(photo by Angeline Swinkels)
(photo by Angeline Swinkels)
Starting September 2025 I'm a postdoc at KdVI, University of Amsterdam working with Maximilian Engel. I did my PhD at TU Eindhoven under the supervision of Mark Peletier and André Schlichting (Uni Ulm). I'm interested in understanding behaviour of machine learning models (and in all kinds of mathematical tools which can help me get the job done). So far the tools included stochastic analysis, aggregation-diffusion PDEs and optimal transport theory.
Email: a [dot] shalova [at] uva [dot] nl
or a [dot] shalova [at] tue [dot] nl
Scholar: link
My PhD thesis: Noisy gradient flows with applications in machine learning, 2025.
Stationary solutions of McKean-Vlasov equation on a high-dimensional sphere and other Riemannian manifolds, (with A. Schlichting). arXiv:2412.14813
Singular-limit analysis of gradient descent with noise injection, (with A. Schlichting and M. Peletier). arXiv:2404.12293
"Nonlinear diffusion limit of non-local interactions on a sphere", Workshop on Random Dynamical Systems and Related Fields, FU Berlin (DE)
Talk at DESY about Transformers and Diffusions, Hamburg (DE)
"Noisy gradient flows in machine learning: two examples" (in two parts), MFO mini-workshop "Probabilistic Perspectives in Neural Network-Based Machine Learning", Oberwolfach (DE)
"Nonlinear diffusion limit of non-local interactions on a sphere", Applied Analysis seminar, University of Ulm (DE)
"Porous medium is the message: variational analysis of toy transformers", European Conference on Computational Optimization 2025, Focus session Optimization and Machine Learning, Alpen-Adria-Universitaet Klagenfurt (Austria)
"Porous medium is the message: variational analysis of toy transformers", 2nd Workshop on Machine Learning in Infinite Dimensions, ETH Zürich (Switzerland)
"Porous medium is the message: variational analysis of toy transformers", FS Stochastische Analysis und Stochastik der Finanzmärkte, Institut für Mathematik, TU Berlin (DE)
"Porous medium is the message: variational analysis of toy transformers", AG Analysis-Probability, MPI for Mathematics in the Sciences, Leipzig (DE)
"Stationary measures of noisy transformers", Oberseminar "Mathematik des Maschinellen Lernens und Angewandte Analysis", University of Würzburg (DE)
"Stationary measures of noisy transformers", Workshop on the Statistical Theory of Neural Networks, Egmond aan Zee (NL)
"Solutions of stationary McKean-Vlasov equation on a sphere", Colloquium of the Institute of Mathematical Statistics and Actuarial Science, University of Bern (Switzerland)
"Solutions of stationary McKean-Vlasov equation on a sphere", seminar "Modern Methods in Applied Stochastics and Nonparametric Statistics", WIAS Berlin (DE)
"Stationary measures of noisy transformers", Physics of AI algorithms 🔥, Les Houches (FR)
"Singular-limit analysis of gradient descent with noise injection", UNRAVEL final workshop, Eindhoven (NL)
"Bifurcations and phase transitions of McKean-Vlasov equation on a sphere", Applied Analysis seminar, University of Ulm (DE)
"Singular-limit analysis of gradient descent with noise injection", Algorithms & Computationally Intensive Inference seminar, University of Warwick (UK)
"Singular-limit analysis of gradient descent with noise injection", workshop Statistical and Probabilistic Analysis of Random Networks and Processes, Université Côte d’Azur (FR)
"Stationary measures of noisy transformers", the Workshop CalcVarNL 2024, Schiermonnicoog (NL)
"Choosing the right noise: regularization properties of noise injection", Mathematics of Data Science seminar, University of Twente (NL)
"Regularization properties of noise injection", MFO workshop "Applied Harmonic Analysis and Data Science", Oberwolfach (DE)
"Regularization properties of dropout", workshop "The Mathematics of Machine Learning", Pisa (IT)
On 17-18 Nov 2025 I was visiting DESY, Hamburg. Thanks to Tim Roith for hosting me!
On 29 Sep - 1 Oct 2025 together with Evelyn Herberg we organized a Focus Session on Optimization and Machine Learning at the European Conference on Computational Optimization 2025 (EUCCO 2025). Thanks to all the speakers for the great talks!
12 September 2025: I am officially a doctor now. We also had a lovely little workshop yesterday. Big thank you to the members of my defense committee for their efforts and interesting talks.
On 25-27 August 2025 I was happy to join the workshop Optimization unplugged at EPFL Bernoulli center, Lausanne. Thanks to the organisers for the invitation.
14-17 July 2025: I was visiting Max Planck Institute for Mathematics in the Sciences. Thanks to Benjamin Gess for the invitation.
On 16-20 June 2025 I had a lot of fun finishing my thesis at the Special Topic School: "Statistical mechanics of spin glasses, neural networks and learning" at the Hausdorff School for Mathematics.
26-29 May 2025: I was visiting University of Würzburg, thanks to Leon Bungert and Eloi Martinet for inviting me.
23 Apr 2025: I was happy to be at NMC.
Together with Evelyn Herberg we organize a Focus Session on Optimization and Machine Learning at the European Conference on Computational Optimization 2025 (EUCCO 2025).
3-8 Mar 2025: I was visiting Andrea Agazzi at the University of Bern.
19 Dec 2024: New preprint "Solutions of stationary McKean-Vlasov equation on a high-dimensional sphere and other Riemannian manifolds" is available online. This is a joint work with André Schlichting.
from 29 Nov to 4 Dec 2024: I was visiting André Schlichting at the University of Ulm
24-29 Nov 2024: I was excited to join Oberwolfach Seminar: Control and Machine Learning
20-22 Nov 2024: I was visiting Matthew Thorpe at the University of Warwick
7 June 2024: Dirk Burgers defended his bachelor thesis titled "Training behavior of 2-layer ReLU neural networks in the condensed regime". I was a co-supervisor of Dirk's final project together with Mark Peletier and Han Slot.
18 Apr 2024: New preprint "Singular-limit analysis of gradient descent with noise injection" is available online. This is a joint work with André Schlichting and Mark Peletier.