CCG Pattern Recognition Seminar

1
CCG Pattern Recognition Seminar

• Lecture 1
• Everything you ever wanted to know about Pattern
  Recognition but were afraid to ask

2
What is Pattern Recognition

• making a decision based on input
• use probability
• if deterministic, nothing to model
• examples
• automated collect calls
• face recognition system
• satellite Image analysis
• movie reviews

3
Making a Collect Call

• Use speech recognition instead of an operator
• Classes
• Yes
• No
• None of the above (NOTA)
• NOTA
• not a well define class
• How good of a particular class in an example

NOTA makes it hard
4
Explore Examples
5
Movie Reviews

• determine if a movie is good
• decision based on 6 critics reviews
• feature vector
• semantics not important

youll love this restaurant might mean you
always hate it

Predictability is importnat
6
Pattern Rec Systems
Input

What kinds of features do you have?
Sensing
Segmentation
feature extraction
classification


post processing
What part of the signal do I look at?
Cost of Failure?
7
Hard Problems

how to describe classes in terms of the features

• choosing features
• choosing model/representation
• determining classes
• building/training/testing a classifier

if not specified, requires expert domain knowledge
how do you optimally select a training test set?
8
What the heck is a training and test set?

• training set
• correctly labeled examples
• used to build models (training)
• test set
• correctly labeled examples
• not used in training
• used to validate models

X
class C
unknown
testing set
X
X
X
X
9
Example Classify Fish by Length

• draw a line separating Bass from Salmon
• all fish to left are Bass, all to right are Salmon

number of fish
length of fish
10
(No Transcript)
11