BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Memento EPFL// BEGIN:VEVENT SUMMARY:Fall 2023 SDSC PhD fellows workshop DTSTART:20231128T102500 DTEND:20231128T162000 DTSTAMP:20260502T134707Z UID:f96876a1ff89fe19dabcd44c780116b480fbb71268b67b92f2dfaabb CATEGORIES:Conferences - Seminars DESCRIPTION:Martin Josifoski\, Laura Manduchi\, Clément Vignac\, Andreas Schlaginhaufen\, Aditya Varre\nSince 2019\, the Swiss Data Science Center (SDSC) has awarded fellowships to 19 PhD students working on various aspec ts of data science theory and methods at EPFL and ETH Zürich\, from stati stics to computer vision\, including machine learning\, statistical signal processing\, and natural language processing.\n\nWe are pleased to invite you to the PhD fellows' workshop of Fall 2023. These workshops are organi zed once per semester\, alternating between EPFL and ETH Zürich\, to prov ide a forum of exchange on methods in data science and an opportunity for five of these fellows to present their work in person. The program will al so feature in the afternoon four short presentations of SDSC collaborative projects. Everyone from the EPFL community is welcome to attend.\n\nDate and Time: Tuesday 28 November 2023\, 10:25am\nRoom: BC 420\n\nProgram (ple ase find the abstract below):\n\n 10:25 - Foreword and session opening  – Dr. Guillaume Obozinski\, Deputy Chief Data Scientist\, SDSC\n 10:30 - Martin Josifoski\, Data Science Lab\, EPFL: \n \n Symbolic Intermediate Representations for Reasoning and Collaborating AI\n \n \n 11:10 - Laura Manduchi\, Medical Data Science group\, ETH Zürich: \n \n Tree Variati onal Autoencoders\n \n \n 12:00 - Lunch\n 13:00 - Clément Vignac\, LTS4\, EPFL: \n \n Discrete denoising diffusion models for graph generation\n \n \n 13:40 - Short presentations of SDSC collaborative projects:\n \n Tr ansformer Models for the Estimation of Transit Time in Watersheds\, Quenti n Duchemin\, SDSC.\n Self-guided ML Algorithms for Real-Time Assimilation \, Interpolation and Rendering of Flow Data\, Victor Cohen\, SDSC.\n Prob abilistic State Space Models for Predicting Turbulence on Pitching Airfoil s\, Christian Donner\, SDSC.\n A Graph Matching Approach to Tracking Neur ons in Freely-Moving C. elegans\, Corinne Jones\, SDSC.\n \n \n 14:30 - C offee break\n 14:50 - Andreas Schlaginhaufen\, Sycamore\, EPFL:\n \n Iden tifiability and Transferability of Rewards in Inverse Reinforcement Learni ng\n \n \n 15:30 - Aditya Varre\, TML\, EPFL\,\n \n On the spectral bias of two-layer linear networks\n \n \n 16:10 - Closing words\n\n\nAbstracts :\n\n\n Tree Variational Autoencoders\, Laura Manduchi\, Medical Data Scie nce group\, ETH Zürich\n\n \n Abstract: In this talk\, I will present t he Tree Variational Autoencoder (TreeVAE)\, a new generative hierarchical clustering model that learns a flexible tree-based posterior distribution over latent variables. It adapts its architecture to discover the optimal tree for encoding dependencies between latent variables and it hierarchica lly divides samples according to their intrinsic characteristics\, sheddin g light on hidden structures in the data.\n \n \n Discrete denoising diffu sion models for graph generation\,  Clément Vignac\, LTS4\, EPFL\n \n A bstract: Graph generation has posed significant challenges in the past yea rs. Traditional methods\, such as Variational Autoencoders (VAEs)\, Genera tive Adversarial Networks (GANs)\, and Normalizing Flows\, have been prima rily limited to generating small graphs\, and often evaluated on molecules composed of no more than 9 atoms. In this talk\, we introduce DiGress\, a denoising diffusion model that scales to substantially larger graphs\, as well as its extension MiDi for molecule generation in 3D. These models le verage a discrete diffusion process progressively edits graphs with noise\ , through the process of adding or removing edges and changing the node fe atures. In contrast to Gaussian noise\, the use of a discrete diffusion pr ocess allows to preserve the sparsity of graphs\, which allows for various architectural improvements. We show experimentally the effectiveness of t his approach\, and propose hypotheses in order to explain why these models outperform GANs and VAEs by such a large margin.\n References:\n \n D iGress: Discrete denoising diffusion for graph generation — (ICLR 2023)\ n MiDi: Mixed Graph and 3D Denoising diffusion for molecule generation — (ECML 2023)\n \n \n \n \n Identifiability and Transferability of Rew ards in Regularized Inverse Reinforcement Learning\, Andreas Schlaginhaufe n\, Sycamore\, EPFL\n \n Abstract: A common reason for favoring inverse r einforcement learning (IRL) over imitation learning is that rewards offer a more succinct and transferable description of a task. However\, due to t he ill-posedness of the inverse problem\, it is widely unclear whether rew ards learned via IRL are indeed transferable to new environments. In this talk\, we focus on regularized IRL and provide sufficient conditions for i dentifiability and transferability to new environments.\n \n \n On the spe ctral bias of two-layer linear networks\, Aditya Varre\, TML\, EPFL\n \n Abstract: In this talk I will present the behaviour of two-layer fully con nected networks with linear activations trained with gradient flow on the square loss. We further discuss how the opti- mization process carries an implicit bias on the parameters that depends on the scale of its initializ ation. The main result is a variational characteriza- tion of the loss min imizers retrieved by the gradient flow for a specific initialization shape . This characterization reveals that\, in the small scale initialization r egime\, the linear neural network’s hidden layer is biased toward having a low-rank struc- ture. To complement our results\, we present a hidden m irror flow that tracks the dynamics of the singular values of the weights matrices and describe their time evolution. We support our findings with n umerical experiments illustrating the phenomena.\n \n \n LOCATION:BC 420 https://plan.epfl.ch/?room==BC%20420 STATUS:CONFIRMED END:VEVENT END:VCALENDAR