►
From YouTube: DevoWorm Summer of Code weekly meeting, 7-5
Description
Meeting for Week 9. Asmit Singh presents the paper "Deep Residual Learning for Image Recognition".
A
C
D
C
A
E
E
The
menial
tourists
will
come,
he
young-hee,
Xiao,
Qian,
Tian,
I'm,
Johnson
and
Esther
was
presented
up,
see
if
you
happen
to
16.
So
Syria
is
the
newest
prestige,
with
a
little
confidence
to
the
same
people
in
the
field
of
computer
vision,
recognition
and
this
fulfills
the
exam
making
in
a
sense
and
and
was
presented
in
the
additional
six
in
addition
of
the
sponsors.
So
I
start
by
giving
it's
a
kind
of
bug,
it's
saying
legs,
so
he
30-minute
challenge.
E
So
he
makes
me
specifically
I
will
not
state
a
state
that
was
created
by
Allah
Stanford
having
sharks.
Would
that
contains
a
huge
amount
of
food
crafts
on
that
made
that
belong
to
be
free
classes,
a
busier
which
has
one
thousand
bunch
of
classes
on
the
training
session,
two
million
in
size
and
the
place
six
hundred
sites.
So
basically
this
and
also
cause
for
me
to
be
started
known,
as
is
you
are
co-related
log
scale
which
the
challenge
so
basically
at
VA?
Oh,
it's
not
a
challenge.
E
They
figured
the
shell
scripts
or
the
algorithms
are
tested
upon
this
data
set
on,
like
they're
priced
to
their
accuracy,
like
the
many
marketers
civets.
So
like
this,
the
units,
the
talent
about
around
Lake,
because
employer
thirty
became
a
matter
of
deepness
of
the
late
nights,
so
basically
like
to
explain
why
ask
me
this.
That
means
something
so
the
moon
is
so
actually
adding
more
less
to
the
network.
Europe's
increasing
its
capability
of
steps
in
those
features
and
more
complex,
which.
E
Because
we
can
understand
the
features,
if
you
have
start,
if
you
have
a
tabular
data
or
structured
data,
there's
nothing
better
than
normal
Shannon.
It
works
for
that
because
we
know
what
features
are
giving
features
are
important
like
you
can
simply
explain
that
will
service
you
just
know
because,
like
we
know
that
always
controlling
it,
but
in
case
of
semi
structured
data
that
eliminates
or
give
you
or
the
sound
like,
not
some
yeah,
maybe
some
we
don't.
E
Shows
it
fall
because
it's
a
vector
it's
system,
rl6,
usados
intensity
of
the
pixels.
We
have
no
idea
about
it,
so
leave
it
upon
the
a
deep,
let
the
maybe
since,
between
deep
networks
and
shadows
got
deep
down,
take
out
its
own
material
and
spent
about
their
image,
features
which
again,
no
one
to
hear
from
them.
So
movie
is
kind
of
cover
more
complex.
E
E
A
problem
that
is
difficult
and
as
we
do,
people
till
this
paper,
as
you
can
see,
has
the
ILS
cucumbers
like
we
couldn't
exceed
in
this
episode.
I'm
certainly
lays
on
thoughts
but
I'm
going
to
delist
the
basement.
So
what
actually
happened
is
an
interesting
one.
It
was
the
cause,
but
it's
not
that
we
suddenly
had
been
jumping
computation
part
process,
an
increase
in
the
sales,
but
there
was
something
over
in
fact,
so
about
the
after
it's
time
he
and
I
talked
to
the
pastor.
This
paper
he
was
working
in
Microsoft
and
correctly.
E
His
sub
scientists
are
faced
with
that.
Physics
of
interests
are
computer,
based
on
deep
learning,
I
love
it
Vogue's
and
he
was
so
into
oxy,
Jackson
economy
defined
as
a
key,
a
framework
of
obtaining
credit
ignites
on
it's,
not
simple,
but
it's
very
cliche,
but
it's
very
sensor
you
can
make
that
feel.
I,
don't
understand
it's
a
state
of
the
art
for
image,
classification,
object,
detection
and
semantic
segmentation.
E
Actually,
the
the
night
period
you
are
going
to
use
in
a
facility
of
could
say
that
state
lab
has
many
other
models,
but
that
we
continually
assess
it
at
the
back.
You
see
the
research
they
have
saying
100
years
150
days
on,
we
have
16%
better
than
the
second
oculus
like
that's.
The
purpose
in
1606
are
interesting
and
makes
do
we
have
to
have
a
challenge
and
there's
without
blame
percent
percentage.
