Search

Our understanding of Solar System evolution is closely tied to interpretations of asteroid composition, particularly the M-class asteroids. These asteroids were initially thought to be the exposed cores of differentiated planetesimals, a hypothesis based on their spectral similarity to iron meteorites. However, recent astronomical observations have revealed hydration on their surface through the detection of 3-$\mu$m absorption features associated with OH and potentially H2O. We present evidence of hydration due mainly to OH on asteroid (16) Psyche, the largest M-class asteroid, using data from the James Webb Space Telescope (JWST) spanning 1.1 - 6.63 $\mu$m. Our observations include two detections of the full 3-$\mu$m feature associated with OH and H2O resembling those found in CY-, CH-, and CB-type carbonaceous chondrites, and no 6-$\mu$m feature uniquely associated with H2O across two observations. We observe 3-$\mu$m depths of between 4.3 and 6% across two observations, values consistent with hydrogen abundance estimates on other airless bodies of 250 - 400 ppm. We place an upper limit of 39 ppm on the water abundance from the standard deviation around the 6-$\mu$m feature region. The presence of hydrated minerals suggests a complex history for Psyche. Exogenous sources of OH-bearing minerals could come from hydrated impactors. Endogenous OH-bearing minerals would indicate a composition more similar to E-or P-class asteroids. If the hydration is endogenous, it supports the theory that Psyche originated beyond the snow line and later migrated to the outer main belt.

Large Language Models (LLMs) tend to be unreliable in the factuality of their answers. To address this problem, NLP researchers have proposed a range of techniques to estimate LLM's confidence over facts. However, due to the lack of a systematic comparison, it is not clear how the different methods compare to one another. To fill this gap, we present a survey and empirical comparison of estimators of factual confidence. We define an experimental framework allowing for fair comparison, covering both fact-verification and question answering. Our experiments across a series of LLMs indicate that trained hidden-state probes provide the most reliable confidence estimates, albeit at the expense of requiring access to weights and training data. We also conduct a deeper assessment of factual confidence by measuring the consistency of model behavior under meaning-preserving variations in the input. We find that the confidence of LLMs is often unstable across semantically equivalent inputs, suggesting that there is much room for improvement of the stability of models' parametric knowledge. Our code is available at (https://github.com/amazon-science/factual-confidence-of-llms)., Comment: accepted on the main track of ACL 2024

Preceding a core-collapse supernova, various processes produce an increasing amount of neutrinos of all flavors characterized by mounting energies from the interior of massive stars. Among them, the electron antineutrinos are potentially detectable by terrestrial neutrino experiments such as KamLAND and Super-Kamiokande via inverse beta decay interactions. Once these pre-supernova neutrinos are observed, an early warning of the upcoming core-collapse supernova can be provided. In light of this, KamLAND and Super-Kamiokande, both located in the Kamioka mine in Japan, have been monitoring pre-supernova neutrinos since 2015 and 2021, respectively. Recently, we performed a joint study between KamLAND and Super-Kamiokande on pre-supernova neutrino detection. A pre-supernova alert system combining the KamLAND detector and the Super-Kamiokande detector was developed and put into operation, which can provide a supernova alert to the astrophysics community. Fully leveraging the complementary properties of these two detectors, the combined alert is expected to resolve a pre-supernova neutrino signal from a 15 M$_{\odot}$ star within 510 pc of the Earth, at a significance level corresponding to a false alarm rate of no more than 1 per century. For a Betelgeuse-like model with optimistic parameters, it can provide early warnings up to 12 hours in advance., Comment: Resubmitted to ApJ. 22 pages, 16 figures, for more information about the combined pre-supernova alert system, see https://www.lowbg.org/presnalarm/

We present labeled morphological segmentation, an alternative view of morphological processing that unifies several tasks. From an annotation standpoint, we additionally introduce a new hierarchy of morphotactic tagsets. Finally, we develop \modelname, a discriminative morphological segmentation system that, contrary to previous work, explicitly models morphotactics. We show that \textsc{chipmunk} yields improved performance on three tasks for all six languages: (i) morphological segmentation, (ii) stemming and (iii) morphological tag classification. On morphological segmentation, our method shows absolute improvements of 2--6 points $F_1$ over the baseline., Comment: CoNLL 2015

