BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Memento EPFL// BEGIN:VEVENT SUMMARY:Addressing Computational and Statistical Gaps with Deep Neural Net works DTSTART:20161007T111500 DTEND:20161007T123000 DTSTAMP:20260413T044442Z UID:db31504552146a046627af8a3322d80fdcf7ee33828e54cb2d49aa4e CATEGORIES:Conferences - Seminars DESCRIPTION:Joan BRUNA    \nMany modern statistical questions are plag ued with asymptotic regimes that separate our current theoretical understa nding with what is possible given finite computational and sample resource s. Important examples of such gaps appear in sparse inference\, high-dimen sional density estimation and non-convex optimization.\n\nIn the former\, proximal splitting algorithms efficiently solve the l1-relaxed sparse codi ng problem\, but their performance is typically evaluated in terms of asym ptotic convergence rates. In unsupervised high-dimensional learning\, a ma jor challenge is how to appropriately combine prior knowledge in order to beat the curse of dimensionality. Finally\, the prevailing dichotomy betwe en convex and non-convex optimization is not adapted to describe the diver sity\nof optimization scenarios faced as soon as convexity fails. In this talk we will illustrate how Deep architectures can be used in order to att ack such gaps. We will first see how a neural network sparse coding model (LISTA\, Gregor & LeCun’10) can be analyzed in terms of a particular mat rix factorization of the\ndictionary\, which leverages diagonalisation wit h invariance of the l1 ball\, revealing a phase transition that is consist ent with numerical experiments. We will then discuss image and texture gen erative modeling and super-resolution\, a prime example of high-dimensiona l inverse problem. In that setting\, we will\nexplain how multi-scale conv olutional neural networks are equipped to beat the curse of dimensionality and provide stable estimation of high frequency information. Finally\, we will discuss recent research in which we explore to what extent the non-c onvexity of the loss surface arising in deep learning problems is hurting gradient descent algorithms\, by efficiently estimating the number of basi ns of attractions.\n\nBio: Joan graduated cum-laude from Universitat Polit ècnica de Catalunya in both Mathematics and Telecommunications Engineerin g\, before graduating in Applied  Mathematics from ENS Cachan (France). H e then became a Sr. Research Engineer in an Image Processing startup\, dev eloping real-time video processing  algorithms. He obtained his PhD in Ap plied Mathematics at École Polytechnique (France). After a postdoctoral s tay at Courant Institute\, NYU\, he became a Postdoctoral fellow at Facebo ok AI Research. He is an Assistant Professor at UC Berkeley\, Statistics D epartment (on leave)\, and since Fall 2016 he is Assistant Professor at Co urant Institute\, NYU (Computer Science\, Center for Data Science and Math ematics (courtesy)).\n\nHis research interests include invariant signal re presentations\, deep learning\, high-dimensional statistics\, and its appl ications to computer vision\, statistical physics and AI. LOCATION:CM 1 105 http://plan.epfl.ch/?request_locale=fr&room=CM+1+105&dom ain=places STATUS:CONFIRMED END:VEVENT END:VCALENDAR