Memahami Deep Learning

Pengantar Deep Learning
9 October 2017
untuk NLP
Alfan F. Wicaksono, FASILKOM UI

Alfan Farizki Wicaksono (alfan@cs.ui.ac.id)
Information Retrieval Lab. Fakultas

Ilmu Komputer Universitas
Indonesia 2017
1
Deep Learning Tsunami
“Deep Learning waves have lapped at the shores of
9 October 2017
computational linguistics for several years now, but 2015
seems like the year when the full force of the tsunami hit the
major Natural Language Processing (NLP) conferences.”
-Dr. Christopher D. Manning, Dec 2015

Christopher D. Manning. (2015). Computational Linguistics and Deep
Learning Computational Linguistics, 41(4), 701–707. 2
Tips
Sebelum mempelajari RNNs dan arsitektur Deep Learning
9 October 2017
yang lainnya, disarankan untuk kita mempelajari beberapa
topik berikut:
• Gradient Descent/Ascent

• Linear Regression
• Logistic Regression
• Konsep Backpropagation dan Computational Graph
• Multilayer Neural Networks
3
Referensi/Bacaan
• Andrej Karpathy’ Blog
9 October 2017
• http://karpathy.github.io/2015/05/21/rnn-effectiveness/
• Colah’s Blog
• http://karpathy.github.io/2015/05/21/rnn-effectiveness/

• Buku Deep Learning Yoshua Bengio
• Y. Bengio, Deep Learning, MLSS 2015
4
Deep Learning vs Machine Learning
• Deep Learning adalah bagian dari isu Machine Learning
9 October 2017
• Machine Learning adalah bagian dari isu Artificial Intelligence

Artificial Intelligence
• Searching
• Knowledge
Representation
Machine Learning
and reasoning
• Planning
Deep Learning
5
Machine Learning : Dirancang manusia
(rule-based) : Di-inferotomatis
9 October 2017
Predicted label: positive

Hand-crafted rules:
if contains(‘menarik’):
return positive
...
6
“Buku ini sangat menarikdan penuh manfaat”
(classical ML) : Di-inferotomatis
Fungsi klasifikasi dioptimasi berdasarkan input (fitur) &
9 October 2017
ouput.

Learn mapping from features to label Classifier
Feature Engineering!
Hand-designedFeatureExtractor:
Contoh: Menggunakan TF-IDF, Representation
informasi syntax dengan POS Tagger, dsb.
7
(Representation Learning) : Di-inferotomatis
Fitur dan fungsi klasifikasi di-optimasi secara bersama-sama.
9 October 2017

Learn Feature Extractor

Contoh: Restricted Boltzman Machine, Representation
Autoencoder, dsb.
8
(Deep Learning) : Di-inferotomatis
DeepLearninglearnsFeatures!
9 October 2017

Fitur yang Lebih High Level
Fitur Kompleks/High-Level
Fitur Sederhana Representation
9
Sejarah
10
Alfan F.Wicaksono, FASILKOM UI 9 October 2017

The Perceptron (Rosenblatt, 1958)
• Sejarah dimulai dimulai dengan Perceptron di akhir tahun 1950.
9 October 2017
• Perceptron terdiri dari 3 layer: Sensory, Association, dan
Response.

http://www.andreykurenkov.com/writing/a-brief-history-of-neural-nets-and-deep-learning/11
Rosenblatt, Frank. “The perceptron: a probabilistic model for information storage and
organization in the brain.” Psychological review 65.6 (1958): 386.
The Perceptron (Rosenblatt, 1958)
9 October 2017
Activation function adalah fungsi non-linier. Dalam kasus perceptron
Rosenblatt, activation function adalah operasi thresholding biasa (step
function).
Learning perceptron menggunakan metode Donald Hebb.
Saat itu, mampu melakukan klasifikasi untuk input Pixel 20x20! 12

http://www.andreykurenkov.com/writing/a-brief-history-of-neural-nets-and-deep-learning/
The organization of behavior: A neuropsychological theory. D. O. Hebb. John Wiley
And Sons, Inc., New York, 1949
The Fathers of Deep Learning(?)
9 October 2017
13
https://www.datarobot.com/blog/a-primer-on-deep-learning/
• Di tahun 2006, ketiga orang tersebut mengembangkan cara
untuk memanfaatkan dan mengatasi masalah training
9 October 2017
terhadap deep neural networks.
• Sebelumnya, banyak orang yang sudah menyerah terkait
manfaat dari neural network, dan cara training-nya.

• Mereka mengatasi masalah terkait Neural Network belum
mampu belajar untuk menemukan representasi yang berguna.
• Geoff Hinton has been snatched up by Google;

• Yann LeCun is Director of AI Research at Facebook;
• Yoshua Bengio holds a position as research chair for
Artificial Intelligence at University of Montreal 14
• Automated learning of data representations and features is
9 October 2017
what the hype is all about!

15
Mengapa sebelumnya “deep learning” tidak sukses?
• Sebenarnya, neural network kompleks sudah banyak
9 October 2017
ditemukan sebelumnya.
• Bahkan Long-Short Term Memory (LSTM) network, yang saat ini
ramai digunakan di bidang NLP, ditemukan tahun 1997
oleh Hochreiter & Schmidhuber.

