150813 research outputs found
Sort by
Savaal: A system for automatically generating high-quality questions from unseen documents
No full textAssessing human understanding through exams and quizzes is fundamental to learning and advancement in both educational and professional settings. However, current solutions to automate the generation of challenging questions from educational materials and documents are insufficient, resulting in superficial or often irrelevant questions. While LLMs have been shown to excel in tasks like question answering, their usage on question generation is underexplored for general domains and at scale. This work presents Savaal, a scalable question-generation system that generates higher-order questions from documents, as well as a real-world system implementation for general use. Savaal accomplishes the following goals and objectives: (i) scalability, capable of generating hundreds of questions from any document (ii) depth of understanding, synthesizing higherorder concepts to test learners’ understanding of the material, and (iii) domain independence, generalizing broadly to any field. Rather than naively providing the entire document in context to an LLM, Savaal breaks down the process of generating questions into a three-stage pipeline. We demonstrate that Savaal outperforms the direct prompting baseline as evaluated by 76 human experts on 71 documents across conference papers and PhD dissertations. We additionally contribute a general system for serving Savaal in real-world scenarios. We demonstrate that our system is scalable, enabling fault-tolerant and horizontal scaling of each individual component in response to fluctuations in usage. Moreover, our architecture enables interactive usage from users and collaboration in groups, reflecting real-world organizations like classrooms or enterprises. We hope that the system enables scalable question generation for educational and corporate use-cases.M.Eng
One-Sided Bounded Noise: Theory, Optimization Algorithms and Applications
No full textCCS ’25, Taipei, TaiwanWe investigate the optimal trade-off between utility and privacy using one-sided perturbation. Unlike conventional privacy-preserving statistical releases, randomization for obfuscating side-channel information is often constrained by infrastructure limitations. In practical scenarios, these constraints may only allow positive and bounded perturbations. For example, extending processing time or sending and storing dummy messages/data is typically feasible. However, implementing modifications in the opposite direction is challenging due to restrictions imposed by hardware capacity, communication protocols, and data management systems. In this paper, we establish the foundation of the positive noise mechanism within three semantic privacy frameworks: Differential Privacy (DP), Maximal Leakage (MaxL), and Probably Approximately Correct (PAC) Privacy. We then present a series of results that characterize or approximate the optimal one-sided noise distribution, subject to a second-moment budget and a bounded maximal magnitude. Building on this theoretical foundation, we develop efficient tools to solve the underlying optimization problems. Through experiments conducted in various scenarios, we demonstrate that existing techniques, such as Truncated Biased Laplace noise, are often suboptimal and result in excessive performance degradation. For instance, in an anonymous communication system with a 250K message budget, our optimized DP noise mechanism achieves a 21× reduction in dummy messages and an 18× reduction in dummy message latency overhead compared to traditional methods
Trapping and Laser Cooling an Ensemble of Ytterbium-171 Atoms for use in an Atomic Clock
No full textOptical lattice clocks require careful preparation of atomic ensembles in order to ensure homogeneous interactions with the clock laser. We demonstrate loading and laser cooling of an ensemble of ytterbium-171 atoms in a 2D optical dipole trap created by an optical cavity. Our loading method ensures that all atoms are located in the intersection of 2 perpendicular dipole traps as verified through absorption imaging. Raman sideband cooling was used to cool the atomic ensemble from 15.7 uK to 6.3 uK as measured through optical sideband spectroscopy on the 578 nm clock transition. Together, these steps improved the transfer of atoms during a Rabi oscillation from the ground to the clock state from approximately 45 percent excitation fraction to 80 percent excitation fraction. The final atomic ensemble preparation is now sufficient for running an atomic clock.S.M
Scaling 3D Scene Perception via Probabilistic Programming
No full textUnderstanding and interpreting the 3D structure of the world is a central challenge in artificial intelligence. Our physical world is 3D, yet our AI systems often “see” that world through pixels and images. In order to build truly intelligent AI systems, we must go beyond pixels and images and build 3D vision systems that can build meaningful and useful 3D representations of the world. This is the problem of 3D scene perception. How do we transform raw visual input into 3D representations of the world? 3D scene perception has numerous applications from robotics to augmented reality. Despite the advances over the last decade, 3D perception remains a major bottleneck in real-world robotics applications. The challenge stems from the immense variability in real-world conditions, e.g. lighting, color, viewpoint, camera properties, object appearance, the incompleteness of visual data due to limited resolution, noise, and occlusions, and the approximations in our models of visual data. Developing more robust and generalizable 3D perception systems would be an important step towards more general-purpose robotics. In this thesis, we explore a probabilistic architecture for 3D perception based on structured generative models and probabilistic programs. We begin with 3DP3, the first iteration of our approach, which infers 3D scene graphs from real-world depth image data. 3DP3 demonstrates that our method could work on real-world benchmarks and correct commonsense errors from deep learning systems. Building on this foundation, we develop Bayes3D, which scaled up these ideas using a GPU-accelerated image likelihood and generative model alongside a parallel coarse-to-fine inference algorithm. Next, we explore two approaches for incorporating RGB image data into generative 3D graphics programs, expanding their applicability. We then introduce DurableVS, which extends inverse-graphics techniques to model scenes involving a robot and multiple cameras, enabling precise control of a robot. Finally, we present Gen3D, which integrates all the key ideas from this thesis into a real-time 3D perception system that uses multi-resolution probabilistic models of 3D matter to enable real-time tracking that is competitive with vision transformers and 3D Gaussian splatting, state-of-the-art methods in computer vision and computer graphics.Ph.D
