blind source separation in the real world
Jul. 11th, 2007 02:09 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslIn the real-life version of the cocktail party problem, there's a time lag between the sources (latent variables) and the receivers (observed variables). So we can't simply apply ICA.
I think this makes the problem harder, since we can't align the data in the two receivers... we may need to make continuity assumptions (i.e. bounded second derivative) in order to get anywhere. I'd like to have a prior over the difference between the lags (one degree of freedom), run ICA over a range of lag-differences, and choose the lag-difference for which the original noise distributions given by ICA make sense.
We have another physical constraint that could help us: the lag is directly proportional to time, which is inversely proportional to the square of the coefficient. I'm wondering if this could make the problem even more tractable than the abstract, lagless version of the problem.
I think this makes the problem harder, since we can't align the data in the two receivers... we may need to make continuity assumptions (i.e. bounded second derivative) in order to get anywhere. I'd like to have a prior over the difference between the lags (one degree of freedom), run ICA over a range of lag-differences, and choose the lag-difference for which the original noise distributions given by ICA make sense.
We have another physical constraint that could help us: the lag is directly proportional to time, which is inversely proportional to the square of the coefficient. I'm wondering if this could make the problem even more tractable than the abstract, lagless version of the problem.