On the afternoon of September 22nd and 23rd, the "HUST-UPSaclay Workshop on Mathematics for Data Science" was held online. Prof. Pascal, Prof. Sheng Yang, Prof. Gramfort, Researcher Chazal, Researcher Chouzenoux, Researcher Mazanti, and Researcher Pfeiffer, all of whom are from UPSaclay, and Prof. Gao Ting from HUST delivered speeches on recent advances in "Mathematical for Data Science".
The workshop, live-streamed both on Zoom and AI TIME, attracted over 40,000 viewers during the two half-days.
The workshop on September 22nd was hosted by Zhenyu Liao from School of Electronic Information and Communications (EIC) of HUST. First, Prof. Qiu Caiming, the Dean of EIC, delivered an opening speech. He extended his warm welcome to the experts from UPSaclay and looked forward to more in-depth academic exchanges and cooperation between the two universities in the field of "Mathematics + Information".
Next, Prof. Pascal gave a presentation on " Robust statistics and clustering - Application to signal and image processing", in which he studied the unsupervised clustering problem in machine learning and proposed a novel and flexible F-EM algorithm for elliptically symmetric distribution families of data. Furthermore, he proved some statistical theoretical properties of the algorithm, and further demonstrated the well performance of the algorithm through numerical experiments. Also, the algorithm was applied to PoLSAR image processing problems.
Prof. Sheng Yang made a speech on “Context-Tree-Based Lossy Compression". He proposed an effective lossy variable compression scheme and designed a quantized compander. Uniform quantization was achieved by introducing a nonlinear transformation, and low-loss index series compression was made based on context-tree.
Gramfort delivered a presentation on "Learning without labels on multivariate bio-signals: From unsupervised to self-supervised learning". He proposed to study label-free multivariate bio-signals with unsupervised and self-supervised learning. He also explored the evolution from the unsupervised learning to the supervised one, and then presented the applications of the proposed methods.
Prof. Ting Gao from HUST gave a presentation on "Identifying, prediction, and control in non-Gaussian stochastic dynamical systems", in which she comprehensively discussed the application of SDEs in machine learning from three aspects: learning the system identifiability for SDEs with Levy noise, discovering the transition phenomenon from data, and predicting with complex datasets by SDEs.
The workshop on September 23rd was chaired by Prof. Yacine Chitour who works at L2S Laboratory of UPSaclay.
Chazal made a presentation themed on "Topological Data Analysis to improve learning models: an introduction and a few examples". In his work, for the machine learning and artificial intelligence, he presented a concrete and effective TDA approach, which is well-grounded in mathematics, and discussed its robustness and interpretability. Meanwhile, he pointed out that the approach can be effectively combined with other machine learning and artificial intelligence techniques.
Chouzenoux delivered a speech on "Unrolled Variational Bayesian Algorithm for Image Blind Deconvolution". Based on the variational Bayesian algorithm and the deep neural network architecture, she proposed a novel blind image deconvolution, the corresponding estimation of which is accurate both in image and the blur kernels. The approach is also competitive in computational efficiency.
Mazanti gave a presentation themed on "Modelling crowd behavior through mean field games". He introduced some recent theoretical advances in crowd movement through mean field games. In contrast to other work, this macroscopic model considers the strategic choices of pedestrians, and predicts the future behavior of other pedestrians based on their experience, which can be used to choose their trajectories. Particularly, in this report, he illustrated the existence of steady states for this kind of mean-field game and described them through a system of partial differential equations.
Pfeiffer delivered a presentation on "Conditional gradient method for mean-field type problems". He proposed an effective generalized conditional gradient algorithm that can be applied to potential mean-field games. Also, this approach can be interpreted as a learning method called fictitious play.
The workshop successfully promotes academic exchanges and cooperation between HUST and UPSaclay, and helps students broaden their horizon.