• Ditambah lagi, orang dahulu percaya bahwa neural
network “can solve everything!”. Tetapi, mengapa mereka tidak
bisa melakukannya dahulu?
16
Mengapa sebelumnya “deep learning” tidak sukses?
• Beberapa alasan, oleh Ilya Sutskever:

• http://yyue.blogspot.co.id/2015/01/a-brief-overview-of-deep-
9 October 2017
learning.html
• Computers were slow. So the neural networks of past were tiny. And
tiny neural networks cannot achieve very high performance on

anything. In other words, small neural networks are not powerful.
• Datasets were small. So even if it was somehow magically possible
to train LDNNs, there were no large datasets that had enough
information to constrain their numerous parameters. So failure was
inevitable.
• Nobody knew how to train deep nets. The current best object recognition
networks have between 20 and 25 successive layers of convolutions.
A 2 layer neural network cannot do anything good on object
recognition. Yet back in the day everyone was very sure that deep
nets cannot be trained with SGD, since that would’ve been too
good to be true 17
The Success of Deep Learning
Salah satu faktor-nya adalah karena saat ini ditemukan cara
9 October 2017
learning yang bekerja secara praktikal.
The success of Deep Learning hinges on a very fortunate fact: that
well-tuned and carefully-initialized stochastic gradient descent
(SGD) can train LDNNs on problems that occur in practice. It is not a

trivial fact since the training error of a neural network as a function
of its weights is highly non-convex. And when it comes to non-
convex optimization, we were taught that all bets are off...
And yet, somehow, SGD seems to be very good at training those large
deep neural networks on the tasks that we care about. The problem
of training neural networks is NP-hard, and in fact there exists a
family of datasets such that the problem of finding the best neural
network with three hidden units is NP-hard. And yet, SGD just solves
it in practice. 18
IlyaSutskever, http://yyue.blogspot.co.at/2015/01/a-brief-overview-of-deep-learning.html
Apa Itu Deep Learning?
19

Apa itu Deep Learning?
• Kenyataannya, Deep Learning = (Deep) Artificial Neural
9 October 2017
Networks (ANNs)
• Dan Neural Networks sebenarnya adalah sebuah
Tumpukan FungsiMatematika

20
Image Courtesy: Google

Secara praktis, (supervised) Machine Learning itu adalah:
9 October 2017
Ekspresikan permasalahan ke dalam sebuah fungsi F (yang
mempunyai parameter θ), lalu secara otomatis cari
parameter θ sehingga fungsi F tepat mengeluarkan output
yang diinginkan.

Predicted label: Y = positive
Y=F(X;θ)
21
X: “Buku ini sangat menarikdan penuh manfaat”
Untuk Deep Learning, fungsi tersebut biasanya terdiri dari
9 October 2017
tumpukan banyak fungsi yang biasanya serupa.
Y  F (F (F ( X ;3 );2 );1 )

Y=positive
F(X;θ3)
Gambar ini sering disebut Tumpukan Fungsi ini
dengan istilah sering disebut dengan
F(X; θ2)
ComputationalGraph TumpukanLayer
F(X; θ1)
22
• Layer yang paling terkenal/umum adalah Fully-Connected
9 October 2017
Layer.
Y  F ( X )  f (W .X  b)
• “weighted sum of its inputs, followed by a non-linear function”

• Fungsi non-linier yang umum digunakan: Tanh (tangent
hyperbolic), Sigmoid, ReLU (Rectified Linear Unit)
W  R MN N  
M unit X  R f w x  b 
  i i 
 
N unit i
b  RM f
w x  b
f (W .X  b)
i i
i 23

X Non-linearity
Mengapa perlu “Deep”?
• Humans organize their ideas and concepts hierarchically
9 October 2017
• Humans first learn simpler concepts and then compose
them to represent more abstract ones
• Engineers break-up solutions into multiple levels of
abstraction and processing

• It would be good to automatically learn / discover these
concepts
24
Y. Bengio, Deep Learning, MLSS 2015, Austin, Texas, Jan 2014
(Bengio & Delalleau 2011)
Neural Networks
Y  f (W1.X  b1 )
9 October 2017
X Y  f (W1.X  b1 )
25
Neural Networks
Y  f (W1.( f (W1.X  b2 ))  b1 )
9 October 2017
Y  f (W2 .H1  b2 )
X H1  f (W1.X  b1 )
26
Neural Networks
Y  f (W1.( f (W2 .( f (W3.X  b3 ))  b2 ))  b1 )
9 October 2017
Y  f (W3.H 2  b3 )
X  
H1  f (W1.X  b1 ) H 2  f (W2 .H1  b2 )
27
Alasan matematis mengapa harus “deep”?
• A neural network with a single hidden layer of enough units
9 October 2017
can approximate any continuous function arbitrarily well.
• In other words, it can solve whatever problem you’re

interested in!
(Cybenko 1998, Hornik 1991)
28
Alasan matematis mengapa harus “deep”?
Akan tetapi...
9 October 2017
• “Enough units” can be a very large number. There are functions
representable with a small, but deep network that would require
exponentially many units with a single layer.
• The proof only says that a shallow network exists, it does not say

how to find it.
• Evidence indicates that it is easier to train a deep network to perform
well than a shallow one.
• A more recent result brings an example of a very large class of
functions that cannot be efficiently represented with a small-
depth network.
(e.g.,Hastadetal.1986,Bengio&Delalleau2011) (Braverman,
2011) 29
Mengapa perlu non-linearity?
 