The field of edge computing has witnessed remarkable growth owing to the increasing demand for real-time processing of data in applications. However, challenges persist due to limitations in performance and power consumption. To overcome these challenges, heterogeneous architectures have emerged that combine host processors with specialized accelerators tailored to specific applications, leading to improved performance and reduced power consumption. However, most of the existing platforms lack the necessary configurability and extendability options for integrating custom accelerators. To overcome these limitations, we introduce in this paper the eXtendible Heterogeneous Energy-Efficient Platform (X-HEEP). X-HEEP is an open-source platform designed to natively support the integration of ultra-low-power edge accelerators. It provides customization options to match specific application requirements by exploring various core types, bus topologies, addressing modes, memory sizes, and peripherals. Moreover, the platform prioritizes energy efficiency by implementing low-power strategies, such as clock-gating and power-gating. We demonstrate the real-world applicability of X-HEEP by providing an integration example tailored for healthcare applications that includes a coarse-grained reconfigurable array (CGRA) and in-memory computing (IMC) accelerators. The resulting design, called HEEPocrates, has been implemented both in field programmable gate array (FPGA) on the Xilinx Zynq-7020 chip and in silicon with TSMC 65nm low-power CMOS technology. We run a set of healthcare applications and measure their energy consumption to demonstrate the alignment of our chip with other state-of-the-art microcontrollers commonly adopted in this domain. Moreover, we present the energy benefits of 4.9x and 4.8x gained by exploiting the integrated CGRA and IMC accelerators compared to running on the host CPU.

Neural graphics primitives are faster and achieve higher quality when their neural networks are augmented by spatial data structures that hold trainable features arranged in a grid. However, existing feature grids either come with a large memory footprint (dense or factorized grids, trees, and hash tables) or slow performance (index learning and vector quantization). In this paper, we show that a hash table with learned probes has neither disadvantage, resulting in a favorable combination of size and speed. Inference is faster than unprobed hash tables at equal quality while training is only 1.2-2.6x slower, significantly outperforming prior index learning approaches. We arrive at this formulation by casting all feature grids into a common framework: they each correspond to a lookup function that indexes into a table of feature vectors. In this framework, the lookup functions of existing data structures can be combined by simple arithmetic combinations of their indices, resulting in Pareto optimal compression and speed., Comment: Project Page: https://research.nvidia.com/labs/toronto-ai/compact-ngp

We investigate the transport properties of a quantum wire of weakly interacting fermions in the presence of a dissipative impurity. We calculate current and conductance in this system due to applied external chemical potential bias that can be measured in experimental realizations of ultracold fermions in quasi one-dimensional traps. Using a Keldysh field theory approach based on the Lindblad equation, we establish a perturbative scheme to study the effect of imbalanced reservoirs. Logarithmically divergent terms are resummed using a renormalization group method, and a novel power-law behavior for the conductance as a function of the potential bias across the wire is found. In contrast to the equilibrium case of a potential barrier in a Luttinger liquid, the conductance exhibits a scaling behavior, which depends on the interaction strength and on the loss probability. Repulsive interactions reduce the conductance of the wire while attractive interactions enhance it. However, perfect reflectivity and transparency are only reached in the absence of dissipation., Comment: 21 pages, 13 figures

The ability of humans to create and disseminate culture is often credited as the single most important factor of our success as a species. In this Perspective, we explore the notion of machine culture, culture mediated or generated by machines. We argue that intelligent machines simultaneously transform the cultural evolutionary processes of variation, transmission, and selection. Recommender algorithms are altering social learning dynamics. Chatbots are forming a new mode of cultural transmission, serving as cultural models. Furthermore, intelligent machines are evolving as contributors in generating cultural traits--from game strategies and visual art to scientific results. We provide a conceptual framework for studying the present and anticipated future impact of machines on cultural evolution, and present a research agenda for the study of machine culture.

