Search

Kinematic lensing (KL) is a new weak lensing technique that reduces shape noise for disk galaxies by including spectroscopically measured galaxy kinematics in addition to photometrically measured galaxy shapes. Since KL utilizes the Tully-Fisher relation, any correlation of this relation with the local environment may bias the cosmological interpretation. For the first time, we explore such a Tully-Fisher environmental dependence (TED) effect as a potential astrophysical systematic for KL. Our derivation of the TED systematic can be described in a similar analytical form as intrinsic alignment for traditional weak lensing. We demonstrate analytically that TED only impacts KL if intrinsic aligment for disk galaxies is non-zero. We further use IllustrisTNG simulations to quantify the TED effect. Our two-point correlation measurements do not yield any additional coherent signals that would indicate a systematic bias on KL, within the uncertainties set by the simulation volume., Comment: 12 pages, 5 figures, published in PRD

Predicting the evolution of spatiotemporal physical systems from sparse and scattered observational data poses a significant challenge in various scientific domains. Traditional methods rely on dense grid-structured data, limiting their applicability in scenarios with sparse observations. To address this challenge, we introduce GrINd (Grid Interpolation Network for Scattered Observations), a novel network architecture that leverages the high-performance of grid-based models by mapping scattered observations onto a high-resolution grid using a Fourier Interpolation Layer. In the high-resolution space, a NeuralPDE-class model predicts the system's state at future timepoints using differentiable ODE solvers and fully convolutional neural networks parametrizing the system's dynamics. We empirically evaluate GrINd on the DynaBench benchmark dataset, comprising six different physical systems observed at scattered locations, demonstrating its state-of-the-art performance compared to existing models. GrINd offers a promising approach for forecasting physical systems from sparse, scattered observational data, extending the applicability of deep learning methods to real-world scenarios with limited data availability.

Accurate vegetation models can produce further insights into the complex interaction between vegetation activity and ecosystem processes. Previous research has established that long-term trends and short-term variability of temperature and precipitation affect vegetation activity. Motivated by the recent success of Transformer-based Deep Learning models for medium-range weather forecasting, we adapt the publicly available pre-trained FourCastNet to model vegetation activity while accounting for the short-term dynamics of climate variability. We investigate how the learned global representation of the atmosphere's state can be transferred to model the normalized difference vegetation index (NDVI). Our model globally estimates vegetation activity at a resolution of \SI{0.25}{\degree} while relying only on meteorological data. We demonstrate that leveraging pre-trained weather models improves the NDVI estimates compared to learning an NDVI model from scratch. Additionally, we compare our results to other recent data-driven NDVI modeling approaches from machine learning and ecology literature. We further provide experimental evidence on how much data and training time is necessary to turn FourCastNet into an effective vegetation model. Code and models will be made available upon publication., Comment: Tackling Climate Change with Machine Learning Workshop @ ICLR 2024

Stochastic Nonlinear Optimal Control (SNOC) involves minimizing a cost function that averages out the random uncertainties affecting the dynamics of nonlinear systems. For tractability reasons, this problem is typically addressed by minimizing an empirical cost, which represents the average cost across a finite dataset of sampled disturbances. However, this approach raises the challenge of quantifying the control performance against out-of-sample uncertainties. Particularly, in scenarios where the training dataset is small, SNOC policies are prone to overfitting, resulting in significant discrepancies between the empirical cost and the true cost, i.e., the average SNOC cost incurred during control deployment. Therefore, establishing generalization bounds on the true cost is crucial for ensuring reliability in real-world applications. In this paper, we introduce a novel approach that leverages PAC-Bayes theory to provide rigorous generalization bounds for SNOC. Based on these bounds, we propose a new method for designing optimal controllers, offering a principled way to incorporate prior knowledge into the synthesis process, which aids in improving the control policy and mitigating overfitting. Furthermore, by leveraging recent parametrizations of stabilizing controllers for nonlinear systems, our framework inherently ensures closed-loop stability. The effectiveness of our proposed method in incorporating prior knowledge and combating overfitting is shown by designing neural network controllers for tasks in cooperative robotics.

