Accueil › Project › OptiMoCSI

OptiMoCSI

OptiMoCSI – Optimization and Monte Carlo Sampling Intertwined

nonsmoothness; scalability; theoretical guarantees; open science 4 society.

 Funded by the French National Research Agency; AAPG 2023 CES48 — starting in Oct 2023 or January 2024, duration 42 months.

The project

  • Its context. The Covid19 pandemic has constituted a massive stress test to numerous aspects of our societies, and still does in various ways. Its massive impact has significantly challenged the scientific communities (and many other compartments of our societies) in their ability to short term reactions and reorganizations. Potential future pandemic maintain the threat at high level and thus the need to develop scientific tools for efficient crisis resolution at the service of the society. During the past three years, OptiMoCSI’s researchers dedicated significant joint and coordinated efforts to design tools for the assessment of the spatio-temporal evolution of the pandemic intensity. Technically, it amounts to estimate a time-varying reproduction number, Rt, from daily new infection counts [1], that are, given the emergency context, of severely limited quality [2]. Beyond publications in major international venues [3–6], the methodological tools we devised permit fully-automated daily updated estimations of Rt for 200+ countries from real Covid19 data, available at the Johns Hopkins University repository, and for the 101 French départements. Furthermore, in an attempt to favor both science committed to society and the conduction of interdisciplinary sciences, interactive animated maps, intended for nonexpert audience, were designed to report on the space-time evolution of the pandemic, made available and automatically updated on a daily basis [6]. Finally, in an open and reproducible science effort, most of the engineered computational tools were shared on author’s repositories e.g., gfort-lab or IXXI. Though focused on the Covid19 pandemic, our joint preliminary works lead us to understand that pandemic monitoring can be envisioned as a representative example of a broad and generic set of complex scientific challenges that modern societies must face. Though promising, the results we achieved brought to light a number of significant and difficult technical deadlocks in i) real-world mechanism/observation joint modeling of highly corrupted nonstationary nonnegative data, and in ii) robust estimation and confidence assessment from combined nonsmooth convex optimization and Monte Carlo (MC) sampling schemes. Investigating these broad methodological and technical challenges forms the core of the present project.
  • Goals and objectives. The overarching goals of OptiMoCSI are to develop methodological contributions
    • to address challenges stemming from the real-world mechanism/observation joint modeling entailed by high societal stake problems, often consisting of corrupted nonstationary nonnegative counts;
    • to devise theoretically-grounded and practically-efficient computational tools permitting to deliver robust estimations of the quantities of interest, based on the models envisioned in the previous step, as well as a compre-hensive and accurate assessment of uncertainties by means of intertwining nonsmooth convex optimization and Monte Carlo sampling schemes;
    • to contribute to open/reproducible science and to easier relations between science and society by devising cartography tools that permit better grasps and thus better uses of uncertainty assessment both for interdisciplinary research deployments and for non scientific general audience communications.
  • Ambitions. Beyond methodological goals, OptiMoCSI intends to produce a set of documented toolboxes, geared towards achieving 3-fold broad ambitions.
    • In collaboration with world-reknown epidemiologists, actual and systematic application to Covid19 pandemic data should permit to assess if and how the developed tools go significantly beyond state-of-the-art tools in computational epidemiology, notably in emergency contexts (few data of limited quality) and as early as at the pandemic outbreak, or for adaptation to other types of pandemic (flue, ebola,. . . ).
    • Beyond epidemiology, methodological tools and toolboxes are (in accordance with ANR E.1 axis goals) aimed to be generically usable for a broad variety of applications in signal and image processing, where an underlying self-exciting process is observed through an imperfect counting device. In neurosciences, accurate estimation of human neural activity from non-invasive electrophysiological modalities (MEG, EEG, . . . ) is a challenging topic in that category (cf. e.g., [7]). Expected outcomes could also extend to a broader variety of cases where a varying intensity function must be estimated in a low-count regime, for instance photonics, scintigraphy and Lidar imaging techniques [8], or dictionary learning in Genomics analysis [9].
    • Furthermore, toolboxes should be ready for use by non statistical signal processing experts, in other interdisciplinary research projects, notably in Social and Human Sciences, where e.g., virality in social networks, or detecting large reshare cascades [10] constitute important problems in online social networks, or geared towards challenges with high societal stakes, decision makers and citizen information.
[1] A. Cori et al. “A new framework and software to estimate time-varying reproduction numbers during epidemics”. In: Am. J. Epidemiol. (2013).
[2] R. K. Nash et al. “Real-time estimation of the epidemic reproduction number: Scoping review of the applications and challenges”. In: PLOS Digital Health 6 (June 2022).
[3] P. Abry, N. Pustelnik, S. Roux, P. Jensen, P. Flandrin, R. Gribonval, C.-G. Lucas, E. Guichard, P. Borgnat and N. Garnier. “Spatial and temporal regularization to estimate COVID-19 reproduction number R (t): Promoting piecewise smoothness via convex optimization”. In: Plos One 8 (2020).
[4] B. Pascal, P. Abry, N. Pustelnik, S. Roux, R. Gribonval and P. Flandrin. “Nonsmooth convex optimization to estimate the Covid-19 reproduction number space-time evolution with robustness against low quality data”. In: IEEE Trans. Signal Process. (2022).
[5] G. Fort, B. Pascal, P. Abry and N. Pustelnik. “Covid19 reproduction number: Credibility intervals by blockwise proximal Monte Carlo samplers”. In: IEEE Trans Signal Process., to appear. (2023).
[6] E. Guichard and P. Abry. “Mesure, estimation et représentations de la Covid-19”. In: Annales des Mines-Responsabilité et environnement. 4. 2022.
[7] C. Allain et al. DriPP: Driven point processes to model stimuli induced patterns in M/EEG signals. Tech. rep. arXiv:2112.06652, 2021.
[8] S. Melidonis et al. Efficient Bayesian computation for low-photon imaging problems. Tech. rep. arXiv:2206.05350, 2022.
[9] J.-M. Arbona et al. Neural network and kinetic modelling of human genome replication reveal replication origin locations and strengths. Tech. rep. bioRxiv:2021.12.15.472795, 2022.
[10] K. Subbian et al. “Detecting large reshare cascades in social networks”. In: Proc. 26th Int. Conf. World Wide Web. 2017.

The team

 

Results

  1. Pandemic Intensity Estimation From Stochastic Approximation-Based Algorithms. P. Abry, J. Chevallier, G. Fort and B. Pascal. In IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing 2023 (CAMSAP) proceedings.
  2. MATLAB codes associated to the paper Pandemic Intensity Estimation From Stochastic Approximation-Based Algorithms.  Codes developed by G. Fort

 

Mentions Légales