Lecture Series on Machine Learning

Ravi Gupta and G. Bharadwaja KumarAU-KBC Research Centre,

MIT Campus, Anna University

Books

Machine Learning by Tom M. Mitchell

Lecture No. 1

Ravi GuptaAU-KBC Research Centre,

MIT Campus, Anna University

Date: 5.03.2008

What is Machine Learning ???

Machine Learning (ML) is concerned with the question of how to construct computer programs that automatically improves with experience.

When Machine Learning ???

• Computers applied for solving a complex problem

• No known method for computing output is present

• When computation is expensive

When Machine Learning ???

• Classification of protein types based amino acid sequences

• Screening of credit applicants (who will default and who will not)

• Modeling a complex chemical reaction, when the precise interaction of the different reactants is unknown

What Machine Learning does???

Decision Function/ Hypothesis

What Machine Learning does???

Training Examples

Decision Function/ Hypothesis

Supervised Learning

Apple

Decision Function/ Hypothesis

Orange

Supervised Classification

Unsupervised Learning

Decision Function/ Hypothesis

Unsupervised Classification

Binary Classification

Decision Function/ Hypothesis

Multiclass Classification

Decision Function/ Hypothesis

Application of Machine Learning• Speech and Hand Writing Recognition)• Robotics (Robot locomotion)• Search Engines (Information Retrieval)• Learning to Classify new astronomical structures• Medical Diagnosis• Learning to drive an autonomous vehicle• Computational Biology/Bioinformatics• Computer Vision (Object Detection algorithms)• Detecting credit card fraud• Stock Market analysis• Game playing• ………………..• ………………….

Search Engines

Sky Image Cataloging and Analysis Tool (SKI CAT) to Classify new astronomical structures

ALVINN (Autonomous Land Vehicle In a Neural Network)

SPHINX - A speech recognizer

Disciplines that contribute to development ML Field

Definition (Learning)

A computer program is said to learn from experience E with respect to some class of tasksT and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. (Tom Mitchell, 1998)

Checker Learning Problem

A computer program that learns to play checkers might improve its performance as measured by its ability to win at the class of tasks involving playing checkers games, through experience obtained by playing games against itself

• Task T : playing checkers• Performance measure P: % of game won against opponents• Training experience E : playing practice game against itself

A Handwritten Recognition Problem

• Task T : recognizing and classifying handwritten words within images

• Performance measure P: % of words correctly classified

• Training experience E : a database of handwritten words with given classifications

A Robot Driving Learning Problem

• Task T : driving on public four-lane highways using vision sensors

Performance measure P: average distance traveled before an error (as judged by human overseer)

• Training experience E : a sequence of images and steering commands recorded while observing a human driver

Designing a Learning System

Checker Learning Problem

• Task T : playing checkers

• Performance measure P: % of game won against opponents

• Training experience E : playing practice game against itself

Checker

Choosing the Training Experience

The type of training experience E available to a system can have significant impact on success or failure of the learning system.

Choosing the Training Experience

One key attribute is whether the training experience provides direct or indirect feedback regarding the choices made by the performance system

Direct Learning

Indirect Learning

Win / LossNext Move

Second attribute is the degree to which the learner controls the sequence of training examples.

Next Move

Another attribute is the degree to which the learner controls the sequence of training examples.

Player Next Move

Teacher / Expert

Player Next Move

How well the training experience represents the distribution of examples over which the final system performance P must be measured.(Learning will be most reliable when the training example follow a distribution similar to that of future test examples)

V. Anand

G. Kasparov

Garry Kasparov playing chess with IBM's Deep Blue. Photo courtesy of IBM.

Kasparov Monkey

Geri’s Game by Pixar in which an elderly man enjoys a game of chess with himself (available at http://www.pixar.com/shorts/gg/theater/index.html)

Won an Academy Award for best animated short film.

Assumptions

• Let us assume that our system will train by playing games against itself.

• And it is allowed to generate as much training data as time permits.

Issues Related to Experience

What type of knowledge/experience should one learn??

How to represent the experience ? (some mathematical representation for experience)

What should be the learning mechanism???

Main objective of the checker playing

Best Next Move

Target Function

Choose Move: B → M

Choose Move is a function

where input B is the set of legal board statesand produces M which is the set of legal moves

Output: M (Moves)

Input: Board State

Target Function

Choose Move: B → M

M = Choose Move (B)

Alternative Target Function

F : B → R

F is the target function, Input B is the board state and Output R denotes a set of real number

Output: M (Moves)

Input: Board State

F = 10

F = 10.8

F = 14.3F = 7.7 F = 10

F = 6.3 F = 9.4

Representation of Target Function

• xl: the number of white pieces on the board• x2: the number of red pieces on the board• x3: the number of white kings on the board• x4: the number of red kings on the board• x5: the number of white pieces threatened by red (i.e.,

which can be captured on red's next turn)• X6: the number of red pieces threatened by white

F'(b) = w0 + w1x1+ w2x2 + w3x3 + w4x4 + w5x5 + w6x6

• Task T: playing checkers• Performance measure P: % of games won in the world

tournament• Training experience E: games played against itself• Target function: F : Board → R• Target function representation

F'(b) = w0 + w1x1+ w2x2 + w3x3 + w4x4 + w5x5 + w6x6

The problem of learning a checkers strategy reduces to theproblem of learning values for the coefficients w0 through w6 inthe target function representation

Adjustment of Weights

( )

2train

,

( ) (F ( ) '( ))train bb F training examples

E Error b F b< >∈

≡ −∑

Generic issues of Machine Learning

• What algorithms exist for learning general target functions fromspecific training examples?

• In what settings will particular algorithms converge to the desired function, given sufficient training data?

