Organizers: Paul Irofti, Laurențiu Leuștean and Andrei Pătraşcu

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

All seminars, except where otherwise indicated, will be Tuesdays between 14:00 and 15.50. The seminars are online.

To receive announcements about the seminar, please send an email to los@fmi.unibuc.ro.

Tuesday, October 19, 2021

Andra Băltoiu (LOS)
Dictionary Learning Heuristics Control Through Adaptive Strategies and Applications to Anomaly Detection

Abstract: Dictionary learning has two model-dependent heuristics: sparsity level and dictionary size. Both are highly dependent on the signals and there are few applications in which an informed, accurate choice is possible. Particular problems, such as classification, entail further parameters that interfere with learning. We present several adaptive strategies that are aimed at finding the optimal values for these heuristics or at avoiding the choice altogether. We also discuss the implications on anomaly detection, which we treat as a special case of classification.

Tuesday, November 2, 2021

Paul Irofti (LOS)
Anomaly Detection ABC

Abstract: We briefly revisit and discuss the standard methods for anomaly detection, such as One Class SVM, SVDD, Isolation Forrest, and Gaussian Mixtures. For each method we analyze the optimization problem, its anomaly detection applications, and the improvements gained by introducing non-linear transformations to the inputs via kernel methods (connecting the anomaly detection field to Hilbert spaces). The papers of interest can be found in the ABC directory here.

Tuesday, November 9, 2021

Andrei Pătraşcu (LOS)
Inexact Proximal Point Algorithm for Convex Optimization under Holderian Growth

Abstract: Proximal Point Algorithm (PPA) gained a long-lasting attraction for both abstract operator theory and numerical optimization communities. Even in modern applications, researchers still use proximal minimization theory to design scalable algorithms that overcome nonsmoothness. Remarkable works established tight relations between the convergence behaviour of PPA and the regularity of the objective function. In this talk we present nonasymptotic iteration complexity of exact and inexact PPA to minimize convex functions under Holderian growth. In particular, we recover well-known results on PPA: finite convergence for sharp minima and linear convergence for quadratic growth, even under presence of inexactness. However, without taking into account the concrete computational effort paid for computing each PPA iteration, any iteration complexity remains abstract and purely informative. Therefore, based on an inner (proximal) gradient/subgradient method subroutine that computes inexact PPA iteration, we secondly present computational complexity bounds of a restarted inexact PPA, that are available when no information on the growth of the objective function is known.

Tuesday, November 16, 2021

Ruxandra Bălucea (LOS)
Low-level Attacks and Software Mitigations on RISC-V

Abstract: RISC-V is a modern open-source architecture developed in 2013 at the University of Berkeley. It maintains the old features of the classic RISC (Reduced Instruction Set Computer) architectures, but tries to remove the vulnerabilities discovered in the past years. In this context, the main goal of RISC-V is to evolve without compromise, improving the performance of the chips only if this is done completely safely. In this context, we introduce some of the most popular attacks like Buffer Overflow, ROP and we present some variants of Spectre, a famous family of attacks reproduced on a speculative RISC-V core named BOOM. We also present a software mitigation for one of these variants and other future directions.

Tuesday, November 23, 2021

Horațiu Cheval (LOS)
Quadratic rates of asymptotic regularity for the Tikhonov-Mann iteration

Abstract: We compute quadratic rates of asymptotic regularity for the Tikhonov-Mann iteration in W-hyperbolic spaces. This iteration is an extension to a nonlinear setting of the modified Mann iteration defined recently by Bot, Csetnek and Meier in Hilbert spaces. We obtain these quantitative results by applying methods from proof mining.
This is joint work with Laurențiu Leuștean.

Tuesday, December 7, 2021

Mihai Adamescu (University of Bucharest)
Interdisciplinarity and transdisciplinary

Abstract: Ecology, sustainability science, drones, sensors, HPC, HCI, ioT, complexity science, digital humanities, mathematics, and sociology. What do all of them have in common? How can we answer even more complex questions that society has? How can we address some of these questions? Why is our thinking so fragmented? How can we find common ground?
A short presentation of the Research Center in Systems Ecology and Sustainability, the Research Infrastructures (eLTER, LifeWatch, ICOS, BIOSCAN) that we are promoting, and the need to cooperate, including possible cooperation themes.

Tuesday, December 14, 2021

Alin Ştefănescu (University of Bucharest)
River: fuzz testing using AI for binaries and IoT

Abstract: TBA

Past seminars