The problem in concentrating on a machine learning or regression model is that one loses some of the descriptive analysis that allows the classification measures themselves to be improved, if not optimized.

Those labs with *successful* application of these ML models appear to be the same ones who are successful in general in their domain and apply the same descriptive analysis as required for more traditional statistical approaches. The era of "you don't need to know the domain, get the data and train the system" is dying - or should be.

I always tell my students to use as much of the apriori information about the problem as possible, then use a ML or CI methods to account for the unknowns or the things you cannot feasibly model or compute.

There are some problems that are just too complex to solve from a first principles, Bayesian point-of-view.

I don't think the Bayesian approach is dead. Indeed, it is lively than ever. If you look at the premier machine learning conferences (ICML, NIPS) you may notice that SVM is no longer as popular as it used to be. With the recent developments in Bayesian nonparametrics the trend is now shifting back to Bayesian models. All of these interesting approaches on Chinese restaurant process, Indian buffett process, latent Dirichlet allocation. nested Chinese restaurant process, dirichlet/gamma/beta processes etc. they are all in the nonparametric Bayesian domain and they will dominate the machine learning literature for another 5 to 10 years IMHO.

I don't think the Bayesian approach is dead. See for example Grifits, T. L., & Tenenbaum, J. B. (2006). Optimal

prediction in everyday cognition. Psychological

Science, 17(9), 767–773. doi:10.1111/j.1467-

9280.2006.01780.x

You argue as if Bayesian approaches and machine learning would be orthogonal approaches. I agree that the vanilla approach to machine learning assumes uniform priors. But more important, classification errors are rarely symmetric, e.g. false positives and false negatives in the dichotomic case do rarely have the same impact on the system performance.

And error models can be built in the penalization function for misclassification in the formulation of the SVM optimization problem, just as an example.

And I agree with James that being able to apply such techniques is beyond vanilla, so if you're understanding what you do there, you are likely to belong to the successful ones already.

My approach is a mixed one: I believe that,as it happens in nature, we should adopt both a deterministic approach & a probabilistic (Bayesian) approach. It depends on the scale of the phenomena that we need to process and/or simulate & the level of complication that we want to introduce.

No Bayesian model based classification is not at all dead. there are lot of works going on various fields based on bayesian approach. By the way lot of modifications/improvements done on the algorithm to achieve the good results.

I'm always greatfull to Bays Theorem applications. He is still alive (the theorem!!!)

@Fausto Galetto: perhaps you would get less down votes if you:

1. you don't use capitalized words everywhere.

2. you provide a relevant answer to the question.

Cheers

No, they are not obsolete. I think many people resorted to sampling methods (Gibbs, Hamiltonian Monte Carlo, etc. ) since integrals were intractable.  However, sampling methods are not fast enough when it comes to practical applications having large/big data. There's a trend of using stochastic variation inference instead.