## Bayesian inference and Gaussian processes

author: Carl Edward Rasmussen, Max Planck Institute
published: Aug. 20, 2007,   recorded: August 2007,   views: 9242
Categories
You might be experiencing some problems with Your Video player.

# Slides

0:41 Slides Bayesian Inference and Gaussian Processes Outline Why Probabilistic Inference in Machine Learning? Interpretations of Probability Notation for Probabilities The Likelihood Function Classical Inference: Maximum Likelihood Bayesian Inference Bayesian Inference - Model Selection A Coin Toss Example: Comparing Two Learners Bernoulli Example: Observations and Predictions Bernoulli Example: Model Comparison

# Report a problem or upload files

If you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.
Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.

# Watch videos:   (click on thumbnail to launch)

Part 1 1:01:15
!NOW PLAYING

Part 2 46:27

Part 3 58:47

Part 4 38:53

Part 5 1:07:20

Part 6 29:31

Would you like to put a link to this lecture on your homepage?
Go ahead! Copy the HTML snippet !

1 Kebi_ww, January 14, 2008 at 6:22 p.m.:

The guy is actually cool. I like this talk. Thanks for uploading.

2 Reader, June 3, 2008 at 7:17 p.m.:

Slide 6.
"Notice: the likelihood function is a probability distribution over observations, not over parameters."

Likelihood function is a function (probability distribution) of a parameter, not the observations:
L(pi|D) or L(pi) is the likelihood function
p(D|pi) is the condition probability
Ref:
1. R. Hogg, A. Craig Introduction to Mathematical Statistics, 4th Ed 1978, p 202.
2. E. Lehmann, G Cassela Theory of Point Estimation (Springer Texts in Statistics) , 2nd Ed, 2003, p. 238
3. http://en.wikipedia.org/wiki/Likelihood

3 Reader, June 3, 2008 at 8:03 p.m.:

Slide 10.
Usage of the Beta distribution with alpha=beta=1 is not correct description of the Leaner B case: probabilities p(pi=0) and p(pi=1) will never be obtained. Therefore, instead of informative prior (Beta distribution with alpha=beta=1) the non-informative prior (Beta distribution with alpha=beta=0) has to be used.

4 Reader, July 28, 2009 at 4:17 a.m.:

Reader #2 is right!!! Come on, how this guy can say such a blunder!?

5 Markus, November 1, 2009 at 1:54 a.m.:

Readers #2 and #4 have a misunderstanding here: the likelihood function really takes two arguments, observed data and model parameters. It will then give you the probability (up to a proportionality constant) of the observed data given the model parameters, i.e. you obtain a probability distribution _over observations_ given the parameters. You do _not_ get a probability distribution over parameters. This is exactly what the slides say and is perfectly consistent with the references that reader #2 provides. Reader #2 conflates "function of a parameter" and "probability distribution of a parameter", which is clearly wrong here.

6 Olivier Mgbra, May 18, 2010 at 1:33 a.m.:

Reader 5 is right; definitely right; Very good video, I enjoyed it !

7 Reader, March 17, 2011 at 3:48 p.m.:

I agree with reader 5. However, coming from the non Bayesian perspective this got me confused as well. The Bayesian approach assumes a prior over the parameters and outputs a distribution over the parameters given the data. Using this a predictive distribution can be evaluated. The problem is less the optimization than solving an integral. The non Bayesian approach optimizes the model and outputs just one set of parameters.