Neural Network Back-propagation - ajou.ac.krcvml.ajou.ac.kr/wiki/images/b/b0/Ch6_1_Backpropagation_1.pdf · 2017-01-30 · Neural Network – Back-propagation ... •Backpropagation

Neural Network –

Back-propagationHYUNG IL KOO

Hidden Layer Representations

• Backpropagation has an ability to discover useful intermediate

representations at the hidden unit layers inside the networks which

capture properties of the input spaces that are most relevant to

learning the target function.

• When more layers of units are used in the network, more complex

features can be invented.

• But the representations of the hidden layers are very hard to

understand for humans.

Basic Math

Optimization

Find 𝒙 that minimizes 𝒇(𝒙)

If 𝒇(𝒙) is differentiable,

But, in many cases, solving the above equation is a still

difficult problem.

Gradient descent

𝛁𝒇(𝒙𝟏)

𝒙𝟏 𝒙𝟐

𝒚 = 𝒇(𝒙)

→

𝛁𝒇(𝒙𝟐)

𝒙𝟑→

Chain rule with a single variable

Chain Rule (multiple variables)

𝑤 = 𝑓 𝑥, 𝑦, 𝑧 =>

Chain Rule

𝑑𝑤

𝑑𝑡=

𝜕𝑓

𝜕𝑥∙𝑑𝑥

𝑑𝑡+

𝜕𝑓

𝜕𝑦∙𝑑𝑦

𝑑𝑡+

𝜕𝑓

𝜕𝑧∙𝑑𝑧

𝑑𝑡

Δ𝑤 ≃𝜕𝑓

𝜕𝑥Δ𝑥 +

𝜕𝑓

𝜕𝑦Δ𝑦 +

𝜕𝑓

𝜕𝑧Δz

Feed-forward neural network

Feed forward network example: 1st layer

The 1st hidden layer

+1

Non-linear function

𝑤11𝑤12𝑤21𝑤22

𝑤31𝑤32

Feed forward network example: 2nd layer

The 2nd hidden layer+1

𝑢11

𝑢12𝑢13

𝑢23

𝑢22

𝑢21

The 1st hidden layer

+1

Forward propagation

The 1st hidden layer The 2nd hidden layer

Soft

Max

Lay

er 1

Lay

er 2

Output

Forward propagation matrix repr.

The 1st hidden layer The 2nd hidden layer

Soft

Max

Lay

er 1

Lay

er 2

Output

Back-propagation algorithm

Parameter update : Gradient Descent

Weight update method

Learning rate

Loss function (𝐿)`

Ground Truth

Dataflow diagram

The 1st hidden layer The 2nd hidden layer

Soft

Max

Output

Lay

er 1

Lay

er 2

VS

Back-propagation step; Loss function

Computing Loss function(𝐿):ex) Cross entropy

Soft

Max

Output Ground Truth

Overview

Lay

er 1

Lay

er 2

The 1st hidden layer The 2nd hidden layer

Soft

Max

Lay

er 1

Lay

er 2

Output Ground Truth

Loss function

1

0

Soft

Max

Back-propagation; 2nd layer

the Layer 2 has to do

Lay

er 2

The 1st hidden layer The 2nd hidden layer

Soft

Max

Output Ground Truth

•Weight update

Error propagation

Back-propagation; 2nd layer

the Layer 2 has to do

Lay

er 2

The 1st hidden layer The 2nd hidden layer

Soft

Max

Output Ground Truth

•Weight update

Error propagation

Error propagation(feed forward network)

The 2nd hidden layerThe 1st hidden layer

+1+1

𝜕𝐿

𝜕𝑧1and

𝜕𝐿

𝜕𝑧2are from its

upper layer.

Weight updates(feed forward network)

The 2nd hidden layerThe 1st hidden layer

+1+1

𝜕𝐿

𝜕𝑧1and

𝜕𝐿

𝜕𝑧2are from its

upper layer.

Lay

er 1

Back propagation; 1st layer

Lay

er 2

The 1st hidden layer The 2nd hidden layer

Soft

Max

Output Ground Truth

the Layer 1 has to do

•Weight update

Error propagation, Input update

???

𝑤𝑖𝑗𝑛𝑒𝑤 = 𝑤𝑖𝑗

𝑜𝑙𝑑 − 𝜇𝜕𝐿

𝜕𝑤𝑖𝑗

The 1st hidden layer

+1

(feed forward network)

