[MUSIK]

so good morning everybody we will continue today to look at different norms

and different optimization problems for pattern for solving pattern recognition related problems

but before we do so let let me briefly summarize the big picture the big picture so

in winter semester we talk about pattern recognition and we

have seen at the beginning of this lecture that we

are basically studying one mapping that maps feature vectors to

class numbers so we want to have a decision rule that takes a feature

vector of fixed dimension D and maps it to a certain class and

there are different ways to set up classifiers we have seen already different approaches and

we have learned about the so called optimal classifier with respect to

the 01 cost function and that's the Bayes classifier the Bayesian

classifier that basically decides for the

class with the maximum posterior probability P

of Y given X and in this context we

have seen that there are different ways to define this posterior

probability so posterior PDF modeling was

one topic and the classifier before

before we talk about the posteriors one comment on

the Bayesian classifier is optimal with respect to what

with respect to the 0 1 loss function right

the 0 1 loss function the 0 1

loss function so 0 1 loss what does it mean it means we

do the right decision it's for free and if we do the wrong

decision we have to pay 1 euro 1 dollar 1 I don't know

what what is you currency rupee so doesn't

matter it's currency independent so posterior PDF

modeling we have seen the generic and

the generic model and the discriminative model and

sorry about my poor hand writing you can live

with it the generic model and the discriminative model

if I talk about a Gaussian classifier does it

fall into the class of discriminative or genetic models

generic models what is Gaussian what is in a

Gaussian classifier the class conditional PDF

and the discriminative modeling approach we have seen which one

was that Gaussian the discriminative model we have

seen Florian you remember that no that was the

generic the discriminative was the direct modeling of

the posterior and we have seen one way to do so the sigmoid function

logistic function right logistic regression logistic function

and we also have seen that we

can associate with an arbitrary decision boundary F

of X is zero in the two dimensional case for instance

we can write down right away the posterior probability by

using this F of X which is basically 1 over

1 plus Y times plus F of X yes sir

ah sorry generative generic generative we

have called it I am so confusing

I'm sorry generative generic I did not write