Modern deep neural networks struggle to transfer knowledge and generalize across domains when deploying to real-world applications. Domain generalization (DG) aims to learn a universal representation from multiple source domains to improve the networ
k generalization ability on unseen target domains. Previous DG methods mostly focus on the data-level consistency scheme to advance the generalization capability of deep networks, without considering the synergistic regularization of different consistency schemes. In this paper, we present a novel Hierarchical Consistency framework for Domain Generalization (HCDG) by ensembling Extrinsic Consistency and Intrinsic Consistency. Particularly, for Extrinsic Consistency, we leverage the knowledge across multiple source domains to enforce data-level consistency. Also, we design a novel Amplitude Gaussian-mixing strategy for Fourier-based data augmentation to enhance such consistency. For Intrinsic Consistency, we perform task-level consistency for the same instance under the dual-task form. We evaluate the proposed HCDG framework on two medical image segmentation tasks, i.e., optic cup/disc segmentation on fundus images and prostate MRI segmentation. Extensive experimental results manifest the effectiveness and versatility of our HCDG framework. Code will be available once accept.
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.
Let $S$ be the special fibre of a Shimura variety of Hodge type, with good reduction at a place above $p$. We give an alternative construction of the zip period map for $S$, which is used to define the Ekedahl-Oort strata of $S$. The method employed is local, $p$-adic, and group-theoretic in nature.
Medical dialogue systems (MDSs) aim to assist doctors and patients with a range of professional medical services, i.e., diagnosis, consultation, and treatment. However, one-stop MDS is still unexplored because: (1) no dataset has so large-scale dialo
gues contains both multiple medical services and fine-grained medical labels (i.e., intents, slots, values); (2) no model has addressed a MDS based on multiple-service conversations in a unified framework. In this work, we first build a Multiple-domain Multiple-service medical dialogue (M^2-MedDialog)dataset, which contains 1,557 conversations between doctors and patients, covering 276 types of diseases, 2,468 medical entities, and 3 specialties of medical services. To the best of our knowledge, it is the only medical dialogue dataset that includes both multiple medical services and fine-grained medical labels. Then, we formulate a one-stop MDS as a sequence-to-sequence generation problem. We unify a MDS with causal language modeling and conditional causal language modeling, respectively. Specifically, we employ several pretrained models (i.e., BERT-WWM, BERT-MED, GPT2, and MT5) and their variants to get benchmarks on M^2-MedDialog dataset. We also propose pseudo labeling and natural perturbation methods to expand M2-MedDialog dataset and enhance the state-of-the-art pretrained models. We demonstrate the results achieved by the benchmarks so far through extensive experiments on M2-MedDialog. We release the dataset, the code, as well as the evaluation scripts to facilitate future research in this important research direction.
In this paper, we propose a novel classification scheme for the remotely sensed hyperspectral image (HSI), namely SP-DLRR, by comprehensively exploring its unique characteristics, including the local spatial information and low-rankness. SP-DLRR is m
ainly composed of two modules, i.e., the classification-guided superpixel segmentation and the discriminative low-rank representation, which are iteratively conducted. Specifically, by utilizing the local spatial information and incorporating the predictions from a typical classifier, the first module segments pixels of an input HSI (or its restoration generated by the second module) into superpixels. According to the resulting superpixels, the pixels of the input HSI are then grouped into clusters and fed into our novel discriminative low-rank representation model with an effective numerical solution. Such a model is capable of increasing the intra-class similarity by suppressing the spectral variations locally while promoting the inter-class discriminability globally, leading to a restored HSI with more discriminative pixels. Experimental results on three benchmark datasets demonstrate the significant superiority of SP-DLRR over state-of-the-art methods, especially for the case with an extremely limited number of training pixels.
In this paper, a concurrent learning framework is developed for source search in an unknown environment using autonomous platforms equipped with onboard sensors. Distinct from the existing solutions that require significant computational power for Ba
yesian estimation and path planning, the proposed solution is computationally affordable for onboard processors. A new concept of concurrent learning using multiple parallel estimators is proposed to learn the operational environment and quantify estimation uncertainty. The search agent is empowered with dual capability of exploiting current estimated parameters to track the source and probing the environment to reduce the impacts of uncertainty, namely Concurrent Learning for Exploration and Exploitation (CLEE). In this setting, the control action not only minimises the tracking error between future agents position and estimated source location, but also the uncertainty of predicted estimation. More importantly, the rigorous proven properties such as the convergence of CLEE algorithm are established under mild assumptions on sensor noises, and the impact of noises on the search performance is examined. Simulation results are provided to validate the effectiveness of the proposed CLEE algorithm. Compared with the information-theoretic approach, CLEE not only guarantees convergence, but produces better search performance and consumes much less computational time.
