To audit the robustness of named entity recognition (NER) models, we propose RockNER, a simple yet effective method to create natural adversarial examples. Specifically, at the entity level, we replace target entities with other entities of the same
semantic class in Wikidata; at the context level, we use pre-trained language models (e.g., BERT) to generate word substitutions. Together, the two levels of attack produce natural adversarial examples that result in a shifted distribution from the training data on which our target models have been trained. We apply the proposed method to the OntoNotes dataset and create a new benchmark named OntoRock for evaluating the robustness of existing NER models via a systematic evaluation protocol. Our experiments and analysis reveal that even the best model has a significant performance drop, and these models seem to memorize in-domain entity patterns instead of reasoning from the context. Our work also studies the effects of a few simple data augmentation methods to improve the robustness of NER models.
Weyl semimetals are well-known for hosting topologically protected linear band crossings, serving as the analog of the relativistic Weyl Fermions in the condensed matter context. Such analogy persists deeply, allowing the existence of the chiral anom
aly under parallel electric and magnetic field in Weyl semimetals. Different from such picture, here we show that, a unique mechanism of the chiral anomaly exists in Weyl semimetals by injecting a spin current with parallel spin polarization and flow direction. The existence of such a chiral anomaly is protected by the topological feature that each Weyl cone can also be a source or drain of the spin field in the momentum space. It leads to measurable experimental signals, such as an electric charge current parallel with an applied magnetic field in the absence of the electric field, and a sharp peak at certain resonant frequency in the injection current in achiral Weyl semimetals through the circular photogalvanic effect. Our work shows that the topological implication of Weyl semimetals goes beyond the link with relativistic Weyl Fermions, and offers a promising scenario to examine the interplay between topology and spin.
We propose an Auto-Parsing Network (APN) to discover and exploit the input datas hidden tree structures for improving the effectiveness of the Transformer-based vision-language systems. Specifically, we impose a Probabilistic Graphical Model (PGM) pa
rameterized by the attention operations on each self-attention layer to incorporate sparse assumption. We use this PGM to softly segment an input sequence into a few clusters where each cluster can be treated as the parent of the inside entities. By stacking these PGM constrained self-attention layers, the clusters in a lower layer compose into a new sequence, and the PGM in a higher layer will further segment this sequence. Iteratively, a sparse tree can be implicitly parsed, and this trees hierarchical knowledge is incorporated into the transformed embeddings, which can be used for solving the target vision-language tasks. Specifically, we showcase that our APN can strengthen Transformer based networks in two major vision-language tasks: Captioning and Visual Question Answering. Also, a PGM probability-based parsing algorithm is developed by which we can discover what the hidden structure of input is during the inference.
We report on the continuum and polarization observations of the Cygnus Loop supernova remnant (SNR) conducted by the Five-hundred-meter Aperture Spherical radio Telescope (FAST). FAST observations provide high angular resolution and high sensitivity
images of the SNR, which will help to disentangle its nature. We obtained Stokes I, Q and U maps over the frequency range of 1.03 - 1.46 GHz split into channels of 7.63 kHz. The original angular resolution is in the range of ~3 arcmin - ~3.8 arcmin, and we combined all the data at a common resolution of 4 arcmin. The temperature scale of the total intensity and the spectral index from the in-band temperature-temperature plot are consistent with previous observations, which validates the data calibration and map-making procedures. The rms sensitivity for the band-averaged total-intensity map is about 20 mK in brightness temperature, which is at the level of confusion limit. For the first time, we apply rotation measure (RM) synthesis to the Cygnus Loop to obtain the polarization intensity and RM maps. The rms sensitivity for polarization is about 5 mK, far below the total-intensity confusion limit. We also obtained RMs of eight extra-galactic sources, and demonstrate that the wide-band frequency coverage helps to overcome the ambiguity of RM determinations.
