HAITHAM BOU AMMAR

MAASTRICHT UNIVERSITY

Transfer for Supervised Learning Tasks

Motivation

Model

€

y

€

x

€

x€

y

Training Data

?

Test Data

Assumptions: 1.Training and Test are from same distribution 2.Training and Test are in same feature space

Not

Tru

e in

man

y re

al-

wor

ld a

pplica

tion

s

Examples: Web-document Classification

Model

?

PhysicsMachine Learning

Life Science

Examples: Web-document Classification

Model

?

PhysicsMachine Learning

Life Science

Content Change !

Assumption violated!

Learn a new model

Learn new Model :

1.Collect new Labeled Data2.Build new model

Reuse & Adapt already learned model !

Examples: Image Classification

Model OneFeatures Task One

Task O

ne

Examples: Image Classification

Cars

Motorcycles

Task Two

Features Task One

Features Task Two

Reuse

Model Two

Traditional Machine Learning vs. Transfer

Source Task

Knowledge

Target Task

Learning System

Different Tasks

Learning System

Learning System

Learning System

Traditional Machine Learning

Transfer Learning

Questions ?

Transfer Learning Definition

Notation: Domain :

Feature Space: Marginal Probability Distribution:

with

Given a domain then a task is :

Labe

l

Space P(Y

|X)

Transfer Learning Definition

Given a source domain and source learning task, a target domain and a target learning task, transfer learning aims to help improve the learning of the target predictive function using the source knowledge, where

or

Transfer Definition

Therefore, if either : Domain Differences

Task Differences

Examples: Cancer Data

Age Smoking

Age Height Smoking

Examples: Cancer Data

Task

Sour

ce: C

lass

ify

into

can

cer o

r no

canc

erTa

sk T

arge

t: C

lass

ify

into

can

cer le

vel o

ne,

canc

er le

vel t

wo,

canc

er le

vel t

hree

Quiz

When doesn’t transfer help ? (Hint: On what should you

condition?)

Questions ?

Questions to answer when transferringW

hat to

Tra

nsfe

r ?

How

to T

ransf

er ?

When

to

Tra

nsf

er ?

Inst

ance

s

?

Mod

el ?

Featu

res ?

Map

M

odel

?

Uni

fy

Feat

ures

?

Wei

ght

Inst

ance

s ?

In w

hich

Situ

atio

ns

Algorithms: TrAdaBoost

Assumptions: Source and Target task have same feature space:

Marginal distributions are different:

Not all source data might be helpful !

Algorithms: TrAdaBoost (Quiz)

Wha

t to

Trans

fer ?

How

to T

rans

fer ?

Inst

ance

s

Wei

ght I

nsta

nces

Algorithm: TrAdaBoost

Idea: Iteratively reweight source samples such that:

reduce effect of “bad” source instances encourage effect of “good” source instances

Requires: Source task labeled data set Very small Target task labeled data set Unlabeled Target data set Base Learner

Algorithm: TrAdaBoost

Weights Initialization

Hypothesis Learning and error calculation

Weights Update

Questions ?

Algorithms: Self-Taught Learning

Problem Targeted is :

1.Little labeled data

2.A lot of unlabeled data

Build a model on the labeled data

Natural scenes

Car

Motorcycl

e

Algorithms: Self-Taught Learning

Assumptions: Source and Target task have different feature space:

Marginal distributions are different:

Label Space is different:

Algorithms: Self-Taught Learning (Quiz)

Wha

t to

Trans

fer ?

How

to T

rans

fer ?

1. D

iscov

er

Featu

res

2. U

nify

Fea

ture

s

Featu

res

Algorithms: Self-Taught Learning

Framework: Source Unlabeled data set:

Target Labeled data set: Natural

scenes

Build classifier for cars and Motorbikes

Algorithms: Self-Taught Learning

Step One: Discover high level features from Source data by

Regularization Term Re-construction Error

Constraint on the Bases

Algorithm: Self-Taught Learning

Unlabeled Data Set

Hig

h Le

vel F

eatu

res

Algorithm: Self-Taught Learning

Step Two: Project target data onto the attained features by

Informally, find the activations in the attained bases such that: 1.Re-construction is minimized 2.Attained vector is sparse

Algorithms: Self-Taught Learning

Hig

h Lev

el F

eatu

res

Algorithms: Self-Taught Learning

Step Three: Learn a Classifier with the new features

Target Task

Source Task

Learn new features (Step 1)

Project target data (Step 2)Learn Model (Step 3)

Conclusions

Transfer learning is to re-use source knowledge to help a target learner

Transfer learning is not generalization

TrAdaBoot transfers instances

Self-Taught learning transfer unlabeled features

hmm