Distance estimation from audio plays a crucial role in various applications, such as acoustic scene analysis, sound source localization, and room modeling. Most studies predominantly center on employing a classification approach, where distances are discretized into distinct categories, enabling smoother model training and achieving higher accuracy but imposing restrictions on the precision of the obtained sound source position. Towards this direction, in this paper we propose a novel approach for continuous distance estimation from audio signals using a convolutional recurrent neural network with an attention module. The attention mechanism enables the model to focus on relevant temporal and spectral features, enhancing its ability to capture fine-grained distance-related information. To evaluate the effectiveness of our proposed method, we conduct extensive experiments using audio recordings in controlled environments with three levels of realism (synthetic room impulse response, measured response with convolved speech, and real recordings) on four datasets (our synthetic dataset, QMULTIMIT, VoiceHome-2, and STARSS23). Experimental results show that the model achieves an absolute error of 0.11 meters in a noiseless synthetic scenario. Moreover, the results showed an absolute error of about 1.30 meters in the hybrid scenario. The algorithm's performance in the real scenario, where unpredictable environmental factors and noise are prevalent, yields an absolute error of approximately 0.50 meters. For reproducible research purposes we make model, code, and synthetic datasets available at https://github.com/michaelneri/audio-distance-estimation., Comment: Accepted for publication in IEEE/ACM Transactions on Audio, Speech, and Language Processing

In recent years, advances in text-to-image (T2I) diffusion models have substantially elevated the quality of their generated images. However, achieving fine-grained control over attributes remains a challenge due to the limitations of natural language prompts (such as no continuous set of intermediate descriptions existing between ``person'' and ``old person''). Even though many methods were introduced that augment the model or generation process to enable such control, methods that do not require a fixed reference image are limited to either enabling global fine-grained attribute expression control or coarse attribute expression control localized to specific subjects, not both simultaneously. We show that there exist directions in the commonly used token-level CLIP text embeddings that enable fine-grained subject-specific control of high-level attributes in text-to-image models. Based on this observation, we introduce one efficient optimization-free and one robust optimization-based method to identify these directions for specific attributes from contrastive text prompts. We demonstrate that these directions can be used to augment the prompt text input with fine-grained control over attributes of specific subjects in a compositional manner (control over multiple attributes of a single subject) without having to adapt the diffusion model. Project page: https://compvis.github.io/attribute-control. Code is available at https://github.com/CompVis/attribute-control., Comment: Project page: https://compvis.github.io/attribute-control

We present SIM-FSVGD for learning robot dynamics from data. As opposed to traditional methods, SIM-FSVGD leverages low-fidelity physical priors, e.g., in the form of simulators, to regularize the training of neural network models. While learning accurate dynamics already in the low data regime, SIM-FSVGD scales and excels also when more data is available. We empirically show that learning with implicit physical priors results in accurate mean model estimation as well as precise uncertainty quantification. We demonstrate the effectiveness of SIM-FSVGD in bridging the sim-to-real gap on a high-performance RC racecar system. Using model-based RL, we demonstrate a highly dynamic parking maneuver with drifting, using less than half the data compared to the state of the art.

Sound Event Detection and Localization (SELD) is a combined task of identifying sound events and their corresponding direction-of-arrival (DOA). While this task has numerous applications and has been extensively researched in recent years, it fails to provide full information about the sound source position. In this paper, we overcome this problem by extending the task to Sound Event Detection, Localization with Distance Estimation (3D SELD). We study two ways of integrating distance estimation within the SELD core - a multi-task approach, in which the problem is tackled by a separate model output, and a single-task approach obtained by extending the multi-ACCDOA method to include distance information. We investigate both methods for the Ambisonic and binaural versions of STARSS23: Sony-TAU Realistic Spatial Soundscapes 2023. Moreover, our study involves experiments on the loss function related to the distance estimation part. Our results show that it is possible to perform 3D SELD without any degradation of performance in sound event detection and DOA estimation., Comment: This paper has been accepted for the 32nd European Signal Processing Conference EUSIPCO 2024 in Lyon

