Advanced Machine Learning

Focus: Learning with matrix parameters

Spring 2009

Project ideas

Projects

Class: MW 5-6:45, Soc.Sc.2 - 137 Office hours: Mo,We 12:30-1:30 - E2-357 Prerequisite: CMPS 242 or my consent some probability theory and linear algebra Rough syllabis Some of my previous classes

Summary of lectures1 Notes 1 Intro to Machine Learning using curve fitting as an example Overfitting, complexity control, regularization Experimental setups w. training, validation and test sets On-line versus batch Definition of regret Halfing algorithm and its bound Weighted Majority algorithm Regret bound for WM via potential function Derivation of additive and multiiplicative on-line updates for linear regression Bug Machine Hedge algorithm, Follow the Leader algorithm Hedge algorithm paper 2 Notes 2 Analysis of Hedge algorithm via potential function Expert framework with a variety of loss functions paper Maple file on how to tune Hedge Talk re. various Share Updates incl. one that induce longterm memory Long term memory paper Original "Tracking the best expert" paper Homework 1 Due We 4-13-09 at the beginning of class. 3 Long-term memory with faces In long-term memory algs old good experts "glow" for a long time Open problem: Other examples of "glowing" Talk re measuring on-lineness Talk w. more details on Disk Spindown Problem Original Disk Spindown paper Notes 3 Information theoretic of relative entropy - motivation of updates and analyses of expert algs Online updates for linear regression - and learning linear threshold functions Motivation for the GD, EG and EGU Visualizations of relatie entropies Maple file 4 Notes 4 Logistic regression Newton updates Linearly Least Squares using the SVD decomposition Newton type algs for logistic regression 5 Recall derivation of GD, EG, and EGU in the case of linear regression (Lecture 3) Notes 5 EG versus GD How to prove the regret bounds The kernel trick and its limits Leaving the span talk Leaving the span paper 6 Notes 6 Made some corrections Optimization - Lagrangians - Duality How applied to Support Vector Machines More on kernels Homework 2 Due Mo 4-27-09 at the beginning of class 7 Bregman divergences, Generalized Pythagorean Thm, Matching Loss, motivation via exponential families Matching loss design Thanks Maya! Matching loss with piece wise linear transfer function 8 Rob Schapire's NIPS 07 tutorial on Boosting web page with video links beautiful paper on Boosting - read by next class Talk focused on Game Theory connetion Entropy Regularized LPBoost 9 Approximate Newton, Newton with constraints Thanks Karen Projection onto the simples 10 Similarity between Hedge algorithm and Bayes Motivation of Bayes rule Hedge and Bayes as Follow the Perturbed Leader Notes 10 KW construction Kalai construction 11 Matrix derivative formulas Notes 11 Online PCA talk Online PCA paper 12 Hw2 solution, part 1: GD variant of Hedge Part 2 Kernelization with matrix instances Homework 3 Due Mo May 25th 5pm 13 PCA, PCA regression, PLS 14 Loss Hedge versus Loss Hedge Notes 14 Tuning of eta for Loss Hedge Tuning of eta for Gain Hedge maple pdf 15 Mo, May 18 - class cancelled 16 We, May 20 - David Helmbold teaches class from 4:30-5:30 in the Machine Learning lab, E2-489 Learning permutations talk paper Fr, May 22 - Meeting with all of you at 3pm in ML lab, E2-489 Discussion of homework and project Notes about Dai Fletcher Dai Fletcher paper 17 Mo, May 25 - Memorial day Homework 3 is due at 5pm 18 Notes 18 The optimal Hedge alg talk paper more notes 19 1-2 Ingo Steinwart gives talk Engineering 2 Building, Room 375 Project discussion at 4pm in ML lab Bring one-page summary of what you are planning to do 20 Bayesian probability calculus for density matrices talk paper Notes 20 EG+- trick for matrices

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