f w x  b 
 

i i
 
9 October 2017
i
f
w x  b
i
i i

Non-linearity
Mengapa perlu fungsi non-linier f?
H1  W1.X  b1 Kalau tanpa f, rangkaian fungsi ini adalah

tetap fungsilinier.
H 2  W2 .H1  b2
Data bisa sangat kompleks, dan terkadang
Y  W3.H 2  b3 hubungan yang ada pada data tidak hanya 30
linier, tetapi bisa non-linier. Perlu
? representasi yang bisa menangkap hal ini.
Training NeuralNetworks
Y  f (W1.( f (W2.( f (W3.X  b3 ))  b2 ))  b1)
9 October 2017
• Secara random, kita inisialisasi semua parameter W1, b1, W2,
b2, W3, b3

• Definisikan sebuah cost function/loss function yang
mengukur seberapa baik fungsi neural network Anda.
• Seberapa jauh nilai yang diprediksi dengan nilai sesungguhnya
• Secara iteratif/berulang-ulang, sesuaikan nilai parameter

sehingga nilai loss function menjadi minimal.
31
• Initialize trainable parameters
9 October 2017
randomly

W(3)
W(2)
W(1)
32
Bukuinisangatbaikdanmendidik

• Initialize trainable parameters
9 October 2017
randomly
• Loop: x = 1 → #epoch:
• Pick a training example

W(3)
W(2)
W(1)
33
x Bukuinisangatbaikdanmendidik

pos neg • Initialize trainable parameters
y’
9 October 2017
True Label 1 0 randomly
• Pick a training example
Pred. Label y 0.3 0.7
(Output) • Compute output by doing feed-

forward process
W(3)
h2
W(2)
h1
W(1)
34

y’
9 October 2017
L
y • Pick a training example
Pred. Label y 0.3 0.7
(Output) • Compute output by doing feed-

forward process
W(3)
• Compute gradient of loss w.r.t.
h2 output
W(2)
h1
W(1)
35

y’
9 October 2017
L
y
Pred. Label • Pick a training example
y 0.3 0.7
(Output) L • Compute output by doing feed-

forward process
W (3)
L • Compute gradient of loss w.r.t.
output
h2
• Backpropagate loss, computing
W(2) gradients w.r.t trainable
parameters. It’s like computing
h1 contribution of error to the
output of each parameter
W(1)
36

y’
9 October 2017
L
y
Pred. Label • Pick a training example
y 0.3 0.7

forward process
W (3)
output
h2
L • Backpropagate loss, computing
W (2) gradients w.r.t trainable
L parameters. It’s like computing
h1 contribution of
error to the output of each
W(1)
parameter
37

y’
9 October 2017
L
Pred. Label y • Pick a training example
y 0.3 0.7

forward process
W (3)
output
h2
L • Backpropagate loss, computing
W (2) gradients w.r.t trainable
L parameters. It’s like computing
h1 contribution of error to the
L output of each parameter
W (1)
38

Gradient Descent(GD)
Take small step in direction of negative gradient!


(tn 1) (tn ) t f ( (t )
)
9 October 2017
n
(t)

39
https://github.com/joshdk/pygradesc
Lebih Detail dalam Hal Teknis …
40

Gradient Descent (GD)
Salah satu framework paling sederhana untuk permasalahan

optimasi (multivariate optimization).
9 October 2017
Digunakan untuk mencari konfigurasi parameter-parameter
sehingga cost function menjadi optimal, dalam hal ini
mencapai local minimum.

GD menggunakan metode iteratif, yang dimulai dari sebuah
titik acak, dan secara perlahan-lahan mengikuti arah negatif
dari gradient sehingga pada akhirnya akan berpindah ke
suatu titik kritikal, yang diharapkan merupakan local
minimum.
41
Tidak dijamin mencapai global minimum !
Problem: carilah nilai x sehingga fungsi f(x) = 2x4 + x3 – 3x2

mencapai titik local minimum.
9 October 2017
Misal, kita pilih x dimulai dari x=2.0:

Algoritma GD konvergen
pada titik x=0.699, yang
merupakan local minimum.
Local minimum
42
Algorithm:
For t  1, 2, ..., Nmax :
9 October 2017
xt 1  xt t f '(xt )
If f '(x t 1 )  then return "converged on critical po int"
If xt  xt 1  then return "converged on x value"

If f (xt 1 )  f (xt ) then return "diverging"
αt : learning rate atau step size pada iterasi ke-t

ϵ: sebuah bilangan yang sangat kecil
Nmax: batas banyaknya iterasi, atau disebut epoch jika iterasi selalu sampai akhir Algoritma
dimulai dengan menebak nilai x1 ! 43
Tips: pilih αt yang tidak terlalu kecil, juga tidak terlalu besar.
Kalau parameter-nya ada banyak ?

