On Learning Sparsity-Promoting Penalties via Implicit Differentiation for Image Regularization and Robust Estimation

In this talk, the speaker discusses their recent work on sparsity-promoting penalties and their applications in image reconstruction and estimation tasks. Sparsity is a key concept in machine learning and computer vision, enabling interpretable and computationally efficient models.

The speaker highlights the challenge of efficiently training neural networks that enforce sparsity penalties. To address this, they introduce the concept of implicit layers and implicit differentiation, which facilitate the training of very deep neural networks. These methods have been successfully deployed in various tasks, demonstrating their practical utility.

The seminar will be held in English.

Докладчик

Стаматис Лефкиммиатис

Head of the Computer Vision Group MTS AI

Оппонент

Андрей Кузнецов

Кандидат технических наук, директор лаборатории FusionBrain AIRI