CMPS 290C Home Page
Advanced Machine Learning


On-line learning
Spring 2007

Manfred K. Warmuth

CLASS PROJECTS

Projects due on F, June 15th
Pls leave printouts of talk and report in box at my office
and link report and talk into the files proj/proj.html

Put both your project and talk file into directory proj

The projects of the previous class will give you an idea
about the expected size and scope of your project and presentation




Syllabis

Organisational
       Class:	TTh 4-5:45, Earth & Marine Sciences B210
Office hours:	Mo,We 11-12, E2-357
Prerequisite:	CMPS 242 - Machine Learning
		or a grad class in Bayesian Statistics
		or consent of instructor
Recommended Textbook by Nicolo Cesa-Bianchi and Gabor Lugosi
"Prediction Learning and Games"

Summary of lectures

1	Notes 1
	On-line versus batch
	Definition of regret
	Halfing algorithm and its bound
	Weighted majority algorithm
	Regret bound via potential functions
 	Bug machine

	Weighted Majority paper
	
2 	Talk re. various Share Updates incl. one that induce longterm memory
 	Long term memory paper
 	Original "Tracking the best expert" paper
 	Talk re measuring on-lineness
 	Talk w. more details on Disk Spindown Problem
 	Original Disk Spindown paper  Updated to journal version
        
 	Homework 1 Due Th April 12, beginning of class
 	Datasets 

3	Talk re building caching strategies based on the shifting expert framework
	Conference paper    Master's thesis w. more details
	Visualizations of relatie entropies    Maple file
	Motivation of relative entropy & relative loss bounds using relative entropy as measure of progress Notes 3 

4 	Combine two priority lists for caching w. Arcing Alg    
	Alternate method based on exponenential weights
  	Attempts to prove relative loss bounds 
	for the disk spindown problem Plots of hormonic weights  Maple file
	and combining heads of lists via exponential weights 
	Weighted Median algorithm for the 2-list case
	Notes 4

 	Homework 2 Due Th April 19, beginning of class  Clarified some of the problems

5 	Combining heads from k lists
	Relative loss bouds for shifting experts
	Intro to co-learning Paper
	Notes 5

6 	More on co-learning
	Lossfunctions other than the square loss paper
	How does Bayesian analysis fit into the expert framework
	Notes 6

7	Learning permutations with exponential weights
   	Intro about predicting the stock market Paper with EG algorithm
	Notes 7
 	Homework 3 Due Tu, May 1, beginning of class 

8	Experimental evaluation of stock market prediction algorithm by Ryan Weber

9	Sols to HW2 & 3
	Status of open problems
	Deriving GD, EG and Newton's update
	Implicit vs Explicit updates
	Notes 9  Made some corrections

10 	Lagrangians and duality
	How used it for proving bounds
	Notes 10 
 	Homework 4 Due Th, May 10, beginning of class 

11	Online PCA talk paper
	Notes 11 

12	Sol HW4 
	More on PCA via learning as well as the best set of experts
	Notes 12  Made some corrections
 	Homework 5 Due Tu, May 15
 	Claimed projects Add yourself

13	Relative loss bounds of GD and EG for linear regression 
	Linear regression with density matrix parameter 
        Tutorial re MRI application
        Paper on finding density matrices with LLS
	Notes 13 

14	SVD LLS, Pseudoinverse,
	Learning disjunctions
       	- inefficient alg with one expert per disjunction
	- Winnow and its bound
	Matrix version of Winnow based on SVD 
        Talk on symmetric to general matrix conversion
	Notes 14 

15	SVMs, kernels
	Leaving the span 
	Notes 15 

16	Rotation invariance
	Notes 16 
	kpca 
	Blessing and curse of multiplicative update

17	Bregman divergences, Generalized Pythagorean Thm, Matching Loss, motivation via exponential families
	Boosting talk w. emphasis on games
	Boosting talk w. emphasis on Bregman divergences 
	TotalBoost paper 

18	Multiplicative updates with kernels talk paper
	Relative loss bounds for Temporal Difference Learning 
	Multiarm Bandit Problem with multiplicative updates 

18'	Gerry Tessauro's lecture on reinforcement learning 

19 	Bayesian probability calculus for density matrices talk paper 

20	Project presentations (about 30 min each):
        Corrie/Anindya
    	Adam
	Jenniffer

Mo     	Rest of presentations in E2-489, 12pm
	Seinjuti
	Jessica
	Ricardo
	David
	Mark/Karen
	Alex

	Projects due end of finals week



---------------------------------------------------------