Carilah θ = θ1, θ2, …, θn sehingga f(θ1, θ2, …, θn) mencapai local minimum !
9 October 2017
while not converged :

 Dimulai dengan menebak
 (t 1)

(t)
t f ( )
(t)
nilai awal θ = θ1, θ2, …, θn

1 1 (t)
1

2(t 1) (t2) t f ((t) )
2(t)

(t 1)

(t)
t f ((t) )
n(t)
n n
44
Take small step in direction of negative gradient!


(tn 1) (tn ) t f ( (t )
)
9 October 2017
n
(t)

45
https://github.com/joshdk/pygradesc
Logistic Regression
Berbeda dengan Linear Regression, Logistic Regression

digunakan untuk permasalahan Binary Classification. Jadi
output yang dihasilkan adalah diskrit ! (1 atau 0), (ya atau
9 October 2017
tidak).
Misal, diberikan 2 hal:

1. Unseen data x yang ingin diprediksi labelnya, yaitu y.
2. Model logistic regression dengan parameter θ0, θ1, …, θn
yang sudah ditentukan.
n
P( y  1| x;)  (0 1 x1  ... n xn )  (0  i xi )
i1
P( y  0 | x;)  1 P( y  1| x;)
46
1
(z) 
1 e z
Logistic Regression
Bisa digambarkan dalam bentuk “single neuron”.

n
P( y  1| x;)  (0 1 x1  ... n xn )  (0  i xi )
9 October 2017
i1
P( y  0 | x;)  1 P( y  1| x;)

1
(z)  x1
1 e z θ1
x2 θ2
θ3
x3
θ0
+1
47
Dengan fungsisigmoid sebagai activationfunction
Logistic Regression
Slide sebelumnya mengasumsikan bahwa parameter θ sudah

ditentukan.
9 October 2017
Bagaimana bila belum ditentukan ?
Kita bisa estimasi parameter θ dengan memanfaatkan
training data {(x(1), y(1)), (x(2), y(2)), …, (y(n), y(n))} yang

diberikan (learning).
x1
θ1
x2 θ2
θ3
x3
θ0
48
+1
Logistic Regression
Learning
Misal,
n
9 October 2017
h(x)  (0 1 x1  ... n xn )  (0  i xi )
i1
Probabilitas masing-masing kelas dapat dituliskan secarasingkat

(persamaan di slide-slide sebelumnya) dengan:

P( y | x;)  (h(x)) y .(1 h(x))1 y y {0,1}
Diberikan m buah training examples, likelihood:

m
L()   P( y(i ) | x ;)
(i )
i1
m
49
 (h(x ))

(i ) y( i ) (i )
.(1 h(x )) 1 y( i )
i1
Logistic Regression
Learning
Dengan MLE, kita akan mencari konfigurasi parameter yang

memberikan nilai likelihood maksimal (= nilai log-likelihood
9 October 2017
maksimal).
l()  log L()

m
  y(i ) log h(x(i ) )  (1 y ) log (1 h(x ))

(i ) (i )
i1
Atau, kita cari parameter yang meminimalkan fungsi negative log-

likelihood (cross-entropy loss). Inilah cost function kita !
m
J ()  y(i ) log h(x(i ) )  (1 y ) log (1 h(x ))
(i ) (i )
i1
50
Logistic Regression
Learning
Agar bisa menggunakan Gradient Descent untuk mencari

parameter yang meminimalkan cost function, kita perlu
9 October 2017
menurunkan:

J ()
i

Dapat ditunjukkan bahwa untuk sebuahtrainingexample(x,y):

J ()  (h (x)  y) x

j
j
Jadi, untuk update nilai parameter di tiap iterasi untuk semua

example:
 :m(h (x (i ) ) y(i ) ) xi 51
j j  j
i1
Logistic Regression
Learning
Batch Gradient Descent untuk learning model

inisialisasi θ1, θ2, …, θn
9 October 2017
 : (h
m
(x (i ) )  y(i ) ) xi

1 1  1
i1
 : m(h (x (i ) )  y(i ) ) xi

2 2  2
i1
 : (h (x (i ) ) y(i ) ) xi

m
n n  n
i1
52
Logistic Regression
Learning
Stochastic Gradient Descent: kita bisa membuat progress dan

update parameter ketika kita melihat sebuah trainingexample.
9 October 2017
inisialisasi θ1, θ2, …, θn

for i := 1 to m do :
 : ( y(i )  h (x (i ) )) xi
1 1  1
 : ( y  h (x )) x
(i ) (i ) i


2 2 2
 :( y(i ) h (x (i ) )) xi

n n  n
53
Logistic Regression
Learning
Mini-Batch Gradient Descent untuk learning model: Daripada kita

gunakan sebuah sample untuk update (seperti online learning), gradient
9 October 2017
dihitung dengan cara rata-rata/sum dari sebuah mini- batch
sample (misal, 32 atau 64 sample).

BatchGradientDescent: gradient dihitung untuk semuasample!
• Untuk sekali step, terlalu besar komputasinya
Online Learning: gradient dihitung untuk sebuah sample!

