Published on 2 May 2024
Gaia and TESS reveal over 58,000 hot pulsating stars
The Gaia and TESS missions have revealed over 58,000 hot pulsating stars of spectral type O, B, A or F. By comparing Gaia's low-cadence light curves to TESS's precision photometry, the pulsation frequencies and nature of variability for these stars is conclusively determined. Four new catalogs provide the confirmed pulsators, frequencies, and priority for future ensemble asteroseismology.
Published on 2 May 2024
Universal LiDAR segmentation
This paper presents M3Net, a framework for multi-task, multi-dataset, multi-modality LiDAR segmentation using a single set of parameters. It combines large-scale driving datasets with different sensors and conducts alignments in data, feature, and label spaces during training. This allows M3Net to train state-of-the-art segmentation models by taming heterogeneous data. Experiments on 12 datasets show it achieves top performance on SemanticKITTI, nuScenes, and Waymo using one shared parameter set.
Published on 2 May 2024
Observation of particle transport to large scales in a turbulent 2D quantum gas
Researchers drive a 2D quantum gas trapped by lasers, exciting waves at a small scale. At long times, particles cascade to larger scales, forming a steady power-law distribution reminiscent of turbulence. Further experiments reveal the particle transport dynamics.
Published on 2 May 2024
An open-source evaluator model
This paper introduces Prometheus 2, an open-source language model specialized for evaluating the quality of text generated by other language models. It demonstrates superior performance in providing scores and rankings that closely match human judgment, while also allowing flexible evaluation based on custom criteria beyond just helpfulness.
Published on 2 May 2024
3D reasoning for autonomous driving
This paper proposes a new framework called OmniDrive to improve autonomous vehicles' 3D situational awareness and planning abilities. It features a novel 3D multimodal language model architecture that leverages sparse queries to encode visual data into a condensed representation before feeding it to a language model. This allows jointly modeling dynamic objects and map elements to align perception and planning. The paper also introduces a more challenging benchmark, OmniDrive-nuScenes, with visual question answering tasks covering scene desc...
Published on 2 May 2024
Improving model performance via concept realignment
This paper proposes a concept intervention realignment module to improve the efficacy of human interventions in concept bottleneck models. It finds that existing approaches often require numerous costly human interventions per image. This is driven by independent treatment of concepts during intervention. To address this, the paper introduces a module to realign concepts based on their relationships, significantly reducing required interventions.
Published on 2 May 2024
Silicon photomultiplier efficiency at liquid nitrogen temperature
Experiments searching for new physics often use silicon photomultipliers (SiPMs) to detect scintillation light at cryogenic temperatures. This paper presents results measuring the photon detection efficiency (PDE) of two SiPMs at 77 K. A significant decrease of over 20% relative to room temperature was found across various operating voltages when detecting 562.5 nm light.
Published on 2 May 2024
Controlling diffraction with photon polarization in atomic gratings
Researchers propose an atomic grating that utilizes Rydberg interactions to control diffraction based on incident photon polarization states. Multiple polarization-dependent diffraction modes are generated without crosstalk. Dynamic optical pulses switch between modes by triggering Rydberg blockade effects in rubidium-87 atoms.
Published on 2 May 2024
Predicting object motion from videos enables diverse robot manipulation
This paper proposes a method to predict how objects should move between an initial and goal scene configuration based on web videos. It then uses these predicted 'tracks' of object motion to generate robot manipulation plans that can successfully manipulate objects in new scenarios not seen during training. A small amount of robot-specific data further refines the open-loop plans into closed-loop policies.
Published on 2 May 2024
Centerless BMS charge algebra
This paper shows that implementing the Wald-Zoupas prescription results in BMS charges that realize the symmetry algebra without anomalies or central extensions at any cut of future null infinity. The authors refine the covariance prescription to improve properties of charge aspects, and introduce an improved, more covariant aspect for Geroch’s supermomentum. For extended BMS symmetries allowing singular vectors, a valid Wald-Zoupas potential exists given a modified symplectic structure with a corner term. The resulting extended BMS Noether ...
Published on 2 May 2024
Factual language model alignment
This paper studies how to align language models to follow instructions while reducing false claims. It finds that standard alignment methods can increase hallucination by training models on unfamiliar data or rewarding very detailed responses. The authors propose methods to make alignment more factual, by eliciting knowledge from the model itself and using separate rewards for factuality and instruction following.
Published on 2 May 2024
Quantifying model generalization capability
This paper introduces a method to evaluate how well deep learning image classifiers can generalize to related but unseen data. After training models on one dataset, their intermediate layers are tested on a different dataset from the same domain. A metric quantifies the degree to which unseen classes form separable clusters in the latent space. This reveals which layers have the most intrinsic generalization ability, with implications for model compression. Surprisingly, high accuracy on seen classes does not imply good generalization. The a...
Published on 2 May 2024
A pathwise analysis of stochastic evolution equations with locally monotone operators
We establish global existence of weak solutions to stochastic evolution equations on a Gelfand triple with locally monotone operators and sufficiently regular driving signals. Our analysis combines monotone operator theory with recent advances in Besov rough path analysis, requiring no probabilistic structure. This allows treating various operators and signals, including p-Laplace, porous medium, shear-thickening fluids, additive/multiplicative Young integrals, and translated integrals arising in regularization by noise.
Published on 2 May 2024
Real-space implementation of linear-response time-dependent Hartree-Fock
This paper implements time-dependent Hartree-Fock (TDHF) theory for calculating molecular excitation energies in real space, overcoming challenges of expensive exact exchange integrals. It achieves practical ground state convergence, benchmarks TDHF against Gaussian orbital results, and shows systematic convergence of spectra.
