sliced Wasserstein estimation of mixtures

A paper by Soheil Kolouri and co-authors was arXived last week about using Wasserstein distance for inference on multivariate Gaussian mixtures. The basic concept is that the parameter is estimated by minimising the p-Wasserstein distance to the empirical distribution, smoothed by a Normal kernel. As the general Wasserstein distance is quite costly to compute, the approach relies on a sliced version, which means computing the Wasserstein distance between one-dimensional projections of the distributions. Optimising over the directions is an additional computational constraint.

“To fit a finite GMM to the observed data, one is required to answer the following questions: 1) how to estimate the number of mixture components needed to represent the data, and 2) how to estimate the parameters of the mixture components.”

The paper contains a most puzzling comment opposing maximum likelihood estimation to minimum Wasserstein distance estimation on the basis that the later would not suffer from multimodality. This sounds incorrect as the multimodality of a mixture model (likelihood) stems from the lack of identifiability of the parameters. If all permutations of these parameters induce exactly the same distribution, they all stand at the same distance from the data distribution, whatever the distance is. Furthermore, the above tartan-like picture clashes with the representation of the log-likelihood of a Normal mixture, as exemplified by the picture below based on a 150 sample with means 0 and 2, same unit variance, and weights 0.3 and 0.7, which shows a smooth if bimodal structure:And for the same dataset, my attempt at producing a Wasserstein “energy landscape” does return a multimodal structure (this is the surface of minus the logarithm of the 2-Wasserstein distance):“Jin et al. proved that with random initialization, the EM algorithm will converge to a bad critical point with high probability.”

This statement is most curious in that the “probability” in the assessment must depend on the choice of the random initialisation, hence on a sort of prior distribution that is not explicited in the paper. Which remains blissfully unaware of Bayesian approaches.

Another [minor mode] puzzling statement is that the p-Wasserstein distance is defined on the space of probability measures with finite p-th moment, which does not make much sense when what matters is rather the finiteness of the expectation of the distance d(X,Y) raised to the power p. A lot of the maths details either do not make sense or seem superfluous.

a well-hidden E step

A recent question on X validated ended up being quite interesting! The model under consideration is made of parallel Markov chains on a finite state space, all with the same Markov transition matrix, M, which turns into a hidden Markov model when the only summary available is the number of chains in a given state at a given time. When writing down the EM algorithm, the E step involves the expected number of moves from a given state to a given state at a given time. The conditional distribution of those numbers of chains is a product of multinomials across times and starting states, with no Markov structure since the number of chains starting from a given state is known at each instant. Except that those multinomials are constrained by the number of “arrivals” in each state at the next instant and that this makes the computation of the expectation intractable, as far as I can see.

A solution by Monte Carlo EM means running the moves for each instant under the above constraints, which is thus a sort of multinomial distribution with fixed margins, enjoying a closed-form expression but for the normalising constant. The direct simulation soon gets too costly as the number of states increases and I thus considered a basic Metropolis move, using one margin (row or column) or the other as proposal, with the correction taken on another margin. This is very basic but apparently enough for the purpose of the exercise. If I find time in the coming days, I will try to look at the ABC resolution of this problem, a logical move when starting from non-sufficient statistics!

empirical Bayes, reference priors, entropy & EM

Klebanov and co-authors from Berlin arXived this paper a few weeks ago and it took me a quiet evening in Darjeeling to read it. It starts with the premises that led Robbins to introduce empirical Bayes in 1956 (although the paper does not appear in the references), where repeated experiments with different parameters are run. Except that it turns non-parametric in estimating the prior. And to avoid resorting to the non-parametric MLE, which is the empirical distribution, it adds a smoothness penalty function to the picture. (Warning: I am not a big fan of non-parametric MLE!) The idea seems to have been Good’s, who acknowledged using the entropy as penalty is missing in terms of reparameterisation invariance. Hence the authors suggest instead to use as penalty function on the prior a joint relative entropy on both the parameter and the prior, which amounts to the average of the Kullback-Leibler divergence between the sampling distribution and the predictive based on the prior. Which is then independent of the parameterisation. And of the dominating measure. This is the only tangible connection with reference priors found in the paper.

The authors then introduce a non-parametric EM algorithm, where the unknown prior becomes the “parameter” and the M step means optimising an entropy in terms of this prior. With an infinite amount of data, the true prior (meaning the overall distribution of the genuine parameters in this repeated experiment framework) is a fixed point of the algorithm. However, it seems that the only way it can be implemented is via discretisation of the parameter space, which opens a whole Pandora box of issues, from discretisation size to dimensionality problems. And to motivating the approach by regularisation arguments, since the final product remains an atomic distribution.

While the alternative of estimating the marginal density of the data by kernels and then aiming at the closest entropy prior is discussed, I find it surprising that the paper does not consider the rather natural of setting a prior on the prior, e.g. via Dirichlet processes.

warp-U bridge sampling

[I wrote this set of comments right after MCqMC 2016 on a preliminary version of the paper so mileage may vary in terms of the adequation to the current version!]