• Terkadang noisy
• Update step sangat kecil
54
Multilayer Neural Network (Multilayer Perceptron)
Logistic Regression sebenarnya bisa dianggap sebagai Neural

Network dengan beberapa unit di layer input, satu unit di
layer output, dan tanpa hidden layer.
9 October 2017
x1
θ1

x2 θ2
θ3
x3
θ0
+1
Dengan fungsi sigmoid sebagai activation function

55
MultilayerNeural Network (MultilayerPerceptron)
Misal, ada 3-layer NN, dengan 3 input unit, 2 hidden unit, dan
2 output unit.
9 October 2017
x1 (1)
𝑊11
(1) (2)
𝑊21 𝑊11

(2)
(1) 𝑊21
𝑊12
x2 (2)
(1) 𝑊12
𝑊22
(2)
(1) 𝑊22
𝑊13
𝑊23(1)
𝑊1
(2) 𝑊2(2)
x3 (1)
(1) 𝑊
𝑊1 2
+1
56
+1 W(1), W(2), b(1), b(2) adalah parameter !
Seandainya parameter sudah diketahui, bagaimana caranya

memprediksi label dari input (klasifikasi) ?
9 October 2017
Dari contoh sebelumnya, ada 2 unit di output layer. Kondisi ini
biasanya digunakan untuk binary classification. Unit
pertama menghasilkan probabilitas untuk pertama, dan unit

kedua menghasilkan probabilitas untuk kelas kedua.
Kita perlu melakukan proses feed-forward untuk

menghitung nilai yang dihasilkan di output layer.
57
Misal, untuk activation function, kita gunakan fungsi

hyperbolic tangent. f (x) tanh(x)
9 October 2017
Untuk menghitung output di hidden layer:
z1(2)  W 11(1) x1  W12(1) x2  W13(1) x3  b1(1)

z(2)
2  W (1)
21 1x  W (1)
22 2 x  W (1)
23 3 x  b2
(1)
a1(2)  f (z1(2) )
a2(2)  f (z (2)
2 ) Ini hanyalah perkalian matrix !
 x1 
W (1)W (1)
W   x   b1(1) 
(1)
13    58
z(2)  W (1) x  b(1)   11(1)
W23(1)   x 
 b(1)
12
2
W 21 W22  2 
(1)
 3 
Jadi, Proses feed-forward secara keseluruhan hingga

menghasilkan output di kedua output unit adalah:
9 October 2017
z (2)  W (1) x  b(1)
a (2)  f (z(2) )

z(3)  W (2) a (2)  b(2)
hW ,b (x) a(3)  softmax (z (3) )
59
Learning
Ada Cost function yang berupa cross-entropy loss: m
9 October 2017
adalah banyaknya training examples.
n 1
m sl sl1
1

l (l) ) 2
J (W , b) 
m
y i, j log( pi, j ) 
2
 (W ji )
i1 j l 1 i1 j 1
Regularization terms
Berarti, cost function untuk satu sample (x,y) adalah:

 
J (W , b; x, y)  y j loghW ,b (x j )  y j logp j 

60
j j
Learning
Namun, kali ini, Cost function kita adalah squared-error: m
9 October 2017
adalah banyaknya training examples.
(i) 2 

1 m  1  nl 1 sl sl1

   ji )
(i) (l) ) 2
J (W , b) m 2 hW ,b (x )  y (W
i1   2 l 1 i1 j 1
Regularization terms
Berarti, cost function untuk satu sample (x,y) adalah:
h
1 1
J (W , b; x, y)   ) y

(i )
h (x ) y(i )
(i ) 2 (x (i ) 2 61
W ,b W ,b j j
2 2 j
Learning
Batch Gradient Descent
9 October 2017
inisialisasi W, b
while not converged : Bagaimana cara menghitung gradient ??


Wij  Wij 
(l ) (l )
J (W , b)
Wij (l )

(l )   J (W , b)
bi  b i
(l )
bi(l )
62
Learning
Misal, (x,y)adalah sebuah training sample, dan dJ(W,b;x,y)

adalah turunan parsial terkait sebuah sample (x,y).
9 October 2017
dJ(W,b;x,y) menentukan overall partial derivative dJ(W,b):

J (W , b)   1 m 

(i ) 
   
(i ) (l )
J (W , b; x , y )  W
W (l )
m W (l ) ij
ij  i1 ij 


 

1
J (W , b) 
m
 J (W , b; x(i ) , y(i ) )
bi(l ) m i1 bi(l )
63
Dihitung dengan teknik BackPropagation

Learning
Back-Propagation
9 October 2017
1. Jalankan proses feed-forward
2. Untuk setiap output unit i pada layer nl (output layer)

i(nl )  J (W , b; x, y)  (ai(nl )  yi ) f '(zi(nl ) )

zi(nl )
3. l = nl - 1, nl - 2, ... 2
Untuk setiap nodes i di layer l
l1
  (W 
(l ) (l ) (l 1)
)  f '(z (l )
)
i ji j i
j 1
4. Finally..
 