Published on 2 May 2024
Vision Transformer for Semantic Image Compression
This paper proposes a novel framework for semantic image communication using vision transformers. It creates an attention mask to prioritize critical image segments for transmission, ensuring reconstruction focuses on key objects. This significantly improves semantic communication quality and bandwidth efficiency by encoding parts proportional to semantic content. Evaluated on TinyImageNet, it succeeds in preserving semantics even when transmitting a fraction of encoded data.
Published on 2 May 2024
AI for manufacturing efficiency and healthcare advancement
This paper reviews applications of AI in manufacturing, focusing on battery management, flow chemistry, additive manufacturing, sensors, and machine vision; and in healthcare, focusing on medical imaging, diagnosis, protein design, and drug discovery. AI promises to enhance productivity, enable new capabilities, and transform industries. But realizing its potential requires addressing challenges around data, algorithms, integration, and responsible development.
Published on 2 May 2024
Qubit-Resonator Interactions Enabled by Driving
This paper develops a theory for multiphoton qubit-resonator interactions that are enhanced when the qubit is driven at specific frequencies. A key result is the realization of qubit-conditional squeezing of the resonator state when the qubit is driven at twice the resonator frequency. This qubit-conditional squeezing can be used to amplify displacements in the resonator, encode qubit states in the resonator, and perform quantum non-demolition measurements of the qubit state. The interactions can be tuned to arbitrary strengths using two-ton...
Published on 2 May 2024
Tools for analyzing least singular values of smoothed random matrices
The paper develops new techniques for lower bounding least singular values of random matrices with limited randomness. The entries depend on polynomials of underlying random variables. This setting captures key challenges in smoothed analysis. The tools involve hierarchical nets and reasoning about higher-order lifts of smoothed matrices. Applications include smoothed analysis guarantees for power sum decompositions, subspace clustering, and certifying robust entanglement.
Published on 2 May 2024
Dependence of spiral galaxy structure on central mass concentration
Researchers developed an algorithm using Galaxy Zoo citizen science data to classify 299 multi-armed and 245 grand design spiral galaxies. They found that for a given stellar mass, grand design galaxies have higher central mass concentrations, smaller half-light radii, earlier Hubble types, and more massive central bulges than multi-armed galaxies. This suggests that a dense central region supports two long spiral arms. Star formation rates were similar for both spiral types when accounting for stellar mass and concentration differences.
Published on 2 May 2024
Deep learning resource allocation in vehicular communications
This paper proposes a deep learning approach to optimize resource allocation in vehicular communications using rate split multiple access. A fractional programming and projected gradient descent based deep unfolding neural network is designed, achieving near optimal performance but with much lower complexity and better resilience to varying conditions.
Published on 2 May 2024
Optimizing fusion power plant maintenance for electricity grid integration
This paper examines optimal maintenance strategies for fusion plants as part of future decarbonized power systems. It finds the value of a plant depends significantly on aligning outage schedules with seasons of low electricity prices. Also, compared to longer outages, frequent short maintenance blocks can increase value despite lower availability.
Published on 2 May 2024
Quantization of the Torus Without Polarizations
This paper introduces a new approach to quantization that does not require polarizations. It is applied to quantize the torus, recovering the noncommutative torus algebra and finite-dimensional representations, without using polarizations. The method unifies other known quantization schemes.
Published on 2 May 2024
Autonomous vehicle performance in wet and dry conditions
This paper evaluates the performance of autonomous vehicles with adaptive cruise control (ACC) in wet and dry conditions using simulation. The ACC system uses sensors and proportional-integral-derivative (PID) control to automatically adjust vehicle speed based on the speed of a leading vehicle and safe following distance. Simulation results show reduced vehicle speeds, longer travel times, and avoidance of rear collisions in wet conditions compared to dry. This demonstrates the importance of testing autonomous vehicle systems under varied w...
Published on 2 May 2024
Localization-guided image editing via cross-attention refinement
This paper proposes a technique called Localization-aware Inversion (LocInv) that uses segmentation maps or bounding boxes to refine cross-attention maps in text-to-image models. This allows for more precise, fine-grained image editing focused on particular objects, while preventing unintended changes to other regions.
Published on 2 May 2024
Library for building force-field-enhanced neural network potentials
FeNNol is a new Python library for easily constructing and training hybrid machine learning potentials that combine neural networks with traditional force field terms. It leverages Jax for fast evaluation and differentiation. The paper shows simulation speeds close to standard force fields.
Published on 2 May 2024
Instruction tuning enables controllable text generation
This paper explores using instruction tuning of large language models as an approach to controllable text generation. The authors introduce an algorithm to automatically generate constraint datasets from only a task dataset and natural language description. They benchmark instruction-tuned models on a new testbed, ConGenBench, finding that prompting outperforms other controllable generation methods, although there are still challenges with structural constraints.
Published on 2 May 2024
Quantum fluid mechanics and trajectories
This paper develops a formalism to describe quantum states in terms of familiar equations from classical mechanics for both fluid flow and particle trajectories. Key results include deriving energy, continuity, and Euler fluid equations analogous to those in classical mechanics, showing their equivalence to the time-dependent many-body Schrodinger equation, and obtaining Lagrangian mechanics and generalized single-particle equations.
Published on 2 May 2024
Impact of dark matter on neutron star structure
This study explores how dark matter cores inside neutron stars affect relations about the stars' mass, radius, and shape changes when rotating. With 5% dark matter, mass, radius, and shape change relations remain accurate within 4%, 2%, and 1.4%. So these universal relations can still reveal dark matter effects on neutron stars.