►
From YouTube: DevoWorm ML: Week 7 (General discussion)
Description
Seventh DevoWormML meeting, October 9. Attendees: Bradly Alicea, Richard Gordon, Jesse Parent, and Abraham Kohrman
A
C
A
B
A
B
A
Image
actually,
like
you,
can't
just
put
it
into
the
algorithm,
get
like
clear,
clean
segmentation.
You
have
to
play
with
the
image
like
the
contrast
and
things
like
that.
That's
a
lot
of
the
tricks
that
people
use
generally
but
I
mean
cuz
the
background.
You
know
it's
like
a
great
grayish
background
and
then
the
organism
has
its
boundary
and
then
it
has
like
an
interior,
actually
the
images
that
he
is
captured
or
a
little
bit
too
clean.
A
B
B
A
B
But
I
like
to
see
even
down
to
why
she
likes
me.
That's
like
one
laser
to
do
that.
The
experiment
I
did
back
in
the
sixty
was
to
zap
cell
number,
two
well,
the
post
Ruby
laser
then
destroy
it.
Yeah
we
published
a
small
abstract
office.
They
said
the
Sun
and
forth
against
the
shards
of
cell
number.
Two,
oh
okay,
okay
and
what's
the
distance
with
frequency
vary,
but
unfortunately,
I've
lost
the
data
since
then
the
so
the
current
was.
B
Looks
like
I'd
have
to
incur
so
what
that
proved
is
that
the
oscillations
were
intrinsic
to
the
individual
cell
and
therefore,
and
the
synchronization
is
entrainment
of
those
oscillators,
okay,
okay
yeah.
So
so
you
know
it
was
a
simple
monthly
experiment,
but
it
took
access
to
this
pulse
would
be
laser
microscope.
E
E
F
E
B
B
A
A
A
D
A
B
B
D
A
D
D
D
B
B
A
D
B
A
Yeah,
we're
working
with
this
he's
a
microscopy
in
Germany,
he's
working
out
of
his
house,
I
think
and
he's
been
collecting
diatoms
from
a
local
river
and
he's
been
putting
him
under
the
microscope,
he's
actually
been
culturing
them
and
then
putting
him
under
the
microscope,
but
he's
gotten
some
really
good.
Video
of
these
colonies.
I
mean
the
stuff
that
they
have
on
YouTube.
Is
you
know
it's
it's
adequate,
but
it's
not
like.
You
know,
really
a
really
good
image,
I
think
so.
Yeah
yeah
on
this
he's
been
extra
collecting
some
pretty
good
videos.
B
B
A
Hi
yeah
glad
you
could
make
it.
We've
been
talking
about
some
some
paper
that
you're
not
involved
in
and
we've
been
talking
about
some
other
things
so
glad
you
could
make
it
I.
Do
you
okay?
What's
new,
and
this
goes
for
you
bring
him
as
well.
A
A
A
So
this
is
a
this
is
building
on
the
conversation
we
had
about
pre
trading
models
two
weeks
ago,
and
so
I
had
the
idea
that
we
should
create
a
blog
post
for
this
topic
and
we
could
publish
it
on
on
the
node,
which
is
the
developmental
biology
blog
and
maybe
get
some
feedback
on
it.
So
it's
latest,
like
I've,
said
I
think
last
week
in
thinking
about
shooting
for
about
1500
words,
so
you
know
the
it
would
be
broken
up
into
a
couple
parts.
A
So
this
would
be
pitched
at
biologists,
but
also
you
know,
give
them
enough
information
to
make
them
think
about,
maybe
using
then
the
problems
for
application
of
biology.
So
you
know
there
are
some
issues
with
applying
it
to
biological
questions
and
again
we
want
to
make
it
attractive
to
biologists
so
that
they
can
think
through
the
implications
of
this,
because
I
think
a
lot
of
biologists
probably
don't
know
about
this
specific
tool.
