Search

Despite substantial progress in multilingual extractive Question Answering (QA), models with high and uniformly distributed performance across languages remain challenging, especially for languages with limited resources. We study cross-lingual transfer mainly focusing on the Generalized Cross-Lingual Transfer (G-XLT) task, where the question language differs from the context language - a challenge that has received limited attention thus far. Our approach seeks to enhance cross-lingual QA transfer using a high-performing multilingual model trained on a large-scale dataset, complemented by a few thousand aligned QA examples across languages. Our proposed strategy combines cross-lingual sampling and advanced self-distillation training in generations to tackle the previous challenge. Notably, we introduce the novel mAP@k coefficients to fine-tune self-knowledge distillation loss, dynamically regulating the teacher's model knowledge to perform a balanced and effective knowledge transfer. We extensively evaluate our approach to assess XLT and G-XLT capabilities in extractive QA. Results reveal that our self-knowledge distillation approach outperforms standard cross-entropy fine-tuning by a significant margin. Importantly, when compared to a strong baseline that leverages a sizeable volume of machine-translated data, our approach shows competitive results despite the considerable challenge of operating within resource-constrained settings, even in zero-shot scenarios. Beyond performance improvements, we offer valuable insights through comprehensive analyses and an ablation study, further substantiating the benefits and constraints of our approach. In essence, we propose a practical solution to improve cross-lingual QA transfer by leveraging a few data resources in an efficient way., Comment: Submitted to the Journal of Artificial Intelligence Research (JAIR)

The determination of seismic event locations with sparse networks or single-borehole systems remains a significant challenge in observational seismology. Leveraging the advantages of the location approach HADES, which was initially developed for locating clustered seismicity recorded at two stations, we present here an improved version of the methodology: HADES-R. Where HADES previously needed a minimum of 4 absolutely located master events, HADES-R solves a least-squares problem to find the relative inter-event distances in the cluster, and uses only a single master event to find the locations of all events, and subsequently applies rotational optimiser to find the cluster orientation. It can leverage iterative station combinations if multiple receivers are available, to describe the cluster shape and orientation uncertainty with a bootstrap approach. The improved method requires P- and S-phase arrival picks, a homogeneous velocity model, a single master event with a known location, and an estimate of the cluster width. The approach is benchmarked on the 2019 Ridgecrest sequence recorded at two stations, and applied to two seismic clusters at the FORGE geothermal test site, including a microseismic monitoring scenario with a DAS in a vertical borehole. Traditional procedures struggle in these settings due to the ill-posed network configuration. The azimuthal ambiguity in this scenario is partially overcome by assuming that all events belong to the same cluster around the master event and a cluster width estimate. We find the cluster shape in both cases, although the orientation remains uncertain. The method's ability to constrain the cluster shape and location with only one well-located event offers promising implications, especially for environments where limited or specialised instrumentation is in use., Comment: This is a pre-copyedited, author-produced version of an article accepted for publication in Geophysical Journal International following peer review. The version of record is available online at: https://academic.oup.com/gji/article/238/2/661/7670630

Optimization-based safety filters, such as control barrier function (CBF) based quadratic programs (QPs), have demonstrated success in controlling autonomous systems to achieve complex goals. These CBF-QPs can be shown to be continuous, but are generally not smooth, let alone continuously differentiable. In this paper, we present a general characterization of smooth safety filters -- smooth controllers that guarantee safety in a minimally invasive fashion -- based on the Implicit Function Theorem. This characterization leads to families of smooth universal formulas for safety-critical controllers that quantify the conservatism of the resulting safety filter, the utility of which is demonstrated through illustrative examples.

