![Page 1: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/1.jpg)
Machine Learning :: Introduction
Part III
Konstantin Tretyakov ([email protected])
MTAT.03.183 – Data Mining
November 19, 2009
![Page 2: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/2.jpg)
In previous episodes
Machine learning :: Introduction :: Part III 19.11.20092
Approaches to data analysis
The general principle is the same, though:
1. Define a set of patterns of interest
2. Define a measure of goodness for the patterns
3. Find the best pattern in the data
![Page 3: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/3.jpg)
In previous episodes
19.11.2009Machine learning :: Introduction :: Part III3
Ad-hoc
Decision trees, forests
Rule induction, ILP
Fuzzy reasoning
…
Objective optimization
Regression models,
Kernel methods, SVM, RBF
Neural networks
…
Instance-based
K-NN, LOWESS
Kernel densities
SVM, RBF
…
Probabilistic models
Naïve Bayes
Graphical models
Regression models
Density estimation
…
Ensemble-learners
Arcing, Boosting,
Bagging, Dagging,
Voting, Stacking
![Page 4: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/4.jpg)
The plan
“Machine learning”
Terminology, foundations, general framework.
Supervised machine learning
Basic ideas, algorithms & toy examples.
Statistical challenges
Learning theory, bias-variance, consistency, …
State of the art techniques
SVM, kernel methods, graphical models, latent variable
models, boosting, bagging, LASSO, on-line learning, deep
learning, reinforcement learning, …
19.11.2009Machine learning :: Introduction :: Part III4
![Page 5: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/5.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
19.11.2009Machine learning :: Introduction :: Part III5
X Y Output
0 0 False
0 1 True
1 0 True
1 1 ?
![Page 6: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/6.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
Is it good or bad?
What should we do to enable learning?
19.11.2009Machine learning :: Introduction :: Part III6
![Page 7: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/7.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
Is it good or bad? Good for cryptographers, bad for data miners
What should we do to enable learning? Introduce assumptions about data (“inductive bias”):
1. How does existing data relate to the future data?
2. What is the system we are learning?
19.11.2009Machine learning :: Introduction :: Part III7
![Page 8: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/8.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
19.11.2009Machine learning :: Introduction :: Part III8
![Page 9: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/9.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
19.11.2009Machine learning :: Introduction :: Part III9
Rule 1: Generalization
will only come through
understanding of
similarity!
![Page 10: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/10.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
19.11.2009Machine learning :: Introduction :: Part III10
![Page 11: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/11.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
19.11.2009Machine learning :: Introduction :: Part III11
![Page 12: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/12.jpg)
The “No Free Lunch” Theorem
Learning purely from data is, in general, impossible
19.11.2009Machine learning :: Introduction :: Part III12
Rule 2: Data can only
partially substitute
knowledge about the
system!
![Page 13: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/13.jpg)
Statistical Learning Theory
What is learning and how to analyze it? There are various ways to answer this question. We’ll just
consider the most popular one.
19.11.2009Machine learning :: Introduction :: Part III13
![Page 14: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/14.jpg)
Statistical Learning Theory
What is learning and how to analyze it? There are various ways to answer this question. We’ll just
consider the most popular one.
Let be the distribution of data.
We observe an i.i.d. sample:
We produce a classifier:
19.11.2009Machine learning :: Introduction :: Part III14
Inductive bias
![Page 15: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/15.jpg)
Statistical Learning Theory
Distribution of data (x – coords, y – color)
19.11.2009Machine learning :: Introduction :: Part III15
![Page 16: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/16.jpg)
Statistical Learning Theory
19.11.2009Machine learning :: Introduction :: Part III16
Sample S and trained classifier g
![Page 17: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/17.jpg)
Statistical Learning Theory
19.11.2009Machine learning :: Introduction :: Part III17
Training error
![Page 18: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/18.jpg)
Statistical Learning Theory
19.11.2009Machine learning :: Introduction :: Part III18
Generalization error
![Page 19: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/19.jpg)
Statistical Learning Theory
19.11.2009Machine learning :: Introduction :: Part III19
Expected generalization error
![Page 20: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/20.jpg)
Statistical Learning Theory
Questions:
What is the generalization error of our classifier?