Validating modeling choices through simulated analyses and quantifying the impact of different systematic effects will form a major computational bottleneck in the preparation for 3$\times$2 analysis with Stage-IV surveys such as Vera Rubin Observatory's Legacy Survey of Space and Time (LSST). We can significantly reduce the computational requirements by using machine learning based emulators, which allow us to run fast inference while maintaining the full realism of the data analysis pipeline. In this paper, we use such an emulator to run simulated 3$\times$2 (cosmic shear, galaxy-galaxy lensing, and galaxy clustering) analyses for mock LSST-Y1/Y3/Y6/Y10 surveys and study the impact of various systematic effects (galaxy bias, intrinsic alignment, baryonic physics, shear calibration and photo-$z$ uncertainties). Closely following the DESC Science Requirement Document (with several updates) our main findings are: {\it a)} The largest contribution to the `systematic error budget' of LSST 3$\times$2 analysis comes from galaxy bias uncertainties, while the contribution of baryonic and shear calibration uncertainties are significantly less important. {\it b)} Tighter constraints on intrinsic alignment and photo-$z$ parameters can improve cosmological constraints noticeably, which illustrates synergies of LSST and spectroscopic surveys. {\it c)} The scale cuts adopted in the DESC SRD may be too conservative and pushing to smaller scales can increase cosmological information significantly. {\it d)} We investigate the impact of photo-$z$ outliers on 3$\times$2 pt analysis and find that we need to determine the outlier fraction to within $5-10\%$ accuracy to ensure robust cosmological analysis. We caution that these findings depend on analysis choices (parameterizations, priors, scale cuts) and can change for different settings., Comment: 16 pages, 10 Figures, To be submitted to PRD. Comments welcome

The next generation of cosmological surveys is expected to generate unprecedented high-quality data, consequently increasing the already substantial computational costs of Bayesian statistical methods. This will pose a significant challenge to analyzing theoretical models of cosmology. Additionally, new mitigation techniques of baryonic effects, intrinsic alignment, and other systematic effects will inevitably introduce more parameters, slowing down the convergence of Bayesian analyses. In this scenario, machine-learning-based accelerators are a promising solution, capable of reducing the computational costs and execution time of such tools by order of thousands. Yet, they have not been able to provide accurate predictions over the wide prior ranges in parameter space adopted by Stage III/IV collaborations in studies employing real-space two-point correlation functions. This paper offers a leap in this direction by carefully investigating the modern transformer-based neural network (NN) architectures in realistic simulated Rubin Observatory year one cosmic shear $\Lambda$CDM inferences. Building on the framework introduced in Part I, we generalize the transformer block and incorporate additional layer types to develop a more versatile architecture. We present a scalable method to efficiently generate an extensive training dataset that significantly exceeds the scope of prior volumes considered in Part I, while still meeting strict accuracy standards. Through our meticulous architecture comparison and comprehensive hyperparameter optimization, we establish that the attention-based architecture performs an order of magnitude better in accuracy than widely adopted NN designs. Finally, we test and apply our emulators to calibrate tension metrics.

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

The nearby Supernova 1987A was accompanied by a burst of neutrino emission, which indicates that a compact object (a neutron star or black hole) was formed in the explosion. There has been no direct observation of this compact object. In this work, we observe the supernova remnant with JWST spectroscopy finding narrow infrared emission lines of argon and sulphur. The line emission is spatially unresolved and blueshifted in velocity relative to the supernova rest frame. We interpret the lines as gas illuminated by a source of ionizing photons located close to the center of the expanding ejecta. Photoionization models show that the line ratios are consistent with ionization by a cooling neutron star or pulsar wind nebula. The velocity shift could be evidence for a neutron star natal kick., Comment: Authors version of manuscript published in Science on 22 Feb 2024

Recent research shows that quasiparticle-induced decoherence of superconducting qubits depends on the superconducting-gap asymmetry originating from the different thicknesses of the top and bottom films in Al/AlO$_x$/Al junctions. Magnetic field is a key tuning knob to investigate this dependence as it can change the superconducting gaps in situ. We present measurements of the parity-switching time of a field-resilient 3D transmon with in-plane field up to 0.41T. At low fields, small parity splitting requires qutrit pulse sequences for parity measurements. We measure a non-monotonic evolution of the parity lifetime with in-plane magnetic field, increasing up to 0.2T, followed by a decrease at higher fields. We demonstrate that the superconducting-gap asymmetry plays a crucial role in the observed behavior. At zero field, the qubit frequency is nearly resonant with the superconducting-gap difference, favoring the energy exchange with the quasiparticles and so enhancing the parity-switching rate. With a higher magnetic field, the qubit frequency decreases and gets detuned from the gap difference, causing the initial increase of the parity lifetime, while photon-assisted qubit transitions increase, producing the subsequent decrease at higher fields. Besides giving a deeper insight into the parity-switching mechanism in conventional transmon qubits, we establish that Al-AlO$_x$-Al JJs could be used in architectures for the parity-readout and manipulation of topological qubits based on Majorana zero modes.

