►
Description
Michael Geer from Humanity.health takes us through a novel set of distributed data challenges. He discusses the potential breakthroughs in health that could result from creative Solutions to federated Machine Learning problems.
A
All
right,
recording
hello,
everyone
listening
on
YouTube:
this
is
the
computer
over
data
working
group.
This
is
our
14th
session
together
and
today,
we're
joined
by
Michael
from
humanity
and
super
interested
to
learn
more
about
his
product
and
what
they're
doing
particularly
to
help
us
think
about.
A
How
can
we
improve
the
quality
of
our
lifespan
as
well
as
sort
of
how
data
fits
into
all
that
in
this
decentralized
data?
Managing
of
data
is
something
that
we're
all
spending
a
lot
of
time
on
as
well.
So
one
quick
one,
quick
promo,
the
Cod
Summit
is
coming
up.
It
has
a
date
now
May
9th
through
10th
and
I'll
just
share
my
screen
briefly.
A
A
Please
go
ahead
and
start
thinking
about
that,
and
as
soon
as
we
have
registration
details
we'll
send
that
out
to
everyone
who
attended
last
year
as
well.
So
that's
all
I
have
Michael.
If
you're
ready,
I
will
stop
sharing
and
and
let
you
take
it
from
there.
B
A
B
Wes
yeah
good
to
be
here.
Let
me
grab
my
slides
here.
C
B
Cool,
so
maybe
a
little
bit
differently
than
some
of
or
maybe
your
more
regular
talks.
This
will
be
a
little
less
technical.
This
will
be
more
on
the
operational
side
of
like
how
do
we,
what
what's
the
what's
the
impact
in
product
side
and
what
are
the
pragmatic
technical
decisions?
That
kind
of
at
least
went
through
our
heads
when
we're
thinking
about
how
to
set
up
the
system
that
I'm
going
to
talk
to
you
about
so
first,
who
am
I
so
I'm
Michael
gear,
I'm,
a
co-founder
of
humanity.
B
This
presentation
throughout
you'll
kind
of
see
will
be
different.
We're
talking
about
two
companies,
because
the
proof
of
concept
that
we're
looking
to
do
is
between
our
two
companies,
and
so
the
other
company
I'll
be
talking
about,
is
wild
thought
AI
and
my
friend
Helene
runs
that
and
so
the
whole
idea
of
this
whole
thing
just
to
say,
you
have
a
framework,
you
know
before
I
dive
into
detail
is
we're
basically
two
consumer
health
companies
and
we
want
to
see.
B
B
So
Helene's
platform
is
another
consumer
health
application,
but
hers
is
very
focused
on
female
sports
performance,
and
so
it
comes
from
that
idea.
That's
okay,
there's
a
lot
of
differences
between
males
and
females
and
we
need
something
that
actually,
especially
when
you're
talking
about
you
know:
Peak
Sports,
Performance,
there's
a
there's,
a
big
difference
and
that's
what
she's
focused
on
and
Humanity
has
a
bit
of
a
broader
Focus
which
is
we're.
B
You
know
someone
think
of
us
as
like
a
longevity
tech
company,
but
we're
very
much
focused
on
increasing
your
health
span
and
so
we're
a
trip
say
a
traffic
navigation
app,
but
for
health,
meaning
we're
taking
in
all
the
data
and
watching
what
everybody's
doing
and
we're
then
guiding
you
to
tell
you
hey.
B
It
looks
like
if
you
do
this
this
and
this
you
particularly
not
just
the
general
public
you'll
you'll,
actually
slow
down
your
aging,
which
means
that
you're
going
to
have
more
healthy
years
and
so
really
to
bring
the
focus
for
this
talk
onto
the
data
and
so
I
like
this
title,
because
it
at
first
sounds
like
hyperbole.
But
then,
when
you
think
about
it
for
two
more
seconds,
it's
like
no
yeah.
B
Actually
it's
doing
that,
and
so
we
all
you
know,
I'll,
be
preaching
to
the
choir
and
probably
actually
the
mainstream
now
with
all
the
AI
kind
of
examples,
but
we
all
know
that
AI,
given
the
right
amount
of
data
and
the
right
data
can
do
amazing
things
and
so
I
think
it's
about
time.
We
all
kind
of
get
over.
That
first
question
was:
if
we
have
more
Health
Data
and
the
AI
works
on
it,
will
it
actually
be
able
to
help
us
be
healthier?
B
I
think
we
should
just
say:
yes,
it
will,
and
so
let's
get
access
to
more
of
that
Health
Data
and
so
really
on
a
day
by
day
basis.
