Oct 27
projects
Discuss projects & share what you did.
continue discussion of Bayesian networks
Last week we started working through chap 14, on Bayesian networks.
Big picture :
This the starting point point in the text for doing "maching learning" -
it's the first "model" that we might use as a framework for getting a computer to "learn",
for example the naive bayes spam filter text classification algorithm.
1. the big ideas
- N (may be large) random variables to represent a system of interest
- sparse network of "dependencies" between variables - conditional probabilities , "causal" connections
- (Note: this is similar to the neural nets we'll do later ...)
- often "train" with some things given, but "ask" for other things
- depending on which variables are given "known" values
- The probability theory follows from Baye's law
- ... but we can also use numerical sampling to get the job done.
- several appproaches ...
- Gibbs sampling is one we'll look at (powerful, simple to implement)
- details: "burn in period", "nth sample to avoid correlations"
- software libraries exist (e.g. PyMC) to do run this stuff ...
2. practice AIMA problems (we did about half of 4.12 in class last week)
- 14.11 nuclear power plan
- 14.12 telescopes
- 14.18 rain/sun monte-carlo sampling
3. write some code :
- In class, write a python program to implement Gibbs sampling for one of the toy problems in the text or wikipedia examples.