How to estimate it?
How to find a classifier with low generalization error?
What is the expected generalization error of our
method?
How to estimate it?
What methods have low expected generalization error?
What methods are asymptotically optimal (consistent)?
When is learning computationally tractable?
19.11.2009Machine learning :: Introduction :: Part III20
![Page 21: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/21.jpg)
Statistical Learning Theory
Some answers: for linear classifiers
Finding a linear classifier with a small training error is a
good idea
(however, finding such a classifier is NP-complete, hence
an alternative method must be used)
19.11.2009Machine learning :: Introduction :: Part III21
![Page 22: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/22.jpg)
Statistical Learning Theory
Some answers: in general
Small training error small generalization error.
But only if you search a limited space of classifiers.
The more data you have, the larger space of classifiers you
can afford.
19.11.2009Machine learning :: Introduction :: Part III22
![Page 23: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/23.jpg)
Overfitting
Why limited space?
Suppose your hypothesis space is just one classifier:
f(x) = if [x > 3] then 1 else 0
You pick first five training instances:
(1 0), (2 0), (4 1), (6 1), (-1 0)
How surprised are you? How can you interpret it?
19.11.2009Machine learning :: Introduction :: Part III23
![Page 24: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/24.jpg)
Overfitting
Why limited space?
Suppose your hypothesis space is just one classifier:
f(x) = if [x > 3] then 1 else 0
You pick first five training instances:
(1 0), (2 0), (4 1), (6 1), (-1 0)
How surprised are you? How can you interpret it?
What if you had 100000 classifiers to start with and
one of them matched? Would you be surprised?
19.11.2009Machine learning :: Introduction :: Part III24
![Page 25: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/25.jpg)
Overfitting
Why limited space?
Suppose your hypothesis space is just one classifier:
f(x) = if [x > 3] then 1 else 0
You pick first five training instances:
(1 0), (2 0), (4 1), (6 1), (-1 0)
How surprised are you? How can you interpret it?
What if you had 100000 classifiers to start with and
one of them matched? Would you be surprised?
19.11.2009Machine learning :: Introduction :: Part III25
Large hypothesis space Small training error
becomes a matter of chance Overfitting
![Page 26: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/26.jpg)
Bias-variance dilemma
So what if the data is scarce?
No free lunch
Bias-variance tradeoff:
The only way out is to introduce a strong yet
“correct” bias (or, well, to get more data).
19.11.2009Machine learning :: Introduction :: Part III26
Bias Variance
Optimal errorYour classifier
error
Optimal error in your
hypothesis class
=“Overfitting”
![Page 27: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/27.jpg)
Summary
Learning can be approached formally
Learning is feasible in many cases
But you pay with data or prior knowledge
19.11.2009Machine learning :: Introduction :: Part III27
![Page 28: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/28.jpg)
Summary
Learning can be approached formally
Learning is feasible in many cases
But you pay with data or prior knowledge
You have to be careful with complex models
Beware overfitting
If data is scarce – use simple models: they are not
optimal, but at least you can fit them from data!
19.11.2009Machine learning :: Introduction :: Part III28
Using complex models with scarce
data is like throwing data away.
![Page 29: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/29.jpg)
Next
“Machine learning”
Terminology, foundations, general framework.
Supervised machine learning
Basic ideas, algorithms & toy examples.
Statistical challenges
Learning theory, bias-variance, consistency…
State of the art techniques
SVM, kernel methods, graphical models, latent variable
models, boosting, bagging, LASSO, on-line learning, deep
learning, reinforcement learning, …
5.11.2009Machine learning :: Introduction29
![Page 30: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/30.jpg)
Linear classification
19.11.2009Machine learning :: Introduction :: Part III30
![Page 31: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/31.jpg)
Linear classification
19.11.2009Machine learning :: Introduction :: Part III31
![Page 32: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/32.jpg)
Linear classification
19.11.2009Machine learning :: Introduction :: Part III32
![Page 33: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/33.jpg)
Linear classification
Learning a linear classifier from data:
Minimizing training error
NP-complete
Minimizing sum of error squares
Suboptimal, yet can be easy and fun: e.g. the perceptron.