Neural radiance fields achieve unprecedented quality for novel view synthesis, but their volumetric formulation remains expensive, requiring a huge number of samples to render high-resolution images. Volumetric encodings are essential to represent fuzzy geometry such as foliage and hair, and they are well-suited for stochastic optimization. Yet, many scenes ultimately consist largely of solid surfaces which can be accurately rendered by a single sample per pixel. Based on this insight, we propose a neural radiance formulation that smoothly transitions between volumetric- and surface-based rendering, greatly accelerating rendering speed and even improving visual fidelity. Our method constructs an explicit mesh envelope which spatially bounds a neural volumetric representation. In solid regions, the envelope nearly converges to a surface and can often be rendered with a single sample. To this end, we generalize the NeuS formulation with a learned spatially-varying kernel size which encodes the spread of the density, fitting a wide kernel to volume-like regions and a tight kernel to surface-like regions. We then extract an explicit mesh of a narrow band around the surface, with width determined by the kernel size, and fine-tune the radiance field within this band. At inference time, we cast rays against the mesh and evaluate the radiance field only within the enclosed region, greatly reducing the number of samples required. Experiments show that our approach enables efficient rendering at very high fidelity. We also demonstrate that the extracted envelope enables downstream applications such as animation and simulation., Comment: SIGGRAPH Asia 2023. Project page: research.nvidia.com/labs/toronto-ai/adaptive-shells/

Magnetic kagome metals, in which topologically non-trivial band structures and electronic correlation are intertwined, have recently emerged as an exciting platform to explore exotic correlated topological phases, that are usually not found in weakly interacting materials described within the semi-classical picture of electrons. Here, via a comprehensive single-crystal neutron diffraction and first-principles density functional theory study of the archetypical topological kagome metal Mn$_3$Sn, which is also a magnetic Weyl fermion material and a promising chiral magnet for antiferromagnetic spintronics, we report the realisation of an emergent spin-density wave (SDW) order, a hallmark correlated many-body phenomenon, that is engineered by the Fermi surface nesting of topological flat bands. We further reveal that the phase transition, from the well-known high-temperature coplanar and non-collinear k = 0 inverse triangular antiferromagnetic order to a double-$k$ non-coplanar modulated incommensurate magnetic structure below $T_1$ = 280 K, is primarily driven by the SDW instability. The double-$k$ nature of this complex low-temperature magnetic order, which can be regarded as an intriguing superposition of a longitudinal SDW with a modulation wavevector k$_L$ and a transverse incommensurate helical magnetic order with a modulation wavevector k$_T$, is unambiguously confirmed by our observation of the inter-modulation high-order harmonics of the type of 2k$_L$+k$_T$. This discovery not only solves a long-standing puzzle concerning the nature of the phase transition at $T_1$, but also provides an extraordinary example on the intrinsic engineering of correlated many-body phenomena in topological matter. The identified multi-$k$ magnetic state can be further exploited for the engineering of the new modes of magnetization and chirality switching in antiferromagnetic spintronics.

Neural surface reconstruction has been shown to be powerful for recovering dense 3D surfaces via image-based neural rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo, which combines the representation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our approach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures., Comment: CVPR 2023, project page: https://research.nvidia.com/labs/dir/neuralangelo

In [Ramanujan J. 52 (2020), 275-290], Romik considered the Taylor expansion of Jacobi's theta function $\theta_3(q)$ at $q=e^{-\pi}$ and encoded it in an integer sequence $(d(n))_{n\ge0}$ for which he provided a recursive procedure to compute the terms of the sequence. He observed intriguing behaviour of $d(n)$ modulo primes and prime powers. Here we prove (1) that $d(n)$ eventually vanishes modulo any prime power $p^e$ with $p\equiv3$ (mod 4), (2) that $d(n)$ is eventually periodic modulo any prime power $p^e$ with $p\equiv1$ (mod 4), and (3) that $d(n)$ is purely periodic modulo any 2-power $2^e$. Our results also provide more detailed information on period length, respectively from when on the sequence vanishes or becomes periodic. The corresponding bounds may not be optimal though, as computer data suggest. Our approach shows that the above congruence properties hold at a much finer, polynomial level., Comment: 54 pages; AmS-LaTeX; added references [7] and [8] that prove periodicity

