Transcript Machine learning and Neural Networks

DEPARTMENT of
COMPUTER
SCIENCE and
INFORMATION
TECHNOLOGIES
CEN 559
Machine Learning
2011-2012 Fall Term
Dr. Abdülhamit Subaşı
[email protected]
Office Hour: Open Door Policy
Class Schedule:Monday 17:00-19:45
Course Objectives
Present the key algorithms and theory
that form the core of machine learning.
Draw on concepts and results from many
fields, including statistics, artifical
intelligence, philosophy, information
theory, biology, cognitive science,
computational complexity, and control
theory.
Textbooks
1. Du and Swamy, Neural Networks in a Softcomputing
Framework, Springer-Verlag London Limited, 2006.
2. Sebe, Cohen, Garg and Huang, Machine Learning in
Computer Vision, Springer, 2005.
3. Chow and Cho, Neural Networks and Computing,
Imperial College Press, 2007.
4. Mitchell T., Machine Learning, McGraw Hill, 1997.
5. T. Hastie,R. Tibshirani, J. Friedman, The Elements of
Statistical Learning, Second Edition, Springer, 2008.
Brief Contents
Introduction
Concept Learning
Decision Tree Learning
Artificial Neural Networks
Evaluation Hypotheses
Bayesian Learning
Computational Learning Theory
Reinforcement Learning
Grading
Midterm Examination
25%
Research & Presentation
25%
Final Examination
50%
Minimum 15 pages word document, related PPT
and presentation
Research Topics:




















Linear Methods for Classification
Linear Regression
Logistic Regression
Linear Discriminat Analysis
Perceptron
Kernel Smoothing Methods Ref5
Kernel Density Estimation and Classification (Naive Bayes)
Mixture Models for Density Estimation and Classification
Radial Basis Function Networks - Ref1
Basis Function Networks for Classification – Ref3
Advanced Radial Basis Function Networks– Ref3
Fundamentals of Machine Learning and Softcomputing –Ref1
Neural Networks Ref5
Multilayer Perceptrons- Ref1
Hopfield Networks and Boltzmann Machines - Ref1
SVM Ref5
KNN Ref5
Competitive Learning and Clustering - Ref1
Unsupervised Learning k means Ref5
Self-organizing Maps– Ref3
Research Topics:





















Principal Component Analysis Networks (PCA, ICA)- Ref1
Fuzzy Logic and Neurofuzzy Systems - Ref1
Evolutionary Algorithms and Evolving Neural Networks (PSO) - Ref1
Discussion and Outlook (SVM, CNN, WNN) - Ref1
Decision Tree Learning Duda&Hart
Random Forest Ref5
PROBABILISTIC CLASSIFIERS-REF2
SEMI-SUPERVISED LEARNING-REF2
MAXIMUM LIKELIHOOD MINIMUM ENTROPY HMM-REF2
MARGIN DISTRIBUTION OPTIMIZATION-REF2
LEARNING THE STRUCTURE OF BAYESIAN NETWORK CLASSIFIERS-REF2
OFFICE ACTIVITY RECOGNITION-REF2
Model Assessment and Selection REF5
Cross-Validation
Bootstrap Methods
Performance ROC, statistic
WEKA Machine Learning Tool
TANGARA Machine Learning Tool
ORANGE Machine Learning Tool
NETICA Machine Learning Tool
RAPID MINER Machine Learning Tool
What is Machine Learning?
Machine learning is the process in which a machine
changes its structure, program, or data in response to
external information in such a way that its expected
future performance improves.
Learning by machines can overlap with simpler processes,
such as the addition of records to a database, but other
cases are clear examples of what is called “learning,”
such as a speech recognition program improving after
hearing samples of a person’s speech.
Components of a Learning Agent
•
Curiosity Element – problem
generator; knows what the agent
wants to achieve, takes risks (makes
problems) to learn from
• Learning Element – changes the
future actions (the performance
element) in accordance with the
results from the performance analyzer
• Performance Element – choosing
actions based on percepts
• Performance Analyzer – judges the
effectiveness of the action, passes info
to the learning element
Why is machine learning
important?
Or, why not just program a computer to know
everything it needs to know already?
Many programs or computer-controlled robots must be
prepared to deal with things that the creator would not
know about, such as game-playing programs, speech
programs, electronic “learning” pets, and robotic
explorers.
Here, they would have access to a range of unpredictable
knowledge and thus would benefit from being able to
draw conclusions independently.
Relevance to AI
•
Helps programs handle new situations based on the input and
output from old ones
• Programs designed to adapt to humans will learn how to
better interact
• Could potentially save bulky programming and attempts to
make a program “foolproof”
• Makes nearly all programs more dynamic and more powerful
while improving the efficiency of programming.
Approaches to Machine
Learning
• Boolean logic and resolution
• Evolutionary machine learning – many algorithms / neural
networks are generated to solve a problem, the best ones
survive
• Statistical learning
• Unsupervised learning – algorithm that models outputs from
the input, knows nothing about the expected results
• Supervised learning – algorithm that models outputs from the
input and expected output
• Reinforcement learning – algorithm that models outputs from
observations
Current Machine Learning
Research
Almost all types of AI are developing machine learning, since it
makes programs dynamic.
Examples:
• Facial recognition – machines learn through many trials what
objects are and aren’t faces
•
Language processing – machines learn the rules of English
through example; some AI chatterbots start with little
linguistic knowledge but can be taught almost any language
through extensive conversation with humans
Future of Machine
Learning
• Gaming – opponents will be able to learn from the player’s
strategies and adapt to combat them
•
Personalized gadgets – devices that adapt to their owner as he
changes (gets older, gets different tastes, changes his modes)
•
Exploration – machines will be able to explore environments
unsuitable for humans and quickly adapt to strange properties
Problems in Machine
Learning
•
Learning by Example:
• Noise in example classification
• Correct knowledge representation
•
Heuristic Learning
• Incomplete knowledge base
• Continuous situations in which there is no absolute answer
•
Case-based Reasoning
• Human knowledge to computer representation
Problems in Machine
Learning
• Grammar – meaning pairs
 new rules must be relearned a number of times to
gain “strength”
• Conceptual Clustering
Definitions can be very complicated
Not much predictive power
Successes in Research
• Aspects of daily life using machine learning
Optical character recognition
Handwriting recognition
Speech recognition
Automated steering
Assess credit card risk
Filter news articles
Refine information retrieval
Data mining