In warp-U bridge sampling, newly arXived and first presented at MCqMC 16, Xiao-Li Meng continues (in collaboration with Lahzi Wang) his exploration of bridge sampling techniques towards improving the estimation of normalising constants and ratios thereof. The bridge sampling estimator of Meng and Wong (1996) is an harmonic mean importance sampler that requires iterations as it depends on the ratio of interest. Given that the normalising constant of a density does not depend on the chosen parameterisation in the sense that the Jacobian transform preserves this constant, a degree of freedom is in the choice of the parameterisation. This is the idea behind warp transformations. The initial version of Meng and Schilling (2002) used location-scale transforms, while the warp-U solution goes for a multiple location-scale transform that can be seen as based on a location-scale mixture representation of the target. With K components. This approach can also be seen as a sort of artificial reversible jump algorithm when one model is fully known. A strategy Nicolas and I also proposed in our nested sampling Biometrika paper.

Once such a mixture approximation is obtained. each and every component of the mixture can be turned into the standard version of the location-scale family by the appropriate location-scale transform. Since the component index k is unknown for a given X, they call this transform a random transform, which I find somewhat more confusing that helpful. The conditional distribution of the index given the observable x is well-known for mixtures and it is used here to weight the component-wise location-scale transforms of the original distribution p into something that looks rather similar to the standard version of the location-scale family. If no mode has been forgotten by the mixture. The simulations from the original p are then rescaled by one of those transforms, which index k is picked according to the conditional distribution. As explained later to me by XL, the random[ness] in the picture is due to the inclusion of a random ± sign. Still, in the notation introduced in (13), I do not get how the distribution Þ [sorry for using different symbols, I cannot render a tilde on a p] is defined since both ψ and W are random. Is it the marginal? In which case it would read as a weighted average of rescaled versions of p. I have the same problem with Theorem 1 in that I do not understand how one equates Þ with the joint distribution.

Equation (21) is much more illuminating (I find) than the previous explanation in that it exposes the fact that the principle is one of aiming at a new distribution for both the target and the importance function, with hopes that the fit will get better. It could have been better to avoid the notion of random transform, then, but this is mostly a matter of conveying the notion.

On more specifics points (or minutiae), the unboundedness of the likelihood is rarely if ever a problem when using EM. An alternative to the multiple start EM proposal would then be to get sequential and estimate the mixture in a sequential manner, only adding a component when it seems worth it. See eg Chopin and Pelgrin (2004) and Chopin (2007). This could also help with the bias mentioned therein since only a (tiny?) fraction of the data would be used. And the number of components K has an impact on the accuracy of the approximation, as in not missing a mode, and on the computing time. However my suggestion would be to avoid estimating K as this must be immensely costly.

Section 6 obviously relates to my folded Markov interests. If I understand correctly, the paper argues that the transformed density Þ does not need to be computed when considering the folding-move-unfolding step as a single step rather than three steps. I fear the description between equations (30) and (31) is missing the move step over the transformed space. Also on a personal basis I still do not see how to add this approach to our folding methodology, even though the different transforms act as as many replicas of the original Markov chain.

non-local priors for mixtures

[For some unknown reason, this commentary on the paper by Jairo Fúquene, Mark Steel, David Rossell —all colleagues at Warwick— on choosing mixture components by non-local priors remained untouched in my draft box…]

Choosing the number of components in a mixture of (e.g., Gaussian) distributions is a hard problem. It may actually be an altogether impossible problem, even when abstaining from moral judgements on mixtures. I do realise that the components can eventually be identified as the number of observations grows to infinity, as demonstrated for instance by Judith Rousseau and Kerrie Mengersen (2011). But for a finite and given number of observations, how much can we trust any conclusion about the number of components?! It seems to me that the criticism about the vacuity of point null hypotheses, namely the logical absurdity of trying to differentiate θ=0 from any other value of θ, applies to the estimation or test on the number of components of a mixture. Doubly so, one might argue, since a very small or a very close component is undistinguishable from a non-existing one. For instance, Definition 2 is correct from a mathematical viewpoint, but it does not spell out the multiple contiguities between k and k’ component mixtures.