We report a comprehensive single-crystal neutron diffraction investigation of the Kitaev quantum magnet $\alpha$-RuCl$_{3}$ under hydrostatic pressure. Utilizing a He-gas pressure cell, we successfully applied an ideal hydrostatic pressure in situ at low temperatures, which allows to effectively eliminate any possible influences from the structural transition occurring between 200 K and 50 K under ambient conditions. Our experiments reveal a gradual suppression of the ziagzag antiferromagnetic order as hydrostatic pressure increases. Furthermore, a reversible pressure-induced structural transition occurs at a critical pressure of $P_d$ = 0.15 GPa at 30 K, as evidenced by the absence of magnetic order and non-uniform changes in lattice constants. The decrease in magnetic transition temperature is discussed in relation to a pressure-induced change in the trigonal distortion of the Ru-Cl octahedra in this compound. Our findings emphasize the significance of the trigonal distortion in Kitaev materials, and provide a new perspective on the role of hydrostatic pressures in the realization of the Kitaev quantum spin liquid state in $\alpha$-RuCl$_{3}$., Comment: 7 pages with 4 figures

Thermal tuning of the optical refractive index in the waveguides to control light phase accumulation is essential in photonic integrated systems and applications. In silicon photonics, microheaters are mainly realized by metal wires or highly doped silicon lines placed at a safe distance (1um) from the waveguide to avoid considerable optical loss. However, this poses a significant limitation for heating efficiency because of the excessive free-carrier loss when a heater is brought closer to the optical path. In this work, we present a new concept of using optically transparent 2D semiconductors (e.g. MoS2) for realizing highly efficient waveguide integrated heaters operating at telecom wavelengths. We demonstrate that a single-layer MoS2 heater with negligible optical absorption in the infrared can be placed in close proximity (only 30nm) to the waveguide and show the best-reported heating efficiency of 15 mW per FSR without sacrificing the optical insertion loss. The heater response time is 25us, limited by Au 1L-MoS2 Schottky contact. Both the efficiency and response time can be further significantly improved by realizing 2D MoS2 heaters with ohmic contacts. Our work shows clear advantages of employing 2D semiconductors for heaters applications and paves the way for developing novel energy-efficient, lossless 2D heaters for on-chip photonic integrated circuits.

Variational quantum algorithms represent a promising approach to quantum machine learning where classical neural networks are replaced by parametrized quantum circuits. However, both approaches suffer from a clear limitation, that is a lack of interpretability. Here, we present a variational method to quantize projective simulation (PS), a reinforcement learning model aimed at interpretable artificial intelligence. Decision making in PS is modeled as a random walk on a graph describing the agent's memory. To implement the quantized model, we consider quantum walks of single photons in a lattice of tunable Mach-Zehnder interferometers trained via variational algorithms. Using an example from transfer learning, we show that the quantized PS model can exploit quantum interference to acquire capabilities beyond those of its classical counterpart. Finally, we discuss the role of quantum interference for training and tracing the decision making process, paving the way for realizations of interpretable quantum learning agents., Comment: 11+8 pages, 6+9 figures, 2 tables. F. Flamini and M. Krumm contributed equally to this work

We present a parallelized optimization method based on fast Neural Radiance Fields (NeRF) for estimating 6-DoF pose of a camera with respect to an object or scene. Given a single observed RGB image of the target, we can predict the translation and rotation of the camera by minimizing the residual between pixels rendered from a fast NeRF model and pixels in the observed image. We integrate a momentum-based camera extrinsic optimization procedure into Instant Neural Graphics Primitives, a recent exceptionally fast NeRF implementation. By introducing parallel Monte Carlo sampling into the pose estimation task, our method overcomes local minima and improves efficiency in a more extensive search space. We also show the importance of adopting a more robust pixel-based loss function to reduce error. Experiments demonstrate that our method can achieve improved generalization and robustness on both synthetic and real-world benchmarks., Comment: ICRA 2023. Project page at https://pnerfp.github.io/

