623509 research outputs found
Sort by
Learning Video Representations without Natural Videos
No full textWe show that useful video representations can be learned from synthetic videos and natural images, without incorporating natural videos in the training. We propose a progression of video datasets synthesized by simple generative processes, that model a growing set of natural video properties (e.g., motion, acceleration, and shape transformations). The downstream performance of video models pre-trained on these generated datasets gradually increases with the dataset progression. A VideoMAE model pre-trained on our synthetic videos closes 97.2\% of the performance gap on UCF101 action classification between training from scratch and self-supervised pre-training from natural videos, and outperforms the pre-trained model on HMDB51. Introducing crops of static images to the pre-training stage results in similar performance to UCF101 pre-training and outperforms the UCF101 pre-trained model on 11 out of 14 out-of-distribution datasets of UCF101-P. Analyzing the low-level properties of the datasets, we identify correlations between frame diversity, frame similarity to natural data, and downstream performance. Our approach provides a more controllable and transparent alternative to video data curation processes for pre-training.Project page: https://unicorn53547.github.io/video_syn_rep
Constraints on local primordial non-Gaussianity with 3d Velocity Reconstruction from the Kinetic Sunyaev-Zeldovich Effect
No full textThe cosmic velocity field is an unbiased probe of the total matter distribution but is challenging to measure directly at intermediate and high redshifts. The large-scale velocity field imprints a signal in the cosmic microwave background (CMB) through the kinetic Sunyaev-Zeldovich (kSZ) effect. We perform the first 3d reconstruction of the large-scale velocity field from the kSZ effect by applying a quadratic estimator to CMB temperature maps and the 3d positions of galaxies. We do so by combining CMB data from the fifth data release of the Atacama Cosmology Telescope (in combination with Planck) and a spectroscopic galaxy sample from the Sloan Digital Sky Survey. We then measure the galaxy-velocity cross-power spectrum and detect the presence of the kSZ signal at a signal-to-noise ratio of 7.2. Using this galaxy-velocity cross-correlation alone, we constrain the amplitude of local primordial non-Gaussianity finding . This pathfinder measurement sets the stage for joint galaxy-CMB kSZ constraints to significantly enhance the information obtained from galaxy surveys through sample variance cancellation
From Green\u27s formula to Derived Hall algebras
No full textThe aim of this note is to clarify the relationship between Green\u27s formula and the associativity of multiplication for derived Hall algebra in the sense of Toën (Duke Math J 135(3):587-615, 2006), Xiao and Xu (Duke Math J 143(2):357-373, 2008) and Xu and Chen (Algebr Represent Theory 16(3):673-687, 2013). Let be a finitary hereditary abelian category. It is known that the associativity of derived Hall algebra implies Green\u27s formula. We show the converse statement holds. Namely, Green\u27s formula implies the associativity of the derived Hall algebra .24page
Different Horses for Different Courses: Comparing Bias Mitigation Algorithms in ML
No full textWith fairness concerns gaining significant attention in Machine Learning (ML), several bias mitigation techniques have been proposed, often compared against each other to find the best method. These benchmarking efforts tend to use a common setup for evaluation under the assumption that providing a uniform environment ensures a fair comparison. However, bias mitigation techniques are sensitive to hyperparameter choices, random seeds, feature selection, etc., meaning that comparison on just one setting can unfairly favour certain algorithms. In this work, we show significant variance in fairness achieved by several algorithms and the influence of the learning pipeline on fairness scores. We highlight that most bias mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization, suggesting that the choice of the evaluation parameters-rather than the mitigation technique itself-can sometimes create the perceived superiority of one method over another. We hope our work encourages future research on how various choices in the lifecycle of developing an algorithm impact fairness, and trends that guide the selection of appropriate algorithms.To appear at AFME@NeurIPS 202
LLM4DS: Evaluating Large Language Models for Data Science Code Generation
No full textThe adoption of Large Language Models (LLMs) for code generation in data science offers substantial potential for enhancing tasks such as data manipulation, statistical analysis, and visualization. However, the effectiveness of these models in the data science domain remains underexplored. This paper presents a controlled experiment that empirically assesses the performance of four leading LLM-based AI assistants-Microsoft Copilot (GPT-4 Turbo), ChatGPT (o1-preview), Claude (3.5 Sonnet), and Perplexity Labs (Llama-3.1-70b-instruct)-on a diverse set of data science coding challenges sourced from the Stratacratch platform. Using the Goal-Question-Metric (GQM) approach, we evaluated each model\u27s effectiveness across task types (Analytical, Algorithm, Visualization) and varying difficulty levels. Our findings reveal that all models exceeded a 50% baseline success rate, confirming their capability beyond random chance. Notably, only ChatGPT and Claude achieved success rates significantly above a 60% baseline, though none of the models reached a 70% threshold, indicating limitations in higher standards. ChatGPT demonstrated consistent performance across varying difficulty levels, while Claude\u27s success rate fluctuated with task complexity. Hypothesis testing indicates that task type does not significantly impact success rate overall. For analytical tasks, efficiency analysis shows no significant differences in execution times, though ChatGPT tended to be slower and less predictable despite high success rates. This study provides a structured, empirical evaluation of LLMs in data science, delivering insights that support informed model selection tailored to specific task demands. Our findings establish a framework for future AI assessments, emphasizing the value of rigorous evaluation beyond basic accuracy measures.11 page
ChatHTTPFuzz: Large Language Model-Assisted IoT HTTP Fuzzing
