All Years Seminars
[INMA] 2022-03-22 (14h) : Learning from pairwise comparisons and gradient representations of neural networks
At Euler building (room a.002)
Speaker :
Bálint Daróczy (UCLouvain (INMA))
Abstract : During the seminar we will consider two separate problems. First, we study of learning problems in which we would like to learn intrinsic values of objects based on pairwise comparisons. We suggest an algorithm and deterime a minimax rate and show that both the upper and the lower error bounds are connected to the trace of the Moore-Penrose inverse of the weighted Laplacian of the comparison graph. In the second part of the seminar, we consider feed-forward neural networks and investigate how the tangent space of the network can be exploited to refine the decision in case of ReLU (Rectified Linear Unit) activations by forming a similarity function with a Riemannian metric.
[INMA] 2022-03-15 (14h) : Certified learning, or learning for verification?
At Euler building (room a.002)
Speaker :
Alessandro Abate (University of Oxford)
Abstract : We are witnessing an increased, inter-disciplinary convergence between areas underpinned by model-based reasoning and by data-driven learning. Work across these areas is not only scientifically justified, but also motivated by industrial applications where access to information-rich data has to be traded off with a demand for safety criticality: cyber-physical systems are exemplar applications. In this talk, I will report on ongoing initiatives in this cross-disciplinary domain. According to the dual perspective in the title of this talk, I will sketch, on the one hand, results where formal methods can provide certificates to learning algorithms, and on the other hand, results where learning can bolster formal verification and strategy synthesis objectives.
[INMA] 2022-03-08 (14h) : Welcome Seminar
At Euler building (room a.002)
Section 1:Towards tight convergence rates for optimization methods on hypoconvex functions
Speaker :
Rotaru, Teodor
Abstract : Using the framework of performance estimation (PEP), we established the first tight convergence rates of the gradient method for smooth hypoconvex (or weakly-convex) functions. These functions’ curvature (i.e., maximum Hessian eigenvalue) belongs to the interval [µ, L], where µ is negative and L is positive. With the help of PEP, we obtained mathematical proofs for a large range of step sizes. As a direct application, we recommend the optimal step size that minimizes the convergence rate.
Section 2:Accelerating large-scale Kernel Support Vector Machines
Speaker :
Sofiane Tanji (UCLouvain)
Abstract : Kernel methods provide an elegant extension to well-known linear statistical learning. Due to their poor scalability in time and memory, they have limited applications in large scale learning. We propose Snacks : a kernel SVM solver which can tackle large-scale datasets. Our approach consists in using an accelerated version of the stochastic subgradient method to solve the primal optimization problem on smaller random subspaces. The computational savings do not lead to any degradation in terms of learning performance and we demonstrate the effectiveness of the proposed algorithm on benchmark datasets.
[INMA] 2022-03-01 (14h) : Exploring geometry for deep learning
At Euler building (room a.002)
Speaker :
Massart, Estelle
Abstract : We propose to use stochastic Riemannian coordinate descent on the orthogonal group for recurrent neural network training. The algorithm rotates successively two columns of the recurrent matrix, an operation that can be efficiently implemented as a multiplication by a Givens matrix. In the case when the coordinate is selected uniformly at random at each iteration, we prove the convergence of the proposed algorithm under standard assumptions on the loss function, stepsize and minibatch noise. In addition, we numerically demonstrate that the Riemannian gradient in recurrent neural network training has an approximately sparse structure. Leveraging this observation, we propose a faster variant of the proposed algorithm that relies on the Gauss-Southwell rule. Experiments on a benchmark recurrent neural network training problem are presented to demonstrate the effectiveness of the proposed algorithm. Time permitting, I will also present another application of geometry in deep learning, namely, exploiting invariant curves in the loss landscape when training Wasserstein generative adversarial neural networks.
[INMA] 2022-02-22 (14h) : Adaptive oscillator as a template for the assistance of cyclical movements.
At Euler building (room a.002)
Speaker :
Ronsse, Renaud
Abstract : In this lecture, I will review the theory of adaptive oscillators and how this tool has been initially developed by generating primitive locomotion movements in robotics. Then I will explain how this framework can be extended to provide different types of assistance in cyclical movements (mainly for the lower-limb, but not exclusively). I will further illustrate the coding of a simple adaptive oscillator on a haptic device that can be manipulated by the participants.
Seminars
Subscribe (ICS)
INMA Contact Info
Mathematical Engineering (INMA)
L4.05.01
Avenue Georges Lemaître, 4
+32 10 47 80 36
secretaire-inma@uclouvain.be
Mon – Fri 9:00A.M. – 5:00P.M.
JOIN US