In the scope of assessing aquifer systems in areas where freshwater is scarce, estimation of transit times is a vital step to quantify the effect of groundwater abstraction. Transit time distributions of different shapes, mean residence times, and contributions are used to represent the hydrogeological conditions in aquifer systems and are typically inferred from measured tracer concentrations by inverse modeling. In this study, a multi-tracer sampling campaign was conducted in the Salalah Plain in Southern Oman including CFCs, SF6, 39Ar, 14C, and 4He. Based on the data of three tracers, a two-component Dispersion Model (DMmix) and a nonparametric model with six age bins were assumed and evaluated using Bayesian statistics. In a Markov Chain Monte Carlo approach, the maximum likelihood parameter estimates and their uncertainties were determined. Model performance was assessed using Bayes factor and leave-one-out cross-validation. Both models suggest that the groundwater in the Salalah Plain is composed of a very young component below 30 yr and a very old component beyond 1,000 yr, with the nonparametric model performing slightly better than the DMmix model. All wells except one exhibit reasonable goodness of fit. Our results support the relevance of Bayesian modeling in hydrology and the potential of nonparametric models for an adequate representation of aquifer dynamics., Comment: 21 pages, 5 figures, 2 tables

Every moment, countless meteoroids enter our atmosphere unseen. The detection and measurement of meteors offer the unique opportunity to gain insights into the composition of our solar systems' celestial bodies. Researchers, therefore, carry out a wide-area-sky-monitoring to secure 360-degree video material, saving every single entry of a meteor. Existing machine intelligence cannot accurately recognize events of meteors intersecting the earth's atmosphere due to a lack of high-quality training data publicly available. This work presents four reusable open source solutions for researchers trained on data we collected due to the lack of available labeled high-quality training data. We refer to the proposed dataset as the NightSkyUCP dataset, consisting of a balanced set of 10,000 meteor- and 10,000 non-meteor-events. Our solutions apply various machine learning techniques, namely classification, feature learning, anomaly detection, and extrapolation. For the classification task, a mean accuracy of 99.1\% is achieved. The code and data are made public at figshare with DOI: 10.6084/m9.figshare.16451625, Comment: 11 Pages, 10 Figures, Accepted for publication in MNRAS Journal

Arcus is a proposed soft x-ray grating spectrometer Explorer. It aims to explore cosmic feedback by mapping hot gases within and between galaxies and galaxy clusters and characterizing jets and winds from supermassive black holes, and to investigate the dynamics of protoplanetary discs and stellar accretion. Arcus features 12 m-focal-length grazing-incidence silicon pore optics (SPO) developed for the Athena mission. Critical-angle transmission (CAT) gratings efficiently disperse high diffraction orders onto CCDs. We report new and improved x-ray performance results for Arcus-like CAT gratings, including record resolving power for two co-aligned CAT gratings. Multiple Arcus prototype grating facets were illuminated by an SPO at the PANTER facility. The facets consist of $32\times32.5$ mm$^2$ patterned silicon membranes, bonded to metal frames. The bonding angle is adjusted according to the measured average tilt angle of the grating bars in the membrane. Two simultaneously illuminated facets show minor broadening of the Al-K$_{\alpha}$ doublet in 18$^{\rm th}$ and 21$^{\rm st}$ orders with a best fit record effective resolving power of $R_G \approx 1.3^{+\infty}_{-0.5}\times10^4$ ($3\sigma$), about 3-4 times the Arcus requirement. We measured the diffraction efficiency of quasi-fully illuminated gratings at O-K wavelengths in orders 4-7 in an Arcus-like configuration and compare results with synchrotron spot measurements. After corrections for geometrical effects and bremsstrahlung continuum we find agreement between full and spot illumination at the two different facilities, as well as with the models used for Arcus effective area predictions. We find that these flight-like gratings meet diffraction efficiency and greatly exceed resolving power Arcus requirements., Comment: 19 pages, 16 figures

Neural approximations of scalar and vector fields, such as signed distance functions and radiance fields, have emerged as accurate, high-quality representations. State-of-the-art results are obtained by conditioning a neural approximation with a lookup from trainable feature grids that take on part of the learning task and allow for smaller, more efficient neural networks. Unfortunately, these feature grids usually come at the cost of significantly increased memory consumption compared to stand-alone neural network models. We present a dictionary method for compressing such feature grids, reducing their memory consumption by up to 100x and permitting a multiresolution representation which can be useful for out-of-core streaming. We formulate the dictionary optimization as a vector-quantized auto-decoder problem which lets us learn end-to-end discrete neural representations in a space where no direct supervision is available and with dynamic topology and structure. Our source code will be available at https://github.com/nv-tlabs/vqad., Comment: SIGGRAPH 2022. Project Page: https://nv-tlabs.github.io/vqad/

