日時(Date) |
2024年 2月 19日 (月) / Mon. Feb. 19th, 2024 3限 (13:30--15:00) / 3rd period (13:30--15:00) |
---|---|

場所(Location) | 本部棟研修ホール. Kenshu Hall （Interdisciplinary Frontier Research Complex No.2） |

司会(Chair) | 佐藤, 大竹 |

講演者(Presenter) | Hugues Talbot (Centre for Visual Computing (CVN), CentraleSupélec, Inria, Université Paris-Saclay) |

題目(Title) | Gauss in the 21st century: learning monotone operators for non-linear blind deconvolution |

概要(Abstract) | One of the most important data analysis tool in all of science is the Gaussian least square regression. It has been revisited many times, but one key ingredient is the model that it is used during its optimisation. Typically the model is given in advance : linear, polynomial, etc, and then fitted to the data. Least-square regression is thus a key ingredient in most inverse problem formulations of variational methods. It is used so much that we think of it as the almost unavoidable and ubiquitous MSE loss in machine learning.Recently, deep learning has been introduced in this context in plug-and-play / deep unrolling methods via learning an image model or a noise model with techniques such as Noise2Noise, or Deep Image priors. But as the name implies, this concerns the regularisation term of variational methods and comes with many caveats.In this talk, we propose to revisit the MSE loss in variational problems, also called the data fidelity model. For this, we propose to learn not only a noise model, but also a data production model, which can include motion or defocus blur, as well as non-linear contrast transforms. For this we train a deep network constrained to act as a monotone operator.We will explain this context, why a monotone operator is useful, how to train a deep network to remain monotone, and applications in blind deconvolution. |

講演言語(Language) | English |

講演者紹介(Introduction of Lecturer) | Hugues Talbot received the PhD in Mathematical Morphology in 1993 from School of Mines, Paris and the Habilitation in 2013 from University Paris East. He was a principal research scientist at CSIRO, Sydney, Australia from 1994 to 2004; then a professor at ESIEE, University Paris-East from 2004 to 2018 ; He was the research dean at ESIEE from 2015 to 2018. He joined CentraleSupelec, University Paris-Saclay in 2018 as first-class professor and became a distinguished professor in 2022. His interests include computer vision, medical imaging, optimisation and machine learning. |