The vast majority of Galactic supernova remnants (SNRs) were detected by their synchrotron radio emission. Recently, the evolved SNR G107.0+9.0 with a diameter of about 3~deg or 75~pc up to 100~pc in size was optically detected with an indication of
faint associated radio emission. This SNR requires a detailed radio study. We aim to search for radio emission from SNR G107.0+9.0 by analysing new data from the Effelsberg 100-m and the Urumqi 25-m radio telescopes in addition to available radio surveys. Radio SNRs outside of the Galactic plane, where confusion is rare, must be very faint if they have not been identified so far. Guided by the H$alpha$ emission of G107.0+9.0, we separated its radio emission from the Galactic large-scale emission. Radio emission from SNR G107.0+9.0 is detected between 22~MHz and 4.8~GHz with a steep non-thermal spectrum, which confirms G107.0+9.0 as an SNR. Its surface brightness is among the lowest known for Galactic SNRs. Polarised emission is clearly detected at 1.4~GHz but is fainter at 4.8~GHz. We interpret the polarised emission as being caused by a Faraday screen associated with G107.0+9.0 and its surroundings. Its ordered magnetic field along the line of sight is below 1~$mu$G. At 4.8~GHz, we identified a depolarised filament along the western periphery of G107.0+9.0 with a magnetic field strength along the line of sight $B{_{||}} sim 15~mu$G, which requires magnetic field compression. G107.0+9.0 adds to the currently small number of known, evolved, large-diameter, low-surface-brightness Galactic SNRs. We have shown that such objects can be successfully extracted from radio-continuum surveys despite the dominating large-scale diffuse Galactic emission.
During October 2019 and March 2020, the luminous red supergiant Betelgeuse demonstrated an unusually deep minimum of its brightness. It became fainter by more than one magnitude and this is the most significant dimming observed in the recent decades.
While the reason for the dimming is debated, pre-phase of supernova explosion, obscuring dust, or changes in the photosphere of the star were suggested scenarios. Here, we present spectroscopic studies of Betelgeuse using high-resolution and high signal-to- noise ratio near-infrared spectra obtained at Weihai Observatory on four epochs in 2020 covering the phases of during and after dimming. We show that the dimming episode is caused by the dropping of its effective temperature by at least 170 K on 2020 January 31, that can be attributed to the emergence of a large dark spot on the surface of the star.
This paper proposes the importance of age and gender information in the diagnosis of diabetic retinopathy. We utilized Deep Residual Neural Networks (ResNet) and Densely Connected Convolutional Networks (DenseNet), which are proven effective on image
classification problems and the diagnosis of diabetic retinopathy using the retinal fundus images. We used the ensemble of several classical networks and decentralized the training so that the network was simple and avoided overfitting. To observe whether the age and gender information could help enhance the performance, we added the information before the dense layer and compared the results with the results that did not add age and gender information. We found that the test accuracy of the network with age and gender information was 2.67% higher than that of the network without age and gender information. Meanwhile, compared with gender information, age information had a better help for the results.
This paper proposes a new deep learning approach to antipodal grasp detection, named Double-Dot Network (DD-Net). It follows the recent anchor-free object detection framework, which does not depend on empirically pre-set anchors and thus allows more
generalized and flexible prediction on unseen objects. Specifically, unlike the widely used 5-dimensional rectangle, the gripper configuration is defined as a pair of fingertips. An effective CNN architecture is introduced to localize such fingertips, and with the help of auxiliary centers for refinement, it accurately and robustly infers grasp candidates. Additionally, we design a specialized loss function to measure the quality of grasps, and in contrast to the IoU scores of bounding boxes adopted in object detection, it is more consistent to the grasp detection task. Both the simulation and robotic experiments are executed and state of the art accuracies are achieved, showing that DD-Net is superior to the counterparts in handling unseen objects.
We propose a new gradient-based approach for extracting sub-architectures from a given large model. Contrarily to existing pruning methods, which are unable to disentangle the network architecture and the corresponding weights, our architecture-pruni
ng scheme produces transferable new structures that can be successfully retrained to solve different tasks. We focus on a transfer-learning setup where architectures can be trained on a large data set but very few data points are available for fine-tuning them on new tasks. We define a new gradient-based algorithm that trains architectures of arbitrarily low complexity independently from the attached weights. Given a search space defined by an existing large neural model, we reformulate the architecture search task as a complexity-penalized subset-selection problem and solve it through a two-temperature relaxation scheme. We provide theoretical convergence guarantees and validate the proposed transfer-learning strategy on real data.
The nonlinear Hall effect is mostly studied as a Berry curvature dipole effect in nonmagnetic materials, which depends linearly on the relaxation time. On the other hand, in magnetic materials, an intrinsic nonlinear Hall effect can exist, which does
not depend on the relaxation time. Here we show that the intrinsic nonlinear Hall effect can be observed in an antiferromagnetic metal: tetragonal CuMnAs, and the corresponding conductivity can reach the order of mA/V$^2$ based on density functional theory calculations. The dependence on the chemical potential of such nonlinear Hall conductivity can be qualitatively explained by a tilted massive Dirac model. Moreover, we demonstrate its strong temperature-dependence and briefly discuss its competition with the second order Drude conductivity. Finally, a complete survey of magnetic point groups are presented, providing guidelines for finding candidate materials with the intrinsic nonlinear Hall effect.
