Skip to document
Welcome to StudocuSign in to access the best study resources
Guest user
0followers
Upload
HomeMy LibraryAsk AI
  • You don't have any recent items yet.
My Library
  • You don't have any courses yet.
  • You don't have any books yet.
  • You don't have any Studylists yet.

6.036 Lectures

Lecture notes for machine learning
Course

Intro to Machine Learning (6.036)

13 Documents
Students shared 13 documents in this course
Academic year: 2021/2022
Uploaded by:
EC
Emma Chabane
Massachusetts Institute of Technology
0followers
8Uploads
4upvotes

Comments

Please sign in or register to post comments.

Students also viewed

Related documents

Preview text

6 LECTURES

LECTURE 1-

Sept

Kim

,

20h

REGRESSION

An

§

Predict

pollution

level

8

Tr a i n i n g

data

####### {

i

n

data

points

É

z for data

point

ie {

1

,

^^

,

n

}

feature vector ✗

"'

= (×

,

"

'

,

.

.-

,

✗ ¥

)

"

c-

Rd

satellite

readings

label

y

"'

c-

IR

What

dewe want?

Good

way

to

predict

new

points

]→

✗

→ he

→

y

Hypothesis

h :

Rd

→

IR

linear

regresses

Hypothesis

class H

:

set of

h

p

hypothesis

isa

hyperplane

!

Linear

regression

hypothesis

{

when d-

l i h (

¥ 0

,

Oo

,

)

= 0 ×-

Of

parameters

when dzl

i

hlx

;

0,

.)

=

Qx,

Odxd

% OR hlx ;o )

= Otx

=

O'

✗

too

6

N.

not same

0 &✗

!!

Good

hypothesis

p

actual

How

good

is

a

regressor

at specific point

: loss

hlg,

a)

guess

learning

a

good regressor

i Recall ✗

→

try

Now

Dn

→

Learning

algorithm

→ h

Tr a i n i n g

error :

square

loss

,

1in

ng ,

extra

"

i

feature

J to)=

In

(

Io

'

(

Io

Ñ

)

Minimize 500 )

:

p

o

an matrix

Gradient

D.

Jlos =

0 = (ETE

)

"

ITI

N

.B

:

If

one

feature exalt

linear combo at the other

,

no

hyperplane

(no

best b)

LECTURE 7

Oct

26th

,

2021

PROBLEM

setup

→ n

NOT rib data

points

BACK PROPAGATION

Derivatives

LECTURE

8- Nov

2nd

,

Lose

Recall

:

①

Fully

connected Nns

inputs

are

completely

connected to

outputs

by

weights

Why

are

F NNS not

good

for

images

?

→

Spatial

locality

( met

all

pixels

matter

,

so we don't

want themallconnected )

  • Translation invariance (

moving

an

image

can't be

reflected

by

ND

i. Convolution neural networks

(

CNN

)

What does

the filter do

?

I

d

b

Filter

  • Reto

picks

out

cases

of 1 surrounded w/

0

¥

(otherwise all would be

0

)

lone

pixel

detector

brewed

d

Padding

Often

designed

so

image

. sire

= convolution .site

Adds unfiltered zeroes

Bias

Adds

bias to

pixels

during

convolution

Can

change

output

of Row

Weights

Overview

Nb of

weights

for CNN

Nbof

weights

F.

NN

Max Pooling

Filter

taking

the max

at its

arguments

  • No

weights

Stride

= nb at

steps

taken btwn calculations

####### *

!

v.

I

:]

[

Cool

too

< ✗

a

<

✗

[

o

,

i ]

X

'

'

pooAHeeosgzhog@wni

.es

¥%sÉ%?É÷¥÷÷÷÷?÷÷¥

÷÷É÷÷

:*

:* :B

μ

*¥☒q ¥R

¥μ÷Eg ,

&÷÷÷%

-0-

H

"

Hairy

,

Was this document helpful?

6.036 Lectures

Course: Intro to Machine Learning (6.036)

13 Documents
Students shared 13 documents in this course

University: Massachusetts Institute of Technology

Was this document helpful?
6.036 LECTURES

Students also viewed

Related documents