Please
regarding
like,
let's
start
with
a
resolution
of
that
so
say:
that's
the
same
thing.
E
I'll
explain
that
eight
days
when
for
PVC
90
days,
the
best
for
me
and
then
I
resonate
152
layer,
state,
then
yeah.
So
like
tell
you
now,
it
seems
exist,
arguing
maybe
decreases
our
accuracy,
but
it's
not
the
case
and
it's
there's
not
monopoly.
It
happens.
So
what?
If
you
just
start
years,
loquacious
increase,
there's
some
political
physics
in
a
plane,
CNN
like
actually
creates
from
twenty
to
fifty
six,
even
the
Creator
is
decrease.
This
is
really
like
shocking,
not
easily
understandable
because,
like
it's
like
city
for
lakhs,.
E
E
A
E
E
A
E
E
E
E
E
E
Hypothesis,
that
for
me,
because
there's
an
optimization
but
not
least
that
may
not
function
is
not
a
tumor
that
will
be
attacked
so
many
lives,
so
the
basement
actually
came
with
a
different
approach
to
this.
By
changing
that
architecture,
the
architecture
is
from
there
also
sourced
on
the
addition
problem.
The
DVD
problem
is
the
problem
decreasing
accuracy
in
the
training
itself.
Then
it's
also
faster
Taming
of
the.
E
The
Griffons
I
don't
know
the
exact
reason,
but
it
has
a
and
complexity
also
it
is,
it
doesn't
actually
crease
the
parameter
feature.
It's
not
actually
increasing
a
lot
in
terms
of
parameters,
and
it's
awfully
have
at
least
as
a
yeah
accuracy.
Okay,
so
that
commonly
method
would
have
been
in
terms
of
many
cases
where
we
would
have
a
very
clear.
We
could
have
an
input,
the
Nakayoshi
function,
which
can
be
a
sigmoid
or
anything,
and
then
we
have
X,
which
is
not
conforming
to
its
own
or
maybe
or
tame
on
this
SFX.
E
E
A
E
E
Serve
the
physician
problem,
it
was
easier
to
play
in
this
lay
under
the
L,
so
that
is
easy
to
play
f
of
X
H
of
X.
If
I
didn't
exact
reason
for
that,
but
I
guess,
like
it
was
easier
to
train
on
X,
was
X
as
some
people
its
effects
on
that.
First
often,
a
optimization
problem
to
go
beyond
so
example.
Make,
for
example,
do
you
know
in
case.
E
At
the
activation
function,
so
the
contributions
of
these
people
solve
the
resident.
It's
shown.
The
residual
learning
is
his
optimization
from
the
nominal
learning
via
capable
learning.
It
was
easy,
the
optimization
they
talk
with
the
problem
depredation.
It
was
decreasing
area
rates
of
deep
networks.
E
Met
which
was
much
bigger
than
its
predecessor
and
it's
actually
the
human
eye,
the
killing
always
around
95
percent
female
5
percent.
It's
actually
the
film
first
time
better
than
the
humans,
and
this
is
okay,
we'll
application
occupations
and
others
music
mission
image
prediction:
localization
segmentation
tasks
they
these
are
the
competition
based
on
these
industries,
like
maybe
animation,
and
save
space
and
saturated
music
solution.
So
this
is
for,
like
the
testing
of
this
benefits
fading,
it's
actually
shows
that
no,
it
is
absolutely
stuff
anyway,
so
infinite.
E
E
Not
eat
overfitting,
but
we
go
dissipating
optimization
and
that
H
of
X
minus
X
is
easy
to
optimize
and
H
of
X
H
of
X
minus
X
is
X
of
x
helps
to
deepen
it.
These
are
the
the
golden
chains
upon
Bailey
Bailey.
The
frosting
is
finished
in
that
first
Angulo
hundred
genes
on
60%
greater
than
the
second,
the
only
thing
I'm
a
huge
increase
in
tons
of
stock
kacie.
Thank
you.
Any
questions.
A
Well,
thank
you
for
presenting
that's
a
very
interesting
paper.
It
gets
us
back
to
a
question.
Someone
asked
a
couple
weeks
ago,
which
is
what
is
the
relevance
of
adding
more
and
more
layers
to
a
network,
and
so
they
answered
the
question
pretty
well
I.
Think
there's
a
question
in
the
chat
from
Rosewall
was
papers
like
tigers,
so
visual
like.
C
E
A
So
I
had
a
question
about
the
lair,
so
you
showed
a
graph
where
you
had
you
were
at
it.
