Search

Contrastive image-text pre-trained models such as CLIP have shown remarkable adaptability to downstream tasks. However, they face challenges due to the high computational requirements of the Vision Transformer (ViT) backbone. Current strategies to boost ViT efficiency focus on pruning patch tokens but fall short in addressing the multimodal nature of CLIP and identifying the optimal subset of tokens for maximum performance. To address this, we propose greedy search methods to establish a "Golden Ranking" and introduce a lightweight predictor specifically trained to approximate this Ranking. To compensate for any performance degradation resulting from token pruning, we incorporate learnable visual tokens that aid in restoring and potentially enhancing the model's performance. Our work presents a comprehensive and systematic investigation of pruning tokens within the ViT backbone of CLIP models. Through our framework, we successfully reduced 40% of patch tokens in CLIP's ViT while only suffering a minimal average accuracy loss of 0.3 across seven datasets. Our study lays the groundwork for building more computationally efficient multimodal models without sacrificing their performance, addressing a key challenge in the application of advanced vision-language models.

In this project, we employ the short-distance factorization to compute the distribution amplitude of the $\eta_c$-meson from Lattice QCD at leading twist. We employ a set of CLS $N_f=2$ ensembles at three lattice spacings and various quark masses to extrapolate the pseudo distribution to the physical point in the isospin limit. We solve the inverse problem modeling the distribution amplitude, and we match our results to the light-cone in the $\overline{\text{MS}}$-scheme. We include a complete error budget, and we compare to two alternative approaches: non-relativistic QCD and Dyson-Schwinger equations, finding good agreement with the latter but not with the former., Comment: 7 pages, 3 figures, contribution to the 41st International Symposium on Lattice Field Theory (LATTICE2024), 28 July - 3 August 2024, Liverpool, UK

The formalism of short-distance factorization, conveyed through the pseudo-distribution approach, connects space-like and light-cone correlators and thus allows for the extraction in lattice QCD of a number of parton distributions. We compute the $t$-dependent parton distribution function of valence quarks in a $\eta_{c}$-meson. After reviewing the main ideas behind the pseudo-distribution formalism, and relying on the analytic structure of Ioffe-time distributions, we come up with a proposal for a model-independent extraction of $t$-dependent parton distribution functions. We present results for the $\eta_{c}$-meson Ioffe time valence $t$-dependent parton distribution function at a renormalization scale of $3~\textrm{GeV}$., Comment: 6 pages, 2 figures, contribution to the 31st International Workshop on Deep Inelastic Scattering (DIS2024), 8-12 April 2024, Grenoble, France

Stacking faults (SF) are important structural defects that play an essential role in the deformation of engineering alloys. However, direct observation of stacking faults at the atomic scale can be challenging. Here, we use the analytical field ion microscopy (aFIM), including density-functional theory informed contrast estimation, to image local elemental segregation at SFs in a creep-deformed solid solution single crystal alloy of Ni-2 at.% W. The segregated atoms are imaged brightly, and time-of-flight spectrometry allows for their identification as W. We also provide the first quantitative analysis of trajectory aberration, with a deviation of approximately 0.4 nm, explaining why atom probe tomography could not resolve these segregations. Atomistic simulations of substitutional W atoms at an edge dislocation in fcc Ni using an analytic bond-order potential indicate that the experimentally observed segregation is due to the energetic preference of W for the center of the stacking fault, contrasting with e.g., Re segregating to partial dislocations. Solute segregation to SF can hinder dislocation motion, increasing the strength of Ni-based superalloys. Yet direct substitution of Re by W envisaged to lower superalloys' costs, requires extra consideration in alloy design since these two solutes do not have comparable interactions with structural defects during deformation.

In this paper, we introduce SaulLM-54B and SaulLM-141B, two large language models (LLMs) tailored for the legal sector. These models, which feature architectures of 54 billion and 141 billion parameters, respectively, are based on the Mixtral architecture. The development of SaulLM-54B and SaulLM-141B is guided by large-scale domain adaptation, divided into three strategies: (1) the exploitation of continued pretraining involving a base corpus that includes over 540 billion of legal tokens, (2) the implementation of a specialized legal instruction-following protocol, and (3) the alignment of model outputs with human preferences in legal interpretations. The integration of synthetically generated data in the second and third steps enhances the models' capabilities in interpreting and processing legal texts, effectively reaching state-of-the-art performance and outperforming previous open-source models on LegalBench-Instruct. This work explores the trade-offs involved in domain-specific adaptation at this scale, offering insights that may inform future studies on domain adaptation using strong decoder models. Building upon SaulLM-7B, this study refines the approach to produce an LLM better equipped for legal tasks. We are releasing base, instruct, and aligned versions on top of SaulLM-54B and SaulLM-141B under the MIT License to facilitate reuse and collaborative research.

