arXiv subjects

Kai Wang

Publications and source records attributed to Kai Wang.

At least 469 records · Page 26Linked to original sources

arXiv2024

CSSL-MHTR: Continual Self-Supervised Learning for Scalable Multi-script Handwritten Text Recognition

Self-supervised learning has recently emerged as a strong alternative in document analysis. These approaches are now capable of learning high-quality image representations and overcoming the limitations of supervised methods, which require a large amount of labeled data. However, these methods are unable to capture new knowledge in an incremental fashion, where data is presented to the model sequentially, which is closer to the realistic scenario. In this paper, we explore the potential of continual self-supervised learning to alleviate the catastrophic forgetting problem in handwritten text recognition, as an example of sequence recognition. Our method consists in adding intermediate layers called adapters for each task, and efficiently distilling knowledge from the previous model while learning the current task. Our proposed framework is efficient in both computation and memory complexity. To demonstrate its effectiveness, we evaluate our method by transferring the learned model to diverse text recognition downstream tasks, including Latin and non-Latin scripts. As far as we know, this is the first application of continual self-supervised learning for handwritten text recognition. We attain state-of-the-art performance on English, Italian and Russian scripts, whilst adding only a few parameters per task. The code and trained models will be publicly available.

cs.CV↗

arXiv2024

A Law of large numbers for vector-valued linear statistics of Bergman DPP

We establish a law of large numbers for a certain class of vector-valued linear statistics for the Bergman determinantal point process on the unit disk. Our result seems to be the first LLN for vector-valued linear statistics in the setting of determinantal point processes. As an application, we prove that, for almost all configurations $X$ with respect to with respect to the Bergman determinantal point process, the weighted Poincaré series (we denote by $d_{h}(\cdot,\cdot)$ the hyperbolic distance on $\mathbb{D}$) \begin{align*} \sum_{k=0}^\infty\sum_{x\in X\atop k\le d_{h}(z,x)<k+1}e^{-sd_{\mathrm{h}}(z,x)}f(x) \end{align*} cannot be simultaneously convergent for all Bergman functions $f\in A^2(\mathbb{D})$ whenever $1<s<3/2$. This confirms a result announced without proof in Bufetov-Qiu's work.

math.PR↗

arXiv2024

IterInv: Iterative Inversion for Pixel-Level T2I Models

Large-scale text-to-image diffusion models have been a ground-breaking development in generating convincing images following an input text prompt. The goal of image editing research is to give users control over the generated images by modifying the text prompt. Current image editing techniques predominantly hinge on DDIM inversion as a prevalent practice rooted in Latent Diffusion Models (LDM). However, the large pretrained T2I models working on the latent space suffer from losing details due to the first compression stage with an autoencoder mechanism. Instead, other mainstream T2I pipeline working on the pixel level, such as Imagen and DeepFloyd-IF, circumvents the above problem. They are commonly composed of multiple stages, typically starting with a text-to-image stage and followed by several super-resolution stages. In this pipeline, the DDIM inversion fails to find the initial noise and generate the original image given that the super-resolution diffusion models are not compatible with the DDIM technique. According to our experimental findings, iteratively concatenating the noisy image as the condition is the root of this problem. Based on this observation, we develop an iterative inversion (IterInv) technique for this category of T2I models and verify IterInv with the open-source DeepFloyd-IF model.Specifically, IterInv employ NTI as the inversion and reconstruction of low-resolution image generation. In stages 2 and 3, we update the latent variance at each timestep to find the deterministic inversion trace and promote the reconstruction process. By combining our method with a popular image editing method, we prove the application prospects of IterInv. The code will be released upon acceptance. The code is available at \url{https://github.com/Tchuanm/IterInv.git}.