They
were
adding
layers
where
they
showed
like
where
they
have
like
20
layers
and
50
some-odd
layers,
and
then
they
showed
the
error
E
and
they
showed
that
it
actually
increases,
which
is,
of
course,
motivates
the
rest
of
the
work.
But
then
what
about
you
know?
Why
is
it
exactly?
The
doesn't
yeah
doesn't
guarantee
you'll
get
better
accuracy.
It's
about.
Let's
see,
yeah.
E
E
Many
features
to
X
star
so
need
some
features.
We
still
easy
if
they
available
t
pain
in
some
sense
like
who's,
easy
to
fit
waits
upon
some
features,
but
some
features
were
not
easy
to
bring
the
weights
upon
tonight.
It
was
causing
a
decree
he
is
not
created
because
it
was
not
properly
taken
in
some
sense,
I'd
I
didn't
know
already.
If
nots,
my
favorites,
of
course
mention
that
not
properly
training
on
it
was
an
optimization
poking.
A
Yeah
yeah
yeah.
We
have
a
lot
of
texts
from
as
well,
so
you
commented
adding
new
layers
doesn't
guarantee
that
you
will
get
better
accuracy,
it's
about
using
residuals
from
the
previous
layer
and
then
the
way
they
used
residual
features
again
in
the
neural
networks
is
something
which
is
groundbreaking.
A
Many
comment,
an
optimization
optimization
means
you
are
optimizing
parameters
to
make
your
model
more
detailed
and
then
covering
more
and
more
features
in
general,
so
yeah
I
mean
that's
the
impression
I
get
that
they
had
a
lot
of
features
and
that
they
were
able
to
get
more
resolution.
But
there
is
a
caveat
where
you
have
more
error
as
well
as
you're,
adding
layers.
So
that's
something
you
have
to
overcome
and
then
they
showed
that
you
can
overcome
them
using
their
methods.
So
I
think
that's
good
I.
Think
that's
a
very
good
paper.
Thank
you.
A
Everyone
for
presenting
your
papers
for
us.
Actually,
if
you
want
to
present
a
paper
you're
welcome
to
do
so.
You
know
if
you
want
to
do
it,
just
send
me
a
paper
selection.
You
can
go
to
the
archive
or
you
can
go
anywhere
to
any
of
the
conference
proceedings
find
a
paper
and
then
present
it
one
week
and
it'll
be
like
in
the
same
format
15
to
20
minutes.
So
you
know
it
if
you
want
to
do
it
that
that's
you're
welcome
to
do
it
so,
okay,
yeah!
A
B
B
A
B
A
Yeah
I
mean
wherever
I
got
in
there
yeah,
but
I
think
that
that
leads
into
some
some
of
the
stuff
in
the
second
and
third
phase,
but
yeah.
That
would
be
good.
If
we
could,
you
know,
do
that
deal
with
that
more
explicitly,
maybe
in
the
meetings
and
in
in
you
know
keeping
an
eye
towards
that
in
the
work
yeah.
B
B
Software
software
class
function
and
I
will
permitted
here
and
also
there
are
two
versions
of
software
Tom
organize
the
last
function
that
I
am
implemented.
So
mr.
I
have
different
combinations
of
these
models
and
last
functions
and
also
parametrically
coded
displayed
that
I
did
this
week,
so
I'm
going
to
show
them
so
show
some
outputs
that
yeah.
B
B
And
for
that,
and
with
that
I've
gotten
I
put
looking
like
this
here,
you
can
see
each
of
this
pixel
is
getting
a
value.
Is
that
in
class,
but
that's
based
on
whether
its
background
or
it's
an
itch.
So
this
is
how
he,
using
this
using
the
model
or
using
the
first
version
of
the
model
like
TF,
not
using
any
after
that
I.
B
B
B
B
Between
TDOT
immunity
under
players
in
we
need
to
specify
the
weights
and
weights
and
biases
for
our.
For
each
of
here,
I
have
to
define
the
weights
in
variances
for
each
of
the
convolutional
layer
or
batch
normalization
is
the
time
in
the
model
that
is,
that
is
the
case
in
it,
but
for
the
players
on
all
of
these
all
of
these
weights.
In
my
SS,
I
don't
have
to
I.
Don't
have
to
specify
that
by
default.
B
A
B
B
A
So
I
guess,
like
the
one
like,
if
you
had
some
that
basically
your
multi
color
maps,
that's
something
you
can
go,
and
you
know
decomposed
by
you
know
counting
the
number
of
pixels
that
are
you
know
of
a
certain
color
within
a
certain
boundary.
