Introduction

-
:

favtomatieally "
how to learn from and make predictions about , data
'
ML :
, - =

CS

a÷:g:!÷;÷mwPb
stats a ← neuro .

→ 's

i t

{
learning from

}iFn÷g÷g÷qg":g
-
labeled finding patterns lstrct .
learning by
examples =

data
"
" make
clustering
a '

regression , "
,

classification
"

dim . reduction ( PCA)

whymlissopopwlarsincen20102.es
of big
era
a -
data "
: NL

needs large datasets

exponential 9 in
→
memory
and computational capabilities
( GPU graphical processing units
:
DL
=

i¥E÷÷*÷÷÷÷÷¥
as
parallel computing for
large scale applications )
-

ML / neural networks :
Bet :

theory started in 60 , ¥01

'

:'
 ML vs classical statistics :
-

=¥e : how to use data

to estimate
=
value of unknown quantity
statistical inference
" "
classically low - dimension
,
ie
,
few parameter to estimate from
data

Mt : how to use data to predict

=

data stat : get the

f
on now unseen
Ex :
that falls grau coast
recordings of apple
.

video
2nd : predict traj .
 redials :
Machine learninging
-

1) observable
-
1 measurable quantities a ERP

features
" "

p = # of
features priori
a
j
'
al assumed
→
Rj ,
. . .

, p
a

important to describe the system of interest

dataset :D ( X y ) { ( Kil ERP gli EIR ) 3in '
=
=
, , ,

i labels liar buts G E Rn

"
' "
" "

inputs X e Rn xp
observations • n

X a [ a
"'
am , . -
n
a
"'
TT n a # data points
, ,
• "
fam poles

x .
fi!!:L I:÷÷÷ ) -
:
•
inhumane.com
learning laws of
motion
.
2) parametrized -
model :

f : se E IRP x O ERM to John) J EY -

-

output alphabet Y J
.

.
real GR :
regression

keassifica-fionms-sfcndojso.tk
weight Cage , height )
discrete 5 E { Oil , - -
-

regs : .
. .
.

fo John J or Ca ) or ) or

L parameters of model
{ fo : ① Erm 3 : hypothesis
class
Function to predict output label from
Wil
-

input sample

µ finding
The
learning pb the A
best "

parameters £ given data D

3) Cowen ( fo )
' "
or
energy :

Ely , f. H1 ) e IR > o

→ used to date the I

performance quality
of the model on the dataset @

based the cost i

learning on
optimizing
-

E
argfionnm.E.elgifegh.im?
-
-

Joon )
'
L l Yi Facial )
,
= ( gi -
squared er .
 assumed
→ noise is
gaussian .

slides vs cats
Dogs
on
keynote
→ discuss all ingredients
"

we 'll understand the

perception learning
'
 M¥6 :

how to
design good predictive
&
a
=

model ? K

:D =D train Diest
o ) direct U

with : D train A @ best =

&
~
90%
-
10%

do the splitting before anything

else and never change that back
,
 even if the results are poor ,

or thermite creates bias .

!
1) led on the
'
training
'
set
o¥y
E'
"

E ecg f. ca 't'd )
argnfin
=

( Dtrain I
with D train =
{ (gcif , htt'd Sia ,

the learnt model

2) Ever on

the test data D test :

"
E out of sample
"
e
.
ooh -
. error
"
generalisation
i error
 '

Four -

757 e ly '
't t , to titis)
l Diest l
wit n'te ) Sj ,
,

|Eat3E
" "
Ein : in -

sample error
"
e
training error

f Yo ff
'
if multiple candidates of model : . .
.

, ,

→
"
beat
"
model is the one that
minimizes the out -

of sample
-
error

→
a
cross - validation "

-
 Sum up
#
-

yesterday 's lecture :

MI :
"
how to learn from and make
,

predictions about , dada

"

so perML : learn from examples

I labeled data
ML
-
ingredients r

N N
←
e : D .
Cx , g)
"
X f IR
"
P "
EIR
→
, y
9
input
.

labels
 3¥79
"'
D= { ( a
, gli )
'

t
data pout label
favuple
seen ERP ?
ya EIR

p
: # of features
'
yci EIR :
regression
C- IN :
classification
2) parametrized model :
-

f : a ERP x a ERM te fo (a) EIR

Fa Ca )
{ Aa :
of
ER
Ya 3 :

t hypothesis class

→ i
model

f¥ :
function
( for
used
new
to

,
predict
unseen
,
"
fresh
"

data points )
"
allow to
→
generation "

=3
)costfunctia a energy functional
( notion of "
optimality "
l " goodness "
off )
Q :
y
EIR
, if a
Cal E IR t IR > o

T
 e Cy✓
, to ex ) )
A
YR N
'
EIRN

=
it e ( y
Cil
,
to Caen ) ) mean
=
square
error
d

squared :
iffy ,
fo ki ) = Mst ( y foot)
,

2
-1¥ ( fo is
g)
"'
Cnc
=
y -

In ,
t

-
learning problem Copt pb ) .
:

afgemjnmii i i iea.ra.mg folacin )

"

Edd train ) =
,
 2
MEE ar g min E (y "'
-
fo Cecily )
Q ERM ly en , nailedtrain
- -

to ) ( O l D train )
got ran
=

g
T
conditioning
" "

" "
given
-

before doing anything else

D=DtrainUD#
{ ⑥ train n D test =
Of
@ train { ( Rei ERP glider ) 3! !Tin
'
a
,
 N N N' test choice
= train + :
your

N train 7 N best
N train ~ O 9 - N
,

This should changes

•
splitting neer be

posteriori a

• the splitting in D= @train U Diest

meet be random ( unbiased )

@train
@ nee learning donne on

=3 -0
 Evaluate your model !
-

"
of sample
"
tout : out error
"
r
generalisation error

EGG )
,f¥
Eat '
-
cab )'

ey * • **

"
Foot > Ein n

.
in -

sample error
"

training
"
" error

)
Ein f
' "
=

y Cali 's
Edwin ,
;
Ein la )
 Ein Ce)
learning argmin
:

often
-