Hasil untuk "Law"

Mehdi Cherti, R. Beaumont, Ross Wightman et al.

Scaling up neural networks has led to remarkable performance across a wide range of tasks. Moreover, performance often follows reliable scaling laws as a function of training set size, model size, and compute, which offers valuable guidance as large-scale experiments are becoming increasingly expensive. However, previous work on scaling laws has primarily used private data & models or focused on uni-modal language or vision learning. To address these limitations, we investigate scaling laws for contrastive language-image pre-training (CLIP) with the public LAION dataset and the open-source OpenCLIP repository. Our large-scale experiments involve models trained on up to two billion image-text pairs and identify power law scaling for multiple downstream tasks including zero-shot classification, retrieval, linear probing, and end-to-end fine-tuning. We find that the training distribution plays a key role in scaling laws as the OpenAI and OpenCLIP models exhibit different scaling behavior despite identical model architectures and similar training recipes. We open-source our evaluation workflow and all models, including the largest public CLIP models, to ensure reproducibility and make scaling laws research more accessible. Source code and instructions to reproduce this study is available at https://github.eom/LAION-AI/sealing-laws-openelip.

Dan Hendrycks, Collin Burns, Steven Basart et al.

We propose a new test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more. To attain high accuracy on this test, models must possess extensive world knowledge and problem solving ability. We find that while most recent models have near random-chance accuracy, the very largest GPT-3 model improves over random chance by almost 20 percentage points on average. However, on every one of the 57 tasks, the best models still need substantial improvements before they can reach expert-level accuracy. Models also have lopsided performance and frequently do not know when they are wrong. Worse, they still have near-random accuracy on some socially important subjects such as morality and law. By comprehensively evaluating the breadth and depth of a model's academic and professional understanding, our test can be used to analyze models across many tasks and to identify important shortcomings.

Leo Gao, John Schulman, Jacob Hilton

In reinforcement learning from human feedback, it is common to optimize against a reward model trained to predict human preferences. Because the reward model is an imperfect proxy, optimizing its value too much can hinder ground truth performance, in accordance with Goodhart's law. This effect has been frequently observed, but not carefully measured due to the expense of collecting human preference data. In this work, we use a synthetic setup in which a fixed"gold-standard"reward model plays the role of humans, providing labels used to train a proxy reward model. We study how the gold reward model score changes as we optimize against the proxy reward model using either reinforcement learning or best-of-$n$ sampling. We find that this relationship follows a different functional form depending on the method of optimization, and that in both cases its coefficients scale smoothly with the number of reward model parameters. We also study the effect on this relationship of the size of the reward model dataset, the number of reward model and policy parameters, and the coefficient of the KL penalty added to the reward in the reinforcement learning setup. We explore the implications of these empirical results for theoretical considerations in AI alignment.

Kenneth R. Ahern, Amy K. Dittmar

E. Schlafly, D. Finkbeiner

We present measurements of dust reddening using the colors of stars with spectra in the Sloan Digital Sky Survey. We measure reddening as the difference between the measured and predicted colors of a star, as derived from stellar parameters from the Sloan Extension for Galactic Understanding and Exploration Stellar Parameter Pipeline. We achieve uncertainties of 56, 34, 25, and 29 mmag in the colors u − g, g − r, r − i, and i − z, per star, though the uncertainty varies depending on the stellar type and the magnitude of the star. The spectrum-based reddening measurements confirm our earlier “blue tip” reddening measurements, finding reddening coefficients different by −3%, 1%, 1%, and 2% in u − g, g − r, r − i, and i − z from those found by the blue tip method, after removing a 4% normalization difference. These results prefer an RV = 3.1 Fitzpatrick reddening law to O'Donnell or Cardelli et al. reddening laws. We provide a table of conversion coefficients from the Schlegel et al. (SFD) maps of E(B − V) to extinction in 88 bandpasses for four values of RV, using this reddening law and the 14% recalibration of SFD first reported by Schlafly et al. and confirmed in this work.

