CS 540: Machine Learning Lecture 1: Introductionarnaud/cs540/lec1_CS540_handouts.pdf ·...
Transcript of CS 540: Machine Learning Lecture 1: Introductionarnaud/cs540/lec1_CS540_handouts.pdf ·...
CS 540: Machine LearningLecture 1: Introduction
AD
January 2008
AD () January 2008 1 / 41
Acknowledgments
Thanks to
Nando de Freitas
Kevin Murphy
AD () January 2008 2 / 41
Administrivia & Announcement
Webpage: http://www.cs.ubc.ca/~arnaud/cs540_2008.html
O¢ ce hours: Tuesday & Thursday 11.30-12.30.
Introductory tutorial Matlab session by Ian Mitchell TODAY, 5pm -7pm, DMP 110
Matlab resources:http://www.cs.ubc.ca/~mitchell/matlabResources.html
AD () January 2008 3 / 41
Pre-requisites
Maths: multivariate calculs, linear algebra, probability.
CS: programming skills, knowledge of data structures and algorithmshelpful but not crucial.
Prior exposure to statistics/machine learning is highly desirable butyou will be ok as long as your math is good.
I will discuss Bayesian statistics at length.
AD () January 2008 4 / 41
(Tentative) grading
The midterm will be sometimes around the end of February and isworth 30%
The �nal project will be in April ?? and is worth 40%.
There will be three (long) assignements worth 30%.
Additional optional exercises will be given.
AD () January 2008 5 / 41
Textbook
C.M. Bishop, Pattern Recognition and Machine Learning, Springer,2006.
Check the webpagehttp://research.microsoft.com/~cmbishop/PRML/
Corrections of exercises available on this webpage.
Each week, there will be required reading, optional reading andbackground reading.
Optional Textbook: Hastie, Tibshirani and Friedman, Elements ofStatistical Learning - Data Mining, Inference and Prediction,Springer-Verlag, 2001.
AD () January 2008 6 / 41
Related classes
CPSC 532L - Multiagent Systems (Leyton-Brown).
CPSC 532M - Dynamic Programming (Ian Mitchell)
Statistics courses.
AD () January 2008 7 / 41
What is Machine Learning?
�Learning denotes changes in the system that are adaptive in thesense that they enable the system to do the task or tasks drawn fromthe same population more e¢ ciently and more e¤ectively the nexttime��Herb Simon.
Machine Learning is an interdisciplinary �eld at the intersection ofStatistics, CS and EE etc.
At the beginning, Machine Learning was fairly heuristic but it has nowevolved and become -in my opinion- Applied Statistics with a CS�avour.
Aim of this course: introducing basic models, concepts andalgorithms.
AD () January 2008 8 / 41
Main Learning Tasks
Supervised learning: given a training set of N input-output pairsfxn, tng 2 X � T , construct a function f : X ! T to predict theoutput bt = f (x) associated to a new input x .- Each input xn is a p-dimensional feature vector (covariates,explanatory variables).- Each output tn is a target variables (response).
Regression corresponds to T =Rd .
Classi�cation corresponds to T = f1, ...,Kg.Aim is to produce the correct output given a new input.
AD () January 2008 9 / 41
Polynomial regression
Figure: Polynomial regression where x 2 R and t 2 R
AD () January 2008 10 / 41
Linear regression
Figure: Linear least square �tting for x 2 R2 and t 2 R. We seek the linearfunction of x that minimizes the sum of square residuals for y .
AD () January 2008 11 / 41
Piecewise linear regression
Figure: Piecewise linear regression function
AD () January 2008 12 / 41
Figure: Linear classi�er where xn 2 R2 and tn 2 f0, 1g
AD () January 2008 13 / 41
Handwritten digit recognition
Figure: Examples of handwritten digits from US postal employes
AD () January 2008 14 / 41
Figure: Can you pick out the tufas?
AD () January 2008 15 / 41
Practical examples of classi�cation
Email spam �ltering (feature vector = �bag of words�)
Webpage classi�cation
Detecting credit card fraud
Credit scoring
Face detection in images
AD () January 2008 16 / 41
Unsupervised learning: we are given training data fxng .Aim is to produce a model or build useful representations for xmodeling the distribution of the data,clustering,data association,dimensionality reduction,structure learning.
AD () January 2008 17 / 41
Hard and Soft Clustering
Figure: Data (left), hard clustering (middle), soft clustering (right)
AD () January 2008 18 / 41
Finite Mixture Models
Finite Mixture of Gaussians
f (x j θ) =k
∑i=1piN
�x ; µi , σ
2i
�
where θ =�
µi , σ2i , pi
i=1,...,k is estimated from some data (x1, ..., xn) .
How to estimate the parameters?
