SS ’15: Machine Learning 2

Lagrange decompositionThe course consists of lectures and practical exercises and concentrates on optimization techniques for machine learning tasks, such as inference and learning for graphical models. The course contains 3 parts: first, we briefly review basic convex optimization techniques widely used in machine learning. The second part  is devoted to inference problem for graphical models: we consider a number of existing algorithms, show their connection and their analysis from optimization point of view. In the final third partwe consider the parameter's learning of graphical models, both probabilistic and discriminative.


News!
- 12.06: In place of the Exercises 15.06, there will be a "Questions & Answers" session for our course. Please prepare your questions (any type, to the theory or exercises), it is a good opportunity to get answers to them.
- 9.06: On 10.06 the lecture will be given by D.Schlesinger.
- 5.06: There will be NO exercises on Monday, 8.06.
- 4.05: There will be no lecture on Wednesday,  May  6, because of Dies academicus.
- 24.04: Exersice sheet 1 is extended to cover the class on 27.04.2015. Table to enter your results added. Installation script updated as well.
- Software installation readme
and PDF-script of the lecture added to the page. Several typos in slides and script from Lecture 1 fixed.
- Lecture time changed:
it will be held on Wed 11:10 instead of 7:30!


Lectures
: Wednesday, 11:10 - 12:40, Room E001, Start: April 15, 2015
Exercises: Monday 9:20-10:50, Room E069, Start:  April 20
Prerequisites: Machine Learning 1, good knowledge of math (linear algebra, optimization), programming (python/C++)
Credits: 2/2/0, oral exam. Enrolment: jExam . Attendees: max. 60
Note: Lectures are held in English.
Script: PDF-Dowload. Script and slides are protected with a password, which reads E001
Software installation readme: Download


Content:


LectureSeminar/Exercises
1 Introduction, motivation, relation to other courses. Graphical models. [slides]
none
2 Inference in graphical models as integer linear program. Linear programming relaxation.[script, Sec. 1] (OpenGM)Introduction to OpenGM inference library, CPLEX as solver: solver parameters, scalability. [exercise][supplement][Table-for-results]

3Convex Optimization. Lagrange duality, complementary slackness.[script, Sec. 1-2; slidesbook-Sec.5]
4Dies academicus (Blackboard) Lagrange duality for linear programs.[exercise][Table-for-results]
5 Dual of LP relaxation. Reparametrization.[slides] (Blackboard) Lagrange duality for quadratic programs. [exercise][Table-for-results]
6 Tree-structured graphical models. Dynamic programming.[slides] (OpenGM)  Dual Of LP Relaxation. Reparametrization. [exercise][supplement][Table-for-results]
7Optimality conditions for dul MAP LP [slides]
 
(Blackboard) Exercises to dynamic programming.[exercise]
8 Relaxation labeling and diffusion algorithms.[slides] none
9 Convex Optimization. First order smooth and non-smooth optimization (gradient, coordinate descent, sub-gradient method) and their application for inference.[script, Sec. 4][slides] Questions and Answers session
10Lagrangian (dual) decomposition. Lagrange decomposition for inference with graphical models.[slides] (OpenGM) Coordinate descent algorithms [exercise][Table-for-results]
11 Sub-modularity. Energy minimization as a min-cut problem.[script, Sec. 6] (OpenGM) Tree decomposition - use of different algorithms.[exercise][supplement][Table-for-results]
12 Graph cut based algorithms. [slides](Whiteboard) Dynamic programming, arc consistency, tree agreement - Exercises.
13 Binary LP relaxation as min-cut. Partial optimality.[script, Sec. 7] (OpenGM) Graph cut based inference.[exercise][Table-for-results]
14 Outlook. What we have learned and advanced topics. [slides](Whiteboard) Binary LP relaxation as min-cut. Partial optimality.



See also the teaching web-page of the Computer Vision Lab Dresden.


image links: crops, lake, tsukubaR, tsukubaL, tsukubaTrueDisp