arXiv subjects
Publications and source records attributed to Kai Zhu.
We have derived the tensor amplitudes for partial wave analysis of $ψ\toΔ\barΔ$, $Δ\to p π$ within the helicity frame, as well as the amplitudes for the other decay sequences with same final states. These formulae are practical for the experiments measuring $ψ$ decaying into $p \bar{p}π^+ π^-$ final states, such as BESIII with its recently collected huge $J/ψ$ and $ψ(2S)$ data samples.
MLP-like models built entirely upon multi-layer perceptrons have recently been revisited, exhibiting the comparable performance with transformers. It is one of most promising architectures due to the excellent trade-off between network capability and efficiency in the large-scale recognition tasks. However, its generalization performance to heterogeneous tasks is inferior to other architectures (e.g., CNNs and transformers) due to the extensive retention of domain information. To address this problem, we propose a novel frequency-aware MLP architecture, in which the domain-specific features are filtered out in the transformed frequency domain, augmenting the invariant descriptor for label prediction. Specifically, we design an adaptive Fourier filter layer, in which a learnable frequency filter is utilized to adjust the amplitude distribution by optimizing both the real and imaginary parts. A low-rank enhancement module is further proposed to rectify the filtered features by adding the low-frequency components from SVD decomposition. Finally, a momentum update strategy is utilized to stabilize the optimization to fluctuation of model parameters and inputs by the output distillation with weighted historical states. To our best knowledge, we are the first to propose a MLP-like backbone for domain generalization. Extensive experiments on three benchmarks demonstrate significant generalization performance, outperforming the state-of-the-art methods by a margin of 3%, 4% and 9%, respectively.
Non-exemplar class-incremental learning is to recognize both the old and new classes when old class samples cannot be saved. It is a challenging task since representation optimization and feature retention can only be achieved under supervision from new classes. To address this problem, we propose a novel self-sustaining representation expansion scheme. Our scheme consists of a structure reorganization strategy that fuses main-branch expansion and side-branch updating to maintain the old features, and a main-branch distillation scheme to transfer the invariant knowledge. Furthermore, a prototype selection mechanism is proposed to enhance the discrimination between the old and new classes by selectively incorporating new samples into the distillation process. Extensive experiments on three benchmarks demonstrate significant incremental performance, outperforming the state-of-the-art methods by a margin of 3%, 3% and 6%, respectively.
Based on $η$-$η'$ mixing analysis, we propose a novel method to extract the physical solutions for the hadronic properties of the $Y(4230)$ resonance from the experimental data. Experimentally, multiple solutions have been reported in the decays of $Y(4230) \to ηJ/ψ$ and $Y(4230) \to η' J/ψ$. Utilizing our method, we determine a unique solution for the process $Y(4230) \to η' J/ψ$. Likewise, two solutions for the process $Y(4230) \to ηJ/ψ$ are preferred among the originally reported three solutions under the assumption that $Y(4230)$ dose not take an $s\bar{s}$ component.
Exemplar-based class-incremental learning is to recognize new classes while not forgetting old ones, whose samples can only be saved in limited memory. The ratio fluctuation of new samples to old exemplars, which is caused by the variation of memory capacity at different environments, will bring challenges to stabilize the incremental optimization process. To address this problem, we propose a novel self-paced imbalance rectification scheme, which dynamically maintains the incremental balance during the representation learning phase. Specifically, our proposed scheme consists of a frequency compensation strategy that adjusts the logits margin between old and new classes with the corresponding number ratio to strengthen the expression ability of the old classes, and an inheritance transfer strategy to reduce the representation confusion by estimating the similarity of different classes in the old embedding space. Furthermore, a chronological attenuation mechanism is proposed to mitigate the repetitive optimization of the older classes at multiple step-wise increments. Extensive experiments on three benchmarks demonstrate stable incremental performance, significantly outperforming the state-of-the-art methods.