Learning from Weak Supervision: Theory, Methods, and Applications
No full textThe growing demand for high-quality labeled data to train machine learning models has driven widespread adoption of weak supervision and synthetic data methods, which use automated models instead of humans for annotation. Large language models (LLMs) have further accelerated this trend because their zero- and few-shot classification performance enables them to serve as effective “synthetic annotators” for various tasks. In practice, the data generated by these weak annotators is imperfect, but it enables the training of strong models. However, theoretical understanding of why training one model on the outputs of another leads to strong performance remains limited, especially when the annotator model exhibits suboptimal performance on the target task. In this thesis, I develop a theoretical framework for learning from weak supervision that captures the key aspects of the problem better than existing approaches in the crowdsourcing and learning-with-noisy-label literature. This framework establishes structural conditions that explain when and why weak supervision can reliably train strong models. Building on these theoretical results, the second part of the thesis introduces methods to improve how models learn from weak supervision and applies these methods to low-labeled-data settings.Ph.D
Rousseau's Freedom as Recognition
No full textTo yearn for freedom is to want to be seen by others as someone. Rousseau, I believe, held such a conception of freedom, alongside his intricate theory of human passions. This essay examines how freedom relates to such passions, and in particular, to the Rousseauian notion of amour-propre. Importantly, the aim here is both interpretive and positive. The essay seeks to locate Rousseau within the old republican tradition in a manner that parts ways with most contemporary readings of Rousseau. But, in doing so, it argues that republican freedom essentially involves a particular status and the recognition of such status by others. On this Rousseauian view, one is free to the extent that others see one as a limit to their arbitrary interference and as entitled to interfere with them non-arbitrarily. Finally, republican freedom, so understood, is shown to be essential to meeting the demands of healthy amour-propre, thereby bringing Rousseau's political and psychological theories closer together
Engineering Winter
No full textAs winters warm and snowfall becomes less reliable, ski resorts worldwide increasingly depend on artificial snow to stay open. Snowmaking, once a stopgap, has become the backbone of entire seasons in a sprawling choreography of pumps and pressurized mist designed to hold trails together. At resorts like Vermont’s Bromley Mountain, snowmakers work through the night, drawing millions of gallons from limited reservoirs and operating within narrowing windows of cold air. What emerges is a portrait of winter in transition: less predictable, more expensive, increasingly manufactured. The efforts to preserve winter recreation carry growing costs in energy, water, and equitable access. Many smaller, independent ski areas struggle to meet the demands of climate adaptation, while larger resorts expand their operations, widening the divide in who can afford to sustain operations. In the American West, where rivers depend heavily on snowpack melt, the spread of snowmaking ties winter recreation to a water system already under immense strain. As artificial snow becomes the norm, winter is increasingly a season bought, built, and rationed, raising the question of whether attempts to keep the season alive are accelerating the changes that threaten to erase it.S.M
Non-Line-of-Sight 3D Object Reconstruction via mmWave Surface Normal Estimation
No full textMobiSys ’25, Anaheim, CA, USAThis paper presents the design, implementation, and evaluation of
mmNorm, a new and highly-accurate method for non-line-of-sight
3D object reconstruction using millimeter wave (mmWave) signals.
In contrast to past approaches for millimeter-wave-based imaging
that perform backprojection for 3D object reconstruction, mmNorm
reconstructs the surface by estimating the object’s surface normals.
To do this, it introduces a novel algorithm that directly estimates
the surface normal vector field from mmWave reflections. By then
inverting the normal field, it can reconstruct structural isosurfaces,
then solve for the exact surface through a novel mmWave optimization framework.
We built an end-to-end prototype of mmNorm using a TI IWR1443
Boost mmWave radar and a UR5e Robotic Arm, and evaluated it
in over 110 real-world experiments across more than 60 different
everyday objects. In a head-to-head comparison with state-of-theart baselines, mmNorm achieves 96% reconstruction accuracy (3D
F-score) compared to 78% for the best-performing baseline. These
results show that mmNorm is capable of high-accuracy mmWave
object reconstruction. The codebase and a video demonstration are
available here: https://github.com/signalkinetics/mmNor
MechStyle: Augmenting Generative AI with Mechanical Simulation to Create Stylized and Structurally Viable 3D Models
No full textSCF ’25, Cambridge, MA, USARecent developments in Generative AI enable creators to stylize 3D models based on text prompts. These methods change the 3D model geometry, which can compromise the model’s structural integrity once fabricated. We present MechStyle, a system that enables creators to stylize 3D printable models while preserving their structural integrity. MechStyle accomplishes this by augmenting the Generative AI-based stylization process with feedback from a Finite Element Analysis (FEA) simulation. As the stylization process modifies the geometry to approximate the desired style, feedback from the FEA simulation reduces modifications to regions with increased stress. We evaluate the effectiveness of FEA simulation feedback in the augmented stylization process by comparing three stylization control strategies. We also investigate the time efficiency of our approach by comparing three adaptive scheduling strategies. Finally, we demonstrate MechStyle’s user interface that allows users to generate stylized and structurally viable 3D models and provide five example applications
Pure Event Semantics
No full textIn a pure event semantics for natural language, the domain of quantification and predication is limited to events and states. I offerpure event semantic analyses of several phenomena, some of which have not been treated before in formal semantics. In the pureevent semantics sketched in the second section, nouns are state predicates, and this provides the starting point for the analyses.The phenomena involve grammatical number, the mass-count distinction, adjectival modification, count adjectives, diminutives,lexical plurals, duals, and mass gender. In the conclusion, there is a brief discussion of potential metaphysical or psychologicalramifications of doing semantics this way