Artificial Intelligence

Lecturer 13 – Machine Learning

School of Information and Communication

1
Introduction of Machine learning
• Definitions of Machine learning…
→ A process by which a system improves its performance [Simon, 1983]
→ Any computer program that improves its performance at some task through
experience [Mitchell, 1997]
→ Programming computers to optimize a performance criterion using example
data or past experience [Alpaydin, 2004]

• Representation of the learning problem [Mitchell, 1997]

2
 Application examples of ML (1)
Web pages filtering problem Web pages categorization problem
• T: to predict which Web pages a given • T: to categorize Web pages in predefined
user is interested in categories
• P: % of Web pages correctly predicted • P: % of Web pages correctly categorized
• E: a set of Web pages identified as • E: a set of Web pages with specified
interested/uninterested for the user categories

Which
cat.?
Interested?

3
Application examples of ML (2)
Handwriting recognition problem Robot driving problem
• T: to recognize and classify • T: to drive on public highways using
handwritten words within images vision sensors
• P: % of words correctly classified • P: average distance traveled before an
• E: a database of handwritten words error (as judged by human overseer)
with given classifications (i.e., labels) • E: a sequence of images and steering
commands recorded while observing a
human driver

Which word? Which steering

we do in the right way Go Move Move Slow Speed

4
Key elements of a ML problem (1)
• Selection of the training examples
• Direct or indirect training feedback
• With teacher (i.e., with labels) or without
• The training examples set should be representative of the future test examples

• Choosing the target function (a.k.a. hypothesis, concept, etc.)

5
Key elements of a ML problem (2)
• Choosing a representation of the target function
• A polynomial function
• A set of rules
• A decision tree
• A neural network
• …

• Choosing a learning algorithm that learns (approximately) the

6
Issues in Machine Learning (1)
• Learning algorithm
• Which algorithms can approximate the target function?
• Under which conditions does a selected algorithm converge (approximately) to
the target function?
• For a certain problem domain and given a representation of examples which
algorithm performs best?

• Training examples
• How many training examples are sufficient?
• How does the size of the training set influence the accuracy of the learned target
function?
• How does noise and/or missing-value data influence the accuracy?

7
Issues in Machine Learning (2)
• Learning process
• What is the best strategy for selecting a next training example? How do
selection strategies alter the complexity of the learning problem?
• How can prior knowledge (held by the system) help?

• Learning capability
• What target function should the system learn?
Representation of the target function: expressiveness vs. complexity
• What are the theoretical limits of learnability?
• How can the system generalize from the training examples?
To avoid the overfitting problem
• How can the system automatically alter its representation?
To improve its ability to represent and learn the target function

8
Types of learning problems
• A rough (and somewhat outdated) classication of
learning problems:
• Supervised learning, where we get a set of training inputs
and outputs
• classication, regression
• Unsupervised learning, where we are interested in capturing
inherent organization in the data
• clustering, density estimation
• Reinforcement learning, where we only get feedback in the
form of how well we are doing (not what we should be
doing)
• Planning