cs.CV↗

arXiv2024

Spatio-temporal Joint Analysis of PM2.5 and Ozone in California with INLA

The substantial threat of concurrent air pollutants to public health is increasingly severe under climate change. To identify the common drivers and extent of spatio-temporal similarity of PM2.5 and ozone, this paper proposed a log Gaussian-Gumbel Bayesian hierarchical model allowing for sharing a SPDE-AR(1) spatio-temporal interaction structure. The proposed model outperforms in terms of estimation accuracy and prediction capacity for its increased parsimony and reduced uncertainty, especially for the shared ozone sub-model. Besides the consistently significant influence of temperature (positive), extreme drought (positive), fire burnt area (positive), and wind speed (negative) on both PM2.5 and ozone, surface pressure and GDP per capita (precipitation) demonstrate only positive associations with PM2.5 (ozone), while population density relates to neither. In addition, our results show the distinct spatio-temporal interactions and different seasonal patterns of PM2.5 and ozone, with peaks of PM2.5 and ozone in cold and hot seasons, respectively. Finally, with the aid of the excursion function, we see that the areas around the intersection of San Luis Obispo and Santa Barbara counties are likely to exceed the unhealthy ozone level for sensitive groups throughout the year. Our findings provide new insights for regional and seasonal strategies in the co-control of PM2.5 and ozone. Our methodology is expected to be utilized when interest lies in multiple interrelated processes in the fields of environment and epidemiology.

physics.ao-ph↗

arXiv2024

Observation of quantum nonlocality in Greenberger-Horne-Zeilinger entanglement on a silicon chip

Nonlocality is the defining feature of quantum entanglement. Entangled states with multiple particles are of crucial importance in fundamental tests of quantum physics as well as in many quantum information tasks. One of the archetypal multipartite quantum states, Greenberger-Horne-Zeilinger (GHZ) state, allows one to observe the striking conflict of quantum physics to local realism in the so-called all-versus-nothing way. This is profoundly different from Bell's theorem for two particles, which relies on statistical predictions. Here, we demonstrate an integrated photonic chip capable of generating and manipulating the four-photon GHZ state. We perform a complete characterization of the four-photon GHZ state using quantum state tomography and obtain a state fidelity of 0.729(6). We further use the all-versus-nothing test and the Mermin inequalities to witness the quantum nonlocality of GHZ entanglement. Our work paves the way to perform fundamental tests of quantum physics with complex integrated quantum devices.

quant-ph↗

arXiv2024

EasyTrack: Efficient and Compact One-stream 3D Point Clouds Tracker

Most of 3D single object trackers (SOT) in point clouds follow the two-stream multi-stage 3D Siamese or motion tracking paradigms, which process the template and search area point clouds with two parallel branches, built on supervised point cloud backbones. In this work, beyond typical 3D Siamese or motion tracking, we propose a neat and compact one-stream transformer 3D SOT paradigm from the novel perspective, termed as \textbf{EasyTrack}, which consists of three special designs: 1) A 3D point clouds tracking feature pre-training module is developed to exploit the masked autoencoding for learning 3D point clouds tracking representations. 2) A unified 3D tracking feature learning and fusion network is proposed to simultaneously learns target-aware 3D features, and extensively captures mutual correlation through the flexible self-attention mechanism. 3) A target location network in the dense bird's eye view (BEV) feature space is constructed for target classification and regression. Moreover, we develop an enhanced version named EasyTrack++, which designs the center points interaction (CPI) strategy to reduce the ambiguous targets caused by the noise point cloud background information. The proposed EasyTrack and EasyTrack++ set a new state-of-the-art performance ($\textbf{18\%}$, $\textbf{40\%}$ and $\textbf{3\%}$ success gains) in KITTI, NuScenes, and Waymo while runing at \textbf{52.6fps} with few parameters (\textbf{1.3M}). The code will be available at https://github.com/KnightApple427/Easytrack.

cs.CV↗

arXiv2024

SciFlow: Empowering Lightweight Optical Flow Models with Self-Cleaning Iterations

Optical flow estimation is crucial to a variety of vision tasks. Despite substantial recent advancements, achieving real-time on-device optical flow estimation remains a complex challenge. First, an optical flow model must be sufficiently lightweight to meet computation and memory constraints to ensure real-time performance on devices. Second, the necessity for real-time on-device operation imposes constraints that weaken the model's capacity to adequately handle ambiguities in flow estimation, thereby intensifying the difficulty of preserving flow accuracy. This paper introduces two synergistic techniques, Self-Cleaning Iteration (SCI) and Regression Focal Loss (RFL), designed to enhance the capabilities of optical flow models, with a focus on addressing optical flow regression ambiguities. These techniques prove particularly effective in mitigating error propagation, a prevalent issue in optical flow models that employ iterative refinement. Notably, these techniques add negligible to zero overhead in model parameters and inference latency, thereby preserving real-time on-device efficiency. The effectiveness of our proposed SCI and RFL techniques, collectively referred to as SciFlow for brevity, is demonstrated across two distinct lightweight optical flow model architectures in our experiments. Remarkably, SciFlow enables substantial reduction in error metrics (EPE and Fl-all) over the baseline models by up to 6.3% and 10.5% for in-domain scenarios and by up to 6.2% and 13.5% for cross-domain scenarios on the Sintel and KITTI 2015 datasets, respectively.

