Search

Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work emphasizes the limitations of RLHF and highlights the importance of a multi-faceted approach to the development of safer AI systems.

Although a superconductor has no DC losses, a superconducting system does have significant losses, especially when it comes to power supply. Here, we study two different power supply systems. The first, a conventional one, consists of a transformer and a diode bridge operating at room temperature, plus current leads that allow the current to flow from the room-temperature medium to the cryogenic medium. The second consists of a transformer with a superconducting secondary winding, combined with a diode bridge operating at cryogenic temperature, thus dispensing with the need for current leads. We are experimentally comparing the performance of conventional and superconducting transformers, as well as the performance of a diode bridge at ambient and cryogenic temperatures. Our results indicate that the prototype superconducting transformer developed has lower winding resistance and secondary leakage inductance than the conventional transformer. In addition, we found that only certain diodes are suitable for operation at cryogenic temperatures. Finally, the diode bridge made from adapted diodes shows reduced losses at cryogenic temperature. This experimental work is the first step towards the realization of a complete power supply system for a superconducting device., Comment: in French language, Symposium de Genie Electrique (SGE 2023), Jul 2023, Lille, France

The effect of two- and three-body interactions on the modulation instability (MI) domain formation of a spin-orbit (SO) and Rabi-coupled Bose-Einstein condensate is studied within a quasi-one-dimensional model. To this aim, we perform numerical and analytical investigations of the associated dispersion relations derived from the corresponding coupled Gross-Pitaevskii equation. The interplay between the linear (SO and Rabi) couplings with the nonlinear cubic-quintic interactions are explored in the mixture, considering miscible and immiscible configurations, with a focus on the impact in the analysis of experimental realizations with general binary coupled systems, in which nonlinear interactions can be widely varied together with linear couplings., Comment: 11 pages, 10 figures

In the anti-Stokes excitation of trivalent rare-earth ions (RE3+), the excitation photons energy is smaller than that of a given absorption transition, and the energy mismatch can be compensated by phonons annihilation from the host lattice. Since the phonon occupation number increases with temperature, heating the system generally increases the efficiency of anti-Stokes excitation. Here, we exploited the intrinsic heating associated with light-to-heat conversion in the interaction of excitation laser light with metallic nanoparticles (Ag or Au) on the surface of submicrometric particles of NdxY1.00-xAl3(BO3)4 (x = 0.10, 0.20, and 1.00) in order to enhance the efficiency of the anti-Stokes excitation at 1064 nm. Several upconversion emissions are observed from 600 nm to 880 nm, the most intense being at 750 nm due to the Nd3+ transition {4F7/2, 4S3/2} - 4I9/2. Giant enhancements are demonstrated, when compared to undecorated NdxY1.00-xAl3(BO3)4 particles. The present results can be expanded to other luminescent materials as well as excitation wavelengths.

Synthetic dimensions (SDs) opened the door for exploring previously inaccessible phenomena in high-dimensional synthetic space. However, construction of synthetic lattices with desired coupling properties is a challenging and unintuitive task, largely limiting the exploration and current application of SD dynamics. Here, we overcome this challenge by using deep learning artificial neural networks (ANNs) to validly design the dynamics in SDs. We use ANNs to construct a lattice in real space that has a predesigned spectrum of mode eigenvalues. By employing judiciously chosen perturbations (wiggling of waveguides), we show experimentally and theoretically resonant mode coupling and tailored dynamics in SDs, which leads to effective transport or confinement of a complex beam profile. As an enlightening example, we demonstrate morphing of light into a topologically protected edge mode in ANN-designed Su-Schrieffer-Heeger photonic lattices. Such ANN-assisted construction of SDs advances towards utopian networks, opening new avenues in fundamental research beyond geometric limitations. Our findings may offer a flexible and efficient solution for mode lasing, optical switching, and communication technologies., Comment: 14 pages and 4 figures

In stopping the spread of infectious diseases, pathogen genomic data can be used to reconstruct transmission events and characterize population-level sources of infection. Most approaches for identifying transmission pairs do not account for the time passing since divergence of pathogen variants in individuals, which is problematic in viruses with high within-host evolutionary rates. This prompted us to consider possible transmission pairs in terms of phylogenetic data and additional estimates of time since infection derived from clinical biomarkers. We develop Bayesian mixture models with an evolutionary clock as signal component and additional mixed effects or covariate random functions describing the mixing weights to classify potential pairs into likely and unlikely transmission pairs. We demonstrate that although sources cannot be identified at the individual level with certainty, even with the additional data on time elapsed, inferences into the population-level sources of transmission are possible, and more accurate than using only phylogenetic data without time since infection estimates. We apply the approach to estimate age-specific sources of HIV infection in Amsterdam MSM transmission networks between 2010-2021. This study demonstrates that infection time estimates provide informative data to characterize transmission sources, and shows how phylogenetic source attribution can then be done with multi-dimensional mixture models.