A
Maybe
they
know
about
machine
learning
a
bit,
but
maybe
not
this
specific
type
of,
and
then
you
know
that
that
involves
talking
a
little
bit
of
input
and
training
data,
the
what
you
know,
the
concept
of
a
feature
space
and
then
talking
about
models
and
biological
problems.
A
So
this
is
more
sort
of
philosophical,
but
you
know
that's
probably
important
as
well,
and
then
the
third
part
would
be
modeling
a
biological
image,
so
the
idea
would
be
model
an
image
or
a
set
of
images
in
this
kind
of
model,
and
there
are
a
number
of
things
to
consider
so
I
think
one
thing
we
pointed
to
is
that
we
want
to
model
a
bunch
of
different
tasks
like
a
very
general
pre-trained
model,
or
do
we
want
to
focus
on
a
single
set
of
attributes
if
we
were
to
create
some
sort
of
you
know
biologically
specific
pre-training
model.
A
Does
this
allow
us
to
have
context
that
cuts
down
on
the
amount
of
training
data
that
we
would
use?
That's
another
point
that
we
had
and
then
this
issue
of
blobs
versus
symbols,
which
are
rival
strategies
for
processing
data
and
I'm
at
last
image.
I've
actually
got
some
resources
and
I've
put
them
on
the
resource
page.
It's
not
this
one.
A
There's
this
debate
in
the
machine
learning
community.
It's
I
think
these
four
are
actually
these
five
references
that
kind
of
talk
about
them.
It's
the
idea
that
you
can
have
either
you
know,
throw
data
at
a
problem
and
extract
features
and
solve
a
problem
that
way
or
I
can
use
a
symbol
system
of
symbols
of
representation
to
model
a
phenomenon.
So
you
know
this
has
been
a
long-standing
debate
in
the
history
of
AI.
Where
we've
had
you
know,
people
like
Gary
Marcus
is
a
cognitive
scientist.
A
Who's
argued
that
we
can
use
symbol,
some
Maalik
systems
to
achieve
the
sort
of
artificial
intelligence
and
then
there's
another
group
of
people
who
say
well.
If
you
throw
a
lot
of
data,
add
a
machine
if
you
throw
just
enough
data,
add
a
machine
and
it
could
be
like
you
know,
terabytes
and
terabytes
of
data,
but
you
that's
actually
better
than
a
symbol,
manipulation
system.
So
there
are
these
five
references
on
this
page
here,
I
kind
of
go
into
this
debate
and
it's
a
little
tedious
at
times,
but
it's
basically
the
idea
of.
A
If
you--if
you
know
anything
about
nature-nurture
in
biology
where
you
know
you
have
on
one
hand,
you
have
machines
that
have
no
prior
knowledge
of
something
and
then
they're
learning
from
data.
That's
presented
to
them.
On
the
other
hand,
you
have
maybe
like
a
machine
with
a
representation
which
would
be
like
a
you
know.
Some
sort
of
you
know
inherent
knowledge
about
something,
but
not
specific
knowledge,
and
then
the
data
would
fill
the
gaps.
A
So
you
know
it's
like
the
difference
between
an
anus
and
the
blank
slate,
and
so
this
paper,
actually
in
particular,
was
really
interesting.
This
is
a
new
one
by
Anthony
zidor,
and
you
know,
he's
he's
been
involved
in
this
debate.
He's
basically
saying
that
we
need
representations
that
have
like
a
genetic
basis.
So
you
know
they're,
using
like
a
innate
encoding
to
aid
their
models
of
intelligence,
so
I
mean
I,
don't
want
to
get
to
doubt
far
down
that
rabbit
hole,
but
I
was
thinking.
A
I
want
to
go
more
into
that
symbolism
versus
data
debate
myself.
This
list
on
slack,
oh
I,
can
put
that
stuff
on
slack
no
problem,
I'll
put
it
up
on
the
open,
worm
slack
and
then
maybe
on
the
orthogonal
slack
and
I
can
send
Abraham
an
email,
or
you
know
something
like
that.