The presence of unreacted lead iodide in organic-inorganic lead halide perovskite solar cells is widely correlated with an increase in power conversion efficiency. We investigate the mechanism for this increase by identifying the role of surfaces and interfaces present between methylammonium lead iodide perovskite films and excess lead iodide. We show how type I and II band alignments arising under different conditions result in either passivation of surface defects or hole injection. Through first-principles simulations of solid-solid interfaces, we find that lead iodide captures holes from methylammonium lead iodide and modulates the formation of defects in the perovskite, affecting recombination. Using surface-sensitive optical spectroscopy techniques, such as transient reflectance and time-resolved photoluminescence, we show how excess lead iodide affects the diffusion and surface recombination velocity of charge carriers in methylammonium lead iodide films. Our coupled experimental and theoretical results elucidate the role of excess lead iodide in perovskite solar cells.
We show in this work that memory intensive computations can result in severe performance problems due to off-chip memory access and CPU-GPU context switch overheads in a wide range of deep learning models. For this problem, current just-in-time (JIT) kernel fusion and code generation techniques have limitations, such as rough fusion plan exploration strategies and limited code generation ability. We propose FusionStitching, a deep learning compiler capable of fusing memory intensive operators, with varied data dependencies and non-homogeneous parallelism, into large GPU kernels to reduce global memory access and context switch overhead automatically. FusionStitching widens the range of operation combinations that fusion can target beyond previous JIT works by introducing data reuse of intermediate values. It explores large fusion spaces to decide optimal fusion plans with considerations of memory access costs, kernel calls and resource usage constraints. FusionStitching tunes the optimal stitching scheme with a domain-specific cost model efficiently. Experimental results show that FusionStitching can reach up to 2.21x speedup compared to state-of-the-art, with 1.45x on average. Besides these experimental results, we integrated our approach into a compiler product and deployed it onto a production cluster for AI workloads with thousands of GPUs. The system has been in operation for more than 4 months and saves 7,000 GPU hours on average for approximately 30,000 tasks per month.
Deep learning-based methods for low-light image enhancement typically require enormous paired training data, which are impractical to capture in real-world scenarios. Recently, unsupervised approaches have been explored to eliminate the reliance on paired training data. However, they perform erratically in diverse real-world scenarios due to the absence of priors. To address this issue, we propose an unsupervised low-light image enhancement method based on an effective prior termed histogram equalization prior (HEP). Our work is inspired by the interesting observation that the feature maps of histogram equalization enhanced image and the ground truth are similar. Specifically, we formulate the HEP to provide abundant texture and luminance information. Embedded into a Light Up Module (LUM), it helps to decompose the low-light images into illumination and reflectance maps, and the reflectance maps can be regarded as restored images. However, the derivation based on Retinex theory reveals that the reflectance maps are contaminated by noise. We introduce a Noise Disentanglement Module (NDM) to disentangle the noise and content in the reflectance maps with the reliable aid of unpaired clean images. Guided by the histogram equalization prior and noise disentanglement, our method can recover finer details and is more capable to suppress noise in real-world low-light scenarios. Extensive experiments demonstrate that our method performs favorably against the state-of-the-art unsupervised low-light enhancement algorithms and even matches the state-of-the-art supervised algorithms.
Many recent machine learning models show dynamic shape characteristics. However, existing AI compiler optimization systems suffer a lot from problems brought by dynamic shape models, including compilation overhead, memory usage, optimization pipeline and deployment complexity. This paper provides a compiler system to natively support optimization for dynamic shape workloads, named DISC. DISC enriches a set of IR to form a fully dynamic shape representation. It generates the runtime flow at compile time to support processing dynamic shape based logic, which avoids the interpretation overhead at runtime and enlarges the opportunity of host-device co-optimization. It addresses the kernel fusion problem of dynamic shapes with shape propagation and constraints collecting methods. This is the first work to demonstrate how to build an end-to-end dynamic shape compiler based on MLIR infrastructure. Experiments show that DISC achieves up to 3.3x speedup than TensorFlow/PyTorch, and 1.8x than Nimble.