Masked Image Modeling (MIM) has emerged as a promising method for deriving visual representations from unlabeled image data by predicting missing pixels from masked portions of images. It excels in region-aware learning and provides strong initializations for various tasks, but struggles to capture high-level semantics without further supervised fine-tuning, likely due to the low-level nature of its pixel reconstruction objective. A promising yet unrealized framework is learning representations through masked reconstruction in latent space, combining the locality of MIM with the high-level targets. However, this approach poses significant training challenges as the reconstruction targets are learned in conjunction with the model, potentially leading to trivial or suboptimal solutions.Our study is among the first to thoroughly analyze and address the challenges of such framework, which we refer to as Latent MIM. Through a series of carefully designed experiments and extensive analysis, we identify the source of these challenges, including representation collapsing for joint online/target optimization, learning objectives, the high region correlation in latent space and decoding conditioning. By sequentially addressing these issues, we demonstrate that Latent MIM can indeed learn high-level representations while retaining the benefits of MIM models.

Audio-visual generalized zero-shot learning is a rapidly advancing domain that seeks to understand the intricate relations between audio and visual cues within videos. The overarching goal is to leverage insights from seen classes to identify instances from previously unseen ones. Prior approaches primarily utilized synchronized auto-encoders to reconstruct audio-visual attributes, which were informed by cross-attention transformers and projected text embeddings. However, these methods fell short of effectively capturing the intricate relationship between cross-modal features and class-label embeddings inherent in pre-trained language-aligned embeddings. To circumvent these bottlenecks, we introduce a simple yet effective framework for Easy Audio-Visual Generalized Zero-shot Learning, named EZ-AVGZL, that aligns audio-visual embeddings with transformed text representations. It utilizes a single supervised text audio-visual contrastive loss to learn an alignment between audio-visual and textual modalities, moving away from the conventional approach of reconstructing cross-modal features and text embeddings. Our key insight is that while class name embeddings are well aligned with language-based audio-visual features, they don't provide sufficient class separation to be useful for zero-shot learning. To address this, our method leverages differential optimization to transform class embeddings into a more discriminative space while preserving the semantic structure of language representations. We conduct extensive experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks. Our results demonstrate that our EZ-AVGZL achieves state-of-the-art performance in audio-visual generalized zero-shot learning.

The purpose of this work is to do a systematic feasibility study of measuring in backward region deeply virtual Compton scattering on the pion in Sullivan processes in the framework of collinear QCD factorization where pion to photon transition distribution amplitudes (TDAs) describe the photon content of the $\pi$ meson. Our approach employs TDAs based on the overlap of light front wave functions, using a previously developed pion light-front wave function and deriving a consistent model for the light front wave functions of the photon. This work is expected to lead us to an estimate of the cross-sections that could be measured in the future U.S. and China's electron-ion colliders. It will also provide a comparison with the forward Sullivan DVCS case, which gives access to pion GPDs and for which a strong signal is expected., Comment: Proceedings for 31st International Workshop on Deep Inelastic Scattering

In this proceeding we determine the distribution amplitude of the $\eta_c$-meson from first principles. This quantity appears as a consequence of factorization theorems, and it is necessary to compute the amplitude of multiple exclusive processes. Since it is defined along a light-cone, its calculation via lattice QCD was impossible until recently, when a generalization to Euclidean metric was proposed, and a connection to the physical limit was established. We briefly explain the method of short distance factorization, which allows us to compute the distribution amplitude, and our lattice setup. After summarizing the steps for the continuum and chiral extrapolation, we present our results and compare them to two alternative determinations, one using non-relativistic QCD and another solving the Dyson-Schwinger equations; we find a large discrepancy with the former., Comment: 6 pages, 3 figures, contribution to the 31st International Workshop on Deep Inelastic Scattering (DIS2024), 8-12 April 2024, Grenoble, France