No full textInternet of Things (IoT) devices offer convenience through web interfaces, web VPNs, and other web-based services, all relying on the HTTP protocol. However, these externally exposed HTTP services resent significant security risks. Although fuzzing has shown some effectiveness in identifying vulnerabilities in IoT HTTP services, most state-of-the-art tools still rely on random mutation trategies, leading to difficulties in accurately understanding the HTTP protocol\u27s structure and generating many invalid test cases. Furthermore, These fuzzers rely on a limited set of initial seeds for testing. While this approach initiates testing, the limited number and diversity of seeds hinder comprehensive coverage of complex scenarios in IoT HTTP services. In this paper, we investigate and find that large language models (LLMs) excel in parsing HTTP protocol data and analyzing code logic. Based on these findings, we propose a novel LLM-guided IoT HTTP fuzzing method, ChatHTTPFuzz, which automatically parses protocol fields and analyzes service code logic to generate protocol-compliant test cases. Specifically, we use LLMs to label fields in HTTP protocol data, creating seed templates. Second, The LLM analyzes service code to guide the generation of additional packets aligned with the code logic, enriching the seed templates and their field values. Finally, we design an enhanced Thompson sampling algorithm based on the exploration balance factor and mutation potential factor to schedule seed templates. We evaluate ChatHTTPFuzz on 14 different real-world IoT devices. It finds more vulnerabilities than SNIPUZZ, BOOFUZZ, and MUTINY. ChatHTTPFuzz has discovered 103 vulnerabilities, of which 68 are unique, and 23 have been assigned CVEs
Learning complexity gradually in quantum machine learning models
No full textQuantum machine learning is an emergent field that continues to draw significant interest for its potential to offer improvements over classical algorithms in certain areas. However, training quantum models remains a challenging task, largely because of the difficulty in establishing an effective inductive bias when solving high-dimensional problems. In this work, we propose a training framework that prioritizes informative data points over the entire training set. This approach draws inspiration from classical techniques such as curriculum learning and hard example mining to introduce an additional inductive bias through the training data itself. By selectively focusing on informative samples, we aim to steer the optimization process toward more favorable regions of the parameter space. This data-centric approach complements existing strategies such as warm-start initialization methods, providing an additional pathway to address performance challenges in quantum machine learning. We provide theoretical insights into the benefits of prioritizing informative data for quantum models, and we validate our methodology with numerical experiments on selected recognition tasks of quantum phases of matter. Our findings indicate that this strategy could be a valuable approach for improving the performance of quantum machine learning models.10 + 2 pages, 6 figures, 3 table
Scalable fabrication of erbium-doped high-Q silica microtoroid resonators via sol-gel coating
No full textThis study explores sol-gel methods for fabricating erbium-doped silica microtoroid resonators, addressing the limitations of conventional doping techniques and enhancing device scalability. We develop a reproducible sol-gel process that yields defect-free films for photonic applications, and detail common defects and troubleshooting strategies. Two fabrication methods are compared: traditional film deposition on substrates and the direct coating of prefabricated resonators. The latter enables the fabrication of larger resonator diameters (up to 450 μm) without buckling, while achieving a high-Q factor and a low lasing threshold of 350 μW. These erbium-doped resonators exhibit multi-mode laser oscillations at 1550 nm, revealing the sol-gel method\u27s potential for realizing scalable, gain-doped photonic devices
Efficient Federated Unlearning with Adaptive Differential Privacy Preservation
No full textFederated unlearning (FU) offers a promising solution to effectively address the need to erase the impact of specific clients\u27 data on the global model in federated learning (FL), thereby granting individuals the ``Right to be Forgotten . The most straightforward approach to achieve unlearning is to train the model from scratch, excluding clients who request data removal, but it is resource-intensive. Current state-of-the-art FU methods extend traditional FL frameworks by leveraging stored historical updates, enabling more efficient unlearning than training from scratch. However, the use of stored updates introduces significant privacy risks. Adversaries with access to these updates can potentially reconstruct clients\u27 local data, a well-known vulnerability in the privacy domain. While privacy-enhanced techniques exist, their applications to FU scenarios that balance unlearning efficiency with privacy protection remain underexplored. To address this gap, we propose FedADP, a method designed to achieve both efficiency and privacy preservation in FU. Our approach incorporates an adaptive differential privacy (DP) mechanism, carefully balancing privacy and unlearning performance through a novel budget allocation strategy tailored for FU. FedADP also employs a dual-layered selection process, focusing on global models with significant changes and client updates closely aligned with the global model, reducing storage and communication costs. Additionally, a novel calibration method is introduced to facilitate effective unlearning. Extensive experimental results demonstrate that FedADP effectively manages the trade-off between unlearning efficiency and privacy protection
Amortized Analysis of Leftist Heaps
No full textLeftist heaps and skew heaps are two well-known data structures for mergeable priority queues. Leftist heaps are constructed for efficiency in the worst-case sense whereas skew heaps are self-adjusting, designed for efficiency in the amortized sense. In this paper, we analyze the amortized complexity of leftist heaps to initiate a full performance comparison with skew heaps. We consider both the leftist heaps originally developed by Crane and Knuth, which are also referred to as rank-biased (or, height-biased) leftist heaps, and the weight-biased leftist heaps introduced by Cho and Sahni. We show how weight-biased leftist heaps satisfy the same exact amortized bounds as skew heaps. With these matching bounds we establish a nice trade-off in which storage of weights is used to limit the worst-case complexity of leftist heaps, without affecting the amortized complexity compared to skew heaps. For rank-biased leftist heaps, we obtain the same amortized lower bounds as for skew heaps, but whether these bounds are tight is left as an open problem.13 pages, 3 figures, full version (incl. Appendix B