Section: New Results

Graphical and Markov models

Structure learning via Hadamard product of correlation and partial correlation matrices

Participants : Sophie Achard, Karina Ashurbekova, Florence Forbes.

Structure learning is an active topic nowadays in different application areas, i.e. genetics, neuroscience. Classical conditional independences or marginal independences may not be sufficient to express complex relationships. This work [39] is introducing a new structure learning procedure where an edge in the graph corresponds to a non zero value of both correlation and partial correlation. Based on this new paradigm, we define an estimator and derive its theoretical properties. The asymptotic convergence of the proposed graph estimator and its rate are derived. Illustrations on a synthetic example and application to brain connectivity are displayed.

Optimal shrinkage for robust covariance matrix estimators in a small sample size setting

Participants : Sophie Achard, Karina Ashurbekova, Florence Forbes, Antoine Usseglio Carleve.

When estimating covariance matrices, traditional sample covariance-based estimators are straightforward but suffer from two main issues: 1) a lack of robustness, which occurs as soon as the samples do not come from a Gaussian distribution or are contaminated with outliers and 2) a lack of data when the number of parameters to estimate is too large compared to the number of available observations, which occurs as soon as the covariance matrix dimension is greater than the sample size. The first issue can be handled by assuming samples are drawn from a heavy-tailed distribution, at the cost of more complex derivations, while the second issue can be addressed by shrinkage with the difficulty of choosing the appropriate level of regularization. In this work [66] we offer both a tractable and optimal framework based on shrinked likelihood-based M-estimators. First, a closed-form expression is provided for a regularized covariance matrix estimator with an optimal shrinkage coefficient for any sample distribution in the elliptical family. Then, a complete inference procedure is proposed which can also handle both unknown mean and tail parameter, in contrast to most existing methods that focus on the covariance matrix parameter requiring pre-set values for the others. An illustration on synthetic and real data is provided in the case of the t-distribution with unknown mean and degrees-of-freedom parameters.

Robust penalized inference for Gaussian Scale Mixtures

Participants : Sophie Achard, Karina Ashurbekova, Florence Forbes.

The literature on sparse precision matrix estimation is rapidly growing. Many strong methods are valid only for Gaussian variables. One of the most commonly used approaches in this case is glasso which aims to minimize the negative L1-penalized log-likelihood function. In practice, data may deviate from normality in various ways, outliers and heavy tails frequently occur that can severely degrade the Gaussian models performance. A natural solution is to turn to heavier tailed distributions that remain tractable. For this purpose, we propose [51] a penalized version of the EM algorithm for Gaussian Scale Mixtures.

Non parametric Bayesian priors for graph structured data

Participants : Florence Forbes, Julyan Arbel, Hongliang Lu.

We consider the issue of determining the structure of clustered data, both in terms of finding the appropriate number of clusters and of modelling the right dependence structure between the observations. Bayesian nonparametric (BNP) models, which do not impose an upper limit on the number of clusters, are appropriate to avoid the required guess on the number of clusters but have been mainly developed for independent data. In contrast, Markov random fields (MRF) have been extensively used to model dependencies in a tractable manner but usually reduce to finite cluster numbers when clustering tasks are addressed. Our main contribution is to propose a general scheme to design tractable BNP-MRF priors that combine both features: no commitment to an arbitrary number of clusters and a dependence modelling. A key ingredient in this construction is the availability of a stick-breaking representation which has the threefold advantage to allowing us to extend standard discrete MRFs to infinite state space, to design a tractable estimation algorithm using variational approximation and to derive theoretical properties on the predictive distribution and the number of clusters of the proposed model. This approach is illustrated on a challenging natural image segmentation task for which it shows good performance with respect to the literature. This work [77] will be presented as a poster at BayesComp2020 in Gainesville, Florida, USA, [78].

Bayesian nonparametric models for hidden Markov random fields on count variables and application to disease mapping

Participants : Julyan Arbel, Fatoumata Dama, Jean-Baptiste Durand, Florence Forbes.

Hidden Markov random fields (HMRFs) have been widely used in image segmentation and more generally, for clustering of data indexed by graphs. Dependent hidden variables (states) represent the cluster identities and determine their interpretations. Dependencies between state variables are induced by the notion of neighborhood in the graph. A difficult and crucial problem in HMRFs is the identification of the number of possible states K. Recently, selection methods based on Bayesian non parametric priors (Dirichlet processes) have been developed. They do not assume that K is bounded a priori, thus allowing its adaptive selection with respect to the quantity of available data and avoiding costly systematic estimation and comparison of models with different fixed values for K. Our previous work [77] has focused on Bayesian nonparametric priors for HMRFs and continuous, Gaussian observations. In this work, we consider extensions to discrete observed data typically issued from counts. We define and implement Bayesian nonparametric models for HMRFs with Poisson distributed observations. As an illustration, we propose a new disease mapping model for epidemiology. The inference is done by Variational Bayesian Expectation Maximization (VBEM). Results on synthetic data sets suggest that our model is able to recover the true number of risk levels (clusters) and to provide a good estimation of the true risk level partition. Application on real data then also shows satisfying results.

As a perspective, Bayesian nonparametric models for hidden Markov random fields could be extended to non-Poissonian models (particularly to account for zero-inflated and over-/under-dispersed cases of application) and to regression models.

Hidden Markov models for the analysis of eye movements

Participants : Jean-Baptiste Durand, Brice Olivier, Sophie Achard.