Distribution amplitudes are functions of non-perturbative matrix elements describing the hadronization of quarks and gluons. Thanks to factorization theorems, they can be used to compute the scattering amplitude of high-energy processes. Recently, new ideas have allowed their computation using lattice QCD, which should provide us with a general, fully relativistic determination. We present the first lattice calculation of the $\eta_c$-meson distribution amplitude at leading twist. Starting from the relevant matrix element in discrete Euclidean space on a set of $N_f=2$ CLS ensembles, we explain the method to connect to continuum Minkowski spacetime. After addressing several sources of systematic uncertainty, we compare to Dyson-Schwinger and non-relativistic QCD determinations of this quantity. We find significant deviations between the latter and our result even at small Ioffe times., Comment: 28 pages, 8 figures, 4 tables

Centaurs, distinguished by their volatile-rich compositions, play a pivotal role in understanding the formation and evolution of the early solar system, as they represent remnants of the primordial material that populated the outer regions. Stellar occultations offer a means to investigate their physical properties, including shape, rotational state, or the potential presence of satellites and rings. This work aims to conduct a detailed study of the centaur (54598) Bienor through stellar occultations and rotational light curves from photometric data collected during recent years. We successfully predicted three stellar occultations by Bienor, which were observed from Japan, Eastern Europe, and the USA. In addition, we organized observational campaigns from Spain to obtain rotational light curves. At the same time, we develop software to generate synthetic light curves from three-dimensional shape models, enabling us to validate the outcomes through computer simulations. We resolve Bienor's projected ellipse for December 26, 2022, determine a prograde sense of rotation, and confirm an asymmetric rotational light curve. We also retrieve the axes of its triaxial ellipsoid shape as a = (127 $\pm$ 5) km, b = (55 $\pm$ 4) km, and c = (45 $\pm$ 4) km. Moreover, we refine the rotation period to 9.1736 $\pm$ 0.0002 hours and determine a geometric albedo of (6.5 $\pm$ 0.5) %, higher than previously determined by other methods. Finally, by comparing our findings with previous results and simulated rotational light curves, we analyze whether an irregular or contact-binary shape, the presence of an additional element such as a satellite, or significant albedo variations on Bienor's surface, may be present.

In this manuscript, we investigate the stochastic thermodynamics of Fisher information (FI), meaning we characterize both the \textit{fluctuations} of FI, introducing a parastatistics of that quantity, and thermodynamic quantities. We introduce two initial conditions: an equilibrium initial condition and a minimum entropy initial condition, both under a protocol that drives the system to equilibrium. Its results indicate a dependence of the average FI on both the initial condition and path taken. Furthermore, the results point the chosen parameter directly affects the FI of thermodynamic quantities such as irreversible work and entropy, along with fluctuations of a stochastic FI. Last, we assess the further role of FI of the distribution of thermal quantities within the context of thermostatistical inequalities., Comment: 6 pages, 2 figures, Supplementary material available; comments welcome!

Here, we present an exploratory study of the potential applications of electrically biased materials that possess a nonreciprocal and non-Hermitian electromagnetic response analogous to the electronic response of field-effect transistors. The most distinctive feature of such materials is their chiral-gain, meaning that their response can be active or dissipative depending on the handedness of the wave polarization. Here, we show how the chiral-gain can be harnessed to develop novel electromagnetic devices with unique properties such as chiral lasers, polarization-dependent mirrors, and coherent-perfect-absorber lasers. Furthermore, it is demonstrated that materials with chiral-gain can bypass a reciprocity constraint that typically limits the external coupling strength, thus facilitating the excitation of cavities with extremely large quality factors., Comment: 39 pages, 10 figures

The uses of machine learning (ML) have snowballed in recent years. In many cases, ML models are highly complex, and their operation is beyond the understanding of human decision-makers. Nevertheless, some uses of ML models involve high-stakes and safety-critical applications. Explainable artificial intelligence (XAI) aims to help human decision-makers in understanding the operation of such complex ML models, thus eliciting trust in their operation. Unfortunately, the majority of past XAI work is based on informal approaches, that offer no guarantees of rigor. Unsurprisingly, there exists comprehensive experimental and theoretical evidence confirming that informal methods of XAI can provide human-decision makers with erroneous information. Logic-based XAI represents a rigorous approach to explainability; it is model-based and offers the strongest guarantees of rigor of computed explanations. However, a well-known drawback of logic-based XAI is the complexity of logic reasoning, especially for highly complex ML models. Recent work proposed distance-restricted explanations, i.e. explanations that are rigorous provided the distance to a given input is small enough. Distance-restricted explainability is tightly related with adversarial robustness, and it has been shown to scale for moderately complex ML models, but the number of inputs still represents a key limiting factor. This paper investigates novel algorithms for scaling up the performance of logic-based explainers when computing and enumerating ML model explanations with a large number of inputs.

