## Die Suche ergab 37 Treffer

5. Jun 2008 10:01
Forum: Statistisches Maschinelles Lernen
Thema: Ex 3 P1
Antworten: 7
Zugriffe: 1113

### Re: Ex 3 P1

I guess I answered your first question in class yesterday. w is a vector (orientation of the linear regressor, i.e. regression function). As far as I can see, there should just be a single solution. Try to first group all terms containing w, which simplify to a Gaussian. That should be unique. The r...
15. Mai 2008 12:33
Forum: Statistisches Maschinelles Lernen
Thema: plot
Antworten: 3
Zugriffe: 726

### Re: plot

I am not sure I understand what you have been trying. With plotyy and plot you are definitely not on the right track. Those plot scalar functions of 1 variable, and in case of plotyy, you can plot 2 scalar functions of 1 variable in the same coordinate system. But what you want to do is to plot a sc...
12. Mai 2008 19:05
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P3
Antworten: 7
Zugriffe: 1344

### Re: EX 2 P3

There will be additional office hours tomorrow (13.05.) at 10am in room 048 (see mailing list).

I will have my office hours tomorrow at 3pm in room 317 as usual.
12. Mai 2008 19:03
Forum: Statistisches Maschinelles Lernen
Thema: p4
Antworten: 1
Zugriffe: 676

### Re: p4

That's a somewhat complicated way of thinking about it, but if you mean inverse in the sense of an inverse function, then yes.

Undoing simply means reversing the effects of the normalization.
11. Mai 2008 20:40
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P3
Antworten: 7
Zugriffe: 1344

### Re: EX 2 P3

(1) if i first estimate the class-conditional densities, it would be more or less of the same pattern with p2, but if it is refered to the direct calculation of posteriors, i got stuck already. On page 42, the 3rd lecture, the formular is K_j/K,so k_j stands for the number of points from different ...
11. Mai 2008 20:32
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

I don't know what to recommend. The instructions are pretty explicit as for how to generate a contour plot like in figure 1. I suggest you come to the office hours of one of us. This will be easier. Nonetheless, we can of course not solve the problem for you. We can only try to clarify the tasks and...
10. Mai 2008 18:11
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

The fluctuations in the training error come from the fact that you are randomly splitting the data into test and training data. If you followed the instructions that is perfectly normal. While the training error will be the smallest for small bandwidths, the test error will have a different behavior...
10. Mai 2008 10:43
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

A(3) ok. according to the slides, page 43 if h is too large, the density estimation would be too smooth if h is too small, it won't be smooth enough. I would fix the problem by myself Exactly, that's the behavior you should see. A(4) i almost solve the contour problem. Just want to ask about one th...
9. Mai 2008 21:06
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

I forgot one thing: The other point of the assignment is to show you hands on that while it may seem great that non-parametric density estimation methods can approximate any distribution, they are not without problems. Namely, bandwidth selection is a difficult problem with these methods for which t...
9. Mai 2008 21:01
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

No, you should choose your own values. The whole point of part A is to make you experiment with the data and the method, so that you get an idea for what those parameters mean for the results. There will be no "right" answer for the bandwidth. But the general idea is that you choose one value that g...
9. Mai 2008 19:17
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

The bandwidth has no relation to the standard deviation of the feature(s). It is the standard deviation of the kernel. This means that it is an arbitrary parameter of the KDE method that controls the amount of "smoothing". The user of the KDE method has to choose it somehow. The larger h, the more t...
9. Mai 2008 18:05
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P 1
Antworten: 3
Zugriffe: 789

### Re: EX 2 P 1

Sehe das Problem nicht so ganz, wieso sollte das Plagiarismus sein, wenn du den Lösungsweg auswendig kennst. Die Lösung zu der Aufgabe ist in deinem Link doch auch nicht vorhanden. Und Quellen gibst du einfach an, fertig. Genauso ist es. Natuerlich ist es voellig in Ordnung, wenn Sie die Loesung sc...
9. Mai 2008 17:52
Forum: Statistisches Maschinelles Lernen
Thema: EX 2 P2 A
Antworten: 16
Zugriffe: 2016

### Re: EX 2 P2 A

In this case particular case, bandwidth refers to the standard deviation of the Gaussian kernel. More generally we assume that we have a kernel function k(u), which we use in the kernel density estimate as k(||x - x_i|| / h) (note that this is not the whole formula, just the relevant part). In this ...
22. Apr 2008 09:36
Forum: Statistisches Maschinelles Lernen
Thema: Ex 1. P3
Antworten: 3
Zugriffe: 938

### Re: Ex 1. P3

Hallo, man muss beweisen, dass integral von e^(-t^2) von a nach unendlich eigentlich erf(a) ist, oder ich kann die Formel direkt benutze? mit der folgender Formel erf(a)-erf(b) = erf(a-b) sieht man schon alles, danach fertig? Nein, Sie sollten die Fehlerfunktion wenn moeglich nicht benutzen, d.h. d...
17. Apr 2008 11:10
Forum: Statistisches Maschinelles Lernen