cs.CV↗

arXiv2024

Demonstration of Lossy Linear Transformations and Two-Photon Interference on a Photonic Chip

Studying quantum correlations in the presence of loss is of critical importance for the physical modeling of real quantum systems. Here, we demonstrate the control of spatial correlations between entangled photons in a photonic chip, designed and modeled using the singular value decomposition approach. We show that engineered loss, using an auxiliary waveguide, allows one to invert the spatial statistics from bunching to antibunching. Furthermore, we study the photon statistics within the loss-emulating channel and observe photon coincidences, which may provide insights into the design of quantum photonic integrated chips.

quant-ph↗

arXiv2024

Robust Classical and Quantum Polarimetry with a Single Nanostructured Metagrating

We formulate a new conceptual approach for one-shot complete polarization state measurement with nanostructured metasurfaces applicable to classical light and multi-photon quantum states, by drawing on the principles of generalized quantum measurements based on positive operator-valued measures (POVMs). Accurate polarization reconstruction from a combination of photon counts or correlations from several diffraction orders is robust with respect to even strong fabrication inaccuracies, requiring only a single classical calibration of metasurface transmission. Furthermore, this approach operates with a single metagrating without interleaving, allowing for the metasurface size reduction while preserving the high transmission efficiency and output beam quality. We theoretically obtained original metasurface designs, fabricated the metasurface from amorphous silicon nanostructures deposited on glass, and experimentally confirmed accurate polarization reconstruction for laser beams. We also anticipate robust operation under changes in environmental conditions, opening new possibilities for space-based imaging and satellite optics.

physics.optics↗

arXiv2024

GNSS Spoofing Detection by Crowdsourcing Double Differential Pseudorange Spatial Distribution

It is widely known that spoofing is a major threat that adversely impacts the reliability and accuracy of GNSS applications. In this study, a crowdsourcing double differential pseudorange spatial (D2SP) random set is constructed and the distribution of the set is derived.Based on the variance of the D2SP set, a tri-level hypothesis detection algorithm is designed to classify spoofing-free, fully-spoofed, and partially-spoofed cases in the region of interest (ROI).It does not require the prior knowledge of the truth positions or relative distances of the receivers.Simulation test results show that the proposed D2SP spoofing detection method has the advantages of lower computational complexity and higher tolerance for multipath errors compared with the generalized likelihood ratio test (GLRT) method that is the current mainstream spoofing detection algorithm based on multiple receivers' differential pseudoranges.Moreover, it also shows better flexibility for different sizes of ROI and numbers of the crowdsourcing receivers.

eess.SP↗

arXiv2024

Assessing the Utility of Large Language Models for Phenotype-Driven Gene Prioritization in Rare Genetic Disorder Diagnosis

Phenotype-driven gene prioritization is a critical process in the diagnosis of rare genetic disorders for identifying and ranking potential disease-causing genes based on observed physical traits or phenotypes. While traditional approaches rely on curated knowledge graphs with phenotype-gene relations, recent advancements in large language models have opened doors to the potential of AI predictions through extensive training on diverse corpora and complex models. This study conducted a comprehensive evaluation of five large language models, including two Generative Pre-trained Transformers series, and three Llama2 series, assessing their performance across three key metrics: task completeness, gene prediction accuracy, and adherence to required output structures. Various experiments explored combinations of models, prompts, input types, and task difficulty levels. Our findings reveal that even the best-performing LLM, GPT-4, achieved an accuracy of 16.0%, which still lags behind traditional bioinformatics tools. Prediction accuracy increased with the parameter/model size. A similar increasing trend was observed for the task completion rate, with complicated prompts more likely to increase task completeness in models smaller than GPT-4. However, complicated prompts are more likely to decrease the structure compliance rate, but no prompt effects on GPT-4. Compared to HPO term-based input, LLM was also able to achieve better than random prediction accuracy by taking free-text input, but slightly lower than with the HPO input. Bias analysis showed that certain genes, such as MECP2, CDKL5, and SCN1A, are more likely to be top-ranked, potentially explaining the variances observed across different datasets. This study provides valuable insights into the integration of LLMs within genomic analysis, contributing to the ongoing discussion on the utilization of advanced LLMs in clinical workflows.