The precession of astrophysical jets produced by active-galactic nuclei is likely related to the dynamics of the accretion disks surrounding the central supermassive black holes (BHs) from which jets are launched. The two main mechanisms that can drive jet precession arise from Lense-Thirring precession and tidal torquing. These can explain direct and indirect observations of precessing jets; however, such explanations often utilize crude approximations of the disk evolution and observing jet precession can be challenging with electromagnetic facilities. Simultaneously, the Laser Interferometer Space Antenna (LISA) is expected to measure gravitational waves from the mergers of massive binary BHs with high accuracy and probe their progenitor evolution. In this paper, we connect the LISA detectability of binary BH mergers to the possible jet precession during their progenitor evolution. We make use of a semi-analytic model that self-consistently treats disk-driven BH alignment and binary inspiral and includes the possibility of disk breaking. We find that tidal torquing of the accretion disk provides a wide range of jet precession timescales depending on the binary separation and the spin direction of the BH from which the jet is launched. Efficient disk-driven BH alignment results in shorter timescales of $\sim 1$ yr which are correlated with higher LISA signal-to-noise ratios. Disk breaking results in the longest possible times of $\sim 10^7$ yrs, suggesting a deep interplay between the disk critical obliquity (i.e. where the disk breaks) and jet precession. Studies such as ours will help to reveal the cosmic population of precessing jets that are detectable with gravitational waves.

We investigate the effect of magnetic field on a photonic-crystal Josephson traveling-wave parametric amplifier (TWPA). We show that the observed change in photonic bandgap and plasma frequency of the TWPA can be modeled by considering the suppression of the critical current in the Josephson junctions (JJs) of the TWPA due to the Fraunhofer effect and closing of the superconducting gap. Accounting for the JJ geometry is crucial for understanding the field dependence. In one in-plane direction, the TWPA bandgap can be shifted by 2 GHz using up to 60 mT of field, without losing gain or bandwidth, showing that TWPAs without SQUIDs can be field tunable. In the other in-plane direction, the magnetic field is perpendicular to the larger side of the Josephson junctions, so the Fraunhofer effect has a smaller period. This larger side of the JJs is modulated to create the bandgap. The field interacts more strongly with the larger junctions, and as a result, the TWPA bandgap closes and reopens as the field increases, causing the TWPA to become severely compromised already at 2 mT. A slightly higher operating limit of 5 mT is found in out-of-plane field, for which the TWPA's response is hysteretic. These measurements reveal the requirements for magnetic shielding needed to use TWPAs in experiments where high fields at the sample are required; we show that with magnetic shields we can operate the TWPA while applying over 2 T to the sample.

We present a new class of machine-learning emulators that accurately model the cosmic shear, galaxy-galaxy lensing, and galaxy clustering real space correlation functions in the context of Rubin Observatory year one simulated data. To illustrate its capabilities in forecasting models beyond the standard $\Lambda$CDM, we forecast how well LSST Year 1 data will be able to probe the consistency between geometry $\Omega^{\rm geo}_\mathrm{m}$ and growth $\Omega^{\rm growth}_\mathrm{m}$ dark matter densities in the so-called split $\Lambda$CDM parameterization. When trained with a few million samples, our emulator shows uniform accuracy across a wide range in an 18-dimensional parameter space. We provide a detailed comparison of three neural network designs, illustrating the importance of adopting state-of-the-art Transformer blocks. Our study also details their performance when computing Bayesian evidence for cosmic shear on three fiducial cosmologies. The transformers-based emulator is always accurate within PolyChord's precision. As an application, we use our emulator to study the degeneracies between dark energy models and growth geometry split parameterizations. We find that the growth-geometry split remains to be a meaningful test of the smooth dark energy assumption., Comment: 16 pages, 6 figures