The Pampilhosa da Serra Space Observatory (PASO) is located in the center of the continental Portuguese territory, in the heart of a certified Dark Sky destination by the Starlight Foundation (Aldeias do Xisto) and has been an instrumental asset to advance science, education and astrotourism certifications. PASO hosts astronomy and Space Situational Awareness (SSA) activities including a node of the Portuguese Space Surveillance \& Tracking (SST) infrastructure network, such as a space radar currently in test phase using GEM radiotelescope, a double Wide Field of View Telescope system, a EUSST optical sensor telescope. These instruments allow surveillance of satellite and space debris in LEO, MEO and GEO orbits. The WFOV telescope offers spectroscopy capabilities enabling light curve analysis and cosmic sources monitoring. Instruments for Space Weather are being considered for installation to monitor solar activities and expand the range of SSA services., Comment: 9 pages, 2 figures, Keywords: Dark Sky, radioastronomy, space, space debris, space situational awareness; Draft from Talk presented at Astronomy and Satellite Constellations: Pathways Forward, IAU Symposium 385, October 2023, Ed. C. Walker, D.Turnshek, P.Grimley, D.Galadi-Enriquez & M.Aub\'e, International Astronomical Union Proceedings Series, Cambridge University Press, 2024

Next-generation cellular networks will play a key role in the evolution of different vertical industries. Low latency will be a major requirement in many related uses cases. This requirement is specially challenging in scenarios with high mobility of end devices, such as vehicular communications. The Multi-Access Edge Computing (MEC) paradigm seeks to satisfy it. In this article we propose the dynamic deployment of anchor point network functions at edge locations and the assignment of terminals to these anchor points with the joint objective of minimizing communications latency and reducing network overhead. We formally define the problem as a multi-objective optimization and also propose a novel heuristic greedy algorithm for approximating the solution. This algorithm compares favorably with baseline and state-of-the-art strategies for latency minimization while reducing the overhead caused by network reconfigurations., Comment: Article published in IEEE Transactions on Vehicular Technology

Current visual generation methods can produce high quality videos guided by texts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio Synchronized Visual Animation (ASVA), a task animating a static image to demonstrate motion dynamics, temporally guided by audio clips across multiple classes. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating dynamic animations guided by audios. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our models superior performance. We further explore AVSyncDs potential in a variety of audio synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation. More videos on project webpage https://lzhangbj.github.io/projects/asva/asva.html., Comment: ECCV 2024

In this paper, we introduce SaulLM-7B, a large language model (LLM) tailored for the legal domain. With 7 billion parameters, SaulLM-7B is the first LLM designed explicitly for legal text comprehension and generation. Leveraging the Mistral 7B architecture as its foundation, SaulLM-7B is trained on an English legal corpus of over 30 billion tokens. SaulLM-7B exhibits state-of-the-art proficiency in understanding and processing legal documents. Additionally, we present a novel instructional fine-tuning method that leverages legal datasets to further enhance SaulLM-7B's performance in legal tasks. SaulLM-7B is released under the MIT License.

SARComm is a novel wireless communication system that enables passive satellite backscatter connectivity using existing spaceborne synthetic aperture radar (SAR) signals. We demonstrate that SAR signals from the European Space Agency's Sentinel-1 satellite, used to image Earth's terrain, can be leveraged to enable low-power ground-to-satellite communication. This paper presents the first cooperative, on-the-ground target that modulates SAR backscatter to send information bits and analyzes how to extract them from publicly available Sentinel-1 datasets. To demonstrate the system, we evaluate the effectiveness of modulating the radar cross section of corner reflectors both mechanically and electronically to encode data bits, develop a deployment algorithm to optimize corner reflector placement, and present a SAR processing pipeline to enable communication., Comment: 13 pages, 19 figures

