## a case for Bayesian deep learnin

Posted in Books, pictures, Statistics, Travel, University life with tags , , , , , , , , , , on September 30, 2020 by xi'an

Andrew Wilson wrote a piece about Bayesian deep learning last winter. Which I just read. It starts with the (posterior) predictive distribution being the core of Bayesian model evaluation or of model (epistemic) uncertainty.

“On the other hand, a flat prior may have a major effect on marginalization.”

Interesting sentence, as, from my viewpoint, using a flat prior is a no-no when running model evaluation since the marginal likelihood (or evidence) is no longer a probability density. (Check Lindley-Jeffreys’ paradox in this tribune.) The author then goes for an argument in favour of a Bayesian approach to deep neural networks for the reason that data cannot be informative on every parameter in the network, which should then be integrated out wrt a prior. He also draws a parallel between deep ensemble learning, where random initialisations produce different fits, with posterior distributions, although the equivalent to the prior distribution in an optimisation exercise is somewhat vague.

“…we do not need samples from a posterior, or even a faithful approximation to the posterior. We need to evaluate the posterior in places that will make the greatest contributions to the [posterior predictive].”

The paper also contains an interesting point distinguishing between priors over parameters and priors over functions, ony the later mattering for prediction. Which must be structured enough to compensate for the lack of data information about most aspects of the functions. The paper further discusses uninformative priors (over the parameters) in the O’Bayes sense as a default way to select priors. It is however unclear to me how this discussion accounts for the problems met in high dimensions by standard uninformative solutions. More aggressively penalising priors may be needed, as those found in high dimension variable selection. As in e.g. the 10⁷ dimensional space mentioned in the paper. Interesting read all in all!

## Why do we draw parameters to draw from a marginal distribution that does not contain the parameters?

Posted in Statistics with tags , , , , , , , on November 3, 2019 by xi'an

A revealing question on X validated of a simulation concept students (and others) have trouble gripping with. Namely using auxiliary variates to simulate from a marginal distribution, since these auxiliary variables are later dismissed and hence appear to them (students) of no use at all. Even after being exposed to the accept-reject algorithm. Or to multiple importance sampling. In the sense that a realisation of a random variable can be associated with a whole series of densities in an importance weight, all of them being valid (but some more equal than others!).

## maximum of a Dirichlet vector

Posted in Books, Statistics with tags , , , , , , , on September 26, 2016 by xi'an

An intriguing question on Stack Exchange this weekend, about the distribution of max{p¹,p²,…}the maximum component of a Dirichlet vector Dir(a¹,a²,…) with arbitrary hyper-parameters. Writing the density of this random variable is feasible, using its connection with a Gamma vector, but I could not find a closed-form expression. If there is such an expression, it may follow from the many properties of the Dirichlet distribution and I’d be interested in learning about it. (Very nice stamp, by the way! I wonder if the original formula was made with LaTeX…)

## uniform correlation mixtures

Posted in Books, pictures, Statistics, University life with tags , , , , , , , , , on December 4, 2015 by xi'an

Kai Zhang and my friends from Wharton, Larry Brown, Ed George and Linda Zhao arXived last week a neat mathematical foray into the properties of a marginal bivariate Gaussian density once the correlation ρ is integrated out. While the univariate marginals remain Gaussian (unsurprising, since these marginals do not depend on ρ in the first place), the joint density has the surprising property of being

[1-Φ(max{|x|,|y|})]/2

which turns an infinitely regular density into a density that is not even differentiable everywhere. And which is constant on squares rather than circles or ellipses. This is somewhat paradoxical in that the intuition (at least my intuition!) is that integration increases regularity… I also like the characterisation of the distributions factorising through the infinite norm as scale mixtures of the infinite norm equivalent of normal distributions. The paper proposes several threads for some extensions of this most surprising result. Other come to mind:

1. What happens when the Jeffreys prior is used in place of the uniform? Or Haldane‘s prior?
2. Given the mixture representation of t distributions, is there an equivalent for t distributions?
3. Is there any connection with the equally surprising resolution of the Drton conjecture by Natesh Pillai and Xiao-Li Meng?
4. In the Khintchine representation, correlated normal variates are created by multiplying a single χ²(3) variate by a vector of uniforms on (-1,1). What are the resulting variates for other degrees of freedomk in the χ²(k) variate?
5. I also wonder at a connection between this Khintchine representation and the Box-Müller algorithm, as in this earlier X validated question that I turned into an exam problem.

## Approximate Integrated Likelihood via ABC methods

Posted in Books, Statistics, University life with tags , , , , , , , , on March 13, 2014 by xi'an

My PhD student Clara Grazian just arXived this joint work with Brunero Liseo on using ABC for marginal density estimation. The idea in this paper is to produce an integrated likelihood approximation in intractable problems via the ratio

$L(\psi|x)\propto \dfrac{\pi(\psi|x)}{\pi(\psi)}$

both terms in the ratio being estimated from simulations,

$\hat L(\psi|x) \propto \dfrac{\hat\pi^\text{ABC}(\psi|x)}{\hat\pi(\psi)}$

(with possible closed form for the denominator). Although most of the examples processed in the paper (Poisson means ratio, Neyman-Scott’s problem, g-&-k quantile distribution, semi-parametric regression) rely on summary statistics, hence de facto replacing the numerator above with a pseudo-posterior conditional on those summaries, the approximation remains accurate (for those examples). In the g-&-k quantile example, Clara and Brunero compare our ABC-MCMC algorithm with the one of Allingham et al. (2009, Statistics & Computing): the later does better by not replicating values in the Markov chain but instead proposing a new value until it is accepted by the usual Metropolis step. (Although I did not spend much time on this issue, I cannot see how both approaches could be simultaneously correct. Even though the outcomes do not look very different.) As noted by the authors, “the main drawback of the present approach is that it requires the use of proper priors”, unless the marginalisation of the prior can be done analytically. (This is an interesting computational problem: how to provide an efficient approximation to a marginal density of a σ-finite measure, assuming this density exists.)

Clara will give a talk at CREST-ENSAE today about this work, in the Bayes in Paris seminar: 2pm in room 18.