ZnSe and related materials like ZnMgSe and ZnCdSe are promising II-VI host materials for optically mediated quantum information technology such as single photon sources or spin qubits. Integrating these heterostructures into photonic crystal (PC) cavities enables further improvements, for example realizing Purcell-enhanced single photon sources with increased quantum efficiency. Here we report on the successful implementation of two-dimensional (2D) PC cavities in strained ZnSe quantum wells (QW) on top of a novel AlAs supporting layer. This approach overcomes typical obstacles associated with PC membrane fabrication in strained materials, such as cracks and strain relaxation in the corresponding devices. We demonstrate the attainment of the required mechanical stability in our PC devices, complete strain retainment and effective vertical optical confinement. Structural analysis of our PC cavities reveals excellent etching anisotropy. Additionally, elemental mapping in a scanning transmission electron microscope confirms the transformation of AlAs into AlOx by post-growth wet oxidation and reveals partial oxidation of ZnMgSe at the etched sidewalls in the PC. This knowledge is utilized to tailor FDTD simulations and to extract the ZnMgSe dispersion relation with small oxygen content. Optical characterization of the PC cavities with cross-polarized resonance scattering spectroscopy verifies the presence of cavity modes. The excellent agreement between simulation and measured cavity mode energies demonstrates wide tunability of the PC cavity and proves the pertinence of our model. This implementation of 2D PC cavities in the ZnSe material system establishes a solid foundation for future developments of ZnSe quantum devices.

Phase winding effects in hollow semiconductor nanowires with superconducting shells have been proposed as a route to engineer topological superconducting states. We investigate GaAs/InAs core/shell nanowires with half-shells of epitaxial aluminium as a potential platform for such devices, where the thin InAs shell confines the electron wave function around the GaAs core. With normal contacts we observed pronounced $h/e$ flux periodic oscillations in the magnetoconductance, indicating the presence of a tubular conductive channel in the InAs shell. Conversely, the switching current in Josephson junctions oscillates with approximately half that period, i.e. $h/2e$, indicating transport via Andreev transport processes in the junction enclosing threading magnetic flux. On these structures, we systematically studied the gate-, field-, and temperature-dependent evolution of the supercurrent. Results indicate that Andreev transport processes can occur about the wire circumference indicating full proximitization of the InAs shell from the half-shell superconducting contacts., Comment: 13 pages, 6 figures

There exist very few mid-infrared (IR) observations of supernovae (SNe) in general. Therefore, SN 1987A, the closest visible SN in 400 years, gives us the opportunity to explore the mid-IR properties of SNe, the dust in their ejecta and surrounding medium, and to witness the birth of a SN remnant (SNR). The James Webb Space Telescope (JWST), with its high spatial resolution and extreme sensitivity, gives a new view on these issues. We report on the first imaging observations obtained with the Mid-InfraRed Instrument (MIRI). We build temperature maps and discuss the morphology of the nascent SNR. Our results show that the temperatures in the equatorial ring (ER) are quite non-uniform. This could be due to dust destruction in some parts of the ring, as had been assumed in some previous works. We show that the IR emission extends beyond the ER, illustrating the fact that the shock wave has now passed through this ring to affect the circumstellar medium on a larger scale. Finally, while sub-mm Atacama Large Millimeter Array (ALMA) observations have hinted at the location of the compact remnant of SN 1987A, we note that our MIRI data have found no such evidence., Comment: 19 pages, 19 figures, 2 tables; Accepted for publication in the Astrophysical Journal (February 2, 2024)

Due to its optimal complexity, the multigrid (MG) method is one of the most popular approaches for solving large-scale linear systems arising from the discretization of partial differential equations. However, the parallel implementation of standard MG methods, which are inherently multiplicative, suffers from increasing communication complexity. In such cases, the additive variants of MG methods provide a good alternative due to their inherently parallel nature, although they exhibit slower convergence. This work combines the additive multigrid method with the multipreconditioned conjugate gradient (MPCG) method. In the proposed approach, the MPCG method employs the corrections from the different levels of the MG hierarchy as separate preconditioned search directions. In this approach, the MPCG method updates the current iterate by using the linear combination of the preconditioned search directions, where the optimal coefficients for the linear combination are computed by exploiting the energy norm minimization of the CG method. The idea behind our approach is to combine the $A$-conjugacy of the search directions of the MPCG method and the quasi $H_1$-orthogonality of the corrections from the MG hierarchy. In the numerical section, we study the performance of the proposed method compared to the standard additive and multiplicative MG methods used as preconditioners for the CG method.