Our
lack
of
the
right
Health
Data
to
train
models
that
then
are
used
by
not
just
consumer
health
companies
but
by
hospitals
and
others,
is
actually
leading
to
more
depth,
leading
to
more
unhealthy
people.
B
The
thing
that
made
me
actually
early
on
even
before
we
launched
Humanity,
so
my
background,
I
I,
launched
one
of
the
biggest
dating
sites
in
the
world
called
Badoo.
My
my
co-founder
launched
one
of
the
first
social
networks,
had
a
million
users
before
Facebook
and
and
got
quite
big,
and
so
we,
our
whole
specialty,
is
getting
things
out
to
hundreds
of
millions
of
users.
B
So
when
we
delved
into
the
science
side
of
things,
it
was
a
lot
of
me
just
spending
time
in
labs
and
just
talking
to
people
and
not
really
having
a
specific
agenda
other
than
I
wanted
to
find
out.
You
know
how
do
we,
you
know?
B
How
can
we
or
you
know,
detect
cancer
earlier,
like
I
started
with
certain
questions,
but
I
got
wide,
pretty
quick
and
what
I
realized
when
talking
with
all
these
people
and
and
luckily
enough,
some
of
the
top
names
that
you
would
see
in
any
particular
field
are
actually
sometimes
quite
generous
with
their
time
and
I'd
be
sitting
with
this
person
who
you
know,
I
knew
their
name,
because
it
was
the
name
on
the
model
kind
of
thing
and
when
they
actually
told
me
their
story,
it
was
like.
B
Oh
wow,
you
spent
like
90
of
your
time,
just
collecting
the
data
and
only
ten
percent
of
the
time,
doing
all
the
analysis
that
you're
known
for-
and
that
seems,
and
when
that
became
a
repetitive
story
when
talking
with
each
of
these
people,
it
was
like.
Oh
well,
you
know,
that's
that's
highly
inefficient,
but
also
the
I
would
say.
B
The
bigger
problem
here
is
not
just
that
you
know
say
like
Steve
Horvath
or
in
the
longevity
space
or
George
Church
needs
to
spend
a
ton
of
time
collecting
data,
it's
that
most
of
the
time
that
data
can
only
be
collected
by
the
likes
of
Steve
Horvath
or
Cynthia,
Kenyon
or
or
George
Church
a
lot
of
times.
The
access
to
the
data
will
just
never
be,
for
the
other.
You
know
eight
billion
of
us,
and
so
that's
that's
a
big.
B
The
access
and
the
time
that
it
takes
even
for
the
top
people
to
get
the
get
the
data
is
huge
and
then
you
know
hearkening
back
to
you
know
very
simple
statements,
but
very
true
is
we
I?
Think
people
that
spend
time
in
the
health
care
space
know
that
it's
quite
a
meme
and
because
it's
actually
very
true
is
that
you
know
a
lot
of
clinical
trials
for
years
and
years
and
years
we're
not
even
having
like
female
data
in
them
for
various
reasons.
None
of
them
that
great.
B
But
you
know:
that's
that's
what
it
was,
and
so
you
can
expect
when
you
build
a
model
or
come
to
an
understanding
based
on
you
know,
people
that
are
not
the
people
that
you're
going
to
apply
that
model
to.
You
could
very
clearly
understand
what
that
would
cause
problems.
Basically,
you
will
never
get.
You
won't
get
the
effect
that
you
want,
because
you
train
on
a
data
set.
That's
not
you
know
not
a
good
reference,
and
so
in
real
terms,
we
need
to
get
access
to
all
the
data.
B
That's
actually
relevant
to
the
cohorts
that
we're
going
to
try
to
help
so
I'm,
going
to
walk
you
through
what
we're
building
I
like
I,
think
this
whole
talk
is
about
because
I
think
one
of
the
biggest
the
greatest
things
about
this
group
and-
and
you
know
the
talks
that
you
guys
have-
is
it
it's
not
about
me
marketing,
specifically
hey
we're
doing
cool
stuff,
it's
more.
Probably
people
watching
this
video
might
actually
be
able
to
help
us.
B
You
know
build
this
cool
stuff,
so
so
I
I'll
try
to
give
you
little
pieces
of
it.
That
might
not
seem
necessary
for
the
storytelling,
but
might
be
necessary
to
see
how
you
might
you
know,
be
able
to
help
out,
and
so
what
we
start
with
is
actually
there's
a
ton
of
Health
Data
out
there
and
so
I
know.
There's
a
lot
of
initiatives
like
how
do
we
get
people
to
donate
data?
