learning optimal summary statistics

“Despite the pursuit of the holy grail of sufficient statistics, most applications will have to settle for the weakest concept of optimal statistics.”Quiz #1: How does Bayes sufficiency [which preserves the posterior density] differ from sufficiency [which preserves the likelihood function]?

Quiz #2: How does Fisher-information sufficiency [which preserves the information matrix] differ from standard sufficiency [which preserves the likelihood function]?

Read a recent arXival by Till Hoffmann and Jukka-Pekka Onnela that I frankly found most puzzling… Maybe due to the Norman train where I was traveling being particularly noisy.

The argument in the paper is to find a summary statistic that minimises the [empirical] expected posterior entropy, which equivalently means minimising the expected Kullback-Leibler distance to the full posterior.  And maximizing the mutual information between parameters θ and summaries t(.). And maximizing the expected surprise. Which obviously requires breaking the sample into iid components and hence considering the gain brought by a specific transform of a single observation. The paper also contains a long comparison with other criteria for choosing summaries.

“Minimizing the posterior entropy would discard the sufficient statistic t such that the posterior is equal to the prior–we have not learned anything from the data.”

Furthermore, the expected aspect of the criterion takes us away from a proper Bayes analysis (and exhibits artifacts as the one above), which somehow makes me question the relevance of comparing entropies under different distributions. It took me a long while to realise that the collection of summaries was set by the user and quite limited. Like a neural network representation of the posterior mean. And the intractable posterior is further approximated by a closed-form function of the parameter θ and of the summary t(.). Using there a neural density estimator. Or a mixture density network.

day one at ISBA 22

Started the day with a much appreciated swimming practice in the [alas warm⁺⁺⁺] outdoor 50m pool on the Island with no one but me in the slooow lane. And had my first ride with the biXi system, surprised at having to queue behind other bikes at red lights! More significantly, it was a great feeling to reunite at last with so many friends I had not met for more than two years!!!

My friend Adrian Raftery gave the very first plenary lecture on his work on the Bayesian approach to long-term population projections, which was recently  a work censored by some US States, then counter-censored by the Supreme Court [too busy to kill Roe v. Wade!]. Great to see the use of Bayesian methods validated by the UN Population Division [with at least one branch of the UN

Stephen Lauritzen returning to de Finetti notion of a model as something not real or true at all, back to exchangeability. Making me wonder when exchangeability is more than a convenient assumption leading to the Hewitt-Savage theorem. And sufficiency. I mean, without falling into a Keynesian fallacy, each point of the sample has unique specificities that cannot be taken into account in an exchangeable model. Nice to hear some measure theory, though!!! Plus a comment on the median never being sufficient, recouping an older (and presumably not original) point of mine. Stephen’s (or Fisher’s?) argument being that the median cannot be recursively computed!

Antonietta Mira and I had our ABC session this afternoon with Cecilia Viscardi, Sirio Legramanti, and Massimiliano Tamborino (Warwick) as speakers. Cecilia linked ABC with normalising flows, in collaboration with Dennis Prangle (whose earlier paper on this connection was presented as the first One World ABC seminar). Thus using past simulations to approximate the posterior by a neural network, possibly with a significant increase in computing time when compared with more rudimentary SMC-ABC methods in larger dimensions. Sirio considered summary-free ABC based on discrepancies like Rademacher complexity. Which more or less contains MMD, Kullback-Leibler, Wasserstein and more, although it seems to be dependent on the parameterisation of the observations. An interesting opening at the end was that this approach could apply to non iid settings. Massi presented a paper coauthored with Umberto that had just been arXived. On sequential ABC with a dependence on the summary statistic (hence guided). Further bringing copulas into the game, although this forces another choice [for the marginals] in the method.

Tamara Broderick talked about a puzzling leverage effect of some observations in economic studies where a tiny portion of individuals may modify the significance or the sign of a coefficient, for which I cannot tell whether the data or the reliance on statistical significance are to blame. Robert Kohn presented mixture-of-Gaussian copulas [not to be confused with mixture of Gaussian-copulas!] and Nancy Reid concluded my first [and somewhat exhausting!] day at ISBA with a BFF talk on the different statistical paradigms take on confidence (for which the notion of calibration seems to remain frequentist).

Side comments: First, most people in the conference are wearing masks, which is great! Also, I find it hard to read slides from the screen, which I presume is an age issue (?!) Even more aside, I had Korean lunch in a place that refused to serve me a glass of water, which I find amazing.

set-valued sufficient statistic

While the classical definition of a statistic is one of a real valued random variable or vector, less usual situations call for broader definitions… For instance, in an homework problem from Mark Schervish’s Theory of Statistics, a sample from the uniform distribution of a ball of unknown centre θ and radius ς is associated with the convex hull of said sample as “sufficient statistic”, albeit the object being a set. Similarly, if the radius ς is known, the set made of the intersection of all the balls of radius ς centred at the observations is sufficient, in that the likelihood is constant for θ inside and zero outside. As discussed in this X validated question, this does not define an optimal estimator of the center θ, while Pitman’s best location equivariant does, while the centre of this sufficient set, but it is not sufficient as a statistic and is not necessarily the MVUE, if unbiased.

sufficient statistics for machine learning

By chance, I came across this ICML¹⁹ paper of Milan Cvitkovic and Günther Koliander, Minimal Achievable Sufficient Statistic Learning,  on a form of sufficiency for machine learning. The paper starts with “our” standard notion of sufficiency albeit in a predictive sense, namely that Z=T(X) is sufficient for predicting Y if the conditional distribution of Y given Z is the same as the conditional distribution of Y given X. It also acknowledges that minimal sufficiency may be out of reach. However, and without pursuing this question into the depths of said paper, I am surprised that any type of sufficiency can be achieved there since the model stands outside exponential families… In accordance with the Darmois-Pitman-Koopman lemma. Obviously, this is not a sufficiency notion in the statistical sense, since there is no likelihood (albeit there are parameters involved in the deep learning network). And Y is a discrete variate, which means that

is a sufficient “statistic” for a fixed conditional, but I am lost at how the solution proposed in the paper, could be minimal when the dimension and structure of T(x) are chosen from the start. A very different notion, for sure!

Concentration and robustness of discrepancy-based ABC [One World ABC ‘minar, 28 April]

Our next speaker at the One World ABC Seminar will be Pierre Alquier, who will talk about “Concentration and robustness of discrepancy-based ABC“, on Thursday April 28, at 9.30am UK time, with an abstract reported below.

Approximate Bayesian Computation (ABC) typically employs summary statistics to measure the discrepancy among the observed data and the synthetic data generated from each proposed value of the parameter of interest. However, finding good summary statistics (that are close to sufficiency) is non-trivial for most of the models for which ABC is needed. In this paper, we investigate the properties of ABC based on integral probability semi-metrics, including MMD and Wasserstein distances. We exhibit conditions ensuring the contraction of the approximate posterior. Moreover, we prove that MMD with an adequate kernel leads to very strong robustness properties.