q-bio.QM↗

arXiv2024

Efficient Unsupervised Community Search with Pre-trained Graph Transformer

Community search has aroused widespread interest in the past decades. Among existing solutions, the learning-based models exhibit outstanding performance in terms of accuracy by leveraging labels to 1) train the model for community score learning, and 2) select the optimal threshold for community identification. However, labeled data are not always available in real-world scenarios. To address this notable limitation of learning-based models, we propose a pre-trained graph Transformer based community search framework that uses Zero label (i.e., unsupervised), termed TransZero. TransZero has two key phases, i.e., the offline pre-training phase and the online search phase. Specifically, in the offline pretraining phase, we design an efficient and effective community search graph transformer (CSGphormer) to learn node representation. To pre-train CSGphormer without the usage of labels, we introduce two self-supervised losses, i.e., personalization loss and link loss, motivated by the inherent uniqueness of node and graph topology, respectively. In the online search phase, with the representation learned by the pre-trained CSGphormer, we compute the community score without using labels by measuring the similarity of representations between the query nodes and the nodes in the graph. To free the framework from the usage of a label-based threshold, we define a new function named expected score gain to guide the community identification process. Furthermore, we propose two efficient and effective algorithms for the community identification process that run without the usage of labels. Extensive experiments over 10 public datasets illustrate the superior performance of TransZero regarding both accuracy and efficiency.

cs.SI↗

arXiv2024

Neural Attributed Community Search at Billion Scale

Community search has been extensively studied in the past decades. In recent years, there is a growing interest in attributed community search that aims to identify a community based on both the query nodes and query attributes. A set of techniques have been investigated. Though the recent methods based on advanced learning models such as graph neural networks (GNNs) can achieve state-of-the-art performance in terms of accuracy, we notice that 1) they suffer from severe efficiency issues; 2) they directly model community search as a node classification problem and thus cannot make good use of interdependence among different entities in the graph. Motivated by these, in this paper, we propose a new neurAL attrIbuted Community sEarch model for large-scale graphs, termed ALICE. ALICE first extracts a candidate subgraph to reduce the search scope and subsequently predicts the community by the Consistency-aware Net , termed ConNet. Specifically, in the extraction phase, we introduce the density sketch modularity that uses a unified form to combine the strengths of two existing powerful modularities, i.e., classical modularity and density modularity. Based on the new modularity metric, we first adaptively obtain the candidate subgraph, formed by the k-hop neighbors of the query nodes, with the maximum modularity. Then, we construct a node-attribute bipartite graph to take attributes into consideration. After that, ConNet adopts a cross-attention encoder to encode the interaction between the query and the graph. The training of the model is guided by the structure-attribute consistency and the local consistency to achieve better performance. Extensive experiments over 11 real-world datasets including one billion-scale graph demonstrate the superiority of ALICE in terms of accuracy, efficiency, and scalability.

cs.SI↗

arXiv2024

Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching

The ultimate goal of Dataset Distillation is to synthesize a small synthetic dataset such that a model trained on this synthetic set will perform equally well as a model trained on the full, real dataset. Until now, no method of Dataset Distillation has reached this completely lossless goal, in part due to the fact that previous methods only remain effective when the total number of synthetic samples is extremely small. Since only so much information can be contained in such a small number of samples, it seems that to achieve truly loss dataset distillation, we must develop a distillation method that remains effective as the size of the synthetic dataset grows. In this work, we present such an algorithm and elucidate why existing methods fail to generate larger, high-quality synthetic sets. Current state-of-the-art methods rely on trajectory-matching, or optimizing the synthetic data to induce similar long-term training dynamics as the real data. We empirically find that the training stage of the trajectories we choose to match (i.e., early or late) greatly affects the effectiveness of the distilled dataset. Specifically, early trajectories (where the teacher network learns easy patterns) work well for a low-cardinality synthetic set since there are fewer examples wherein to distribute the necessary information. Conversely, late trajectories (where the teacher network learns hard patterns) provide better signals for larger synthetic sets since there are now enough samples to represent the necessary complex patterns. Based on our findings, we propose to align the difficulty of the generated patterns with the size of the synthetic dataset. In doing so, we successfully scale trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation for the very first time. Code and distilled datasets are available at https://gzyaftermath.github.io/DATM.