I
mean
basically
that's
what
we
want
kind
of
a
map
that
shows
the
boundaries
pretty
clearly
and
then
shows
the
interior
of
the
cell.
You
know
so
you
have
like.
C
A
Image
on
the
right,
Alex
also
has
I
think
the
boundary
might
be
a
little
bit
more
defined
in
that
one
left
hand
side.
Yes,
it's
yeah,
it's
gonna
be
a
little
bit
tricky
to
get
like
I
guess.
The
next
step,
for
that
would
be
to
kind
of,
like
maybe
see
what
kind
of
statistics
you
can
get
out
of
the
image
like
if
you
were
to
then
white
count
like
you
know,
find
out
where
the
boundaries
are
what's
inside
the
boundaries.
So
if
you
were
doing
like
a.
A
A
So
if
you
want
to
do
that,
if
you
want
to
try
to
wake
try
to
segment
these
images
more
simply
just
by
using
like
a
you
know,
just
like
some
sort
of
centroid
method,
then
you
can
actually
get
like
numbers
where
you
have
like
the
location
of
different
cells
or
blobs
in
their
area,
and
that
might
be,
let's
see
how
much
you
can
get
out
of
that.
That
makes
sense.
I
mean
I.
Can
more
I
can
work
with
you
on
that?
A
I
mean
and
then
it
would
just
be
test
them
and
see
which
one
looks
you
know,
which
ones
look
better
because
it's
hard
to
say
which
ones
actually
look
better
just
by
looking
at
them.
The
criterion
should
be
something
quantitative,
but
I,
don't
know
what
what
it's
gonna
do.
I
mean
I,
don't
know
what
it's
actually
gonna
pull
out
of
the
images.
B
Yes,
it's
about
this
idea
of
decay
after
this
model
has
outputted
this
image.
Basically,
if
you
do
some
some
other
unsupervised
segmentation,
like
they
still
a
clustering,
algorithm,
RK
k
and
k
needs
a
thrusting
algorithm,
basically
that
if
you
do
something
like
that,
that
actually
depends
on
centroids
and
I
think
maybe
we
could
get
something
useful
like
that.
B
The
the
same
boundary
in
the
cell
is
having
a
different
color
and
is
having
different
color
and
the
bagging
had
a
different
color.
So
after
that,
after
that,
then
then
I
had
a
limitation
in
mind
because,
after
after
the
model
after
the
model,
outputs
this
that
k-means
clustering
will
be
a
post-processing
to
this
image
right.
So
after
meeting
input,
this
this
results
in
a
j
which
is
written
in
java.
So
it's
missing
very
easy
to
do.
B
K-Means,
clustering
in
python,
and
also
I
don't
know
how
well
it
is
a
replicator
where
if
we
have
a
high
place,
it
comes
the
same
in
Java.
So
we
say
that
had
had
that
in
mind,
but
I'm
willing
to
pay
out.
That
looks
like
a
call
that
sounds
like
a
police
in
the
idea
and
yeah
it.
You
know
the
results
after
playing
out
paintings.
A
So
basically,
what
you've
done
is
you've,
taken
the
image
and
you've
kind
of
defined
a
skeleton
and
then
that
skeleton
now
we're
gonna,
take
it
apart
and
we're
gonna
see
which
one's
more
accurate
or
which
skeletons
are
more
accurate
for
the
purposes
of
the
next
step,
I
mean
there
is
no
so
vague
about,
like
the
algorithm
is
because,
of
course,
there
are
a
number
of
algorithms.
You
could
use
I.
Think
k-means.
Clustering
is
probably
good.
A
You
know
it's,
you
know
it's
something
that
uses
a
good
criterion
for
finding
the
edges
and
defining
the
areas.
And
then
you
know,
if
you
have
a
feature
like
the
background,
it's
pretty
easy
to
identify
it.
You
can
just
simply
say
you
know
it's
the
largest
part.
Okay,
we
know
if
the
features
over,
maybe
like
six
thousand
pixels.
We
know
you
can
throw
it
out
because
it's
you
know
obviously
wrong.
So
we
know
kind
of
the
size
scale
and
then
so
we
know
the
size
scale.
Oh
there's
some
feedback.
A
Yeah,
okay,
so
yeah
I
mean
that
and
then
yeah.
Then
there
are
other
tricks
like
you
can
invert
the
image
or
you
know,
do
some
other
basic
tricks
to
sort
of
augment
the
skeleton
here
so
I
mean
you
know,
depending
on
how
the
k-means
algorithm
is
picking
up
on
the
edges
you
can
play
around.