This research theme is supported by a LabEx PERSYVAL-Lab project-team grant.

Joint work with: Anne Guérin-Dugué (GIPSA-lab) and Benoit Lemaire (Laboratoire de Psychologie et Neurocognition)

In the last years, GIPSA-lab has developed computational models of information search in web-like materials, using data from both eye-tracking and electroencephalograms (EEGs). These data were obtained from experiments, in which subjects had to decide whether a text was related or not to a target topic presented to them beforehand. In such tasks, reading process and decision making are closely related. Statistical analysis of such data aims at deciphering underlying dependency structures in these processes. Hidden Markov models (HMMs) have been used on eye-movement series to infer phases in the reading process that can be interpreted as strategies or steps in the cognitive processes leading to decision. In HMMs, each phase is associated with a state of the Markov chain. The states are observed indirectly though eye-movements. Our approach was inspired by Simola et al. (2008) [86], but we used hidden semi-Markov models for better characterization of phase length distributions (Olivier et al., 2017) [85]. The estimated HMM highlighted contrasted reading strategies, with both individual and document-related variability. New results were obtained in the standalone analysis of the eye-movements. A comparison between the effects of three types of texts was performed, considering texts either closely related, moderately related or unrelated to the target topic.

Then, using the restored state values, statistical characteristics of EEGs were compared according to strategies, brain wave frequencies and EEG channels (i.e., location on scalp). Differences in variance and correlations related to strategy changes were highlighted. Dependency graphs interpreted as maps of functional brain connectivity were estimated for each strategy and frequency and their changes were interpreted.

These results were published in Brice Olivier's PhD manuscript [12]. Although the approach was sufficient to highlight significant discrimination of strategies, it suffered from somewhat overlapping eye-movement characteristics over strategies. As a result, high uncertainty in the phase changes arose, which could induce underestimation of EEG and eye movement abilities to discriminate strategies.

This is why we developed integrated models coupling EEG and eye movements within one single HMM for better identification of strategies. Here, the coupling incorporated some delay between transitions in both EEG and eye-movement state sequences, since EEG patterns associated to cognitive processes occur lately with respect to eye-movement state switches. Moreover, EEGs and scanpaths were recorded with different time resolutions, so that some resampling scheme had to be added into the model, for the sake of synchronizing both processes. An associated EM algorithm for maximum likelihood parameter estimation was derived.

Our goal for this coming year is to implement and validate our coupled model for jointly analyzing eye-movements and EEGs in order to improve the discrimination of reading strategies.

Comparison of initialization strategies in the EM algorithm for hidden Semi-Markov processes

Participants : Jean-Baptiste Durand, Brice Olivier.

This research theme is supported by a LabEx PERSYVAL-Lab project-team grant.

Joint work with: Anne Guérin-Dugué (GIPSA-lab)

In Subsection 7.3.6, hidden semi-Markov models (HSMMs) were used to infer reading strategies from eye-movement and EEG time series. Model parameters were estimated by the EM algorithm. Its principle is to build a sequence of parameters with increasing likelihood values, starting from a starting point. The impact of this starting point has not been investigated in the case of HSMMs; this is why we aimed at developing and assessing an initialization method based on the available sequence lengths [48]. This consists in randomly choosing a number of transitions and then, uniformly-distributed transition times given the number of transitions. These transition times break the sequences into segements and assign uniformly-distributed states to each segment with the constraint that two consecutive states should be different.

The method was compared to other initialization strategies and was shown to be efficient on several data sets with multiple categorical sequences.

Lossy compression of tree structures

Participant : Jean-Baptiste Durand.

Joint work with: Christophe Godin and Romain Azaïs (Inria Mosaic)

The class of self-nested trees presents remarkable compression properties because of the systematic repetition of subtrees in their structure. The aim of our work is to achieve compression of any unordered tree by finding the nearest self-nested tree. Solving this optmization problem without more assumptions is conjectured to be an NP-complete or NP-hard problem. In [40], we firstly provided a better combinatorial characterization of this specific family of trees. In particular, we showed from both theoretical and practical viewpoints that complex queries can be quickly answered in self-nested trees compared to general trees. We also presented an approximation algorithm of a tree by a self-nested one that can be used in fast prediction of edit distance between two trees.

Our goal for this coming year is to apply this approach to quantify the degree of self-nestedness of several plant species and extend first results obtained on rice panicles stating that near self-nestedness is a fairly general pattern in plants.

Bayesian neural networks

Participants : Julyan Arbel, Mariia Vladimirova.

Joint work with: Pablo Mesejo from University of Granada, Spain, Jakob Verbeek from Inria Grenoble Rhône-Alpes, France.

We investigate in [45] deep Bayesian neural networks with Gaussian priors on the weights and ReLU-like nonlinearities, shedding light on novel sparsity-inducing mechanisms at the level of the units of the network, both pre- and post-nonlinearities. The main thrust of the paper is to establish that the units prior distribution becomes increasingly heavy-tailed with depth. We show that first layer units are Gaussian, second layer units are sub-Exponential, and we introduce sub-Weibull distributions to characterize the deeper layers units. Bayesian neural networks with Gaussian priors are well known to induce the weight decay penalty on the weights. In contrast, our result indicates a more elaborate regularisation scheme at the level of the units. This result provides new theoretical insight on deep Bayesian neural networks, underpinning their natural shrinkage properties and practical potential.