B
How
do
we,
you
know
I
I
personally,
think
they're
unnecessary,
because
there's
a
ton
of
data
out
there,
even
in
just
the
consumer
health
space,
just
a
ton
of
data,
the
the
question
is
more
and
probably
the
quickest
path
forward.
At
least
that's
what
we've
came
come
to
is
how
do
we
unleash
that
data?
B
How
do
we
open
up
that
data
to
model
training
while
at
the
same
time
keeping
it
completely
private
and
in
the
past
that
was
a
trade-off
and
I
think
that
trade-off
doesn't
exist
anymore,
I
think
we
can
do
both
and
so
what
you're
seeing
on
the
map
here?
So
what
we
have
is
Humanity,
which
is
a
consumer
app,
as
I
said,
we
have
right
now
like
155
000
users,
so
we're
getting
billions
of
data
points
biomarkers.
You
know.
B
Movement
patterns
actions
all
that
sort
of
stuff,
so
all
that
real
data,
wild
dead,
AI,
slightly
less
users
but
but
similar
kind
of
data
pouring
in
and
then
I
just
have
the
kind
of
like
representative,
like
there's
a
clinical
trial
or
there's
a
research
study.
You
know
as
well
just
to
show
that
this
isn't
just
for
consumer
health
companies.
This
is
you
know,
name
your
group,
your
hospital
group
or
whichever
group
might
be
collecting
a
lot
of
Health
Data.
B
Allowing
them
to
come
into
the
system
will
be
helpful
for
every
single
player,
and
so
we
want
to
keep
that
private
right.
So
in
the
past
you
kind
of
have
anonymization
and
then
these
kind
of
bespoke
you
know
bilateral
deals
where
okay
we're
going
to
anonymize
this
data,
then
we're
gonna,
you
know,
share
it
but
owning
with
this
group
or
we're
going
to
share
it.
B
But
only
in
this
you
know
this
secure
data
environments,
but
the
big
big
problem
with
that
is
it
severely
limits
access
to
that
data,
meaning
the
person,
maybe
even
one
of
you
watching
that
might
not
have
like
all
the
connections
in
the
world.
B
You
know
the
person
that
has
a
brilliant
idea
and
some
machine
learning
skills
if
they're
not
able
to
access
the
data
that
really
an
idea
will
never
turn
into
a
model
that
then
saves
lives
and
so
having
any
kind
of
restriction
on
who
can
train
their
model
on
on
the
data
from
step,
one
as
a
big
impact
change,
much
like
if
only
a
few
of
us
had
access
to
the
internet,
I
had
the
pleasure
of
getting
to
hang
out
with
Tim
bernersley
when
I
was
running
a
consumer
VPN
in
a
different
lifetime,
and
you
know
if
he
hadn't
created
the
world
wide
web,
he
didn't
create
the
internet.
B
He
created
the
World
Wide
Web.
The
impact
wouldn't
have
been,
you
know
even
100th,
of
what
it
ended
up
being,
probably
one
thousands,
so
we're
keeping
the
data
private.
So
what
are
we?
What
are
we
doing
we're
creating
synthetic
data
in
its
early
days,
but
we're
creating
synthetic
data
from
our
real
user
data
and
the
main
reason
for
that
is
because
we
we
want
to
really
open
up
the
access
on
the
far
end
of
the
system.
So
once
all
this
data
is
opened
up
and
there's
Federated
learning
across
it,
we
don't
want
a
big
limitation.
B
You
know
on
on
the
access
to
be
able
to
train
a
model,
and
so
taking
that
decision
to
do
synthetic
data
as
opposed
to
Federated
learning
using
differential
privacy
and
possibly
homographic
encryption.
You
know
people
watching
this
probably
know
three
other.
You
know
ways
that
you
can
add
privacy
into
that.
Just
didn't
seem
like
it
was
going
to
be
good
enough
to
not
have
any
or
almost
no
restrictions
on
the
on
the
far
side
of
the
system
where
people
are
training
stuff,
and
so
we
have
a
real
user
data.
B
We
create
synthetic
data
and
then
the
last
piece
of
this
POC
is
you
have
a
Federated
model,
sorry
that
this
is
the
one
slide
in
this
that
my
designer
didn't
do
so
I'm
sure
you
can
recognize
that
very
clearly
the
you
have
different
people
and
it
doesn't
have
to
be
this
just
the
people
that
have
inputted
their
data.
You
have
different
people,
training,
whatever
model
so
like
in
Humanities
case,
we'll
be
training
a
model
to
figure
out
the
impacts
of
different
actions.