The paper starts with a comprehensive coverage of l’état de l’art… When using a Bayes factor to compare a k-component and an h-component mixture, the behaviour of the factor is quite different depending on which model is correct. Essentially overfitted mixtures take much longer to detect than underfitted ones, which makes intuitive sense. And BIC should be corrected for overfitted mixtures by a canonical dimension λ between the true and the (larger) assumed number of parameters  into

2 log m(y) = 2 log p(y|θ) – λ log O(n) + O(log log n)

I would argue that this purely invalidates BIG in mixture settings since the canonical dimension λ is unavailable (and DIC does not provide a useful substitute as we illustrated a decade ago…) The criticism about Rousseau and Mengersen (2011) over-fitted mixture that their approach shrinks less than a model averaging over several numbers of components relates to minimaxity and hence sounds both overly technical and reverting to some frequentist approach to testing. Replacing testing with estimating sounds like the right idea.  And I am also unconvinced that a faster rate of convergence of the posterior probability or of the Bayes factor is a relevant factor when conducting

As for non local priors, the notion seems to rely on a specific topology for the parameter space since a k-component mixture can approach a k’-component mixture (when k'<k) in a continuum of ways (even for a given parameterisation). This topology seems to be summarised by the penalty (distance?) d(θ) in the paper. Is there an intrinsic version of d(θ), given the weird parameter space? Like one derived from the Kullback-Leibler distance between the models? The choice of how zero is approached clearly has an impact on how easily the “null” is detected, the more because of the somewhat discontinuous nature of the parameter space. Incidentally, I find it curious that only the distance between means is penalised… The prior also assumes independence between component parameters and component weights, which I think is suboptimal in dealing with mixtures, maybe suboptimal in a poetic sense!, as we discussed in our reparameterisation paper. I am not sure either than the speed the distance converges to zero (in Theorem 1) helps me to understand whether the mixture has too many components for the data’s own good when I can run a calibration experiment under both assumptions.

While I appreciate the derivation of a closed form non-local prior, I wonder at the importance of the result. Is it because this leads to an easier derivation of the posterior probability? I do not see the connection in Section 3, except maybe that the importance weight indeed involves this normalising constant when considering several k’s in parallel. Is there any convergence issue in the importance sampling solution of (3.1) and (3.3) since the simulations are run under the local posterior? While I appreciate the availability of an EM version for deriving the MAP, a fact I became aware of only recently, is it truly bringing an improvement when compared with picking the MCMC simulation with the highest completed posterior?

The section on prior elicitation is obviously of central interest to me! It however seems to be restricted to the derivation of the scale factor g, in the distance, and of the parameter q in the Dirichlet prior on the weights. While the other parameters suffer from being allocated the conjugate-like priors. I would obviously enjoy seeing how this approach proceeds with our non-informative prior(s). In this regard, the illustration section is nice, but one always wonders at the representative nature of the examples and the possible interpretations of real datasets. For instance, when considering that the Old Faithful is more of an HMM than a mixture.

MDL multiple hypothesis testing

“This formulation reveals an interesting connection between multiple hypothesis testing and mixture modelling with the class labels corresponding to the accepted hypotheses in each test.”

After my seminar at Monash University last Friday, David Dowe pointed out to me the recent work by Enes Makalic and Daniel Schmidt on minimum description length (MDL) methods for multiple testing as somewhat related to our testing by mixture paper. Work which appeared in the proceedings of the 4th Workshop on Information Theoretic Methods in Science and Engineering (WITMSE-11), that took place in Helsinki, Finland, in 2011. Minimal encoding length approaches lead to choosing the model that enjoys the smallest coding length. Connected with, e.g., Rissannen‘s approach. The extension in this paper consists in considering K hypotheses at once on a collection of m datasets (the multiple then bears on the datasets rather than on the hypotheses). And to associate an hypothesis index to each dataset. When the objective function is the sum of (generalised) penalised likelihoods [as in BIC], it leads to selecting the “minimal length” model for each dataset. But the authors introduce weights or probabilities for each of the K hypotheses, which indeed then amounts to a mixture-like representation on the exponentiated codelengths. Which estimation by optimal coding was first proposed by Chris Wallace in his book. This approach eliminates the model parameters at an earlier stage, e.g. by maximum likelihood estimation, to return a quantity that only depends on the model index and the data. In fine, the purpose of the method differs from ours in that the former aims at identifying an appropriate hypothesis for each group of observations, rather than ranking those hypotheses for the entire dataset by considering the posterior distribution of the weights in the later. The mixture has somehow more of a substance in the first case, where separating the datasets into groups is part of the inference.

variational Bayes for variable selection

Xichen Huang, Jin Wang and Feng Liang have recently arXived a paper where they rely on variational Bayes in conjunction with a spike-and-slab prior modelling. This actually stems from an earlier paper by Carbonetto and Stephens (2012), the difference being in the implementation of the method, which is less Gibbs-like for the current paper. The approach is not fully Bayesian in that, not only an approximate (variational) representation is used for the parameters of interest (regression coefficient and presence-absence indicators) but also the nuisance parameters are replaced with MAPs. The variational approximation on the regression parameters is an independent product of spike-and-slab distributions. The authors show the approximate approach is consistent in both frequentist and Bayesian terms (under identifiability assumptions). The method is undoubtedly faster than MCMC since it shares many features with EM but I still wonder at the Bayesian interpretability of the outcome, which writes out as a product of estimated spike-and-slab mixtures. First, the weights in the mixtures are estimated by EM, hence fixed. Second, the fact that the variational approximation is a product is confusing in that the posterior distribution on the regression coefficients is unlikely to produce posterior independence.