We illustrate how purpose-specific, graphical modeling enables application experts with different levels of expertise to collaboratively design and then produce complex applications using their individual, purpose-specific modeling language. Our illustration includes seven graphical Integrated Modeling Environments (IMEs) that support full code generation, as well as four browser-based applications that were modeled and then fully automatically generated and produced using DIME, our most complex graphical IME. While the seven IMEs were chosen to illustrate the types of languages we support with our Language-Driven Engineering (LDE) approach, the four DIME products were chosen to give an impression of the power of our LDE-generated IMEs. In fact, Equinocs, Springer Nature's future editorial system for proceedings, is also being fully automatically generated and then deployed at their Dordrecht site using a deployment pipeline generated with Rig, one of the IMEs presented. Our technology is open source and the products presented are currently in use., Comment: 43 pages, 30 figures

The current study investigates the asymptotic spectral properties of a finite difference approximation of nonlocal Helmholtz equations with a Caputo fractional Laplacian and a variable coefficient wave number $\mu$, as it occurs when considering a wave propagation in complex media, characterized by nonlocal interactions and spatially varying wave speeds. More specifically, by using tools from Toeplitz and generalized locally Toeplitz theory, the present research delves into the spectral analysis of nonpreconditioned and preconditioned matrix-sequences. We report numerical evidences supporting the theoretical findings. Finally, open problems and potential extensions in various directions are presented and briefly discussed., Comment: 28 pages, 10 figures. arXiv admin note: text overlap with arXiv:2206.05171 by other authors

Odd Radio Circles (ORCs) are a class of low surface brightness, circular objects approximately one arcminute in diameter. ORCs were recently discovered in the Australian Square Kilometre Array Pathfinder (ASKAP) data, and subsequently confirmed with follow-up observations on other instruments, yet their origins remain uncertain. In this paper, we suggest that ORCs could be remnant lobes of powerful radio galaxies, re-energised by the passage of a shock. Using relativistic hydrodynamic simulations with synchrotron emission calculated in post-processing, we show that buoyant evolution of remnant radio lobes is alone too slow to produce the observed ORC morphology. However, the passage of a shock can produce both filled and edge-brightnened ORC-like morphologies for a wide variety of shock and observing orientations. Circular ORCs are predicted to have host galaxies near the geometric centre of the radio emission, consistent with observations of these objects. Significantly offset hosts are possible for elliptical ORCs, potentially causing challenges for accurate host galaxy identification. Observed ORC number counts are broadly consistent with a paradigm in which moderately powerful radio galaxies are their progenitors., Comment: 25 pages, 15 figures. Accepted for publication in PASA

We study a generalization of classical active learning to real-world settings with concrete prediction targets where sampling is restricted to an accessible region of the domain, while prediction targets may lie outside this region. We analyze a family of decision rules that sample adaptively to minimize uncertainty about prediction targets. We are the first to show, under general regularity assumptions, that such decision rules converge uniformly to the smallest possible uncertainty obtainable from the accessible data. We demonstrate their strong sample efficiency in two key applications: active fine-tuning of large neural networks and safe Bayesian optimization, where they achieve state-of-the-art performance., Comment: accepted in NeurIPS 2024. arXiv admin note: text overlap with arXiv:2402.15441

We study the question: How can we select the right data for fine-tuning to a specific task? We call this data selection problem active fine-tuning and show that it is an instance of transductive active learning, a novel generalization of classical active learning. We propose ITL, short for information-based transductive learning, an approach which samples adaptively to maximize information gained about the specified task. We are the first to show, under general regularity assumptions, that such decision rules converge uniformly to the smallest possible uncertainty obtainable from the accessible data. We apply ITL to the few-shot fine-tuning of large neural networks and show that fine-tuning with ITL learns the task with significantly fewer examples than the state-of-the-art.