
is,
you
know,
are
other
people
buying
and
drinking
the
kool-aid
believe
I.
B
C
D
C
B
I
found
it
just
this
part
of
that
verse.
It
says
an
event
could
be
a
spike
of
birth.
We
also
used
for
is
in
our
model,
and
he
says
the
event
rate
can
communicate,
Purab
signals
and
the
best
way
to
communicate
top-down
signals,
and
you
can
update
the
weights
using
the
difference
in
the
first
probability
between
time
steps,
so
the
gist
of
it.
They
gets.
E
C
F
C
C
C
C
Is
a
this
basic
attitude
existed
back?
We
you
know
in
the
old
days
a
variant
a
way
back
when
there's
always
been
this
sort
of
tension
between
neuroscience
and
AI
and
when
I,
when
I
applied
to
MIT
AI
lab
I
mean
that's
what
I
ran
up
against.
This
is
before
neural
networks.
This
is
the
classic
AI
and
you
know
they
basically
said
that
you
know
you
don't
need
to
study
the
brain
and
here's.
Why?
Because
our
algorithms
already
capture
that
and
the
brain,
it's
just
a
messy
version
of
those
algorithms.
C
C
C
B
E
E
C
Take
a
one
level
up
further
I
mean
the
big
question
here
is
talking
about
you
know
or
what
they
doing
today.
What's
going
on
in
the
brain
doesn't
have
to
be,
but
the
real
question
is:
do
we
need
to
know
what's
going
on
the
brain
to
move
forward
and
that's
the
broad
question
as
its
backdrop
or
reinforcement,
learning
or
whatever
you
want
to
call
it
all
those
other
terms
you
know,
are
they
you
know?
Do
you
need
to
know
the
brain
to
make
progress
you
get
to
those
really.
The
big
question
I
think.
B
C
B
F
B
A
C
B
B
Actual
being
good,
and
what
is
next,
in
the
blurring,
it's
a
really
good
talk.
His
lights
are
on
line.
I
forget
something.
So
he
said,
the
long-term
goal
is
to
learn
representations
that
disentangle
causal
features.
He
talked
a
lot
about
magaling,
abstract
space
that
goes
back
to
that
consonant
prior
I'm.
C
B
C
B
C
G
B
C
B
You
have
two
things
where
I
would
have
felt
like
quitting
interesting.
Is
this,
for
example,
this
thing
about
imagine
abstract
space
and
one
of
the
ideas
that
language
can
be
the
size
abstract
space,
and
so
they
are
working
on
this
ground
and
language
understanding
problems.
It's
similar
to
that
clever.
One
like
book,
mags
put
the
blue
key
next
to
the
green
ball.
It's
like
a
very
simple
one,
but
you.
B
C
D
B
His
idea
of
system
1
versus
system,
which
goes
back
to
the
communist
priority,
that
you
have
abstract
concepts
and
you
can
do
counterfactual
reasoning
and
generalization
on
top
of
abstract
concepts,
but
that's
and
then,
when
you
need
you,
you
have
to
ground
them
by
the
system,
one
like
that's
the
the
first
inference,
classifying
image,
etc.
So
we
need
to
have
two
systems
over
one
another.
D
B
About
doing
self
supervised,
learning
latent
space
rather
than
a
race
pace
that
we
don't
reason
pixels
with
anything
in
pictures
and
doesn't
make
sense
that
we
were
doing
like
self
supervised
learning
on
top
of
pictures
and
think
CPC
there's
a
little
bit
of
that,
and
so
the
question
is
prior
I.
Think
when
we
talk
about
that
discovering
good
disentangle
representations,
it's
a
one
big
deal
that
goes
straight
toward
the
continual
learning
problem.
Right
now,
our
representations,
our
own
thing
up,
and
if
you
change
one
thing
you
just
mess
up
your
whole
network.
B
He
talked
about
the
work
being
done
by
Berkeley,
also
provides
agents
and
twin-stick
rewards
that
the
agents
define
themselves.
What
is
the
goal?
What
is
it
reward
and
they
can
learn
by
playing.
So
when
you
assign
then
a
downswing
to
test,
they
already
know
some
stuff,
so
the
agent
can,
for
example,
see
a
ball
and
decide.
B
Oh
I'm,
just
gonna
interact
with
the
ball
and
then
we
can
learn
things
like
throwing
the
ball
grabbing
the
ball
and
when
you
give
a
Dom
string
test
like
okay
put
the
ball
in
the
table,
I
already
learned
like
a
lot
of
things
just
by
playing
and
that
it's
an
analogy
how
humans
learn
talk
about
going
beyond
iid
like
a
lot
of
FIFA
soccer
and
having
to
mean
the
main
thing
I
got
from
his
talk
is
he's
really
he
really
wants.
We,
the
research
is
to
bring
go
Phi
like
good
old-fashioned.
C
B
B
C
E
C
B
C
D
G
B
So
yeah
it
does
activity
last
night,
so
this
presentation
is
on
the
Google,
Drive
I
think
it's
it's
activated
a
Yasir
manger
deck
wheeshes,
it's
not,
and
the
other
ones
are
more
on
in
general.
If
someone
wants
like
a
good
introduction
to
narrow
networks,
I
really
like
you
go
along
here
presentation.
He
had
like
this
online
I
like.