QuoteOriginally posted by: chiral3So lets say that you go our a few steps and the errors are small - you get a good fit. Go out farther and you start to get some very positive error terms. So the update makes the derivative a bit more positive in that region. OK, its feeding of the error terms, but it wouldn't be called a neural net on it's own.smells like a training a neural net chiral.

I'll concede. If you automate the process, and call the error terms feedback, it is neural net-like. I don't know too much about nets. I used one once in an experiment, I read about Kurtzweil, but they're just like so 90's

QuoteOriginally posted by: chiral3I'll concede. If you automate the process, and call the error terms feedback, it is neural net-like. I don't know too much about nets. I used one once in an experiment, I read about Kurtzweil, but they're just like so 90's I don't either other than this...you got a bunch of nodes with connections. the connections have weights, there are input nodes, output nodes, and internal nodes. connections can have loops. you feed input, compare to output, and adjust the weights till you settle down to an answer. this is "learning". there is a gradient decent method for adjusting weights that boils down to regression in the case of really really simple nets with no internal nodes.

there is a bookKalman Filtering and Neural Networks (Wiley)edited by Simon Haykinbuy it

In a nutshell, a KF is an MLE that makes smart bombs smarter.

QuoteOriginally posted by: rezathere is a bookKalman Filtering and Neural Networks (Wiley)edited by Simon Haykinbuy itDear Evil One,One of my problems is that I always re-prove things...."Everything has been thought of before" -smart Polish colleague of Nonius.Regards,Pedro.

>> In a nutshell, a KF is an MLE that makes smart bombs smarter. for real?

for real? Absolutely for real. KF is a good example of dual use technology.

QuoteIn a nutshell, a KF is an MLE that makes smart bombs smarter. For correcting errors in GPS positioning, but I don't think that is the "nutshell" textbook definition

For correcting errors in GPS positioning, but I don't think that is the "nutshell" textbook definition Not quite. Those babies move fast, so prediction of trajectory is the "nutshell". GPS uses particle filters, no?

On a completely unrelated topic: I recently bought this book Spiking Neuron Models (for $10 at strand). I wanted to work through it to get e feel.

I am only aware of KF for missile navigation. I don't know how robust they are compared to using lasers to paint a target, or how successful they are. I think that it is literally Hamiltonian mechanics with corrections. I don't know though.

Kalman Filtering in a nutshell gives you an optimal estimate with respect to least square error in the presence of white noise. It is a feedback type of mechanism but the similarity to a neural network ends there.More complexity follows obviously. Neural nets allow you to sort of come up with an estimator without assuming a model (e.g. Guassian.) Neural nets are not optimal with respect to anything and it is difficult to characterize the error in estimates. There is some formal mathematical work done with support vector machines of which radial basis function neural nets are a subset of so perhaps the estimates are put more in a context.

Reza,QuoteDidn't these guys ever hear of the Yule-Walker equation?? I know I haven't any references plaese? articles ...Hamilton, Time Series Analysis p 59Ahh. I understand, you weren't one of the janitors at MIT.

Quotepresence of white noiseI don't think that it has anything to do with white noise. If the word "Gaussian" is creeping in, then fine. However, QuoteKalman Filtering in a nutshell gives you an optimal estimate with respect to least square error makes the most sense as a nutshell description