Zofia Trstanova

Stephane Redon, Gabriel Stoltz, Zofia Trstanova

We consider Langevin dynamics associated with a modified kinetic energy vanishing for small momenta. This allows us to freeze slow particles, and hence avoid the re-computation of inter-particle forces, which leads to computational gains. On the other hand, the statistical error may increase since there are a priori more correlations in time. The aim of this work is first to prove the ergodicity of the modified Langevin dynamics (which fails to be hypoelliptic), and next to analyze how the asymptotic variance on ergodic averages depends on the parameters of the modified kinetic energy. Numerical results illustrate the approach, both for low-dimensional systems where we resort to a Galerkin approximation of the generator, and for more realistic systems using Monte Carlo simulations.

Zofia Trstanova, Nadir El Manouzi, Maryline Chen et al.

In today's world, the presence of online disinformation and propaganda is more widespread than ever. Independent publishers are funded mostly via digital advertising, which is unfortunately also the case for those publishing disinformation content. The question of how to remove such publishers from advertising inventory has long been ignored, despite the negative impact on the open internet. In this work, we make the first step towards quickly detecting and red-flagging websites that potentially manipulate the public with disinformation. We build a machine learning model based on multilingual text embeddings that first determines whether the page mentions a topic of interest, then estimates the likelihood of the content being malicious, creating a shortlist of publishers that will be reviewed by human experts. Our system empowers internal teams to proactively, rather than defensively, blacklist unsafe content, thus protecting the reputation of the advertisement provider.

Frederik Heber, Zofia Trstanova, Benedict Leimkuhler

With the advent of GPU-assisted hardware and maturing high-efficiency software platforms such as TensorFlow and PyTorch, Bayesian posterior sampling for neural networks becomes plausible. In this article we discuss Bayesian parametrization in machine learning based on Markov Chain Monte Carlo methods, specifically discretized stochastic differential equations such as Langevin dynamics and extended system methods in which an ensemble of walkers is employed to enhance sampling. We provide a glimpse of the potential of the sampling-intensive approach by studying (and visualizing) the loss landscape of a neural network applied to the MNIST data set. Moreover, we investigate how the sampling efficiency itself can be significantly enhanced through an ensemble quasi-Newton preconditioning method. This article accompanies the release of a new TensorFlow software package, the Thermodynamic Analytics ToolkIt, which is used in the computational experiments.