Ekaterina Iakovleva

I am currently a postdoctoral researcher working on fairness and maching unlearning with Enzo Tartaglione. I am a member of Multimedia team at Télécom Paris, Institut Polytechnique de Paris. Before that I worked in the same team with Stéphane Lathuilière on zero-shot application of LLM reasoning to image editing diffusion models.

My main research interests are efficient and trustworthy AI. The questions that intrigue me most are what internal knowledge do neural networks actually learn and how does their behavior change when we act on them. I study these questions from various points of view: adaptability of pre-trained neural networks for generalisation or debiasing purposes, presence of subnetworks that might preserve undesirable knowledge about biases or unsafe data, or influence of hyperparameters on the presence of bias. In my search for possible solutions I prefer to focus on methods that are computationally and/or data efficient, such as learning small network adaptors or compressing neural networks by pruning.

Previously, I was a doctoral student in THOTH team at Inria Grenoble. I recieved my Ph.D. in Mathematics and Informatics at Université Grenoble Alpes in 2022, advised by Jakob Verbeek and Karteek Alahari. My thesis focuses on Transfer Learning in the presence of limited and non-stationary data, proposing computationally-efficient and data-efficient methods for adapting pre-trained neural networks to novel tasks and data domains.

Before my doctoral studies, I recieved a master’s degree in Industrial and Applied Mathematics (“mention très bien”) at Institut Polytechnique de Grenoble in 2018 supervised by Jakob Verbeek, when I worked on factorising autoregressive generative image models. I also recieved a double master’s degree (“with distinction”) in Computer Science at Skoltech and in Applied Mathematics and Physcis at Moscow Institute of Physcis and Technology in 2017 supervised by Dmitry Vetrov, when I worked on applications of Bayesian methods to deep learning.