How to estimate the number of components k?
AD () January 2008 19 / 41
AD () January 2008 20 / 41
AD () January 2008 21 / 41
Modelling dependent data
AD () January 2008 22 / 41
AD () January 2008 23 / 41
Hidden Markov Models
In this case, we have
xn j xn�1 � f ( �j xn�1) ,yn j xn � g ( �j xn) .
For example, stochastic volatility model
xn = αxn�1 + σvn where vn � N (0, 1)
yn = β exp (xn/2)wn where wn � N (0, 1)
where the process (yn) is observed but (xn, α, σ, β) are unknown.
AD () January 2008 24 / 41
Tracking Application
AD () January 2008 25 / 41
Hierarchical Clustering
Figure: Dendogram from agglomerative hierarchical clustering with averagelinkage to the human microarray data
AD () January 2008 26 / 41
Dimensionality reduction
Figure: Simulated data in three classes, near the surface of a half-sphere
AD () January 2008 27 / 41
Figure: The best rank-two linear approximation to the half-sphere data. The rightpanel shows the projected points with coordinates given by the �rst to principalcomponents of the data
AD () January 2008 28 / 41
Manifold learning
Figure: Data lying on a low-dimensional manifold
AD () January 2008 29 / 41
Structure learning
Figure: Graphical model
AD () January 2008 30 / 41
Ockham�s razor
How can we predict the test set if it may be arbitrarily di¤erent fromthe training set?We need to make an assumption that the future will be like thepresent in statistical termsAny hypothesis that agrees with all is as good as any other, but wegenerally prefer �simpler�ones.
AD () January 2008 31 / 41
Active learning
Active learning is a principled way of integrating decision theory withtraditional statistical methods for learning models from data
In active learning, the machine can query the environment. That is, itcan ask questions.
Decision theory leads to optimal strategies for choosing when andwhat questions to ask in order to gather the best possible data.
�Good�data is often better than a lot of data
AD () January 2008 32 / 41
Active learning and surveillance
In network with thousands of cameras, which camera views should bepresented to the human operator?
AD () January 2008 33 / 41
Active learning and sensor networks
How doe we optimally choose among a subset of sensors in order toobtain the best understanding of the world while minimizing resourceexpenditure (power, bandwidth, distractions)?
AD () January 2008 34 / 41
Active learning example
AD () January 2008 35 / 41
Decision theoretic foundations of machine learning
Agents (humans, animals, machines) always have to make decisionsunder uncertainty
Uncertainty arises because we do not know the �true� state of theworld; e.g. because of limited �eld of view, because of �noise�, etc.
The optimal way to behave when faced with uncertainty is as follows
Estimate the state of the world, given the evidence you have seen upuntil now (this is called your belief state).Given your preferences over possible future states (expressed as a utilityfunction), and your beliefs about the e¤ects of your actions, pick theaction that you think will be the best.
AD () January 2008 36 / 41
Decision making under uncertainty
You should choose the action with maximum expected utility
a�t+1 = argmaxat+12A
∑xPθ (Xt+1 = x j a1:t+1, y1:t )Uφ (x)
where
Xt represents the �true� (unknown) state of the world at time ty1:t = y1, . . . , yt represents the data I have seen up until nowa1:t are the actions that I have taken up until nowU(x) is a utility function on state xθ are parameters of your world modelφ are parameters of your utility function
AD () January 2008 37 / 41
Decision making under uncertainty
We will later justify why your beliefs should be represented usingprobability theory.
It can also be shown that all your preferences can be expressed usinga scalar utility function (shocking, but true).
Machine learning is concerned with estimating models of the world(θ) given data, and using them to compute belief states (i.e.,probabilistic inference).
Preference elicitation is concerned with estimating models (φ) ofutility functions (we will not consider this issue).
AD () January 2008 38 / 41
Decision making under uncertainty: examples
Robot navigation: Xt=robot location, Yt = video image, At =direction to move, U(X ) = distance to goal.
Medical diagnosis: Xt = whether the patient has cancer, Yt =symptoms, At = perform clinical test, U(X ) = health of patient.
Financial prediction: Xt = true state of economy, Yt = observedstock prices, At = invest in stock i , U(X ) = wealth.
AD () January 2008 39 / 41
Tentative topics
Review of probability and Bayesian statistics
Linear and nonlinear regression and classi�cation
Nonparametric and model-based clustering
Finite mixture and hidden Markov models, EM algorithms.
Graphical models.
Variational methods.
MCMC and particle �lters.
PCA, Factor analysis
Boosting.
AD () January 2008 40 / 41
Project
A list of potential projects will be listed shortly.
You can propose your own idea.
I expect the �nal project to be written like a conference paper.
AD () January 2008 41 / 41