You
know
you
can
reduce
the
number
of
colors
to
two
or
something
like
that.
A
A
A
And
then
we
can
put
a
assignment
one,
and
then
you
know
you
can
just
work.
We
can
work
together
on
unselect
on
that.
That's
you
just
send
me
like
results.
If
you
find
something
like
exciting
a
river
yeah
yeah,
so
then
that's
like.
Are
there
any
other
issues
for
that,
or
do
you
think
that's
good
for
now.
B
B
A
A
B
Basically,
Heidi
with
that
4gb
system
is
that,
after
even
after
trading,
the
more
likes
in
Cagayan
neck
fat
downloaded
the
model
I'm
unable
to
run
that
just
because
this
is
all
it
has
about
18
combination
layers
and
some
some
heavy
planing.
So
my
memory
was
not
sufficient
to
load
that
model
and
credit
passing
image
asset
estimates
to
that
model
so
that
I
get
a
prediction.
So
basically,
for
that
purpose,
I
needed
some
RAM
and
now
I
think
it's
it's.
It's
pretty
good.
Okay,.
A
A
A
And
if
you
follow
that
and
you
go
to
the
design
repo
and
you
go
to
the
projects
tab
you
go,
you
will
see
the
digital
battle
area
board
and
then
we
now
have
quite
a
few
things
and
done,
which
is
good.
We
have
quite
a
few
things
in
progress
and
that's
where
we'll
start
number
17
is
still
I
was
to
working
on
this
adding
information
about
the
biology
of
a
square
yeah.
We
had.
A
D
A
A
It's
basically
the
paper
that
they
published
on
I
think
it
was
on
the
just
a
very
basic
overview
of
measurement
techniques
that
they
were
using.
And
so,
if
you
go
to
the
methods
of
that
paper,
you'll
see
some
of
it
we'll
go
through
that
in
more
detail.
I
think
later
on,
maybe
not
this
week,
but
you
know
just
just
so:
you've
come
up
it
over
and
get
yourself
familiarize
with
it
and
then
well
yeah
and
then
review
someone
who
that's
number
thirty.
A
This
is
what
I
was
talking
about:
documentation
of
procedure
number
thirty-six,
that's
actually
where
we're
actually
documenting
what
we're
doing
on
on
the
training,
so
we're
creating
a
training
set,
and
then
you
know
make
sure
that
you're
documenting
everything
so
that
we
can
go
back,
and
you
know
we're
gonna
turn
that
into
like
a
methods
section.
So
we
want
to
have
good
documentation
and
that's
just
a
reminder.
D
A
A
So
welcome
to
Pross
and
we'll
talk
about
a
presentation
so
we'll
put
that
in
progress,
and
then
you
can
send
me.
The
paper
model
features
within
in-between
frames.
Number
28,
I
think
that's
for
a
little
bit
later
on
mathematical
model
development.
Again,
that's
actually
something
out.
I
have
been
I
got
some
stuff
from
Thomas
and
really
I've
reviewed
it
a
little
bit
but
I.
A
Well,
I
mean
that'll,
probably
be
like
for
later
we'll
just
yeah,
we'll
just
we'll
be
able
to
actually
test
it
later,
but
okay,
so
I
mean
that
sounds
good.
That
you
have
at
least
been
thinking
about.
A
wall
well
address
this
more
in
in
coming
weeks,
but
I
just
wanted
to
see
where
you
it,
maybe
we'll
put
that
in
progress
since
kind
of
already
thinking
about
it
and
we're
not
quite
there
but
we'll
get
there.
A
Then
creepy
do
resources
for
vessel
where
you
biology,
number
21,
that's
still
kind
of
out
there
in
the
coming
weeks
and
then
define
features,
know
that
in
between
frames,
that's
gonna
require
28,
so
that's
gonna
be
left
for
later.
So
before
we
go,
are
there
any
issues
that
you
can
think
of
in
the
next
week
that
you
want
to
put
on
the
board.
A
Okay,
all
right
I
think
that's
good!
Thank
you
for
the
updates.
It
sounds
like
you
guys
are
making
progress.
Next
week
we
will
have
a
presentation
from
the
digital
bas,
o
area
trio,
on
their
progress
and
and
that's
alright
for
all
I
think.
That's
all
for
now
again,
you
can
slack
me
during
the
week
that
were
email
me
and
thanks
to
everyone
for
your
efforts
and
talk
to
you
next
week.