SERVING THE QUANTITATIVE FINANCE COMMUNITY

• 1
• 2

Cuchulainn
Posts: 59458
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

### Re: Machine Learning and the physical sciences

maestro,

http://www.physics.ox.ac.uk/phystat05/p ... stat05.pdf

Initial results are promising .. 10 hours for 10,000 samples.

Do you have an opinion yourself? Enlighten us. What's new, apart from the cute name?

I bet you won't give a technical answer.

katastrofa
Posts: 7971
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

### Re: Machine Learning and the physical sciences

Small dogs bark the loudest. No wonder everyone ignores you. Why don't you say something righteous and hopeful for a change?
Old crazy dogs try to bite the air. With no teeth.

katastrofa
Posts: 7971
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

### Re: Machine Learning and the physical sciences

maestro,

http://www.physics.ox.ac.uk/phystat05/p ... stat05.pdf

Initial results are promising .. 10 hours for 10,000 samples.

Do you have an opinion yourself? Enlighten us. What's new, apart from the cute name?

I bet you won't give a technical answer.
In Bayesian inference you can estimate not only the parameters P of your hypothesised model M, but also perform the model comparison (selection). The full Beyes' rule us p(P | data, M) = p(data | P, M) * p(P | M) / p(data | M), where p(data | M) is called "model evidence". Yo usolve the inverse problem p(M_i | data) = p(data | M_i) / \sum_j p(data | M_j) to find the best model, namely the one with the highest p(M_i | data) (which translates to the highest evidence given data).
Bayesian *neural* networks implement the above procedure in addition to the standard parameter fitting.

xx

Cuchulainn
Posts: 59458
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

### Re: Machine Learning and the physical sciences

You seem to be using the usual Bayes' theorem for learning. I recall

$p(P | data, M) = p(data | P, M) * p(P|M) / p(data | M)$

Bhat and Prosper have a slightly different rule (see their equation) wrt prior

$p(P | data, M) = p(data | P, M) * p(P) / p(data | M)$

Does this trick warrant coining yet another name? It feels like NN++ overloading?

Bayesian *neural* networks implement the above procedure in addition to the standard parameter fitting.
Who uses this and where? I would be interested in an explanation in addition to this general description.

Very few articles describe the 'how to' step-by-step process/algorithm from input to output.

katastrofa
Posts: 7971
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

### Re: Machine Learning and the physical sciences

If the prior in the same in all models, then you can use the formula from those guys' work, obvsly.
I can't answer the question about the ML nomenclature. It amazes me too.

The AI craze seems to use BNNs to calculate the uncertainty of the estimated weights. I use it for model selection/testing strategies/etc., as stated above. I used the ML methods twice in the last 2 years. I'm not sure if what I'm doing isn't called ML now, though.

I'll name the longest maturing cheese in my fridge in your honour - Sniffy Irish...