This article focuses on the prediction of the vibration frequency response of handheld probes. A novel approach that involves machine learning and readily available data from probes was explored. Vibration probes are efficient and affordable devices that provide information about testing airborne sound insulation in building acoustics. However, fixing a probe to a vibrating surface downshifts sensor resonances and underestimates levels. Therefore, the calibration response of the sensor included in a probe differs from the frequency response of that same probe. Simulation techniques of complex mechanical systems may describe this issue, but they include hardly obtainable parameters, ultimately restricting the model. Thus, this study discusses an alternative method, which comprises different parts. Firstly, the vibration frequency responses of 85 probes were measured and labelled according to six features. Then, Linear Regression, Decision Tree Regression and Artificial Neural Networks algorithms were analysed. It was revealed that decision tree regression is the more appropriate technique for this data. The best decision tree models, in terms of scores and model structure, were fine-tuned. Eventually, the final suggested model employs only four out of the six original features. A trade-off solution that involved a simple structure, an interpretable model and accurate predictions was accomplished. It showed a maximum average deviation from test measurements ranging from 0.6 dB in low-frequency to 3 dB in high-frequency while remaining at a low computational load. This research developed an original and reliable prediction tool that provides the vibration frequency response of handheld probes.

This paper investigates the optimization of the exhaust gas composition by applying a velocity-controlled Throttle-by-Wire-System on modern 50 cc scooters (Euro 5). Nowadays combustion-powered scooters are still inefficiently restricted, resulting in an unreasonably high fuel consumption and unfavorable exhaust emissions. The velocity control prevents restriction by negatively shifting the ignition timing and regulates the throttle valve opening instead. Injection quantity, engine speed, ignition timing, cylinder wall temperature, exhaust gas temperature, oxygen sensor data, crankshaft position and in-cylinder pressure were acquired to measure engine parameters. At the same time, vehicle data on the CAN bus, such as throttle opening angle, the rider's acceleration command and vehicle velocity were recorded. For determination of the exhaust gas composition, five probes were sensing CO, CO2, NOx, O2 and HC in addition to the temperature and mass flow. A Peugeot Kisbee 50 4T (Euro 5) serves as test vehicle. The original and the optimized restriction were subjected to various gradients on a roller dynamometer at top speed. Thus, a statement can be made about all operating points of restriction. The resistance parameters required, were previously determined in a coast down test. When driving on level ground, a difference of 50% in the throttle opening leads to a 17% improvement in fuel economy. By measuring the engine parameters, optimum ignition timing could be proven with increasing internal cylinder pressure. Further, 17% reduction in exhaust gas flow was demonstrated. CO emissions decreased by a factor of 8.4, CO2 by 1.17 and HC by 2.1 while NOx increased by a factor of 3.

Context - Around the year 2000, Triton's south pole experienced an extreme summer solstice that occurs every about 650 years, when the subsolar latitude reached about 50{\deg}. Bracketing this epoch, a few occultations probed Triton's atmosphere in 1989, 1995, 1997, 2008 and 2017. A recent ground-based stellar occultation observed on 6 October 2022 provides a new measurement of Triton's atmospheric pressure which is presented here. Aims- The goal is to constrain the Volatile Transport Models (VTMs) of Triton's atmosphere that is basically in vapor pressure equilibrium with the nitrogen ice at its surface. Methods - Fits to the occultation light curves yield Triton's atmospheric pressure at the reference radius 1400 km, from which the surface pressure is induced. Results - The fits provide a pressure p_1400= 1.211 +/- 0.039 microbar at radius 1400 km (47 km altitude), from which a surface pressure of p_surf= 14.54 +/- 0.47 microbar is induced (1-sigma error bars). To within error bars, this is identical to the pressure derived from the previous occultation of 5 October 2017, p_1400 = 1.18 +/- 0.03 microbar and p_surf= 14.1 +/- 0.4 microbar, respectively. Based on recent models of Triton's volatile cycles, the overall evolution over the last 30 years of the surface pressure is consistent with N2 condensation taking place in the northern hemisphere. However, models typically predict a steady decrease in surface pressure for the period 2005-2060, which is not confirmed by this observation. Complex surface-atmosphere interactions, such as ice albedo runaway and formation of local N2 frosts in the equatorial regions of Triton could explain the relatively constant pressure between 2017 and 2022., Comment: 8 pages, 4 figures, accepted for publication in Astronomy and Astrophysics