Cloud is critical for planetary climate and habitability, but it is also one of the most challenging parts of studying planets in and beyond the solar system. Previous simulations using global general circulation models (GCMs) found that for 1:1 tida
lly locked (i.e., synchronously rotating) terrestrial planets with oceans, strong convergence and convection produce optically thick clouds over the substellar area. One obvious weakness of these studies is that clouds are parameterized based on the knowledge on Earth, and whether it is applicable to exoplanetary environment is unknown. Here we use a cloud-resolving model (CRM) with high resolution (2 km) in a two-dimensional (2D) configuration to simulate the clouds and circulation on tidally locked aqua-planets. We confirm that the substellar area is covered by deep convective clouds, the nightside is dominated by low-level stratus clouds, and these two are linked by a global-scale overturning circulation. We further find that a uniform warming of the surface causes the width of convection and clouds to decrease, but a decrease of day-night surface temperature contrast or an increase of longwave radiative cooling rate causes the width of convection and clouds to increase. These relationships can be roughly interpreted based on some simple thermodynamic theories. Comparing the results between CRM and GCM, we find that the results are broadly similar although there are many significant differences. Future work is required to use 3D CRM(s) with realistic radiative transfer and with the Coriolis force to examine the clouds and climate of tidally locked planets.
PandaX-4T is a dark matter direct detection experiment located in China jinping underground laboratory. The central apparatus is a dual-phase xenon detector containing 4 ton liquid xenon in the sensitive volume, with about 500 photomultipliers instru
mented in the top and the bottom of the detector. In this paper we present a completely new system of readout electronics and data acquisition in the PandaX-4T experiment. Compared to the one used in the previous PandaX dark matter experiments, the new system features triggerless readout and higher bandwidth. With triggerless readout, dark matter searches are not affected by the efficiency loss of external triggers. The system records single photelectron signals of the dominant PMTs with an average efficiency of 96%, and achieves the bandwidth of more than 450 MB/s. The system has been used to successfully acquire data during the commissioning runs of PandaX-4T.
The frustrated magnetism on the Kondo lattice system motivates intriguing Kondo-breakdown beyond the traditional Doniach scenario. Among them, the fractionalized Fermi liquid (FL*) has drawn a particular interest by virtue of its fractionalized natur
e. Here, we study the phase diagram of $J_{1}$-$J_{2}$ Kondo-Heisenberg model on a honeycomb lattice at a quarter filling. Employing the slave-fermion mean-field theory with $d pm id$ spin liquid ansatz and exact diagonalization, we discuss the emergence of partial Kondo screening in the frustrated regime with comparable $J_{1}$ and $J_{2}$, and the fractionalized superconductor (SC*) which is superconductor analogy of the FL*. Due to the larger number of local spin moments than itinerant electrons, the magnetic fluctuation is still significant even in the strong-coupling limit, which influences the thermodynamic and transport properties qualitatively. In particular, we estimate the thermal conductance to probe the low-energy excitation and show the anomalous behaviour in the SC* phase contrast to the conventional superconductors.
Spatial heterogeneity and temporal variability are general features in planetary weather and climate, due to the effects of planetary rotation, uneven stellar flux distribution, fluid motion instability, etc. In this study, we investigate the asymmet
ry and variability in the transmission spectra of 1:1 spin--orbit tidally locked (or called synchronously rotating) planets around low-mass stars. We find that for rapidly rotating planets, the transit atmospheric thickness on the evening terminator (east of the substellar region) is significantly larger than that of the morning terminator (west of the substellar region). The asymmetry is mainly related to the spatial heterogeneity in ice clouds, as the contributions of liquid clouds and water vapor are smaller. The underlying mechanism is that there are always more ice clouds on the evening terminator, due to the combined effect of coupled Rossby--Kelvin waves and equatorial superrotation that advect vapor and clouds to the east, especially at high levels of the atmosphere. For slowly rotating planets, the asymmetry reverses (the morning terminator has a larger transmission depth than the evening terminator) but the magnitude is small or even negligible. For both rapidly and slowly rotating planets, there is strong variability in the transmission spectra. The asymmetry signal is nearly impossible to be observed by the James Webb Space Telescope (JWST), because the magnitude of the asymmetry (about 10 ppm) is smaller than the instrumental noise and the high variability further increases the challenge.