J (W , b; x, y)  J (W , b; x, y) 64
a(lj )(li1)  (li1)
Wij(l ) bi(l )
Learning (3)
(2) 𝑊1
𝑊11 𝑊1(3)
Back-Propagation
(2)
𝑊21
9 October 2017
(2)
Contoh hitung gradient di output ... 𝑊12 (3)
𝑊2
(2)
𝑊22 𝑊2(3)
J (W , b; x, y) 
1
a (3)
y  2

1
a (3)
y  2
(2) 𝑊(2)
1 1 2 2 𝑊1 2

2 2
J (3)  y 
a1(3)
 a1 1
+1
(3) 
  f z   f ' z (3)   J  a z(3)
a (3) (3)
J (W , b; x, y)   1
 1

1 1
z (3)
z1
1
W (2)
a (3)
z(3)
W (2)
 
(3) 12 12
1 1 1
 a  y  f (3)
(3)
z  W  a W  a  b
(3) (2) (2) (2) (2) (2)
1 1 '(z 1 )  a2(2)
1 11 1 12 2 1
z(3)  (2) 65
1
a2
W (2)
12
Sensitivity – Jacobian Matrix
The Jacobian J is the matrix of partial derivatives of the
9 October 2017
network output vector y with respect to the input vector x.
. . . .
. . . . yk
J 

 J k ,i 
. . . . xi
 
. . . .
These derivatives measure the relative sensitivity of the outputs

to small changes in the inputs, and can therefore be used, for
example, to detect irrelevant inputs.
66
Alex Graves, Supervised Sequence Labelling with Recurrent Neural Networks

9 October 2017
More Complex Neural Networks

(Neural Network Architectures)
Recurrent Neural Networks (Vanilla RNN, LSTM, GRU)

Attention Mechanisms
Recursive Neural Networks
67
Recurrent Neural Networks
O1 O2 O3 O4 O5
9 October 2017
h1 h2 h3 h4 h5
X1 X2 X3 X4 X5
68
One of the famous Deep Learning Architectures in the NLP community
Recurrent Neural Networks (RNNs)
Kita biasanya menggunakan RNNsuntuk:

• Memproses Sequences
9 October 2017
• Menghasilkan Sequences
• …
• Intinya … ada Sequences

Sequences biasanya:
• Urutan kata-kata
• Urutan kalimat-kalimat
• Signal
• Suara
• Video (Sequence of Images)
• … 69
Not RNNs Sequence Input

(Vanilla Feed-Forward NNs) (e.g. Sentence Classification)
9 October 2017
SequenceInput/Output
(e.g. Machine Translation) 70
Sequence Output
(e.g. Image Captioning)
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
9 October 2017
RNNs combine the input vector with their state vector with a fixed

(but learned) function to produce a new state vector.
This can, in programming terms, be interpreted as running a fixed

program with certain inputs and some internal variables.
In fact, it is known that RNNs are Turing-Complete in the sense that

they can simulate arbitrary programs (withproperweights).
71
http://binds.cs.umass.edu/papers/1995_Siegelmann_Science.pdf
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
http://colah.github.io/posts/2015-08-Understanding-LSTMs/
Misal, ada I input unit, K output unit, dan H hidden unit (state).
Y1 Y2 Komputasi RNNs untuk satu sample:
9 October 2017
ht  R H 1
xt  RI 1 yt  RK 1
W (hy ) W ( xh)  R H I W (hh)  R H H

s1 s2
W (hy )  R K H
W (hh)
W ( xh) ht  W ( xh)  xt  W (hh)  st 1
st tanh ht 


X1 X2
yt  W (hy )  st 72
h0  0
Back Propagation Through Time (BPTT)

The loss function depends on the activation
Y1 Y2
of the hidden layer not only through its
9 October 2017
influence on the output layer.
( y ) (h)
 K
(hy ) ( y)
H
(hh) (h)

i,t    Wi, j j ,t   Wi,n n,t 1   f 'h 
i,t
 

s1 s2
 j 1 n1
 
(h)
Lt
Di output: i,t  y
( y)
i,t
Disetiapstep,kecuali paling
kanan:
Lt
X1 X2 i,t(h)  i,T(h)1  0 73
hi,t

The same weights are reused at every
Y1 Y2
timestep, we sum over the whole sequence to
9 October 2017
get the derivatives with respect to the
network weights.
L 
 (h)j ,t  si,t 1
(hy )
W T


s1 s2
Wi,(hh)
j t 1
W (hh) L T
W ( xh) (j y,t )  si,t

Wi,(hy
j
)
t 1
L T
X1 X2
( xh)  
Wi, j
  j ,t  xi,t
(h)
74
t 1
9 October 2017
75
Pascanu et al., On the difficulty of training Recurrent Neural Networks, 2013


Misal, untuk parameter antar state: Term-term ini disebut temporal
contribution: bagaimana W(hh)
9 October 2017
pada step k mempengaruhi cost
pada step-step setelahnya (t>k)
L L h   h
t
t
  t t k
W (hh) k 1 ht hk W (hh)

1 k t
Diputus sampai k step ke belakang.
Di sini artinya “immediate derivative”,
yaitu hk-1 dianggap konstan terhadap
W(hh).
ht ht ht 1 hk 2 hk 1
  
hk ht 1 ht 2 hk 1 hk

 h 
 W  ( xh)
x  W (hh)
s  76
t  st 1
k t 1
W (hh)
W (hh)
Vanishing & Exploding Gradient Problems