cs.CV↗

arXiv2024

High-energy Neutrinos from Outflows Powered by Kicked Remnants of Binary Black Hole Mergers in AGN Accretion Disks

Merging of stellar-mass binary black holes (BBH) could take place within the accretion disk of active galactic nuclei (AGN). The resulting BH remnant is likely to accrete the disk gas at a super-Eddington rate, launching a fast, quasi-spherical outflow (wind). Particles will be accelerated by shocks driven by the wind, subsequently interacting with the shocked disk gas or radiation field through hadronic processes and resulting in the production of high-energy neutrinos and potential electromagnetic (EM) emissions. This study delves into the intricate evolution of the shock driven by the remnant BH wind within AGN disks. Subsequently, we calculated the production of neutrinos and the expected detection numbers for a single event, along with their contributions to the overall diffuse neutrino background. Our analysis, considering various scenarios, reveals considerable neutrino production and possible detection by IceCube for nearby events. The contribution of the remnant BH winds on the diffuse neutrino background is minor due to the low event rate density, but it can be improved to some extent for some optimistic parameters. We also propose that there could be two neutrino/EM bursts, one originating from the premerger BBH wind and the other from the remnant BH wind, with the latter typically having a time gap to the GW event of around tens of days. When combined with the anticipated gravitational waves (GW) emitted during the BBH merger, such a system emerges as a promising candidate for joint observations involving neutrinos, GWs, and EM signals.

astro-ph.HE↗

arXiv2024

MLLMs-Augmented Visual-Language Representation Learning

Visual-language pre-training has achieved remarkable success in many multi-modal tasks, largely attributed to the availability of large-scale image-text datasets. In this work, we demonstrate that Multi-modal Large Language Models (MLLMs) can enhance visual-language representation learning by establishing richer image-text associations for image-text datasets. Our approach is simple, utilizing MLLMs to extend multiple diverse captions for each image. To prevent the bias introduced by MLLMs' hallucinations and monotonous language styles, we propose "text shearing" to maintain the quality and availability of extended captions. In image-text retrieval, without introducing additional training cost, our method consistently obtains 5.6 ~ 35.0 and 16.8 ~ 46.1 improvement on Recall@1 under the fine-tuning and zero-shot settings, respectively. Notably, we obtain zero-shot results that are comparable to fine-tuning on target datasets, which encourages more exploration of the versatile use of MLLMs.

cs.CV↗

arXiv2024

A Markovian regime-switching stochastic SEQIR epidemic model with governmental policy

In this paper, a stochastic SEQIR epidemic model with Markovian regime-switching is proposed and investigated. The governmental policy and implement efficiency are concerned by a generalized incidence function of the susceptible class. We have the existence and uniqueness of the globally positive solution to the stochastic model by using the Lyapunov method. In addition, we study the dynamical behaviors of the disease, and the sufficient conditions for the extinction and persistence in mean are obtained. Finally, numerical simulations are introduced to demonstrate the theoretical results.

math.PR↗

arXiv2024

Difference Learning for Air Quality Forecasting Transport Emulation

Human health is negatively impacted by poor air quality including increased risk for respiratory and cardiovascular disease. Due to a recent increase in extreme air quality events, both globally and locally in the United States, finer resolution air quality forecasting guidance is needed to effectively adapt to these events. The National Oceanic and Atmospheric Administration provides air quality forecasting guidance for the Continental United States. Their air quality forecasting model is based on a 15 km spatial resolution; however, the goal is to reach a three km spatial resolution. This is currently not feasible due in part to prohibitive computational requirements for modeling the transport of chemical species. In this work, we describe a deep learning transport emulator that is able to reduce computations while maintaining skill comparable with the existing numerical model. We show how this method maintains skill in the presence of extreme air quality events, making it a potential candidate for operational use. We also explore evaluating how well this model maintains the physical properties of the modeled transport for a given set of species.

cs.LG↗