B
You
can
take
throughout
the
day
on
the
particular
person
that
you
are
on
the
end
point,
which
is
increasing
your
health
span,
so
decreasing
your
your
chance
of
future
disease,
and
so
that's
our
model,
but
then,
while
that
I
I
might
have
different
models
right,
they
they
might
be
like
okay.
How
do
we
basically
figure
out
the
the
peak
the
peak
times
and
the
peak
actions
for
a
female?
B
That's
in
this
particular
part
of
her
cycle
to
you
know
what
what
is
your
training
plan
today,
and
so
everybody
will
have
different
models
and
then
a
research
study,
the
sky's,
the
limit
right.
You
know
any
any
question
that
we
have
about
preventative
health
for
this
particular
type
of
data.
You
know
you'll
be
able
to
answer
it
because
you'll
be
able
to
you'll
get
access
because
we
can
get
pretty.
You
know.
Wide
Open,
Access
you'll
be
able
to
actually
train
your
models.
B
Where
you
have
either
you
know
the
group
name,
you
either
have
this
kind
of
distributed.
You
know,
compute,
you
have
the
you
know
distributed
storage.
All
these
kind
of
things
are
open
game
for
making
this
a
viable
and
a
robust
system,
and
so
this
is
just
like
a
again
poor
visualization,
but
kind
of
like
all
of
us
have
probably
gone
down
through.
You
know
some
chat,
GPT
or
you
know
Bing
AI
rabbit
holes.
The
last
few
weeks,
I
think
the
thing
that
really
struck
me.
B
Finally,
with
all
that
is
that
it
understands
me,
you
know
it
no
matter
how
you
know,
I
try
not
to
be
rude
to
it
anymore.
I
think
I
was
a
little
rude,
the
first
couple
times,
but
you
know
it
gives
me
an
answer
to
my
you
know
my
my
prompt.
It's
not
quite
right,
no
matter
what
inane
way
I
come
up
with
kind
of
you
know,
trying
to
direct
it
to
get
closer
to
what
I
had
in
mind.
B
It
almost
always
understands
me,
which
is
just
an
insanely
different
experience
than
you
know,
my
sequel
or
you
know
anything
else
you
might
have
used
in
the
past
or
or
keyword.
You
know
kind
of
systems
like
like
search,
and
so
you
could
imagine
if
you
have
all
this
data
within
access
to
tune.
You
know
a
large
language
model
system.
You
can
imagine
that
not
just
with
anybody
with
a
certain
skill
set
can
then
interact
with
that
system
and
come
up
with
you
know,
answers
or
different
research
or
better
models
for
people.
B
Really.
Anybody
now
can
interact
with
it,
because
because
large
language
models,
kind
of
like
change
the
game
in
that
interaction,
you
know
place.
B
This
was
when
I
so
I
gave
this
part
of
this
talk
at
the
centralized
science
in
London,
which
was
a
great
conference
and
but
I
think
chat
GPT.
This
was
you
know
the
first
prompt
I
gave
I
gave
it.
You
know
really
pointed
out
the
most
interesting
stuff
was.
It
increases
the
diversity
of
perspectives?
B
That's
also
that's
also
the
on
the
data
side,
so
the
the
reference
data,
but
also
obviously
the
the
research
that's
done
and
then
more
accessible
to
researchers
and
communities
that
are
traditionally
been
underrepresented
in
science,
and
you
can
name
a
few
that
are
definitely
underrepresented,
but
I
think
you
can
really
widen
that
out
to
like,
if
there's
a
million
people
doing
science,
research,
there's
eight
billion
people
not
really
having
that
much.
You
know
access
to
it
and
I.
B
B
So
you
have
female
of
the
of
a
higher
higher
BMI
female,
the
lower
BMI
male
of
a
higher
BMI
male,
the
lower
BMI,
and
then
the
different
actions
that
can
take-
and
this
was
just
a
sampling
that
we
took
out
and
I-
think
the
one
thing
you
can
take
away
from
this
chart
is
that
the
bars
are
different.
What
we've
seen
in
our
real
data
is
that
the
impact
of
different
actions
for
different
types
of
people
is
different.
B
B
Personalized
guidance
in
health
makes
an
impact
and
that
impact
is
judged
against
what
we're
seeing
in
our
models
at
end
point
of
predictive
future
health
and
so
get
access
to
more
data.
You
get
better,
you
know
you
get
better
guidance
just
again
for,
like
you
know
more
of
like
describing
our
concrete
system,
so
I'll
walk
you
through.
You
know
kind
of
a
user
flow
here
right.
So
we
one
of
our
one
of
our
friends
and
investors
is
Jane
Metcalf.
She
was
the
co-founder
of
Wired
Magazine
and
so
I
use
her
as
an
example.