Bengio et al., (1994) said that “the exploding gradients problem refers to the
large increase in the norm of the gradient during training. Such events are
9 October 2017
caused by the explosion of the long term components, which can grow
exponentially more then short term ones.”
And “The vanishing gradients problem refers to the opposite behaviour, when

long term components go exponentially fast to norm 0, making it impossible for
the model to learn correlation between temporally distant events.”
Kokbisaterjadi?Cobalihatsalahsatutemporalcomponentdarisebelumnya:
ht ht ht 1 hk 2 hk 1

  
hk ht 1 ht 2 hk 1 hk
In the same way a product of t - k real numbers can shrink to zero or explode to
77
infinity, so does this product of Matrices. (Pascanu et al.,)
Bengio, Y., Simard, P., and Frasconi, P. (1994). Learning long-term dependencies with gradient
descent is difficult. IEEE Transactions on Neural Networks
Sequential Jacobian biasa digunakan
Recurrent Neural Networks (RNNs) untuk analisis penggunaan
konteks pada RNNs.
Vanishing& Exploding Gradient Problems
9 October 2017
78

Solusi untuk Vanishing Gradient Problem

1) Penggunaan non-gradient based training algorithms (Simulated
Annealing, Discrete Error Propagation, etc.) (Bengioetal.,1994)
9 October 2017
2) Definisikan arsitektur baru di dalam RNN Cell!, seperti Long-Short Term
Memory(LSTM)(Hochreiter&Schmidhuber,1997).

3) Untuk metode yang lain, silakan merujuk (Pascanu et al., 2013).
Pascanu et al., On the difficulty of training Recurrent Neural Networks, 2013
S. Hochreiter and J. Schmidhuber. Long Short-Term Memory. Neural Computation,9(8):1735

1780, 1997 79
Y. Bengio, P. Simard, and P. Frasconi. Learning Long-Term Dependencies with Gradient

Descent is Difficult. IEEE Transactions on Neural Networks, 1994
Variant: Bi-Directional RNNs
Alfan F. Wicaksono, FASILKOM UI 9 October 2017

80

Sequential Jacobian (Sensitivity) untuk Bi-DirectionalRNNs

81

Long-Short Term Memory (LSTM)
1. The LSTM architecture consists of a set of recurrently

connected subnets, known as memory blocks.
9 October 2017
2. These blocks can be thought of as a differentiable version
of the memory chips in a digital computer.
3. Each block contains:

1. Self-connected memory cells
2. Three multiplicativeunits (gates)
1. Input gates (analogue of write operation)
2. Output gates (analogue of read operation)
3. Forget gates ((analogue of reset operation))
82

1780, 1997
017
2
r
ob
e
ct
O
9

83
1780, 1997
The multiplicative gates allow LSTM

memory cells to store and access
9 October 2017
information over long periods of time,
thereby mitigating the vanishing gradient
problem.

For example, as long as the input gate
remains closed (i.e. has an activation near
0), the activation of the cell will not be
overwritten by the new inputs arriving in
the network, and can therefore be made
available to the net much later in the
sequence, by opening the output gate.
84
1780, 1997
Long-Short Term Memory (LSTM) Visualisasi
laindarisatucelldiLSTM
9 October 2017
85

1780, 1997
Komputasi diLSTM
9 October 2017
86

1780, 1997
Preservation of Gradient Information padaLSTM

87
1780, 1997
Example: RNNs for POS Tagger
(Zennaki, 2015)

PRP VBD TO JJ NN
h1 h2 h3 h4 h5
88
I went to west java
LSTM + CRF for Semantic Role Labeling
(Zhou and Xu, ACL 2015)
9 October 2017
89
Attention Mechanism
A potential issue with this encoder–decoder approach is that a neural
9 October 2017
network needs to be able to compress all the necessary information of
a source sentence into a fixed-length vector. This may make it difficult
for the neural network to cope with long sentences, especially those
that are longer than the sentences in the training corpus.

Dzmitry Bahdanau, et al., Neural machine translation by jointly
learning to align and translate, 2015
90
Attention Mechanism
Neural Translation Model TANPA Attention Mechanism
9 October 2017
Sutkever, Ilya et al., Sequence to Sequence Learning with Neural
Networks, NIPS 2014.
91
https://blog.heuritech.com/2016/01/20/attention-mechanism/
Attention Mechanism
Neural Translation Model TANPA Attention Mechanism
9 October 2017
Sutkever, Ilya et al., Sequence
to Sequence Learning with
Neural Networks, NIPS 2014.

92
Attention Mechanism
Mengapa Perlu Attention Mechanism?
9 October 2017
• Each time the proposed model generates a word in a translation, it
(soft-)searches for a set of positions in a source sentence where the
most relevant information is concentrated. The model then predicts
a target word based on the context vectors associated with these

source positions and all the previous generated target words.
• … it encodes the input sentence into a sequence of vectors and
chooses a subset of these vectors adaptively while decoding the
translation. This frees a neural translation model from having to
squash all the information of a source sentence, regardless of its
length, into a fixed-length vector.
Dzmitry Bahdanau, et al., Neural machine translation by jointly