A kinetic system has an absolute concentration robustness (ACR) for a molecular species if its concentration remains the same in every positive steady state of the system. Just recently, a condition that sufficiently guarantees the existence of an ACR in a rank-one mass-action kinetic system was found. In this paper, it will be shown that this ACR criterion does not extend in general to power-law kinetic systems. Moreover, we also discussed in this paper a necessary condition for ACR in multistationary rank-one kinetic system which can be used in ACR analysis. Finally, a concept of equilibria variation for kinetic systems which are based on the number of the system's ACR species will be introduced here.

We prove that in any rank one symmetric space of non-compact type $M\in\{\mathbb{R} H^n,\mathbb{C} H^m,\mathbb{H} H^m,\mathbb{O} H^2\}$, geodesic spheres are uniformly quantitatively stable with respect to small $C^1$-volume preserving perturbations. We quantify the gain of perimeter in terms of the $W^{1,2}$-norm of the perturbation, taking advantage of the explicit spectral gap of the Laplacian on geodesic spheres in $M$. As a consequence, we give a quantitative proof that for small volumes, geodesic spheres are isoperimetric regions among all sets of finite perimeter., Comment: 24 pages

A principled understanding of generalization in deep learning may require unifying disparate observations under a single conceptual framework. Previous work has studied \emph{grokking}, a training dynamic in which a sustained period of near-perfect training performance and near-chance test performance is eventually followed by generalization, as well as the superficially similar \emph{double descent}. These topics have so far been studied in isolation. We hypothesize that grokking and double descent can be understood as instances of the same learning dynamics within a framework of pattern learning speeds. We propose that this framework also applies when varying model capacity instead of optimization steps, and provide the first demonstration of model-wise grokking., Comment: ML Safety Workshop, 36th Conference on Neural Information Processing Systems (NeurIPS 2022)

The railway industry is searching for new ways to automate a number of complex train functions, such as object detection, track discrimination, and accurate train positioning, which require the artificial perception of the railway environment through different types of sensors, including cameras, LiDARs, wheel encoders, and inertial measurement units. A promising approach for processing such sensory data is the use of deep learning models, which proved to achieve excellent performance in other application domains, as robotics and self-driving cars. However, testing new algorithms and solutions requires the availability of a large amount of labeled data, acquired in different scenarios and operating conditions, which are difficult to obtain in a real railway setting due to strict regulations and practical constraints in accessing the trackside infrastructure and equipping a train with the required sensors. To address such difficulties, this paper presents a visual simulation framework able to generate realistic railway scenarios in a virtual environment and automatically produce inertial data and labeled datasets from emulated LiDARs and cameras useful for training deep neural networks or testing innovative algorithms. A set of experimental results are reported to show the effectiveness of the proposed approach., Comment: Under review

Research in Fairness, Accountability, Transparency, and Ethics (FATE) has established many sources and forms of algorithmic harm, in domains as diverse as health care, finance, policing, and recommendations. Much work remains to be done to mitigate the serious harms of these systems, particularly those disproportionately affecting marginalized communities. Despite these ongoing harms, new systems are being developed and deployed which threaten the perpetuation of the same harms and the creation of novel ones. In response, the FATE community has emphasized the importance of anticipating harms. Our work focuses on the anticipation of harms from increasingly agentic systems. Rather than providing a definition of agency as a binary property, we identify 4 key characteristics which, particularly in combination, tend to increase the agency of a given algorithmic system: underspecification, directness of impact, goal-directedness, and long-term planning. We also discuss important harms which arise from increasing agency -- notably, these include systemic and/or long-range impacts, often on marginalized stakeholders. We emphasize that recognizing agency of algorithmic systems does not absolve or shift the human responsibility for algorithmic harms. Rather, we use the term agency to highlight the increasingly evident fact that ML systems are not fully under human control. Our work explores increasingly agentic algorithmic systems in three parts. First, we explain the notion of an increase in agency for algorithmic systems in the context of diverse perspectives on agency across disciplines. Second, we argue for the need to anticipate harms from increasingly agentic systems. Third, we discuss important harms from increasingly agentic systems and ways forward for addressing them. We conclude by reflecting on implications of our work for anticipating algorithmic harms from emerging systems., Comment: Accepted at FAccT 2023

We provide novel constraints on the parameters defining the universal pressure profile (UPP) within clusters of galaxies, and explore their dependence on the cluster mass and redshift, from measurements of Sunyaev-Zel'dovich Compton-$y$ profiles. We employ both the $\textit{Planck}$ 2015 MILCA and the ACT-DR4 $y$ maps over the common $\sim 2,100\,\text{deg}^2$ footprint. We combine existing cluster catalogs based on KiDS, SDSS and DESI observations, for a total of 23,820 clusters spanning the mass range $10^{14.0}\,\text{M}_{\odot}