B
Sometimes
so
Jane
she
joins.
So
this
is
a
user
flow
and
then
I'll
show
you
what's
happening
in
the
back
end
Jane.
She
joins
Humanity.
We
pull
in
on
her
digital
biomarkers.
That's
things
like
movement
pattern,
second,
by
second
heart
rate
pattern.
If
she
has
wearables,
which
Jane
definitely
does
second
by
second
and
that's
getting
you
back
this
biological
age
for
the
user,
really
their
next
step
is
they're
just
getting
guidance
and
they
have
points
assigned
to
those
actions.
B
So
10
minutes
of
moderate
intensity
activity
will
get
you
six
six
points,
but
what
happens
in
our
system
is
possibly
for
me
as
opposed
to
Jane.
Maybe
I'll
get
10
points
you
know.
Maybe
James
doesn't
need
as
much
moderate
activity.
Maybe
she
needs
more
sleep
than
I
do,
and
so
that's
that's
how
the
user
identifies
with
that.
Is
they
basically
get?
You
know
different
amounts
of
points
depending
on
who
they
are
and
every
day
she's
just
trying
to
get
her
score
up
to
100.
B
Just
like
you
know
all
the
other
kind
of
gamification
you've
seen
in
the
past
and
what
we've
done
in
our
in
our
companies
with
the
dating
site
and
social
network
and
then
really
the
the
end
goal
is
to
slow
down.
Your
aging
is
basically
to
have
a
lower
biological
age,
which
basically
means
that
you
have
the
health
risk
factors
of
a
person
younger
than
your
chronological
age.
If
you
have
it
lower
and
that's
the
whole
goal
of
the
thing.
B
So
after
a
week
and
a
half,
maybe
two
weeks,
you'll
start
to
see
your
rate
of
Aging
change
if
you're,
following
the
guidance
and
keeping
your
score.
So
importantly,
though,
what's
happening
in
the
back
end,
so,
like
I
said
she
joins,
we
take.
We
take
all
of
Jane's
kind
of
a
digital
biomarkers
from
Apple
health
kit.
In
this
case,
what
we're
doing
is
we've
trained
a
model
on
external
longitudinal
data
sets
and
so
again
access
to
data.
So
these
longitudinal
data
sets
are
fantastic.
B
That
means
they've
been
watching
the
same
people,
so
the
same
participants
for
decades
and
so
UK
bioback's
been
watching
for
about
two
decades.
Now
they
have
about
500
000
people
they've
been
watching
for
two
decades
different
biobacks
Estonian
biobank
in
the
US.
You
have
Framingham,
you
have
Jackson
heart
study,
you
have
it,
you
have
a
bunch
of
them
around
the
world.
The
Dunedin
you'll
hear
a
lot
about.
B
The
Dunedin
has
a
good
biobank
out
in
New
Zealand,
and
so
basically
you
train
the
models,
the
predictive
models
on
that
longitudinal
data,
which
basically
says
this
is
what
we've
measured
on
people
in
the
past
these
participants,
and
then
this
is
what
actually
happened
to
them:
health-wise
for
the
next
10
20s,
in
some
cases,
30
40
years,
depending
on
the
buy
Bank
and
then
what
we
do.
Is
we
pull
in
Jane's
those
similar
markers?
You
know
the
same
ones,
maybe
so
the
UK
biobank
were
lucky.
B
A
hundred
thousand
people
18
years
ago
had
accelerometers
and
heart
rate
monitors
stuck
onto
them
for
for
about
eight
weeks,
I
believe,
and
so
that
gives
you
a
pattern.
You
have
future
health,
so
we've
taken
her
digital
biomarkers
and
we
basically
compared
to
that
pattern.
We
give
a
biological
age,
and
so
that's
great
now
we
have
an
end
point
that
endpoint
is
we're
trying
to
push
off
disease
for
Jane.
B
Then
the
other
side
is
very
much
like
ways.
The
traffic
navigation
system
is
you're.
Looking
at
these,
you
know
action
actions
that
you
can
take
during
the
day
and
we're
monitoring
what
actions
she's
taking
and
we're
basically
saying.
Okay,
this
group
of
actions
and
all
the
attributes
that
make
Jane
Jane
goes
into
that
action
navigation
model.
We
then
tag
all
those
actions
with.
Did
it
improve
that
endpoint?
B
Did
her
biological
age
go
up
or
down,
and
you
know
rinse
and
repeat
every
single
day
and
you
start
to
actually
get
different
coefficients
different
weightings
for
different
actions
for
the
type
of
person.
That
Jane
is,
of
course,
as
you
get
100
chains.