93
learning to align and translate, 2015
Attention Mechanism
Neural Translation Model – dengan Attention Mechanism
9 October 2017
94
Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio, Neural Machine Translation by Jointly Learning to Align and
Translate, arXiv:1409.0473, 2016
Attention Mechanism
Neural Translation Model – dengan Attention Mechanism
9 October 2017
95
Cell merepresentasikan bobot attention, terkait translation.
Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio, Neural Machine Translation by Jointly Learning to Align and
Translate, arXiv:1409.0473, 2016
Attention Mechanism
Simple Attention Networks untuk Sentence Classification
9 October 2017
96
Colin Raffel, Daniel P. W. Ellis, F EED -F ORWARD NETWORKS WITH ATTENTION CAN
S OLVE S OME L ONG-T ERM MEMORY P ROBLEMS, Workshop track - ICLR 2016
Attention Mechanism
Hierarchical Attention Networks untuk Sentence Classification
9 October 2017
97
Yang, Zichao, et al., Hierarchical Attention Networks for Document Classification, NAACL 2016
Attention Mechanism
Hierarchical Attention Networks untuk Sentence Classification

Task: Prediksi Rating Dokumen

98
Yang, Zichao, et al., Hierarchical Attention Networks for Document Classification, NAACL 2016
Attention Mechanism

99
Attention Mechanism

100
Xu, Kelvin, et al. « Show, Attend and Tell: Neural Image Caption
Generation with Visual Attention (2016).
Attention Mechanism

101
Xu, Kelvin, et al. « Show, Attend and Tell: Neural Image Caption
Generation with Visual Attention (2016).
Attention Mechanism
Attention Mechanism untuk Textual Entailment

Diberikan pasangan premis-hipotesis, tentukan apakah 2 tersebut
kontradiksi, tidak berhubungan, atau logically entail.
9 October 2017
Attention Model digunkan
Sebagai contoh:
untuk menghubungkan
• Premis: “A wedding party taking pictures“ kata-kata di premis dan

• Hipotesis: “Someone got married“ hipotesis.
102
Tim Rocktaschel et al., REASONING ABOUT ENTAILMENT WITH NEURAL ATTENTION, ICLR 2016
Attention Mechanism
Attention Mechanism untuk Textual Entailment

Diberikan pasangan premis-hipotesis, tentukan apakah 2 tersebut
kontradiksi, tidak berhubungan, atau logically entail.
9 October 2017
103
Tim Rocktaschel et al., REASONING ABOUT ENTAILMENT WITH NEURAL ATTENTION, ICLR 2016
Recursive NeuralNetworks

104
R. Socher, C. Lin, A. Y. Ng, and C.D. Manning. 2011a. Parsing Natural Scenes and Natural Language with
Recursive Neural Networks. In ICML

p1 gW .b; c bias 105

Socher et al., Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank, EMNLP
2013

106
Socher et al., Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank, EMNLP
2013
Convolutional Neural Networks (CNNs) for Sentence Classification
(Kim, EMNLP 2014)
9 October 2017
107
Recursive Neural Network for SMT Decoding.
(Liu et al., EMNLP 2014)
9 October 2017
108

Memahami Deep Learning

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Memahami Deep Learning

Uploaded by

Copyright:

Available Formats

Pengantar Deep Learning

Alfan F. Wicaksono, FASILKOM UI

Information Retrieval Lab. Fakultas

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

(classical ML) : Di-inferotomatis

Fungsi klasifikasi dioptimasi berdasarkan input (fitur) &

Alfan F. Wicaksono, FASILKOM UI

(Representation Learning) : Di-inferotomatis

Fitur dan fungsi klasifikasi di-optimasi secara bersama-sama.

Alfan F. Wicaksono, FASILKOM UI

Learn Feature Extractor

(Deep Learning) : Di-inferotomatis

Alfan F. Wicaksono, FASILKOM UI

Fitur yang Lebih High Level

Fitur Sederhana Representation

Alfan F.Wicaksono, FASILKOM UI 9 October 2017

Alfan F. Wicaksono, FASILKOM UI

Learning perceptron menggunakan metode Donald Hebb.

Saat itu, mampu melakukan klasifikasi untuk input Pixel 20x20! 12

Alfan F. Wicaksono, FASILKOM UI

• Geoff Hinton has been snatched up by Google;

Alfan F. Wicaksono, FASILKOM UI

• Sebenarnya, neural network kompleks sudah banyak

Alfan F. Wicaksono, FASILKOM UI

• Beberapa alasan, oleh Ilya Sutskever:

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F.Wicaksono, FASILKOM UI 9 October 2017

Alfan F. Wicaksono, FASILKOM UI

Image Courtesy: Google

Alfan F. Wicaksono, FASILKOM UI

Y  F (F (F ( X ;3 );2 );1 )

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

• A neural network with a single hidden layer of enough units

• In other words, it can solve whatever problem you’re

Alfan F. Wicaksono, FASILKOM UI

(Cybenko 1998, Hornik 1991)

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Mengapa perlu fungsi non-linier f?

H1  W1.X  b1 Kalau tanpa f, rangkaian fungsi ini adalah

Alfan F. Wicaksono, FASILKOM UI

• Secara iteratif/berulang-ulang, sesuaikan nilai parameter

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Alfan F. Wicaksono, FASILKOM UI

Take small step in direction of negative gradient!

Alfan F. Wicaksono, FASILKOM UI

Alfan F.Wicaksono, FASILKOM UI 9 October 2017

Salah satu framework paling sederhana untuk permasalahan

Alfan F. Wicaksono, FASILKOM UI

Problem: carilah nilai x sehingga fungsi f(x) = 2x4 + x3 – 3x2