going Bayesian

[gusl]

I'm now trying to show that ICA is totally useless for the Gaussian case. The idea being that ICA does not test for conditional independence.

Apparently, Peter has been imagining otherwise, since in the Gaussian case, a regression coefficient being close to 0 implies conditional independence.

I really pushed for my idea of Bayesianizing ICA, in order to deal with the more general case in which the structure of the problem is a general multi-level DAG, instead of the Cocktail Party Problem (CPP). He finally agreed that I could be right. To make sure, I'm now going to evaluate ShimizuSearch against a new method I'll name StupidSearch, which just spits out a random graph (OR an empty graph OR a random full DAG).

However, I'm still intending to follow his suggestion of studying multiple-indicator graphs and whether it's possible to get a unique maximum-likelihood latent values for each data point; and to read economics literature that uses time series.

Some ideas of where to look for linear systems:
* electrical circuits (we could apply this to debugging of circuits)
* traffic flow at large temporal scale (e.g. > 1 day)
* water flow in rivers and streams
* economics (no reason to be linear, but no-one will complain, since they've always been making the linearity assumption)

Comments

[altamira16.livejournal.com]

I was just at a lecture today, and some things that I was only vaguely paying attention to were discussed. One point of interest was that Kalman filters are used when dealing with linear Gaussian systems. Particle filters are used when you can't assume the system is Gaussian. (I am not sure about linearity.

ICA

[gustavolacerda.livejournal.com]

There was a comment asking what ICA was, which apparently got deleted. I imagine you've found the wikipedia article? You can see my poster, which also has references to some good papers. Thanks for the pointer on Kalman filters. It sounds like something I ought to learn.

Re: ICA

[altamira16.livejournal.com]

I went back and looked at some of your earlier posts to see if I had somehow missed something you had written about ICA. I had!

[gustavolacerda.livejournal.com]

what lecture was this?

[altamira16.livejournal.com]

My boss is giving evening lectures at work on topics relevant to anti-submarine warfare.