Hi,
We have decided to cover Probabilistic Graphical Models from the next session onwards. I have tried to summarize and present the bigger picture just to make everyone understand the bigger picture! :P
Two BIG Machine Learning Themes:
Any problem in ML can be approached from two different perspectives, which we can also term as two themes of Machine Learning:
-
Kernel Based Methods
- Graphical Models
We have already discussed a bit about Kernel Based methods in the previous sessions(more on Kernels maybe a little later, once we are comfortable with Convex Optimization). Its time we discuss something about Graphical Models.
Definition:
GMs provide a unifying framework for capturing complex dependencies among random variables, and building large-scale multivariate statistical models.
Why Probabilistic Graphical Models?
- Probabilistic models allow convenient way to represent knowledge as a joint probability distributions.
- Joint probability distributions support inferences that output probabilities that are useful for decision making.
- But using joint probability distribution tables does not scale. Probabilistic graphical models (PGMs) make them scalable.
- We can learn PGMs from data.
Basic approach towards understanding Graphical Models include understanding the following sub-tasks:
- Representation: Understanding how to represent data in a Graphical Model.
- Inference: Understanding how to draw conclusions/make inferences given a Graphical Model.
- Learning
- Learning model parameters
- Learning model structure
Some of the famous Graphical Models finding huge applications include:
- Markov Random Fields
- Conditional Random Fields
- Bayesian Networks (Hidden Markov Models can be considered as simplest Bayesian Networks)
We can divide these among ourselves and discuss each of them the next time we meet.
I am taking up MRFs. :)
--
Rishabh.