Organizers: Paul Irofti, Laurențiu Leuștean and Andrei Pătraşcu

The LOS Seminar is the working seminar of the LOS research center.

The seminars are online. All seminars, except where otherwise indicated, will be Tuesdays at 14:00.

To receive announcements about the seminar, please send an email to

Tuesday, May 11, 2021

Andrei Pătraşcu (LOS)
On complexity of the first-order algorithms for convex optimization

Abstract: The potential of the first-order algorithms in the big data context has been confirmed repeatedly in the last decades, in fields like machine learning, signal processing and computer science. In this seminar we will introduce and analyze the main first-order methods successfully used to solve many convex optimization models under various smoothness and convexity features. Our analysis will provide insights on their iterative behaviour and iteration complexity for deterministic and stochastic (noisy) settings. Finally, some concrete applications will be discussed.

Tuesday, April 20, 2021

Cristi Rusu (LOS)
Learning orthonormal dictionaries as efficient as the Fast Fourier Transform

Abstract: Dictionary learning is a popular technique in signal processing and machine learning used to build simple or efficient representations of data. In this talk, we will look at this problem when the dictionary we construct is constrained to be orthonormal. We discuss why is property is desirable and how to introduce additional properties such as numerical efficiency. The goal is to show how to learn transformations from data that are at least as efficient as the Fast Fourier Transform (FFT). We will discuss applications and extensions of this work.

Tuesday, April 6, 2021

Florin Stoican (Polytechnic University of Bucharest and LOS)
About the use of B-spline functions in motion planning

Abstract: B-spline functions are a popular choice for describing trajectories associated with nonlinear dynamics and for imposing continuous-time constraint validation. Exploiting their properties (local support, convexity and positivity) usually leads to sufficient conditions which are conservative. In this talk I will present two methods which reduce the conservatism: the first makes use of sum-of-squares polynomials to provide a linear matrix inequality-type formulation giving necessary and sufficient conditions and the second makes use of knot refinement strategies to improve arbitrarily -well on the sufficient conditions. In both cases, the approaches are validated for a motion planning problem where off-line trajectories with obstacle(s) avoidance guarantees are generated.