• Which algorithms perform best for which types of problems and representations?

• How much training data is sufficient?

• What is the best way to reduce the learning task to one or more function approximation problems?

Generic issues of Machine Learning

• When and how can prior knowledge held by the learner guide the process of generalizing from examples?

• Can prior knowledge be helpful even when it is only approximately correct?

• What is the best strategy for choosing a useful next training experience, and how does the choice of this strategy alter the complexity of the learning problem?

• How can the learner automatically alter its representation to improve its ability to represent and learn the target function?

Concept Learning

Concept Learning

A task of acquiring a potential hypothesis (solution) thatbest fits the training examples

Concept Learning Task

Objective is to learn EnjoySport

{Sky, AirTemp, Humidity, Wind, Water, Forecast} → EnjoySport

Tom enjoys his favorite water sports

Concept Learning Task

Objective is to learn EnjoySport

{Sky, AirTemp, Humidity, Wind, Water, Forecast} → EnjoySport

Output

<x1, x2, x3, x4, x5, x6> → <y>

Input variables

Notations

NotationsInstances (X) Target Concept (C)

Training examples (D)

Concept Learning

Acquiring the definition of a general category from a given set ofpositive and negative training examples of the category.

A hypothesis h in H such that h(x) = c(x) for all x in X

instance

Target concept

Instancehypothesis

Instance Space

Suppose that the target hypothesis that we want to learn for the current problem is represented as a conjunction of all the attributes

Sky = .... AND AirTemp = …. AND Humidity= …. AND Wind = …. ANDWater= …. AND Forecast= …. THEN EnjoySport= ….

Instance Space

Possible distinct instances = 3 * 2 * 2 * 2 * 2 * 2 = 96

Suppose the attribute Sky has three possible values, and that AirTemp, Humidity, Wind, Water, and Forecast each have two possible values

InstanceSpace

Hypothesis Space

Hypothesis Space: A set of all possible hypotheses

Possible syntactically distinct Hypotheses for EnjoySport

= 5 * 4 * 4 * 4 * 4 * 4 = 5120

• Sky has three possible values• Fourth value don’t care (?)• Fifth value is empty set Ø

Hypothesis Space

h = <Sunny, Warm, ?, Strong, Warm, Same>

Normal / High

Hypothesis Space

h = <Ø, Warm, ?, Strong, Warm, Same>

Concept Learning as Search

Concept learning can be viewed as the task of searching through a large space of hypothesis implicitly defined by the hypothesisrepresentation.

The goal of the concept learning search is to find the hypothesis that best fits the training examples.

General-to-Specific Learning

Most General Hypothesis: h = <?, ?, ?, ?, ?, ?>

Every day Tom his enjoy i.e., Only

positive examples.

Most Specific Hypothesis: h = < Ø, Ø, Ø, Ø, Ø, Ø>

General-to-Specific Learning

Consider the sets of instances that are classified positive by hl and by h2. h2 imposes fewer constraints on the instance, thus it classifies more instances as positive.

Any instance classified positive by hl will also be classified positive by h2. Therefore, we say that h2 is more general than hl.

Definition

Given hypotheses hj and hk, hj is more_general_than_or_equal_tohk if and only if any instance that satisfies hk also satisfies hj.

We can also say that hj is more_specific_than hk when hk is more_general_than hj.

More_general_than relation

FIND-S: Finding a Maximally Specific Hypothesis

Step 1: FIND-S

h0 = <Ø, Ø, Ø, Ø, Ø, Ø>

Step 2: FIND-S

h0 = <Ø, Ø, Ø, Ø, Ø, Ø>

x1 = <Sunny, Warm, Normal, Strong, Warm, Same>

a1 a2 a3 a4 a5 a6

Iteration 1

h1 = <Sunny, Warm, Normal, Strong, Warm, Same>

x2 = <Sunny, Warm, High, Strong, Warm, Same>

h2 = <Sunny, Warm, ?, Strong, Warm, Same>

h1 = <Sunny, Warm, Normal, Strong, Warm, Same>

Iteration 2

Iteration 3 Ignore h3 = <Sunny, Warm, ?, Strong, Warm, Same>

x4 = < Sunny, Warm, High, Strong, Cool, Change >

h4 = <Sunny, Warm, ?, Strong, ?, ?>

h3 = < Sunny, Warm, ?, Strong, Warm, Same >

Iteration 4

Step 3

Output

Key Properties of FIND-S Algorithm

• Guaranteed to output the most specific hypothesis within H that is consistent with the positive training examples.

• Final hypothesis will also be consistent with the negative examples provided the correct target concept is contained in H, and provided the training examples are correct.

Unanswered Questions by FIND-S

• Has the learner converged to the correct target concept? Although FIND-S will find a hypothesis consistent with the training data, it has no way to determine whether it has found the only hypothesis in H consistent with the data (i.e., the correct target concept

• Why prefer the most specific hypothesis?In case there are multiple hypotheses consistent with the training examples, FIND-S will find the most specific. It is unclear whether we should prefer this hypothesis over, say, the most general, or some other hypothesis of intermediate generality.

Unanswered Questions by FIND-S

• Are the training examples consistent?

In most practical learning problems there is some chance that the training examples will contain at least some errors or noise. Such inconsistent sets of training examples can severely mislead FIND-S, given the fact that it ignores negative examples. We would prefer an algorithm that could at least detect when the training data is inconsistent and, preferably, accommodate such errors.

Unanswered Questions by FIND-S

• What if there are several maximally specific consistent hypotheses?

In the hypothesis language H for the EnjoySport task, there is always a unique, most specific hypothesis consistent with any set of positive examples. However, for other hypothesis spaces (which will be discussed later) there can be several maximally specific hypotheses consistent with the data.