Well,
there's
only
one
Jane,
but
100
people
like
Jane.
You
know
a
thousand
people
like
Jane.
Her
guidance
will
become
more
and
more
accurate
and
on
the
wild
dot.
Ai
sides.
You
know
the
mission
just
so,
let's
go
back
to
the
missions.
B
Is
you
know
you
want
to
provide
them
with
actually
personalized
guidance
to
help
them
in
the
different
stages
of
their
of
their
training
and
in
ours,
very
clear
Mission.
We
want
to
add
a
billion
years
of
health
healthy
years
to
people
before
the
end
of
the
decade
and
we've
given
back
about
50
000
now,
according
to
the
models,
so
basically
added
50,
000
healthy
years
to
people's
lives,
and
none
of
this
is
possible
without
data.
The
data,
the
great
thing
is,
we
have
more
and
more
sensors
out
there.
B
We
have
more
and
more
data
being
collected.
How
do
we
open
up
access
to
that
data
while
keeping
it
private
and
hopefully
our
proof
of
concept
which
we'd
love
to
have
you
know?
People
help
with
is
is
one
big
kind
of
step
and
one
big
Beacon
that
we
can
show
that
it's
possible.
A
Fantastic,
thank
you
so
much
Michael.
What
a
an
incredible
vision
of
bringing
the
data
together
and
you've
hit
on
a
lot
of
topics
that
we've
talked
about
in
the
community.
More
recently,
I
mean
panels,
not
ml
folks
were
on
and
they
were
building
large
language
models
in
the
gesturian
way.
So
this
is
really
good.
Botter
and
I
appreciate
that
the
aspect
that
not
only
are
we
talking
about
like
underlying
tools
in
the
community,
but
also
having
real
use
cases
that
makes
unique
value
from
these
tools.
A
Could
you
say
anything
else
about
from
the
analytics
perspective
you
know
just
as
sort
of
as
a
Layman
doing
traditional
analytics.
I
might
say
you
know
this
person
ID
one.
Two
three
four
exists
in
a
couple:
different
databases:
I'm
going
to
try
to
find
people
that
have
this
scenario.
A
These
sets
of
Sarah
and
outcomes,
and
then,
when
you're
thinking
about
behaviors
across
different
synthetic
data
sets,
do
you
still
need
that
correlation
back
to
an
individual
for
for
analytics
purposes
or,
if
or
or
could
you
kind
of
achieve
the
outcomes
without
that
kind
of
correlation
back
to
a
single
person?
Even
an
anonymized
person.
B
Yeah,
it's
a
great
question.
I've
spent
most
of
my
time
and
the
folks
I'm
working
with
I've
spent
most
of
our
time
and
I
always
forget,
which
is
vertical,
which
is
horizontal.
Let's
let's
say:
we've
spent
most
of
our
time
on
the
vertical
Federated
learning
from
getting
the
terms
mixed
up.
Let
me
know
but
the
basically,
that
means
that
we
don't
need
the
same
person
to
be.
B
We
don't
need
to
know
that
the
same
person
is
in
two
of
the
nodes
in
the
Federated
Learning
System,
so
we're
not
trying
to
combine
different
types
of
data
on
the
same
person
across
the
Federated
learning,
where
we're
basically
trying
to
increase
the
amount
of
people
that
we're
able
to
train
across
so
you
know,
doesn't
necessarily
need
to
have
overlap.
It
doesn't
hurt
too
much,
even
if
there
is,
but
basically
we
don't
need
to
know
that
it's
the
same
person
across
each
there's,
definitely
maybe
to
go
a
little
bit
deeper
into
that.
B
But
then
this
is
getting
more
into
Conte.
Conjecture
is
there's,
definitely
a
lot
of
learnings
that
can
be
had
that
then
can
be
put
back
together.
So,
basically,
synthetic
data
is
you're,
doing
you're,
doing
deep
learning
on
a
data
set.
What
that
deep
learning
is
just
much
better
at
today
than
it
was
10
years
ago.
Is
it
can
pull
out
all
the
relationships
between?
You
know
multiple,
multiple
variables
right
relationships
that
we
might
not
have.
You
know
done
with
our
analysis
in
the
past.
B
Those
relationships
can
be
if
you
have
the
same
type
of
data,
so
not
the
same
person
but
the
same
type
of
data.
You
can
almost
sometimes
use
that
as
a
link,
so
you're,
basically
mapping
between
different
learning
groups
of
you
know
how
different
variables
work
together.
I'm
I'm,
anytime,
I
I
get
an
explanation
from
a
physicist.