The weak mixing angle is a probe of the vector-axial coupling structure of electroweak interactions. It has been measured precisely at the $Z$-pole by experiments at the LEP and SLD colliders, but its energy dependence above $M_Z$ remains unconstrained. In this contribution we propose to exploit measurements of Neutral-Current Drell Yan at large invariant dilepton masses at the Large Hadron Collider, to determine the scale dependence of the weak mixing angle in the $\overline{MS}$ renormalisation scheme, $\sin^2 \theta_w^{\overline{MS}}(\mu)$. Such a measurement can be used to test the Standard Model predictions for the $\overline{MS}$ running at TeV scales, and to set model-independent constraints on new states with electroweak quantum numbers. To this end, we present an implementation of $\sin^2 \theta_w^{\overline{MS}}(\mu)$ in the POWHEG-BOX Monte Carlo event generator, which we use to explore the potential of future analyses with the LHC Run~3 and High-Luminosity datasets. In particular, the impact of the higher order corrections and of the uncertainties due to the knowledge of parton distribution functions are studied., Comment: Presented at DIS2023 - XXX International Workshop on Deep-Inelastic Scattering and related subjects; minor modifications to the references included

In balancing safety with the nominal control objectives, e.g., stabilization, it is desirable to reduce the time period when safety filters are in effect. Inspired by traditional spacecraft maneuvers, and with the ultimate goal of reducing the duration when safety is of concern, this paper proposes an event-triggered control framework with switching state-based triggers. Our first trigger in the scheme monitors safety constraints encoded by barrier functions, and thereby ensures safety without the need to alter the nominal controller--and when the boundary of the safety constraint is approached, the controller drives the system to the region where control actions are not needed. The second trigger condition determines if the safety constraint has improved enough for the success of the first trigger. We begin by motivating this framework for impulsive control systems, e.g., a satellite orbiting an asteroid. We then expand the approach to more general nonlinear system through the use of safety filtered controllers. Simulation results demonstrating satellite orbital maneuvers illustrate the utility of the proposed event-triggered framework., Comment: 8 pages, submitted to IEEE Conference on Decision and Control, 2023

A common assumption on the deployment of safeguarding controllers on the digital platform is that high sampling frequency translates to a small violation of safety. This paper investigates and formalizes this assumption through the lens of Input-to-State Safety. From this perspective, and leveraging control barrier functions (CBFs), we propose an alternative solution for maintaining safety of sample-and-hold controlled systems without any violation to the original safe set. Our approach centers around modulating the sampled control input in order to guarantee a more robust safety condition. We analyze both the time-triggered and the event-triggered sample-and-hold implementations, including the characterization of sampling frequency requirements and trigger conditions. We demonstrate the effectiveness of our approach in the context of adaptive cruise control through simulations., Comment: Submitted to IEEE American Control Conference 2024

An upgrade of the near detector of the T2K long baseline neutrino oscillation experiment is currently being conducted. This upgrade will include two new Time Projection Chambers, each equipped with 16 charge readout resistive Micromegas modules. A procedure to validate the performance of the detectors at different stages of production has been developed and implemented to ensure a proper and reliable operation of the detectors once installed. A dedicated X-ray test bench is used to characterize the detectors by scanning each pad individually and to precisely measure the uniformity of the gain and the deposited energy resolution over the pad plane. An energy resolution of about 10% is obtained. A detailed physical model has been developed to describe the charge dispersion phenomena in the resistive Micromegas anode. The detailed physical description includes initial ionization, electron drift, diffusion effects and the readout electronics effects. The model provides an excellent characterization of the charge spreading of the experimental measurements and allowed the simultaneous extraction of gain and RC information of the modules.

Measurements of neutral current Drell-Yan production at large invariant dilepton masses can be used to test the energy scale dependence (running) of the electroweak mixing angle. In this work, we make use of a novel implementation of the full next-to-leading order electroweak radiative corrections to the Drell-Yan process using the $\overline{\mathrm{MS}}$ renormalization scheme for the electroweak mixing angle. The potential of future analyses using proton-proton collisions at $\sqrt{s}=13.6~\mathrm{TeV}$ in the Run 3 and High-Luminosity phases of the LHC is explored. In this way, the Standard Model predictions for the $\overline{\mathrm{MS}}$ running at $\mathrm{TeV}$ scales can be probed.

