DF22400
Machine Learning (3p)
Introduction and Application for Automated Performance Tuning

HT/2012

Recommended for
Graduate (CUGS, CIS, ...) students interested in the application of machine learning techniques to advanced system performance optimization, as in compiler construction, library generation, runtime systems, parallel computing, software engineering, system simulation and optimization.
Interested undergraduate students are also welcome, as long as there are free seats left.

Organization
Lectures (ca. 15h), optional theoretical exercises for self-assessment, student projects and/or presentations. Written exam.

The course was last given
This is a new course.

Goals
The course introduces fundamental techniques of machine learning and considers case studies for its application in automated system performance tuning, such as auto-tuning library generators, compilers, and runtime systems.

Prerequisites
Linear algebra. Discrete mathematics. Data structures and algorithms. Some basic knowledge of computer architecture is assumed. For the case study presentations, some background in at least one application area, such as compiler construction, library generation, signal processing software, runtime systems, or software composition, is required.

Contents
"[Machine] learning is the process of [automatically] constructing, from training data, a fast and/or compact surrogate function that heuristically solves a decision, prediction or classification problem for which only expensive or no algorithmic solutions are known. It automatically abstracts from sample data to a total decision function."
- [Danylenko et al., Comparing Machine Learning Approaches..., SC'2011, LNCS 6708]

Schedule
Lecture block (2.5 days, room Donald Knuth), presentation session (0.5 days, room Donald Knuth) and exam (0.5 days, room von Neumann) in Linköping.

Lecture notes

Exercises collection

DayTimeRoom LecturerTopic
Monday
22/10
2012
13:15-17:00Donald
Knuth
C. Kessler Organization and overview
Supervised Learning
Decision Tree Learning
Tuesday
23/10
2012
09:15-16:00Donald
Knuth
C. Kessler Neural Networks: Biological Neural Networks; McCulloch-Pitts units; Perceptron; Competitive Learning; Feed-Forward Networks; Backpropagation Algorithm
Application Area: Automated Performance Tuning
Adaptive sampling and decision tree learning for optimized composition
Demo: C4.5 (Lu Li)
Selection of papers or projects for student presentations
Friday
26/10
2012
10:15-16:00Donald
Knuth
W. Löwe Decision diagrams
Bayesian classifiers
SVM
Demo rapid miner
Monday
26/11
2012
13:00-17:00Donald
Knuth
Student presentations
Tuesday
4/12
2012
13:00 v.Neumann Written exam
Remark: Breaks are not shown in this schedule, and will be set as appropriate.

Literature

If you prefer working with a textbook, you may e.g. use the ones mentioned below as additional reading. Note that most books cover much more than what we can go through in our introductory course.

As a good introductory textbook we recommend

Note that a part of our lecture material will be based on this book.

Other useful books: (e.g. for further reading)

General articles: Articles on decision trees: Articles on decision graphs (BDDs): Articles on SVM: Articles on Bayesian Classifiers:

Further literature references will be added later.

Teachers

Examination

TEN1: Written exam (Welf, Christoph) 1.5p.

UPG1: Small project with presentation (possibly done in groups of 2), or (individual) presentation of a research paper, 1.5p

Credit
3p if both examination moments are fulfilled. Admission to the exam requires attendance in 50% of the lectures and lessons.

Comments
New course 2012.

Related Courses
Machine Learning course at Control Engineering, ISY, Linköping University, given every other year. More mathematically than algorithmically oriented, and more details compared to our more light-weight course.
... list to be continued ...


This page is maintained by Christoph Kessler (chrke \at ida.liu.se)