E. Rashedi, H. Nezamabadi-pour, S. Saryazdi

Jingqing Han

J. Cardelli, G. Clayton, J. Mathis

J. Murphy, J. P. Riley

P. Kroupa

A universal initial mass function (IMF) is not intuitive, but so far no convincing evidence for a variable IMF exists. The detection of systematic variations of the IMF with star-forming conditions would be the Rosetta Stone for star formation. In this contribution an average or Galactic-field IMF is defined, stressing that there is evidence for a change in the power-law index at only two masses: near 0.5 M⊙ and near 0.08 M⊙. Using this supposed universal IMF, the uncertainty inherent in any observational estimate of the IMF is investigated by studying the scatter introduced by Poisson noise and the dynamical evolution of star clusters. It is found that this apparent scatter reproduces quite well the observed scatter in power-law index determinations, thus defining the fundamental limit within which any true variation becomes undetectable. The absence of evidence for a variable IMF means that any true variation of the IMF in well-studied populations must be smaller than this scatter. Determinations of the power-law indices α are subject to systematic errors arising mostly from unresolved binaries. The systematic bias is quantified here, with the result that the single-star IMFs for young star clusters are systematically steeper by Δα≈0.5 between 0.1 and 1 M⊙ than the Galactic-field IMF, which is populated by, on average, about 5-Gyr-old stars. The MFs in globular clusters appear to be, on average, systematically flatter than the Galactic-field IMF (Piotto & Zoccali; Paresce & De Marchi), and the recent detection of ancient white-dwarf candidates in the Galactic halo and the absence of associated low-mass stars (Ibata et al.; Mendez & Minniti) suggest a radically different IMF for this ancient population. Star formation in higher metallicity environments thus appears to produce relatively more low-mass stars. While still tentative, this is an interesting trend, being consistent with a systematic variation of the IMF as expected from theoretical arguments.

L. Onsager

E. Mansfield

J. Bekenstein

In previous work we introduced the concept of black-hole entropy, which we identified with the surface area of the black hole in question expressed in units of the Planck length squared. We suggested that the appropriate generalization of the second law for a region containing a black hole is that the black-hole entropy plus the common entropy in the black-hole exterior never decreases. Here we establish the validity of this law for the infall of an entropy-bearing system into a much larger and more massive generic stationary black hole. To do this we determine a general lower bound for the increase in black-hole entropy, and an upper bound for the entropy of the system, while allowing for quantum effects at each stage. In passing we show that the generalized second law is a statistical law which becomes over-whelmingly probable in the limit of a macroscopic system. We also consider briefly more general situations. Finally, we give two simple examples of predictions made by the generalized second law for black-hole formation processes.

R. Cooter, T. Ulen

J. Brundage

W. Landes, R. Posner

David R. Johnson, David G. Post

Eugenio Giannelli, Stacey Law, Eoghan McDowell

We show that the $p$-part of the degree of an irreducible character of a symmetric group is completely determined by the set of vanishing elements of $p$-power order. As a corollary we deduce that the set of zeros of prime power order controls the degree of such a character. The same problem is analysed for alternating groups, where we show that when $p=2$ this data can only be determined up to two possibilities. We prove analogous statements for the defect of the $p$-block containing the character and for the $p$-height of the character.

Bim Gustavsson, Stacey Law

Let $p$ be any prime. We determine precisely those irreducible characters of symmetric groups which contain at most $p$ distinct linear constituents in their restriction to a Sylow $p$-subgroup, answering a question of Giannelli and Navarro. Moreover, we identify all of the linear constituents of such characters, and in the case $p = 2$ explicitly calculate a new class of Sylow branching coefficients for symmetric groups indexed by so-called almost hook partitions.

Ayda Mussa Yousif Abdulrahman, Rafiduraida binti Abdul Rahman

This research aims to investigate the current operational status of the Ajman Police, focusing on identifying elements and issues that affect operational excellence. Using change management models, including Kotter's 8 Step Model and the ADKAR Model, the paper critically examines the hierarchical structure of the Ajman Police, its specialist groups, and their performance indicators. The problem statement highlights the negative impact of traditional and rigid organizational structures on innovation, responsiveness, and the limitations of implementing effective public safety measures, prevention, and community policing. The research design adopted is a qualitative methodology, and a sample of senior police officers was interviewed to record their views on the issues of operation and preparedness to change. In conducting the study, Semi-structured interviews were conducted with 10 participants. Results indicate that the Ajman Police has already ventured into technological advancements and civil policing. However, there are still gaps in continuous development, innovation, and the implementation of modern change management practices. The research proposes a culturally, operationally, and technologically oriented framework for change management, specifically tailored to the context of the Ajman Police. The study makes a significant research contribution to both the practice and theory fields by providing a guideline for a change management roadmap for the Ajman Police and other similar agencies, ensuring operational excellence in fast-changing environments.