Marina Costantini
I am a PhD candidate conducting my research at Eurecom and student of Sorbonne University. My advisor is Thrasyvoulos Spyropoulos, and my research is partially funded by the Futur & Ruptures scholarship, granted by the Fondation Mines-Télécom.
My research interests span many aspects of optimization theory, with particular focus on decentralized multi-agent optimization in networks. You can find more details below.
For my complete experience, here is the link to my CV.
Feel free to connect with me on LinkedIn and Twitter; I am also on Google Scholar.
If you want to send me an email, please write to marina.costantini@eurecom.fr.
NEW: I am looking for a research internship for 2023, if you are interested in my profile or know someone who may be, ping me!
Short Bio
I got my Engineering degree from the National University of La Plata, Argentina. I did my graduation project at the Group of Control Applications, with Hernán De Battista and Fabricio Garelli as my advisors.
After graduating I joined the Center for Computational Imaging and Simulation Technologies in Biomedicine, after which I pursued a Master's degree at Eurecom, where I am now a PhD candidate.
My Research
My current main topic of research concerns the development and analysis of decentralized optimization algorithms.
Given a network of agents, each of them holding a private function, these algorithms find the minimizer of the sum of the functions by letting the nodes exchange optimization values (parameters, gradients...) with their neighbors, but not the functions themselves.
In a typical example, the function of each node is the empirical risk given by
The loss of a machine learning model that all nodes have agreed to learn
The data of each node, which must remain private
The GIF on the right shows how the particular class of gossip decentralized algorithms works: only one pair of nodes linked by an edge need to be active to complete an iteration.
News (and "once were" news)
Oct. 2022: I will be attending NeurIPS in December, and participating in the Women in Machine Learning Workshop (WiML) on Monday 28/11.
July 2022: Our paper "Pick Your Neighbor: Local Gauss-Southwell Rule for Fast Asynchronous Decentralized Optimization" has been accepted at CDC.
May 2022: I'll be presenting our poster “Spread gossip faster and be happy about it!” at the 6th edition of MOMI.
Sep. 2021: We are offering a semester project to work on the impact of graph structure on the convergence of decentralized SGD.
May 2020: Our paper "Impact of Popular Content Relational Structure on Joint Caching and Recommendation Policies" has been accepted at CCDWN of WiOpt.
Jan. 2020: Our paper "Approximation Guarantees for the Joint Optimization of Caching and Recommendation" has been accepted at ICC.
Sep. 2019: I got my Master's degree together with the rest of the amazing promo 2019 of Eurecom. Starting the PhD next month!
Aug. 2019: As part of the Seeds for the Future program of Huawei, I have been interviewed by Le Figaro about my experience at the program, my own career path and my views on the engagement of girls and women with STEM disciplines. The complete article (in French) is found here.
June 2019: Thrilled to share that I have been selected for the 2019 edition of the Seeds for the Future program organized by Huawei. Cannot wait!
Mar. 2019: Amazing news! I have been granted the Futur & Ruptures scholarship by the Fondation Mines-Télécom to partially cover my PhD studies!
Papers
Publications
M. Costantini, N. Liakopoulos, P. Mertikopoulos, and T. Spyropoulos, “Pick your neighbor: Local Gauss-Southwell rule for fast asynchronous decentralized optimization,” To appear in the 61st IEEE Conference on Decision and Control (CDC), 2022. [pdf] [code]
M. Costantini, T. Spyropoulos, T. Giannakas, and P. Sermpezis, “Approximation guarantees for the joint optimization of caching and recommendation,” in IEEE International Conference on Communications (ICC), 2020, pp. 1–7. [pdf]
M. Costantini and T. Spyropoulos, “Impact of popular content relational structure on joint caching and recommendation policies,” in IEEE 18th International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOPT), 2020, pp. 1–8. [pdf]
M. Mozumder, J. M. Pozo, S. Coelho, M. Costantini, J. Simpson, J. R. Highley, P. G. Ince, and A. F. Frangi, “Quantitative histomorphometry of capillary microstructure in deep white matter,” NeuroImage: Clinical, vol. 23, p. 101 839, 2019. [pdf]
S. Coelho, J. M. Pozo, M. Costantini, J. R. Highley, M. Mozumder, J. E. Simpson, P. G. Ince, and A. F. Frangi, “Histological data of axons, astrocytes, and myelin in deep subcortical white matter populations,” Data in brief, vol. 23, p. 103 762, 2019. [pdf]
S. Coelho, J. M. Pozo, M. Costantini, J. R. Highley, M. Mozumder, J. E. Simpson, P. G. Ince, and A. F. Frangi, “Local volume fraction distributions of axons, astrocytes, and myelin in deep subcortical white matter,” Neuroimage, vol. 179, pp. 275–287, 2018. [pdf]
Awards & Distinctions
2019
Scholarship Futur & Ruptures - Institut Mines Télécom; Fondation Mines-Télécom; Carnot Télécom & Société numérique, France
Selected for the Seeds for the Future program, French delegation (12 students) - Huawei Technologies, China
2017
LABEX Master Grant - UCN@Sophia Laboratory of Excellence, France
2015
Award to the Best Engineering Graduates from Argentine Universities - National Academy of Engineering of the Argentine Republic
Award to the Best Graduate of Electronic Engineering - Faculty of Engineering of the National Univ. of La Plata (FiUNLP), Argentina
Joaquín V. González Award to the Graduate with the Best Score from the FiUNLP - Municipality of La Plata, Argentina
2014
Scientific Vocation Incentive Scholarships 2014 (EVC) - Interuniversity National Council (CIN), Argentina
External Resources
Here is a personal selection of a few corners of the Internet that I find useful, interesting and fun. Enjoy!
Maybe I'm stating the obvious, but you should check the blogs of Francis Bach, Sébastien Bubeck and Andrej Karpathy for great posts on optimization and ML.
For posts on a plethora of topics including maths, programming and computer science check the blog of Jeremy Kun.
Follow Gabriel Peyré to enrich your daily newsfeed with "one tweet a day on computational mathematics".
Don't miss this article on Why Momentum Really Works, by Gabriel Goh.
If you have doubts about the non-convexity of Neural Networks, check this loss visualization tool, by Ankur Mohan.
And you probably want to spend come time playing with this Interactive Tutorial on Numerical Optimization, by Ben Frederickson.
The picture in the banner was taken by me and is from my hometown, San Carlos de Bariloche
Last updated in October 2022