Scribd
Upload
40 views

Statistical Methods in Artificial Intelligence CSE471 - Monsoon 2015: Lecture 02

This document provides an overview of a course on statistical methods in artificial intelligence. It outlines the course content, which includes topics like linear classification, neural networks, Bayesian classifiers, clustering, and decision trees. It lists reference materials and the expected assessment breakdown. The remainder of the document summarizes the key points from the first lecture, including an introduction to nearest neighbor classification and the KNN algorithm. Evaluation methods like cross-validation are also introduced.

Uploaded by

srikanth.mujjiga
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
40 views

Statistical Methods in Artificial Intelligence CSE471 - Monsoon 2015: Lecture 02

This document provides an overview of a course on statistical methods in artificial intelligence. It outlines the course content, which includes topics like linear classification, neural networks, Bayesian classifiers, clustering, and decision trees. It lists reference materials and the expected assessment breakdown. The remainder of the document summarizes the key points from the first lecture, including an introduction to nearest neighbor classification and the KNN algorithm. Evaluation methods like cross-validation are also introduced.

Uploaded by

srikanth.mujjiga
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 26

Statistical Methods in Artificial Intelligence

CSE471 - Monsoon 2015 : Lecture 02

Avinash Sharma
CVIT, IIIT Hyderabad

Course Content

Introduction
Linear Classification
Neural Networks
Probability Densities
Bayesian Classifiers
Dimensionality Reduction
Support Vector Machines
Kernel Methods
Clustering Techniques
Decision Tree/Graphical Models

Reference Material
Books
Pattern Classification by Duda, Hart & Stork
The Elements of Statistical Learning by Hastie, Tibshirani and
Friedman
Machine Learning : A probabilistic Perspective by Kevin P. Murphy

Pre-requisite
Basics of Linear Algebra, Probability Theory and Statistics.
Programming in Matlab and C/C++.

Course Website

http://courses.iiit.ac.in

Online Courses/Tutorials and Research Papers

Assessment & Attendance


Policy
Assessment
27% Project
18% Assignments (Best of 6 out of 7)
30% Two Mid-semester Exams (15% each)
25% Final Exam
Zero Marks in assignment to both parties in case
a copy is detected
Cooperation with TAs is expected

Zero Tolerance for institute attendance policy

Expected Outcome
This course would enable students to
understand pattern recognition techniques in
detail.
We will ensure that both theoretical as well as
practical aspects are learnt simultaneously.
The project deliverables are expected to be
working systems attached to some practical
application.

Lecture 02: Plan


Nearest Neighbor (NN) Classifier
KNN Classifier
Classifier Evaluation
Assignment
Basic Linear Algebra Operations
Linear Discriminant Functions

Nearest Neighbor (NN)


Classifier
Two data points (samples) from the same
class should have similar features/attributes.
Similarity in feature space should be aligned
with similarity between among data points in
the real-world scenarios.
The easiest way to classify a data point in
the test data is to find a very similar data
point in the training data.

Nearest Neighbor (NN) Classifier


Eager v/s Lazy learning
Eager learning: Learn a classifier using training
data before receiving a test data sample to classify.
Lazy learning: Simply stores training data and waits
until it is given a test data sample.

NN Classifier
A lazy learning approach where each data sample is
represented as point in a Euclidean space.
Any new test data sample is assigned with the label
of closest data point in the training data using
Euclidean distance metric.

Nearest Neighbor (NN)


Classifier
Class A
Class B

Test #3

Test #2

Test #1

Nearest Neighbor (KNN) Classifier


NN rule leads to partition of the Euclidean
space into cells (Vornoi cells)

K-Nearest Neighbor (KNN) Classifier


Class A
Class B

Test #3

Test #2

Test #1

Practical Aspects
K should be chose empirically and preferably odd to
avoid tie situation.
KNN can have both discrete-value and continuousvalue target functions.
Weighted contributions from different neighbors can
be used to compute final label.
Distance based weighting can be used for giving
higher importance to closer data points.
Performance of NN classification typically degrades
when data is high-dimensional.
This can be avoided by assigning feature weights
inside Euclidean distance computation.

Effect of K on decision
boundaries
1 NN

5 NN

20 NN

KNN Classifier
Advantages:
Learn complex target functions
Training is very fast
Zero loss of information

Disadvantages:
Classification cost for new instances can be
very high
Major computation takes place at
classification time

Classifier Evaluation
Cross Validation is an important means of
evaluating classifiers. Types of cross
validation techniques are:
Random Subsampling
K-fold Cross-Validation
Leave-one-out Cross-validation

Assignment
Iris data or other standard dataset
KNN classifier
Matlab/C++ pipeline
K-fold cross-validation
A 2-3 page report on setup & experimental
evaluation

Basic Linear Algebra


Operations
Vector
Vector Operations
Scaling
Transpose
Addition
Subtraction
Dot Product

Equation of a Plane

Vector
Operations

Transpose

Scaling: Only Magnitude Changes

Vector
Operations

Vector
Operations

Dot Product (Inner Product) of two vectors is a scalar.

Dot product if two perpendicular vectors is 0

Equation of a Plane

Linear Discriminant
Functions
Assumes
a 2-class classification setup

Decision boundary is represented explicitly in


terms of components of .
Aim is to seek parameters of a linear discriminant
function which minimize the training error.
Why Linear ?
Simplest possible
Generalized

Linear Discriminant Functions &


Decision Surfaces

Class A
Class B

The perceptron

Perceptron Decision
Boundary

Perceptron Summary

Decision boundary surface (hyperplane)


divides feature space into two regions
Orientation of the surface is decided by the
normal vector
Location of the surface is determined by the
bias term

is proportional to distance of
from

the surface

positive side,
negative

side

You might also like