B
I
understand
it,
so
you
might
catch
a
little
bit
like
body's
in
Motion
in
this
in
a
confined
space
is
like
how
I
think
about
everything,
and
so
I
think
you
do
get
those
learnings
can
be
applicable,
even
if
you
weren't
able
to
identify
that
the
same
person
had
them
as
long
as
you
have
enough
people
that
have
different
forms
of
overlap
between
the
different
types
of
variable.
Maybe
that
wouldn't
make
sense,
but
that's
at
least
how
my
mind
works.
A
B
I
think
there's
this
system
is
definitely
best
suited
for
coming
up
with
learnings.
That
then
can
be
applied
as
long
as
you
know,
you
know
which
strata
to
to
put
people
into
I
think
if
it's
trying
to
be
like
that
first
step
of
almost
like
combining
data,
that's
been
separated
on
the
same
person,
I
think
I.
Think
Federated,
learning
just
on
its
own
is
fine.
Doing
that.
B
I
I
think
the
learning
step
in
the
modeling
step
can
be
even
if
that's
the
first
step,
The
Next
Step
could
be
the
system
again
once
you
once
you've
combined
it
with
the
Federated
learning.
A
Very
good,
all
right
and
Alexander
I
know
you
joined
us
as
well.
I
want
to
give
you
a
chance
if
you
have
any
questions
to
give
you
some
space
as
well.
C
Yeah,
so
that
was
that
was
great
I'm
still
thinking
see
if
I
have
any
questions
or
any
thoughts
to
share,
but
otherwise
it
was
a
great
representation
and,
and
it's
a
great
subject,
I
I
really
like
the
the
health
area.
So
that's.
B
Nice,
what
do
you?
What
do
you
work
on
today.
C
I'm
mainly
working
with
open
ecosystems,
coordination
and
and
tooling,
so
helping
companies
and
and
projects
grow
their
open
source
communities
and
projects,
and
and
so
but
I've
I've
researched
a
lot
in
in
the
neuroscience
field.
So
so
I
I
guess
if
you
could
elaborate
a
bit
on
the
relationship.
C
Why
do
you
think
about
about
having
the
relationships
be
kind
of
interoperable,
so
you
can,
like
you
said
you
could,
exchange
or
use
the
same
relationships
based
on
if
they,
if
they
have
the
same
relationships
and
it,
how
would
you
deal
with
the
with
the
nodes
and
and
how
would
you
pack
that
with
the
relationship
so
so
they
can
be
shared
across
different
data
sets
using
so
you
you
would
change
nodes,
but
you
you
keep
the
same
relationships
so
maybe
optimize
the
the
models
and
whatnot.
So
how
do
you
think
about
about
that?
C
B
You
saying
just
let
me
try
to
understand
your
question.
It
sounds
like
a
good
one.
The
are
you
saying
that
okay,
you
train
a
model
against
a
few
of
these
nodes
data.
B
Would
you
it
would
then
be
optimal
to
then
take
that
model
and
train
it
against
or
test
it
against.
A
few
of
the
other
nodes
is
that
is
that
the
direction
yeah
yeah
yeah
I,
think
that's
I,
think
a
lot
I
I,
don't
focus
on
it
at
this
high
level,
but
I
think
there's
yeah,
there's
a
lot
of
functionality
in
Federated
systems
where
such
as
okay,
these
are
going
to
be
by
my
test.
B
These
are
my
training
nodes,
the
test
notes
and
and
or
or
just
simply
knowing
the
metadata
there
so
that
you
can
make
all
those
different
decisions
like
include
these
nodes
when
you,
when
you
train
the
model,
yeah
I,
think
that's
super
super
important,
because
people
do
care
where
the
where
the
data
comes
from,
they
do
want
to
be
able
to.
In
many
cases,
do
that
assessment
themselves.
I
I,
don't
know
how
there's
probably
others.
You
know,
people
that
are
working
on.
B
You
know
kind
of
rating
systems
for
nodes,
and
things
like
that,
and
there
probably
are
things
that
should
be
there,
but
I
think
just
having
the
metadata
and
being
able
to
take
notes
in
and
out
of
your
model
training
people
will
be
able
to
do
a
lot
with
that
to
figure
out
you
know.
B
Well,
if
this
data
just
doesn't
seem
to
kind
of
seems
to
be
really
pushing
our
our
motto
in
some
weird
Direction,
you
know:
where
did
it
come
from
at
least
give
me
the
you
know
the
name
of
the
company
or
the
or
the
source
and
I
think
that's
that
should
be
I.
Don't
think
there
needs
to
be
like
a
full
anonymization
of
who,
actually,
there
probably
shouldn't,
be
a
full
anonymization
of
the
nodes
that
are
in
this.