The elliptical power-law (EPL) model of the mass in a galaxy is widely used in strong gravitational lensing analyses. However, the distribution of mass in real galaxies is more complex. We quantify the biases due to this model mismatch by simulating and then analysing mock {\it Hubble Space Telescope} imaging of lenses with mass distributions inferred from SDSS-MaNGA stellar dynamics data. We find accurate recovery of source galaxy morphology, except for a slight tendency to infer sources to be more compact than their true size. The Einstein radius of the lens is also robustly recovered with 0.1% accuracy, as is the global density slope, with 2.5% relative systematic error, compared to the 3.4% intrinsic dispersion. However, asymmetry in real lenses also leads to a spurious fitted `external shear' with typical strength, $γ_{\rm ext}=0.015$. Furthermore, time delays inferred from lens modelling without measurements of stellar dynamics are typically underestimated by $\sim$5%. Using such measurements from a sub-sample of 37 lenses would bias measurements of the Hubble constant $H_0$ by $\sim$9%. Although this work is based on a particular set of MaNGA galaxies, and the specific value of the detected biases may change for another set of strong lenses, our results strongly suggest the next generation cosmography needs to use more complex lens mass models.
Over the past decade, the number of wildfire has increased significantly around the world, especially in the State of California. The high-level concentration of greenhouse gas (GHG) emitted by wildfires aggravates global warming that further increases the risk of more fires. Therefore, an accurate prediction of wildfire occurrence greatly helps in preventing large-scale and long-lasting wildfires and reducing the consequent GHG emissions. Various methods have been explored for wildfire risk prediction. However, the complex correlations among a lot of natural and human factors and wildfire ignition make the prediction task very challenging. In this paper, we develop a deep learning based data augmentation approach for wildfire risk prediction. We build a dataset consisting of diverse features responsible for fire ignition and utilize a conditional tabular generative adversarial network to explore the underlying patterns between the target value of risk levels and all involved features. For fair and comprehensive comparisons, we compare our proposed scheme with five other baseline methods where the former outperformed most of them. To corroborate the robustness, we have also tested the performance of our method with another dataset that also resulted in better efficiency. By adopting the proposed method, we can take preventive strategies of wildfire mitigation to reduce global GHG emissions.
An adaptive interpolation scheme is proposed to accurately calculate the wideband responses in electromagnetic simulations. In the proposed scheme, the sampling points are first carefully divided into several groups based on their responses to avoid the Runge phenomenon and the error fluctuations, and then different interpolation strategies are used to calculate the responses in the whole frequency band. If the relative error does not satisfy the predefined threshold in a specific frequency band, it will be refined until the error criteria is met. The detailed error analysis is also presented to verify the accuracy of the interpolation scheme. At last, two numerical examples including the antenna radiation and the filter simulation are carried out to validate its accuracy and efficiency.
The outstanding performance of organic-inorganic metal trihalide solar cells benefits from the exceptional photo-physical properties of both electrons and holes in the material. Here, we directly probe the free-carrier dynamics in Cs-doped FAPbI3 thin films by spatiotemporal photoconductivity imaging. Using charge transport layers to selectively quench one type of carriers, we show that the two relaxation times on the order of 1 microsecond and 10 microseconds correspond to the lifetimes of electrons and holes in FACsPbI3, respectively. Strikingly, the diffusion mapping indicates that the difference in electron/hole lifetimes is largely compensated by their disparate mobility. Consequently, the long diffusion lengths (3 ~ 5 micrometers) of both carriers are comparable to each other, a feature closely related to the unique charge trapping and de-trapping processes in hybrid trihalide perovskites. Our results unveil the origin of superior diffusion dynamics in this material, crucially important for solar-cell applications.
Few-shot class-incremental learning is to recognize the new classes given few samples and not forget the old classes. It is a challenging task since representation optimization and prototype reorganization can only be achieved under little supervision. To address this problem, we propose a novel incremental prototype learning scheme. Our scheme consists of a random episode selection strategy that adapts the feature representation to various generated incremental episodes to enhance the corresponding extensibility, and a self-promoted prototype refinement mechanism which strengthens the expression ability of the new classes by explicitly considering the dependencies among different classes. Particularly, a dynamic relation projection module is proposed to calculate the relation matrix in a shared embedding space and leverage it as the factor for bootstrapping the update of prototypes. Extensive experiments on three benchmark datasets demonstrate the above-par incremental performance, outperforming state-of-the-art methods by a margin of 13%, 17% and 11%, respectively.