Maximizing the margin
Doable and well-founded by statistical learning theory
19.11.2009Machine learning :: Introduction :: Part III33
![Page 34: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/34.jpg)
The margin
19.11.2009Machine learning :: Introduction :: Part III34
![Page 35: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/35.jpg)
The margin
For any point x, its distance to the hyperplane:
Assuming all points are classified correctly:
The margin is then:
19.11.2009Machine learning :: Introduction :: Part III35
![Page 36: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/36.jpg)
Maximal margin training
Find (w, b) such that Margin is maximal, can be
shown to be equivalent with:
subject to
This is doable using efficient optimization
algorithms.
19.11.2009Machine learning :: Introduction :: Part III36
![Page 37: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/37.jpg)
Soft margin training
If data is not linearly separable:
subject to
It is called the Support Vector Machine (SVM).
It can be generalized to regression tasks.
19.11.2009Machine learning :: Introduction :: Part III37
![Page 38: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/38.jpg)
Soft margin training
In more general form:
where
is the hinge loss.
19.11.2009Machine learning :: Introduction :: Part III38
![Page 39: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/39.jpg)
Regularization
There are many algorithms which essentially
look as follows:
For given data D find a model M, which minimizes
Error(M,D)+ Complexity(M)
An SVM is a linear model, which minimizes
Hinge loss + l2-norm penalty
19.11.2009Machine learning :: Introduction :: Part III39
![Page 40: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/40.jpg)
Going nonlinear
But a linear classifier is so linear!
19.11.2009Machine learning :: Introduction :: Part III40
![Page 41: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/41.jpg)
Solution: a nonlinear map
Instead of classifying points x, we’ll classify
points in a higher-dimensional space.
19.11.2009Machine learning :: Introduction :: Part III41
![Page 42: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/42.jpg)
Kernel methods (not density!)
For nearly any linear classifier:
Trained on a dataset
The resulting vector w can be represented as:
Which means:
19.11.2009Machine learning :: Introduction :: Part III42
![Page 43: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/43.jpg)
Kernel methods (not density!)
19.11.2009Machine learning :: Introduction :: Part III43
![Page 44: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/44.jpg)
Kernel methods (not density!)
Function K is called a kernel, it measures
similarity between objects.
The computation of is unnecessary.
You can use any type of data.
Your method is nonlinear.
Any linear method can be kernelized.
Kernels can be combined.19.11.2009Machine learning :: Introduction :: Part III44
![Page 45: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/45.jpg)
Summary
SVM:
A maximal margin linear classifier.
A linear model, which minimizes
Hinge loss + l2-norm penalty
Kernelizable
Kernel methods:
An easy and elegant way of “plugging-in”
nonlinearity
different data types
19.11.2009Machine learning :: Introduction :: Part III45
![Page 46: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/46.jpg)
Summary summary
“Machine learning”
Terminology, foundations, general framework.
Supervised machine learning
Basic ideas, algorithms & toy examples.
Statistical challenges
Learning theory, bias-variance, consistency,…
State of the art techniques
SVM, kernel methods.
5.11.2009Machine learning :: Introduction46
![Page 47: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/47.jpg)
Machine learning is important
19.11.2009Machine learning :: Introduction :: Part III47
![Page 48: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/48.jpg)
Machine learning is the future
19.11.2009Machine learning :: Introduction :: Part III48
![Page 49: Machine Learning :: Introduction Part IIIIn previous episodes 2 Machine learning :: Introduction :: Part III 19.11.2009 Approaches to data analysis The general principle is the same,](https://reader033.vdocument.in/reader033/viewer/2022042812/5faf9d4335fbd30f4a11beea/html5/thumbnails/49.jpg)
Questions?
19.11.2009Machine learning :: Introduction :: Part III49