Here, we present an in-depth theoretical analysis of the linear electrooptic effect in low-symmetry three-dimensional (3D) conductive materials with large Berry curvature dipoles. Our study identifies two distinct kinetic contributions to the linear electrooptic effect: a gyrotropic Hermitian (conservative) piece and a non-Hermitian term that can originate optical gain. We concentrate on the study of 3D materials belonging to the 32 ($D_3$) point group subject to a static electric bias along the trigonal axis. Our investigation shows that doped trigonal tellurium has promising properties, with its gyrotropic electrooptic response offering the potential for realizing electrically-biased electromagnetic isolators and inducing significant optical dichroism. Most notably, it is demonstrated that under sufficiently large static electric bias, tellurium's non-Hermitian electrooptic response may lead to optical gain. Using first-principles calculations, it is shown that n-doped tellurium is particularly promising, as it can host significantly larger Berry curvature dipoles than the more common p-doped tellurium., Comment: 38 pages, 9 figures

Relying on the polynomiality property of generalized parton distributions, which roots on Lorentz covariance, we prove that it is enough to know them at vanishing- and low-skewness within the DGLAP region to obtain a unique extension to their entire support up to a D-term. We put this idea in practice using two methods: Reconstruction using artificial neural networks and finite-elements methods. We benchmark our results against standard models for generalized parton distributions. In agreement with the formal expectation, we obtain a very accurate reconstructions for a maximal value of the skewness as low as 20% of the longitudinal momentum fraction. This result might be relevant for reconstruction of generalized parton distribution from experimental and lattice QCD data, where computations are for now, restricted in skewness., Comment: 20 pages, 10 figures

For a vector field on a compact manifold we consider the invariant measure for the stochastic perturbation as well as the corresponding Ma\~n\'e lagrangian. We show that the limit of the invariant measure of the perturbated system, when the perturbation tends to zero, determines a particular weak KAM solution of the lagrangian., Comment: 9 pages

Optical microscopy is an indispensable tool in life sciences research, but conventional techniques require compromises between imaging parameters like speed, resolution, field-of-view, and phototoxicity. To overcome these limitations, data-driven microscopes incorporate feedback loops between data acquisition and analysis. This review overviews how machine learning enables automated image analysis to optimise microscopy in real-time. We first introduce key data-driven microscopy concepts and machine learning methods relevant to microscopy image analysis. Subsequently, we highlight pioneering works and recent advances in integrating machine learning into microscopy acquisition workflows, including optimising illumination, switching modalities and acquisition rates, and triggering targeted experiments. We then discuss the remaining challenges and future outlook. Overall, intelligent microscopes that can sense, analyse, and adapt promise to transform optical imaging by opening new experimental possibilities., Comment: 7 pages, 4 figures, review

Motivated by our recent findings in [Phys. Rev. Lett. 128, 013902, 2022], which introduces a new class of electromagnetic bulk materials whose response is similar to conventional semiconductor transistors, here we propose a one-dimensional (1D) version of such a material based on transmission lines coupled with FET isolators. We demonstrate that the response of this 1D system is nonreciprocal and non-Hermitian, analogous to the idealized transistor-metamaterial, and is also characterized by a broken time-reversal symmetry. We analyze the wave propagation in the system and find that the interaction between the eigenmodes can either lead to gain or loss depending on the propagation distance. Furthermore, it is also shown that the system may be operated at an exceptional point, wherein the response of the structure is singular, and the power gain is maximized. Finally, we demonstrate that the exceptional point coincides with the point of operation of typical microwave amplifiers, such as the distributed amplifier.

Humans possess a remarkable ability to integrate auditory and visual information, enabling a deeper understanding of the surrounding environment. This early fusion of audio and visual cues, demonstrated through cognitive psychology and neuroscience research, offers promising potential for developing multimodal perception models. However, training early fusion architectures poses significant challenges, as the increased model expressivity requires robust learning frameworks to harness their enhanced capabilities. In this paper, we address this challenge by leveraging the masked reconstruction framework, previously successful in unimodal settings, to train audio-visual encoders with early fusion. Additionally, we propose an attention-based fusion module that captures interactions between local audio and visual representations, enhancing the model's ability to capture fine-grained interactions. While effective, this procedure can become computationally intractable, as the number of local representations increases. Thus, to address the computational complexity, we propose an alternative procedure that factorizes the local representations before representing audio-visual interactions. Extensive evaluations on a variety of datasets demonstrate the superiority of our approach in audio-event classification, visual sound localization, sound separation, and audio-visual segmentation. These contributions enable the efficient training of deeply integrated audio-visual models and significantly advance the usefulness of early fusion architectures.