We propose novel triangle relations, not the well-known triangle singularity, for better understanding of the exotic XYZ states. Nine XYZ resonances, X(3872), Y(4230), Zc(3900), X(4012), Y(4360/4390), Zc(4020), X(4274), Y(4660), and Zcs(3985) have been classified into triples to construct three triangles based on the assumption that they are all tetra-quark states. We also suggest some channels deserving search for with high priority based on this hypothesis, as well as predictions of a few production/decay rates of these channels. We hope further experimental studies of the XYZ states will benefit from our results.
Few-shot segmentation aims at assigning a category label to each image pixel with few annotated samples. It is a challenging task since the dense prediction can only be achieved under the guidance of latent features defined by sparse annotations. Existing meta-learning method tends to fail in generating category-specifically discriminative descriptor when the visual features extracted from support images are marginalized in embedding space. To address this issue, this paper presents an adaptive tuning framework, in which the distribution of latent features across different episodes is dynamically adjusted based on a self-segmentation scheme, augmenting category-specific descriptors for label prediction. Specifically, a novel self-supervised inner-loop is firstly devised as the base learner to extract the underlying semantic features from the support image. Then, gradient maps are calculated by back-propagating self-supervised loss through the obtained features, and leveraged as guidance for augmenting the corresponding elements in embedding space. Finally, with the ability to continuously learn from different episodes, an optimization-based meta-learner is adopted as outer loop of our proposed framework to gradually refine the segmentation results. Extensive experiments on benchmark PASCAL-$5^{i}$ and COCO-$20^{i}$ datasets demonstrate the superiority of our proposed method over state-of-the-art.
We investigate the on-average properties for 28 star-forming ultra-diffuse galaxies (UDGs) located in low-density environments, by stacking their spectra from the Sloan Digital Sky Survey. These relatively-isolated UDGs, with stellar masses of $\log_{10}(M_*/M_{\odot})\sim 8.57\pm0.29$, have the on-average total-stellar-metallicity [M/H]$\sim -0.82\pm0.14$, iron-metallicity [Fe/H]$\sim -1.00\pm0.16$, stellar age $t_*\sim5.2\pm0.5$ Gyr, $α$-enhancement [$α$/Fe]$\sim 0.24\pm0.10$, and oxygen abundance 12+log(O/H)$\sim 8.16\pm0.06$, as well as central stellar velocity dispersion $54\pm12$ km/s. On the star-formation rate versus stellar mass diagram, these UDGs are located lower than the extrapolated star-forming main sequence from the massive spirals, but roughly follow the main sequence of low-surface-brightness dwarf galaxies. We find that these star-forming UDGs are not particularly metal-poor or metal-rich for their stellar masses, as compared with the metallicity-mass relations of the nearby typical dwarfs. With the UDG data of this work and previous studies, we also find a coarse correlation between [Fe/H] and magnesium-element enhancement [Mg/Fe] for UDGs: [Mg/Fe]$\simeq-0.43(\pm0.26)$[Fe/H]$-0.14(\pm0.40)$.
The strong, electromagnetic and mixed strong-electromagnetic amplitudes of the $ψ(2S)$ decays into baryon-anti-baryon pairs have been obtained by exploiting all available data sets in the framework of an effective Lagrangian model. We observed that at the $ψ(2S)$ mass the QCD regime is not completely perturbative, as can be inferred by the relative strength of the strong and the mixed strong-electromagnetic amplitudes. Recently a similar conclusion has been reached also for the $J/ψ$ decays. The relative phase between the strong and the electromagnetic amplitudes is $φ= (58\pm 8)^\circ$, to be compared with $φ= (73\pm 8)^\circ$ obtained for the $J/ψ$. On the other hand, in the case of the $ψ(2S)$ meson, different values of the ratio between strong and mixed strong-electromagnetic amplitudes are phenomenologically required, while for the $J/ψ$ meson only one ratio was enough to describe the data. Finally, we also observed a peculiar behavior of the mixed strong-electromagnetic amplitudes of the decays $ψ(2S)\toΣ^+ \overline Σ^-$ and $ψ(2S)\toΣ^- \overline Σ^+$.