That's
just
a
copy-paste
thing
for
me,
so
I
can
just
you
know,
but
yeah
I
mean,
and
so
let
me
I
just
wanted
to
give
you
an
idea
of
well
what
I
was
thinking
for
this
post.
A
Let
me
maybe
this
week,
I
can
get
a
chance
to
flesh
it
out
a
bit
more
and
I
can
send
you
guys
some
of
these
references
for
the
final
part,
you
know
I
mean
I,
can
come
up
with
some
basic
sort
of
I
have
an
outline
now,
but
like
some
basic
text
that
we
can
play
around
with
more
and
then,
if
you
have
any
ideas
you
can
add
in
but
we're
you
know
we're
getting
towards
1500
words,
so
you
know
we're
gonna
try
to
get
it.
A
A
Unless
you
make
it,
you
know
super
compelling,
but
and
then,
like
I
said
we
can
post
it
at
the
node
I,
don't
know
if
Jesse's
familiar
with
the
node,
it's
a
developmental
biology
blog
that,
like
people
can
submit
content
to
so
like
you
know,
you
just
submit
a
post
and
it's
posted.
You
know
you
have
all
sorts
of
people
posting
different
things
there,
but
it'll
give
some
exposure
to
that
community
and
that's
that's
kind
of
what
we
want.
A
A
So
what
do
you
guys
think
about
that?
Do
you
think?
No,
maybe
over
the
next
few
weeks,
we'll
kind
of
work
on
it,
some
more
and
get
it
flushed
out
in
terms
of
complete
sentences.
I
mean
I'm.
Just
taking
this
outline
from
the
conversation
we
had.
So
you
know
it's
like
we're
all
kind
of
contributing
Abraham
and
myself.
We
were
in
in
this
conversation
and
then
a
little
bit
with
Vinay
and
then
Jesse's
perfectly
welcome
to
add
things
in
yeah.
That's
fine
I
mean
especially,
you
know.
If
there's
like
the
I
mean
I.
A
Okay,
I,
don't
know
I
happen
Abraham,
he
dropped
out
of
the
meeting,
but
his
connection
wasn't
very
good
but
and
then
I
guess
for
Jesse
see
the
other
thing.
I
had
and
I
get
I
don't
know.
Well,
I'll
mention
it
now.
I've
been
talking
with
someone
another
collaborator.
He
was
here
before
Eden
Roc
and
we
were
talking
about
doing
like
sort
of
a
distill
for
developmental
biology.
A
A
It's
you
know,
it's
actually
a
journal
that
they
have
this,
so
people
publish
articles
on
here,
and
you
know
it's
eight
different
topics
and
machine
learning
and
the
Ohio
brand
welcome
back.
I
was
just
showing
Jesse
about
this
journal
that
they
have
a
machine,
learning
called
distill
and
okay
good
and
it's
unique
because
I
mean
these
are
like
journal
articles
of
people,
publish
and
submit
to
the
journal.
Well,
what's
unique
about
this
type
of
journal?
Is
that
you
can,
you
know,
publish
articles
with
interactive
features
in
it.
A
So
here's
one
where
you
have
like
this
animation
in
the
paper
and
then
the
paper
is,
you
know
it's
like
a
regular
academic
paper,
but
you
can
embed
demos
in
the
middle
or
animations.
You
can
obviously
put
in
your
equations
and
all
that.
But
the
advantage
of
this
is
that
it's
not
sort
of
cast
in
PDF,
so
people
can
interact
with
the
the
article
and
you
know
it's
a
lot
more.
A
A
A
D
A
We'll
talk
about
it
more
in
coming
weeks.
He
already
has
the
infrastructure
he
has
a
blog,
for
example,
where
he's
kind
of
done
this
sort
of
thing
before
so
I
mean
the
idea
is
to
have
a
venue
where
you
know
it
can
showcase
ideas
and
it's
kind
of
a
niche,
because
that
sort
of
niche
isn't
really
filled
right
now,
with
a
journal,
we
have,
you
know
some
journals
and
developmental
biology.