Cosmological simulations are crucial tools in studying the Universe, but they typically do not directly match real observed structures. Constrained cosmological simulations, on the other hand, are designed to match the observed distribution of galaxies. Here we present constrained simulations based on spectroscopic surveys at a redshift of z~2.3, corresponding to an epoch of nearly 11 Gyrs ago. This allows us to 'fast-forward' the simulation to our present-day and study the evolution of observed cosmic structures self-consistently. We confirm that several previously-reported protoclusters will evolve into massive galaxy clusters by our present epoch, including the 'Hyperion' structure that we predict will collapse into a giant filamentary supercluster spanning 100 Megaparsecs. We also discover previously unknown protoclusters, with lower final masses than typically detectable by other methods, that nearly double the number of known protoclusters within this volume. Constrained simulations, applied to future high-redshift datasets, represents a unique opportunity for studying early structure formation and matching galaxy properties between high and low redshifts., Comment: Submitted: 10 November 2021; Accepted: 28 April 2022 in Nature Astronomy https://www.nature.com/articles/s41550-022-01693-0 32 pages, 9 Figures

We present a large-scale synthetic dataset for novel view synthesis consisting of ~300k images rendered from nearly 2000 complex scenes using high-quality ray tracing at high resolution (1600 x 1600 pixels). The dataset is orders of magnitude larger than existing synthetic datasets for novel view synthesis, thus providing a large unified benchmark for both training and evaluation. Using 4 distinct sources of high-quality 3D meshes, the scenes of our dataset exhibit challenging variations in camera views, lighting, shape, materials, and textures. Because our dataset is too large for existing methods to process, we propose Sparse Voxel Light Field (SVLF), an efficient voxel-based light field approach for novel view synthesis that achieves comparable performance to NeRF on synthetic data, while being an order of magnitude faster to train and two orders of magnitude faster to render. SVLF achieves this speed by relying on a sparse voxel octree, careful voxel sampling (requiring only a handful of queries per ray), and reduced network structure; as well as ground truth depth maps at training time. Our dataset is generated by NViSII, a Python-based ray tracing renderer, which is designed to be simple for non-experts to use and share, flexible and powerful through its use of scripting, and able to create high-quality and physically-based rendered images. Experiments with a subset of our dataset allow us to compare standard methods like NeRF and mip-NeRF for single-scene modeling, and pixelNeRF for category-level modeling, pointing toward the need for future improvements in this area., Comment: ECCV 2022 Workshop on Learning to Generate 3D Shapes and Scenes. Project page at http://www.cs.umd.edu/~mmeshry/projects/rtmv

Author profiling classifies author characteristics by analyzing how language is shared among people. In this work, we study that task from a low-resource viewpoint: using little or no training data. We explore different zero and few-shot models based on entailment and evaluate our systems on several profiling tasks in Spanish and English. In addition, we study the effect of both the entailment hypothesis and the size of the few-shot training sample. We find that entailment-based models out-perform supervised text classifiers based on roberta-XLM and that we can reach 80% of the accuracy of previous approaches using less than 50\% of the training data on average.

Recent advances in natural language processing (NLP) have led to strong text classification models for many tasks. However, still often thousands of examples are needed to train models with good quality. This makes it challenging to quickly develop and deploy new models for real world problems and business needs. Few-shot learning and active learning are two lines of research, aimed at tackling this problem. In this work, we combine both lines into FASL, a platform that allows training text classification models using an iterative and fast process. We investigate which active learning methods work best in our few-shot setup. Additionally, we develop a model to predict when to stop annotating. This is relevant as in a few-shot setup we do not have access to a large validation set.

We study the problem of building text classifiers with little or no training data, commonly known as zero and few-shot text classification. In recent years, an approach based on neural textual entailment models has been found to give strong results on a diverse range of tasks. In this work, we show that with proper pre-training, Siamese Networks that embed texts and labels offer a competitive alternative. These models allow for a large reduction in inference cost: constant in the number of labels rather than linear. Furthermore, we introduce label tuning, a simple and computationally efficient approach that allows to adapt the models in a few-shot setup by only changing the label embeddings. While giving lower performance than model fine-tuning, this approach has the architectural advantage that a single encoder can be shared by many different tasks., Comment: ACL 2022