Aprendizagem Automática / Machine Learning

2010/2011

1st semester


News

This section contains news regarding the operation of the course. Please consult it regularly.
 

  • The results of the Special Epoch exam have been entered at the Administrative Office.


Presentation

The Machine Learning course is offered by DEEC, in the 1st semester of every year, to the students of the Master's programs in Electrical and Computer Engineering and in Biomedical Engineering and of the PhD / DEA program in Information Security. It is also open to students of other programs and to voluntary students.

The course's objective is to transmit the theoretical basis of machine learning, to introduce various types of learning systems, including neural networks, support vector machines, decision trees and some unsupervised learning systems, and to give lab practice on the use of several of these systems.

The course is open to foreign students. All essential course materials are in English. The classes will be given in English if there is at least one student who requires it.


Faculty

Note: The e-mail addresses given above are in a form that is not directly usable. To use them you should replace " - at - " with "@".


Syllabus

  • Concept of learning. Supervised and unsupervised learning
  • The linear learning unit (adaline). The LMS algorithm and gradient methods.
  • Multilayer perceptrons.
    • Rosenblatt's perceptron (brief reference).
    • Training criteria. Gradient-based optimization. The backpropagation algorithm.
    • Fast training methods: Momentum term and adaptive step sizes.
    • Deterministic and stochastic training. Ljung's convergence conditions.
  • Statistical aspects of supervised learning.
    • Regression problems: Estimation of conditional means, medians and percentiles.
    • Classification problems: The Bayes classifier. The Naive Bayes classifier. Estimation of posterior probabilities by means of supervised-learning systems.
  • The generalization problem.
    • Concept of generalization.
    • Stopped training with cross-validation.
    • Concept of regularization. The most common regularization functionals.
  • Support vector machines.
    • Concept of maximal-margin linear classifier. The associated optimization problem.
    • Non-separable problems. Nonlinear maximal-margin classifiers and kernel methods.
  • Decision trees.
    • Concept of decision tree.
    • Learning in decision trees. The ID3 algorithm.
    • The generalization problem in decision trees. Pruning.
  • Unsupervised learning.
    • Clustering and vector quantization. The K-Means algorithm.
    • Estimation of probability densities, and its applications. The EM algorithm for Gaussian mixture densities.
    • Principal components analysis. Concept of principal components and its usefulness. Estimation of the principal components by algebraic methods and by the Oja/Sanger rule.

Labs

Registration in the lab

The registration is made through the fenix system. The registration will open on Friday 17 September (time to be announced). At that time, the lab (room 5.15, on the 5th floor of the North Tower) will be open, so that the students can use the lab's computers for registration. A faculty member will be present to assist with any special registration needs.

To register, log in on the fenix system, open this course's page and select, in the left-hand menu, "Agrupamentos" (Groups). Then click the group that corresponds to the lab shift that you want to register in.

Students who wish to use the previous year's lab grades shouldn't register for the lab in the current semester. Otherwise their previous grades will automatically be disregarded.

The lab groups are formed by two students each. Students should arrange their own grouping and register in pairs. Students who register individually may have another student assigned to their group, and may find that their work schedules are not very compatible.


Generic information

Lab classes take place in room 5.15 (also called LSDC1), on the 5th floor of the North Tower.

The lab assignments will be published, with an advance of about a week, in the Labs Page.

You should prepare your lab work at home, by getting acquainted with the assignment and by answering the questions that don't require actual lab work.

The lab reports should be handed in at the end of the corresponding lab sessions.

In most cases the reports are made by just filling the appropriate blank spaces in the assignment. You shouldn't exceed the blank spaces that are given in the assignment. The purpose of this rule is to limit the size of the reports. Reports that exceed the space that is given will have their grades penalized.

Lab works are performed using the Matlab software. A free quasi-clone of Matlab is Octave.


Doubts Sessions

To avoid a waste of faculty time, the students who wish to attend a doubts session must send an e-mail message to the corresponding professor until 19:00 of the day before the session, and must be present at the time appointed for the session.

  • Prof. Luís B. Almeida - Thursdays at 14:00 at his office, room 9.19 of the North Tower.
    Ask the security officer, at the entrance hall of the North Tower, to announce you, to get access to the 9th floor.
     
  • Prof.ª Margarida Silveira - Tuesdays at 14:00 in room 5.15, 5th floor of the North Tower.

Grading

  • The course's grade is the average of the grades of the lab, with a weight of 50%, and of the exam, also with a weight of 50%. The grades of both the lab and the exam must be at least 9.5.
     
  • The exam has a duration of three hours. The students may take a sheet of paper, of size A4, with contents of their choice. Only non-alphanumeric calculators are allowed. Use of phones or of any other communication equipment is not allowed, not even for use as calculators or as clocks.
     
  • The students may use, in the current semester, the lab grades from the previous year (but not from earlier years). Students who wish to use the previous year's lab grades shouldn't register for the lab in the current semester. Otherwise their previous grades will automatically be disregarded.


Grades:


Bibliography

  • Basic bibliography
    The basic bibliography is available through the link above. Access requires authentication. The username and password will be given in the course lectures.
     

  • Complementary bibliography (which is not indispensable for this course):
    • Tom M. Mitchell, Machine Learning, McGraw-Hill, 1997 or later. 
    • Simon Haykin, Neural Networks: A Comprehensive Foundation, Prentice-Hall, 1999 or later.
    • Jorge Marques, Reconhecimento de Padrões: Métodos Estatísticos e Neuronais, IST press, 1998.

  • You can find here a set of slides, prepared by Prof. Fernando Silva, corresponding to a previous version of this course. These slides cover a large part of the course, although they don't exactly correspond to the present course. The slides are in Portuguese only. Access requires authentication, as for the basic bibliography.

Problems and exams

Current semester's exams

Previous semesters' exams

Problems


This page is maintained by Luís Borges de Almeida - Email: luis.almeida - at - lx.it.pt.