Some
journals
in
you
know
computational
biology
and
machine
learning,
but
nothing
in
that
specific
area.
A
A
A
Okay,
here's
my
here's
the
blogpost,
so
this
is
a
you
know,
getting
it
down
to
about
1500
words
and
we
did
have
like
three
sections
to
it.
So
the
first
one
is:
what
is
a
pre
train
model,
which
is
a
you
know,
defining
what
this
is
for
a
biological
audience,
so
we'd
have
a
short
description
of
what
it
is.
A
You
know
that
kind
of,
like
a
very
brief
survey
of
the
current
models
that
exist.
So
you
know
you
have
specialized
models
for
language.
You
have
general
models
for
just
like
you
know,
objects,
but
nothing
for
biology
specifically,
and
then
you
know
what
is
what
is
of
specific
interest
to
biologists,
about
pre-trade
models?
A
I
mean
I
like
to
think
about
it
as
like
another,
like
maybe
it's
a
part
of
deep
learning,
but
it's
also,
maybe
you
know
a
method
like
any
other
type
of
machine
learning
that
might
help
people
analyze
data
and
create
a
model
for
understanding
they're.
You
know
generalizing
to
different
types
of
similar
data
and
things
like
that.
Then,
like
a
section
that
would
be
a
problems
for
application
of
biology.
So
you
know
we
talked
about
in
our
discussion
the
amount
of
input
in
training
data.
A
So
when
a
preacher
and
model
reduces
the
amount
of
data
you
would
need
to
train
a
model.
Then
we
can
talk
about
what
the
features
space
in
terms
of
biology.
So
you
know
is
the
feature
space.
You
know,
what
does
it
look
like
we're
talking
about?
Like
you
know,
genes
are
we
talking
about
phenotypes
and
then
a
discussion
about
models
and
that's
more
philosophical
as
I
was
telling
Jessie.
A
So
there's
this
idea
of
models
being
useful,
but
not
you
know,
but
not
like
the
actual
phenomena
and
then
like
this
idea
of
that.
Maybe
you
have
to
customize
a
model
to
biological
problems
in
order
to
make
them
more
useful,
but
that
I
mean
that's
I
know
it's
philosophical.
We
want
to
make
that
very
brief
and
then
the
final
section
is
modeling
a
biological
image.
So
this
is
like
you
know
the
idea
that
you
have
we're
taking
a
biological
image
and
we're
doing
things
to
it.
A
A
Biology
so
I,
you
know
like
I,
said
before
you
know
this
is
something
I'm
gonna
try
to
flesh
it
out
a
bit
more
in
terms
of
complete
sentences,
but
hopefully
you
know
we
can
work
out
over
the
next
few
weeks.
Maybe
we
can
work
out
in
the
meeting
or
you
know
whenever
a
of
time
and
then
we
can
get
it
flushed
out.
A
A
H
H
H
H
H
G
A
H
H
A
G
A
A
A
C
C
A
Yeah
thanks
so
yeah
I
can
send
that
information
out
again
in
an
email,
and
then
you
know,
like
I,
said
that
the
blog
post
thing
can
work.
I
can
work
out
some
of
the
complete
sentences
and
add
in
what
we
just
talked
about
with
examples,
and
then
we
can
just
keep
working
on
it
over
the
next
couple
weeks.
So
and
then
next
week
I
was
gonna
present.
A
Sometimes
their
outputs
are
like
the
output,
things
that
look
like
patterns
but
they're,
not
and
so
that
resembles
something
in
human
perception
called
pareidolia,
which
is
like,
where
you
see
things
like
faces
and
objects
where
there's
no
face,
but
it
looks
like
a
face,
and
so
like
I
have
some
interesting
things
to
talk
about
there,
but
I'll
keep
that
for
next
week
and
then
so
you
know,
if
anyone
has
any
other
question.