# IS22 Prize Spotlight

Congratulations to Lénaïc Chizat, the 2022 recipient of the SIAM Activity Group on Imaging Science Best Paper Prize, and Nicolas Keriven, the 2022 recipient of the SIAM Activity Group on Imaging Science Early Career Prize. Both prizes will be presented at the 2022 SIAM Conference on Imaging Science (IS22) to be held in a virtual format on March 21 – 25, 2022.

Lénaïc Chizat

Lénaïc Chizat is the 2022 recipient of the SIAM Activity Group on Imaging Science Best Paper Prize which will be awarded at the 2022 SIAM Conference on Imaging Science (IS22). Chizat will give a talk at the conference titled, “Sparse Optimization on Measures with Over-parameterized Gradient Descent” on Wednesday, March 23 at 7:00 a.m. ET.

The prize was awarded to Chizat for his paper, "Sparse optimization on measures with over-parameterized gradient descent” for laying out the mathematical foundations of particles-based methods for off-the-grid sparse regularization.

The SIAM Activity Group on Imaging Science awards this prize every two years to the author(s) of the most outstanding paper on mathematical and computational aspects of imaging published within the four calendar years preceding the year prior to the award year.

The term “imaging” can be broadly interpreted to include image formation, inverse problems in imaging, image processing, image analysis, image interpretation and understanding, computer graphics, and visualization.

Chizat is a tenure track assistant professor at EPFL in the Institute of Mathematics, where he leads the DOLA chair (Dynamics of Learning Algorithms). Until 2021, he was a CNRS researcher at Laboratoire de mathématiques d’Orsay in France. In his research, he studies and develops optimization algorithms for machine learning and signal processing. In particular, he has contributed to the following fields: optimization in the space of measures, theory of optimal transport, and the analysis of artificial neural networks.

**Q: Why are you excited to receive the SIAG/IS Best Paper Prize?**

**A:** I have worked hard on this problem for months, while enjoying the structure that was unfolding before my eyes, so it is awesome to realize that some fellow researchers have also appreciated the output! Above all, I am grateful that the SIAM community gives me the occasion to share this work to a large and diverse audience during the conference.

**Q: Could you tell us a bit about the research that won you the prize?**

**A:** Suppose a telescope takes a blurry and noisy picture of a cluster of stars. One wishes to recover the exact number, position and intensity of the stars from this picture. Statisticians have studied this problem and shown that if the stars are not too close to each other, the ground truth is simply given by the configuration of stars which would lead to a similar observation while not being too complex (e.g., not too many stars).

In the paper, I show that this optimization problem can be solved by starting from a large population of particles (i.e., fictional stars) and run gradient descent on their position and intensity: as the algorithm runs, these particles will coalesce and recover the correct configuration of stars.

Here I presented the work in a specific context, but this idea and its analysis apply in many settings. In particular, it helps understanding certain effects observed when training artificial neural networks. The common mathematical problem behind these seemingly unrelated situations is that of minimizing a convex functional over the space of measures.

**Q: What does your work mean to the public?**

**A:** I am interested in algorithms that extract structure from data. I like to focus my attention toward those algorithms that are so natural that they are often already used as heuristics in practical contexts. Through a theoretical analysis, we understand when and how these algorithms work. Often, this suggests simple modifications of the heuristic which may have a strong practical impact. For instance, in this work, the analysis suggests two ideas: the use of overparameterization and of multiplicative updates; given these two modifications, we switch from a wobbly heuristic to an algorithm with convergence and efficiency guarantees.

**Q: What does being a member of SIAM mean to you?**

**A:** SIAM plays a central role in the animation of the scientific activity around my research themes. In particular, SIAM journals and conferences are unique venues to stay in touch with the international community and exchange ideas. To me, it is only natural to support this society, which is part of what makes our research field active and alive.

Nicolas Keriven

Nicolas Keriven is the 2022 recipient of the SIAM Activity Group on Imaging Science Early Career Prize, which will be awarded at the 2022 SIAM Conference on Imaging Science (IS22). Keriven will give a talk at the conference titled, “Graph Neural Networks on Large Random Graphs: Convergence, Stability, Universality” on Wednesday, March 23 at 3:00 p.m. ET.

The prize was awarded to Keriven for his novel and deep work in signal processing, imaging, optimization, learning, and data science.

The SIAM Activity Group on Imaging Science awards this prize every two years to one individual in their early career, in the field of imaging science, for distinguished contributions to the field in the three calendar years prior to the award year.

Keriven is a CNRS researcher at GIPSA-lab in Grenoble, France. He studied at Ecole Polytechnique in Palaiseau and ENS Cachan, before completing his Ph.D. with Rémi Gribonval in IRISA Rennes. From 2017 to 2019, he was a post-doctorate with Gabriel Peyré at ENS Paris, on the CFM-ENS "Laplace" chair in Data Science. He received the Best Paper Award at SPARS2017 and is currently the PI of ANR project GRandMa on Random Graphs in Machine Learning.

**Q: Why are you excited to receive the SIAG/IS Early Career Prize?**

**A:** I feel honored and humbled to receive this year's SIAG/IS Early Career Prize, as I have no doubt that all the nominated candidates were brilliant. I feel particularly grateful that this recognition comes after two years of pandemic, during which doing research and interacting with other researchers was difficult. I want to thank my fantastic collaborators and students for working together through these difficult times.

**Q: Could you tell us a bit about the research that won you the prize?**

**A:** I am interested in exploring how the structure of the data influences machine learning algorithms, in particular, with the objective of identifying generalized notions of low-dimensional models or regularity and exploiting them to improve the algorithms in various ways. For the last few years, I have particularly focused on "non-Euclidean" data such as graphs and manifolds. Part of my nomination was a series of two NeurIPS papers exploring the relationship between Graph Neural Networks, which are state-of-the-art deep learning architectures on graphs, and models of random graphs, which are fundamental tools in graph theory often used in social network or mesh analysis. It helps understand their properties on large graphs and exploit the underlying random structures to improve them.

**Q: What does your work mean to the public?**

**A:** Deep learning architectures are increasingly part of many real-world systems. Born primarily from the imaging sciences, they have the potential to be applied to many types of data. My work consists in studying their relationship with exotic data structures, to better understand different use cases, what they are capable of, and reduce their energy footprint with minimal impact on performance.

**Q: What does being a member of SIAM mean to you?**

**A:** In a time where we realize that connecting with other researchers is more than ever an essential part or our activities, active communities and organizations such as SIAM are crucial. SIAM has always offered a wealth of great research and brilliant people to connect with, and I am grateful and humbled by this prize.