Machine Learning 09/10

achoohomelessΤεχνίτη Νοημοσύνη και Ρομποτική

14 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

87 εμφανίσεις

REGENT: ALEXANDRE BERNARDI NO
Machine Learning 09/10
Motivation
Outline
￿
This Year’s Module
￿
Goals
￿
Organization
￿
Contents
￿
Bibliography
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
A Brief Introduction to Machine Learning
￿
What is Machine Learning ?
￿
What problems can be tackled by Machine Learning ?
￿
Examples.
￿
What we know and not know on Machine Learning.
￿
Human vs Machine Learning
Goals
￿
The course presents an introduction to Learning
Theory:the study and modeling of systems that can
learn from past experience and find useful
patterns
in
the
data
.
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
patterns
in
the
data
.
￿
Laboratory sessions will allow the students to
consolidate the learned concepts through computer
simulations (Matlab).
Organization
￿
Lectures :
￿
Tuesdays and Thursdays, 15:30h-17:00h, EA3
￿
Labs/Problem Sessions:
￿
Tuesdays or Thursdays, 17:00h-18:30h,
LSDC1
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
LSDC1
￿
Assessment:
￿
Final Exam (50%)
￿
Lab Work (50%) – 5 out of 6 works are
evaluated
￿
Web page
￿
http://www.isr.ist.utl.pt/~alex/aauto0910
Contents
Mathematical and statistical models of learning systems:
- Supervised and unsupervised learning.
- Optimization methods
- The problem of generalization.
- The statistical view of machine learning.
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
Computational Models:
- Artificial Neural Networks
- Support Vector Machines
- Decision Trees
- Clustering and Vector Quantization
- Principal Component Analysis
Biological Aspects:
- Biological Motivation
- The human brain
- Models of biological neurons:
-McCulloch & Pitts Neuron
-Rosenblatt’s Perceptron
-Widrow & Hoff Adaline
Bibliography
￿
Online Contents
￿
Slides Handouts
￿
Essential and Auxiliary Papers
￿
Problems, Past Exams
￿
Books
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Books
￿
[Mitchell99] Machine Learning, Tom
Mitchell, McGraw-Hill, 1999.
￿
[Haykin95] Neural Networks, Simon
Haykin, MacMillan, 1995.
￿
[Marques05] Reconhecimento de Padrões:
Métodos Estatísticos e Neuronais, Jorge
Marques, IST-Press, 2005.
A Brief Introduction to
Machine Learning
“How can we build computer systems that
automatically improve with experience, and what
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
automatically improve with experience, and what
are the fundamental laws that govern all learning
processes?”
A Machine Learning System
Observations
(External)
Experience
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
Learning System
Models
(Internal)
Input/Query Output/Decision
Why Machine Learning ?
￿
Huge Impact in Society, Economy, Health.
￿
Recent progresses in theory and algorithms.
￿
Computational power available.
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Computational power available.
￿
Growing amount of available data.
￿
Emerging Industries
￿
Study Human and Animal Learning
Problems Tackled by ML Techniques
￿
Data Mining: historical records to improve
performance.
￿
Medical Records -> Medical Diagnosis
￿
Credit Records -> Credit Risk Analysis
￿
Systems that cannot be programmed by hand.
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Systems that cannot be programmed by hand.
￿
Autonomous Driving
￿
Face Recognition
￿
Systems that have to adapt to the environment
￿
Newsreader that learns user interests
￿
Adaptive Control Systems
Example: Datamining
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Given:
￿
9714 patient records, each describing a pregnancy and birth.
￿
Each patient record contains 215 features.
￿
Learn to predict:
￿
Classes of future patients at high risk for Emergency Cesarean Section
Example: Datamining (cont.)
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
One of the 18 learned rules ->
Over training data: 26/41 = .63,
Over test data: 12/20 = .60
If
No previous normal delivery,
and
Abnormal 2nd trimester ultrasound
and
malpresentation at admission
Then
Probability of Emergency C-Section is 0.6
Example: Object Recognition
￿
Object Recognition is very hard to program problem. Learning
approaches are more easily tackled:
http://www.youtube.com/watch?v=lUUxGvDqok4
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
Example: Robot Control Learning
￿
Learning to balance a pole.
￿
Must adapt the control law if the pole height
changes.
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
Example: Robot Cognitive Learning
￿
Robot learning from experience and human
interaction.
http://www.youtube.com/watch?v=lguIxvLzm3k
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
Other Application Areas
￿
Human Computer Interfaces:
￿
Speech, Face Recognition
￿
Postal Automation:
￿
Handwritten digit recognition.
￿
Autonomous Driving:
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Learning road patterns.
￿
Bioinformatics:
￿
Learn models of gene expression.
￿
Disease Evolution:
￿
Learn disease dynamical models
￿
Medical Imaging:
￿
Detect anomalies in x-ray images, fMRi, PET-scan, TAC, etc ...
What do we know ?
￿
Excellent algorithms for pure induction
￿
SVM’s, decision trees, graphical models, neural nets, ...
￿
Algorithms for dimensionality reduction
￿
PCA, ICA, compression algorithms, ...
￿
Fundamental information theoretic bounds relate data and
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Fundamental information theoretic bounds relate data and
biases to probability of successful learning
￿
PAC learning theory, statistical estimation, grammar induction, ...
￿
Active learning by querying teacher is much more data-
efficient than random observation
￿
Algorithms to learn from delayed feedback (reinforcement)
￿
Temporal difference learning, Q learning, policy iteration, ...
￿

What we do not know?
￿
Skill transfer
￿
How can skills learned in one domain be used in other ?
￿
Co-training
￿
Can learning using multiple modalities simultaneously help the
learning process ?
￿
Never
-
ending learning
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Never
-
ending learning
￿
Continuously updating the knowledge.
￿
Learning from instruction:
￿
lectures, discussion
￿
Role of emotions:
￿
motivation, forgetting, curiosity, fear, boredom, ...
￿
Implicit (unconscious) versus explicit (deliberate)
learning
Human vs Machine Learning
￿
Can Human Learning Theories help deriving better
Machine Learning Algorithms ?
￿
Can Machine Learning Theory help understanding
better Human Learning ?
￿
Some analogies are being found between Machine
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Some analogies are being found between Machine
and Human Learning:
￿
Reinforcement (TD) Learning <-> Dopamin system in the
brain.
￿
Co-training <-> Intersensory Redundancy Hypothesis
￿
Dimensionality Reduction <-> Optimal sparse codes yield
Gabor filters, as found in the visual cortex.
A Multidisciplinary Area
￿
Artificial Intelligence
￿
Optimization
￿
Philosophy
￿
Statistics
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
Statistics
￿
Psychology and Neurobiology
￿
Control Theory
￿
Information Theory
￿
Computational Complexity Theory
￿
...
Course Planning
￿
Fundamentals of Machine Learning
￿
Basis of Optimization
￿
Neural Networks
￿
The Generalization Problem
Alexandre Bernardino, alex@isr.ist.utl.pt Machine Learning, 2009/2010
￿
The Generalization Problem
￿
Support Vector Machines
￿
Decision Trees
￿
Unsupervised Learning
￿
Probabilistic Methods
￿
Seminars