In this work, we evaluate the energy usage of fully embedded medical diagnosis aids based on both segmentation and classification of medical images implemented on Edge TPU and embedded GPU processors. We use glaucoma diagnosis based on color fundus images as an example to show the possibility of performing segmentation and classification in real time on embedded boards and to highlight the different energy requirements of the studied implementations. Several other works develop the use of segmentation and feature extraction techniques to detect glaucoma, among many other pathologies, with deep neural networks. Memory limitations and low processing capabilities of embedded accelerated systems (EAS) limit their use for deep network-based system training. However, including specific acceleration hardware, such as NVIDIA's Maxwell GPU or Google's Edge TPU, enables them to perform inferences using complex pre-trained networks in very reasonable times. In this study, we evaluate the timing and energy performance of two EAS equipped with Machine Learning (ML) accelerators executing an example diagnostic tool developed in a previous work. For optic disc (OD) and cup (OC) segmentation, the obtained prediction times per image are under 29 and 43 ms using Edge TPUs and Maxwell GPUs, respectively. Prediction times for the classification subsystem are lower than 10 and 14 ms for Edge TPUs and Maxwell GPUs, respectively. Regarding energy usage, in approximate terms, for OD segmentation Edge TPUs and Maxwell GPUs use 38 and 190 mJ per image, respectively. For fundus classification, Edge TPUs and Maxwell GPUs use 45 and 70 mJ, respectively., Comment: 37 pages, 20 figures. Accepted manuscript. Published in "Engineering Applications of Artificial Intelligence" (Elsevier), ISSN 0952-1976

We report on the first lattice determination of the pseudoscalar meson $\eta_c$ light-cone distribution amplitude, using a set of three CLS $N_f=2$ ensembles at a pion mass $m_{\pi} \sim 270~\text{MeV}$ and lattice spacings $a \sim 0.076~\text{fm}$, $0.066~\text{fm}$ and $0.049~\text{fm}$. Employing Short Distance Factorization, we extract the pseudo-DA on the lattice for Ioffe times $\nu \leq 4.5$, and the various lattice spacings allow us to take the continuum limit. We employ a basis of Jacobi polynomials to parametrize the distribution amplitude, which allows to express the matching to the pseudo distribution in closed form, and we observe a strong effect which we attribute to the heavy charm-quark mass., Comment: Contribution to the 40th International Symposium on Lattice Field Theory (Lattice 2023), July 31st - August 4th, 2023. Fermi National Accelerator Laboratory

This paper investigates the performance and fuel-saving effect of a velocity control algorithm on modern 50 cc scooters (Euro 5). The European Parliament has adopted major CO$_2$ emission reductions by 2030. But modern combustion- powered scooters are inefficiently restricted and emit unnecessary amounts of CO$_2$. Replacing the original restriction method with the system presented in this paper, the engine's operating point is being improved significantly. Therefore, a Throttle-by-Wire-System senses the rider's throttle command and manipulates the throttle valve. A redundant wheel speed sensor measures the precise vehicle velocity using the magneto-resistive principle. The entire system is managed by a central ECU, executing the actual velocity control, fail-safe functions, power supply and handling inputs/outputs. For velocity control, an adaptive PI-controller has been simulated, virtually tuned and implemented, limiting the max. velocity regulated by legal constraints (45 km/h). In this way, the environmentally harmful restrictors used today can be bypassed. By implementing a human-machine interface, including a virtual dashboard, the system is capable of interfacing with the rider. For evaluation purposes a measurement box has been developed, logging vehicle orientation, system/control variables and engine parameters. A Peugeot Kisbee 50 4T (Euro 5) is serving as test vehicle. Finally, the system has been evaluated regarding performance and fuel efficiency both through simulation and road testing. Fuel savings of 13.6 % in real-world test scenarios were achieved while maintaining vehicle performance.

Brownian fluctuations arise for any quantity that depends on the stochastic variables of a Brownian particle. In this study, we explore the Brownian fluctuations of a bidimensional quadratic potential that exhibits two regimes: a confining regime and a non-confining regime. We divide the total potential into two contributions and analyze the mean, variance, skewness, excess kurtosis, and their distributions for each contribution as well as for the total potential. Our analysis offers an understanding into the statistical behavior of each quantity., Comment: 21 pages, 11 figures