Weight updates

𝜕𝐿

𝜕𝑦1, 𝜕𝐿

𝜕𝑦2and

𝜕𝐿

𝜕𝑦3are

from its upper layer

𝑤𝑖𝑗𝑛𝑒𝑤 = 𝑤𝑖𝑗

𝑜𝑙𝑑 − 𝜇𝜕𝐿

𝜕𝑤𝑖𝑗

Block-based perspective

Basic Math

Y= 𝑿⊤𝑨𝑿

Y+𝛥𝑌 ≃ 𝑋 + Δ𝑋 ⊤𝐴 𝑋 + Δ𝑋= 𝑋⊤𝐴𝑋 + Δ𝑋⊤𝐴𝑋 + 𝑋⊤Δ𝑋 + Δ𝑋⊤𝐴ΔX≈ 𝑋⊤𝐴𝑋 + 𝑋⊤ 𝐴 + 𝐴⊤ ΔX

=𝜕𝑌

𝜕𝑋

𝜕𝑌

𝜕𝐴⇒

Y+𝛥𝑌 ≈ 𝑋⊤ 𝐴 + Δ𝐴 𝑋= 𝑋⊤𝐴𝑋 + 𝑋⊤Δ𝐴𝑋

Δ𝑌 = 𝑋⊤Δ𝐴𝑋= 𝑡𝑟 𝑋⊤Δ𝐴𝑋= 𝑡𝑟(𝑋𝑋⊤Δ𝐴)

∴𝜕𝑌

𝜕𝐴= 𝑋𝑋⊤

Block-based representation

𝑋 𝑌 𝑍𝑌 = 𝑊𝑋 + 𝑏 𝑍 = 𝜑(𝑌)

𝜕𝐿

𝜕𝑍

𝜕𝐿

𝜕𝑌𝜕𝐿

𝜕𝑋

𝜕𝐿

𝜕𝑋

⊤

= 𝑊⊤𝜕𝐿

𝜕𝑌

⊤ 𝜕𝐿

𝜕𝑌

⊤

= diag 𝜑′ 𝑍𝜕𝐿

𝜕𝑍

⊤

𝜕𝐿

𝜕𝑊

⊤

=𝜕𝐿

𝜕𝑌

⊤

𝑋⊤

𝜕𝐿

𝜕𝑏

⊤

=𝜕𝐿

𝜕𝑌

⊤

Forward propagation

(block-based representation)

Output

Soft

Max

Layer 1

Layer 1

Layer 2

Layer 2

Backward propagation; 2nd layer

Ground

Truth

VS

Output

Soft

Max

Layer 1 Layer 2

• Error propagation

Backward propagation; 2nd layer

VS

Ground

Truth

VS

Output

Soft

Max

Layer 1 Layer 2

• Error propagation

Backward propagation; 2nd layer

• Weight update• Error propagation

Ground

Truth

VS

Output

Soft

Max

Layer 1 Layer 2

Backward propagation; 1st layer

• Error propagation

Ground

Truth

VS

Output

Soft

Max

Layer 1 Layer 2

Backward propagation; 1st layer

• Error propagation

Ground

Truth

VS

Output

Soft

Max

Layer 1 Layer 2

• Weight update

Input Optimization while

fixing all weights

Input update

The 1st hidden layer

+1

(feed forward network)

𝜕𝐿

𝜕𝑦1, 𝜕𝐿

𝜕𝑦2and

𝜕𝐿

𝜕3are from

its upper layer.

Output Loss Building

Output Loss Building

𝐼𝑛𝑒𝑤 = 𝐼𝑜𝑙𝑑 − 𝜇𝜕𝐿

𝜕𝐼

Output Loss Building

Inceptionism: Going Deeper into Neural Networks

https://research.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html

Inceptionism: Going Deeper into Neural Networks

DEEP INSIDE CONVOLUTION

NETWORKS

Inputs maximizing class score

Inputs maximizing class score

Objective Function

(to be maximized)

K. Simonyan, A. Vedaldi, A. Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps”, ICLR Workshop 2014

𝑆𝑐 𝐼 − 𝜆|| 𝐼 ||2

Inputs maximizing class score

Objective Function

(to be maximized)

Goose class

𝑆𝑐 𝐼 − 𝜆|| 𝐼 ||2

Maximizing class score

Objective Function

(to be maximized)

Goose class

𝑆𝑐 𝐼 − 𝜆|| 𝐼 ||2

Inputs maximizing class score

dumbbell

bell pepper

cup

lemon husky

dalmatian

Inputs maximizing class score

computer keyboard

Washing machine

kit fox

goose ostrich

limousine

DEEP INSIDE CONVOLUTION

NETWORKS

Saliency visualization

Saliency visualization

• Linear score model for class c:

𝑤 : importance of corresponding pixels of 𝐼 for class 𝑐

Saliency visualization

Objective Function

Dog Class

Saliency visualization

Objective Function

Dog Class

Differentiation

𝐼0 =

𝑆𝑐(𝐼0)

Saliency visualization

saliency map

𝐼0 =

yacht dog monkey

buildingcowwashing machine

A NEURAL ALGORITHM OF

ARTISTIC STYLE

Loss

Gatys, Leon A., Alexander S. Ecker, and Matthias Bethge. "A neural algorithm of artistic style." arXiv preprint arXiv:1508.06576 (2015).

Loss

Artistic style

Artistic style

Backups

Vector-by-scalar

Scalar-by-vector

Vector-by-vector

Scalar-by-matrix

Why 𝜕𝐿

𝜕𝑊= 𝑋

𝜕𝐿

𝜕𝑌?

• We want to find 𝜕𝐿

𝜕𝑊satisfying:

• Δ𝐿 = tr𝜕𝐿

𝜕𝑊Δ𝑊 .