B
That
should
just
be
the
Privacy
for
the
user
data,
but
I
think
yeah,
there's
a
lot
of,
as
you
mentioned,
I
think,
there's
a
lot
of
great
reasons
why
you
would
want
to
know
and
be
able
to
take
nodes
in
and
out
of
a
training,
I
think
the
other
thing
actually
that
that
brings
up
even
another
point
on
the
on
the
more
detailed
side,
but
I
think
important
side
is
another
reason
for
say,
because
the
question
sometimes
comes
up
like
okay,
why
don't
you
you
create
the
synthetic
data
and
just
help
put
it,
you
know
centralize,
it
there's
a
big
reason.
B
There's
a
couple
big
reasons.
Not
to
do
that.
One
is
a
lot
of
this.
Data
is
well
almost
all
the
data
that
I
can
think
of.
That
would
be
useful
in
health.
A
lot
of
it
is
going
to
be
dynamic,
as
in
you're
going
to
be
getting
new
data
types
and
just
new
people
into
the
system.
I
mean
case
in
point
the
ones
here
are
you
know
we
get
thousands
of
users
every
day
kind
of
thing,
so
you
would
want
to
be
able
to
have
some
I.
B
B
A
second
synthetic
data
set
I
think
that
ability
will
actually
be
quite
functional
and
keeps
us
from
the
need
to
predict
every
single
way
that
someone's
going
to
want
to
train
on
the
data.
And,
what's
because
it's
very
hard
to
test
for
whether
that
would
work,
we
could
definitely
test
if,
if
our
models
that
we're
building
will
work
the
same
on
our
real
data
and
our
synthetic
data.
And
we
can
do
that
with
a
couple
close
Partners.
B
A
Very
good
Michael.
Thank
you
so
much
for
the
content.
The
next
question
I
have
last
question
I
have
for
you
is:
how
can
folks
get
involved?
How
can
they
reach
out
to
you
if
they
have
ideas
for
Solutions?
How
can
they
get
involved
in
and
using
the
app
contributing
the
ecosystem
Community,
any
anything
they
can
do
to
get
involved.
B
Yeah
great
question
so
I
think
yeah
anybody
can
with
an
iPhone
right
now,
can
download
the
app
and
then
start
using
it
and
and
that
in
and
of
itself
you
know
adds
to
the
data
set
so
and
and
we'd
like
to
think
it's
going
to
make
you
healthier.
B
We
will
have
Androids
at
least
a
beta
version
in
a
couple
months.
The
main
components
here
is,
and
they
and
they
can
actually
be
kind
of
done
separately.
It
doesn't
need
to
all
live
within
a
particular
system.
Is
the
Federated
learning
framework.
So
anybody
running
that
I
know
that
you
know
you
know,
there's
a
there's
a
few
there's
one
on
this
call
and
there's
a
few
others
that
that
we
both
know.
But
anybody
that's
interested
in
doing
that.
It
doesn't
need
to
be
a
particular.
B
It
doesn't
have
to
be
a
company
necessarily,
and
the
creation
of
the
synthetic
data
is
something
that
we're
internally
doing
tests
on
our
data
now,
but
would
love
collaboration
there
and
then
anybody
that
would
actually
be
want
to
be
one
of
the
partners.
So
a
data
custodian
that
has
some
data
now
and
would
want
to
you
know,
put
in
a
little
bit
of
time
to
become
one
of
these.
You
know
first
proof
of
concept,
Partners
I
think
it
will
definitely
speed
the
you
know.
Everybody
has
different.
B
You
know
road
maps
and
development
Cycles.
So
if
we,
if
we
get
a
couple
more
partners
that
come
into
this,
it
just
speeds
us
getting
to
at
least
two
of
us
getting
to
the
point
where
the
proof
of
concept
is,
you
know
full
circle,
so
those
are
the
main
ones.
I
can
think
of
to
reach
me
quite
easy.
Hopefully
so
it's
mg,
my
name,
is
Michael
gear,
so
mg
at
humanity.email
reach
out.
B
A
You
that's
an
excellent
plug
we're
hoping
to
get
some
of
the
folks
from
groups
like
fatimelda,
Ai
and
flock,
and
a
couple
other
folks
are
doing
as
well.
So
this
is
all
keeping
in
a
good
okay,
good
topic
space
for
us
all,
right,
well,
I
think,
that's
all
for
now!
David!
Listen
to
miss
anything
else.
Thank
you
so
much
for
for
presenting
today
and
we'll
we'll
look
forward
to
the
next
session.
In
a
couple
weeks,
cool.