In this paper, we propose to exploit the side-tuning framework for multimodal document classification. Side-tuning is a methodology for network adaptation recently introduced to solve some of the problems related to previous approaches. Thanks to this technique it is actually possible to overcome model rigidity and catastrophic forgetting of transfer learning by fine-tuning. The proposed solution uses off-the-shelf deep learning architectures leveraging the side-tuning framework to combine a base model with a tandem of two side networks. We show that side-tuning can be successfully employed also when different data sources are considered, e.g. text and images in document classification. The experimental results show that this approach pushes further the limit for document classification accuracy with respect to the state of the art., Comment: 2020 25th International Conference on Pattern Recognition (ICPR)

Knowledge of the state of plastic deformation in metallic structures is vital to prevent failure. This is why non-destructive acoustic tests based on the measurement of first order elastic constants have been developed and intensively used. However, plastic deformations, which are usually heterogeneous in space, may be invisible to these methods if the variation of the elastic constants is too small. In recent years, digital image correlation techniques, based on measurements carried out at the surface of a sample, have been successfully used in conjunction with finite element modeling to gain information about plastic deformation in the sample interior. Acoustic waves can penetrate deep into a sample and offer the possibility of probing into the bulk of a plastically deformed material. Previously, we have demonstrated that nonlinear acoustic methods are far more sensitive to changes in dislocation density than linear ones. Here, we show that the nonlinear Second Harmonic Generation method (SHG) is sensitive enough to detect different zones of von Mises stress as well as effective plastic strain in centimeter-size aluminium pieces. This is achieved by way of ultrasonic measurements on a sample that has undergone a three-point bending test. Because of the relatively low stress and small deformations, the sample undergoes plastic deformation by dislocation proliferation. Thus, we conclude that the nonlinear parameter measured by SHG is also sensitive to dislocation density. Our experimental results agree with numerical results obtained by Finite Element Method (FEM) modeling. We also support the acoustic results by X-Ray Diffraction measurements (XRD). Although intrusive and less accurate, they also agree with the acoustic measurements and plastic deformations in finite element simulations.

In this paper we describe the performance of a prototype of the High Angle Time Projection Chambers (HA-TPCs) that are being produced for the Near Detector (ND280) upgrade of the T2K experiment. The two HA-TPCs of ND280 will be instrumented with eight Encapsulated Resistive Anode Micromegas (ERAM) on each endplate, thus constituting in total 32 ERAMs. This innovative technique allows the detection of the charge emitted by ionization electrons over several pads, improving the determination of the track position. The TPC prototype has been equipped with the first ERAM module produced for T2K and with the HA-TPC readout electronics chain and it has been exposed to the DESY Test Beam in order to measure spatial and dE/dx resolution. In this paper we characterize the performances of the ERAM and, for the first time, we compare them with a newly developed simulation of the detector response. Spatial resolution better than 800 ${\mu \rm m}$ and dE/dx resolution better than 10% are observed for all the incident angles and for all the drift distances of interest. All the main features of the data are correctly reproduced by the simulation and these performances fully fulfill the requirements for the HA-TPCs of T2K.

Social media is being increasingly weaponized by state-backed actors to elicit reactions, push narratives and sway public opinion. These are known as Information Operations (IO). The covert nature of IO makes their detection difficult. This is further amplified by missing data due to the user and content removal and privacy requirements. This work advances the hypothesis that the very reactions that Information Operations seek to elicit within the target social systems can be used to detect them. We propose an Interval-censored Transformer Hawkes (IC-TH) architecture and a novel data encoding scheme to account for both observed and missing data. We derive a novel log-likelihood function that we deploy together with a contrastive learning procedure. We showcase the performance of IC-TH on three real-world Twitter datasets and two learning tasks: future popularity prediction and item category prediction. The latter is particularly significant. Using the retweeting timing and patterns solely, we can predict the category of YouTube videos, guess whether news publishers are reputable or controversial and, most importantly, identify state-backed IO agent accounts. Additional qualitative investigations uncover that the automatically discovered clusters of Russian-backed agents appear to coordinate their behavior, activating simultaneously to push specific narratives.