• from• 𝑌 = 𝑊𝑋

• Δ𝐿 =𝜕𝐿

𝜕𝑌Δ𝑌.

• [Intuitive derivation]

• Δ𝑌 = Δ𝑊𝑋

• Δ𝐿 =𝜕𝐿

𝜕𝑌Δ𝑌 =

𝜕𝐿

𝜕𝑌Δ𝑊𝑋 = tr

𝜕𝐿

𝜕𝑌Δ𝑊𝑋 = tr 𝑋

𝜕𝐿

𝜕𝑌Δ𝑊

•𝜕𝐿

𝜕𝑊= 𝑋

𝜕𝐿

𝜕𝑌

𝑋 𝑌 𝑍 V

𝑌 = 𝑊𝑋 𝑍 = 𝑌 + 𝑏 𝑉 = 𝜙(𝑍)

𝜕𝐿

𝜕𝑉

𝜕𝐿

𝜕𝑍

𝜕𝐿

𝜕𝑌𝜕𝐿

𝜕𝑋

𝜕𝐿

𝜕𝑋

⊤

= 𝑊⊤𝜕𝐿

𝜕𝑌

⊤𝜕𝐿

𝜕𝑌

⊤

=𝜕𝐿

𝜕𝑍

⊤ 𝜕𝐿

𝜕𝑍

⊤

= diag 𝜙′ 𝑍𝜕𝐿

𝜕𝑉

⊤

𝜕𝐿

𝜕𝑊

⊤

=𝜕𝐿

𝜕𝑌

⊤

𝑋⊤𝜕𝐿

𝜕𝑏

⊤

=𝜕𝐿

𝜕𝑍

⊤

Proof

• Element-wise operation

…

𝑦1

…𝑦2

𝑦𝑛

𝑧1 = 𝜑(𝑦1)𝑧2 = 𝜑(𝑦2)

𝑧𝑛 = 𝜑(𝑦𝑛)

…

𝜕𝐿

𝜕𝑦1…

𝜕𝐿

𝜕𝑦2𝜕𝐿

𝜕𝑦𝑛

𝜕𝐿

𝜕𝑧1𝜕𝐿

𝜕𝑧2𝜕𝐿

𝜕𝑧𝑛

𝜕𝐿

𝜕𝑌

⊤

= diag 𝜑′ 𝑍𝜕𝐿

𝜕𝑍

⊤

𝜕𝐿

𝜕𝑦𝑖=

𝜕𝐿

𝜕𝑧𝑖𝜑′ 𝑧𝑖

⇒ ∆𝐿 ≅𝜕𝐿

𝜕𝑌𝑊∆𝑋

=𝜕𝐿

𝜕𝑋

Proof

∆𝐿 ≅𝜕𝐿

𝜕𝑌∆𝑌

∆𝑌 ≅ 𝑊∆𝑋

∴𝜕𝐿

𝜕𝑋

⊤

= 𝑊⊤𝜕𝐿

𝜕𝑌

⊤

∆𝑦 = ∆𝑊𝑋

∆𝐿 ≈ tr𝜕𝐿

𝜕𝑊∆𝑊

∆𝑦 ≅ ∆𝑊𝑋

∆𝐿 ≈𝜕𝐿

𝜕𝑌∆𝑌

∆𝐿 ≅𝜕𝐿

𝜕𝑌∆𝑊𝑋 = tr

𝜕𝐿

𝜕𝑌∆𝑊𝑋

= tr 𝑋𝜕𝐿

𝜕𝑌∆𝑊

∴𝜕𝐿

𝜕𝑊

⊤

=𝜕𝐿

𝜕𝑌

⊤

𝑋⊤

∆𝑌 = 𝑊∆𝑋

New Layer Design

New layer addition

The 1st layer The Nth layer

So

ft M

ax

Lay

er 1

Lay

er N

OutputA newly added layer

…

…

…

…

…

…

𝛽1

𝛽3

𝛽2

Lay

er m

𝛼1

𝛼3

𝛼2

𝑑𝐿

𝑑𝛽3

𝑑𝐿

𝑑𝛽2

New layer design

• Forward pass

• Compute output from input

Backward pass

𝑑𝐿

𝑑𝛽1

Lay

er m

𝑑𝐿

𝑑𝛼3

𝑑𝐿

𝑑𝛼2

𝑑𝐿

𝑑𝛼1

Forward pass

Lay

er m

𝛽1

𝛽3

𝛽2

𝛼1

𝛼3

𝛼2

• Backward pass

• Compute the derivatives w.r.t. data

• Update weights ??

Max

pooli

ng

lay

er

Example: Max pooling

5 6 3 9

2 1 4 11

17 14 3 21

11 13 12 9

6 11

17 21

Derivatives of max

𝑑𝐿

𝑑𝑧

Backward passm

ax

𝑑𝐿

𝑑𝑥=𝑑𝐿

𝑑𝑧

𝑑𝑧

𝑑𝑥

Forward pass

max

x

y

z

• For forward pass • For backward pass

𝑑𝐿

𝑑y=𝑑𝐿

𝑑𝑧

𝑑𝑧

𝑑𝑦