Jing Peng

Communication poses a dominant bottleneck in distributed data parallel training with synchronous stochastic gradient descent, whereas the traditional AllReduce collective used for gradient synchronization limits the efficient utilization of communication compression strategies. In this paper, we propose a low-bit and sparsified gradient communication algorithm with convergence guarantees, called QTopKA2A, which enables effectively overlapping communications with both feed-forward and backpropagation computations through a decomposed communication framework. Specifically, top-k sparsified gradients are communicated via AlltoAll in the backward phase, while low-bit quantized gradients are exchanged via AllGather in the forward phase. Residual-based error feedback mechanisms are employed in both sparsification and quantization stages to compensate for compression errors and guarantee convergence with significantly higher compression ratios. In addition, we leverage an effective tensor fusion to reduce communication startup overhead. Experimental results demonstrate that QTopKA2A preserves dense-level accuracy while achieving up to 4.74 × end-to-end training speedup over existing dense and compressed baselines on a 32-GPU cluster.

This study applies a modality-bridging framework to automatically construct Chain-of-Thought (CoT) reasoning from meteorological images, reducing the need for expert annotation. The proposed pipeline integrates semantic extraction, Pseudo-CoT generation, and logical fusion to produce structured reasoning chains. Using the WeatherQA benchmark, we build datasets under single-image, 3-image, and 20-image settings—with automated and Expert-Guided variants—and evaluate performance on Areas Affected and Conditional Concern tasks. The results show near-expert spatial reasoning and more compact, well-aligned CoTs with reduced-image inputs. Multi-image settings reveal challenges in integrating dense visual cues, while semantic classification remains difficult due to label ambiguity. Overall, modality-bridging offers a scalable, interpretable, and low-cost approach for multimodal meteorological reasoning.

To address the issue of missing near-surface air temperature data caused by the uneven distribution of ground meteorological observation stations, we propose a method for near-surface air temperature estimation based on an improved conditional generative adversarial network (CGAN) framework. Leveraging the all-weather coverage advantage of Fengyun meteorological satellites, Fengyun-4A (FY-4A) satellite remote sensing data are utilized as conditional guiding information for the CGAN, helping to direct and constrain the near-surface air temperature estimation process. In the proposed network model of the method based on the conditional generative adversarial network structure, the generator combining a self-attention mechanism and cascaded residual blocks is designed with U-Net as the backbone, which extracts implicit feature information and suppresses the irrelevant information in the Fengyun satellite data. Furthermore, a discriminator with multi-level and multi-scale spatial feature fusion is constructed to enhance the network’s perception of details and the global structure, enabling accurate air temperature estimation. The experimental results demonstrate that, compared with Attention U-Net, Pix2pix, and other deep learning models, the method presents significant improvements of 68.75% and 10.53%, respectively in the root mean square error (RMSE) and Pearson’s correlation coefficient (CC). These results indicate the superior performance of the proposed model for near-surface air temperature estimation.

Synchronous stochastic gradient descent (S-SGD) with data parallelism has become a de-facto approach in training large-scale deep neural networks (DNNs) on multi-GPU systems. However, S-SGD requires iteratively synchronizing gradients from all workers, which incurs excessive communication costs and limits the scaling efficiency of GPU clusters. Gradient sparsification like top-k sparsification techniques has been shown to be potentially effective in reducing the communication volume. Yet, existing approaches with top-k sparsification still suffer from high communication complexity in that they often need a very low density to achieve better performance while easily sacrificing the model accuracy. To this end, we propose a novel sparse communication approach called TopKA2A, which integrates top-k sparsification with AlltoAll communication to exchange sparse tensors among GPUs, resulting in a notable reduction in communication complexity. With rigorous theoretical analysis on the conditions that TopKA2A can be applied, we design a simple yet effective tensor fusion algorithm based on binary search. We perform in-depth analysis and evaluation of communication efficiency by comparing TopKA2A with state-of-the-art solutions over popular models without compromising model accuracy. Experimental results demonstrate that TopKA2A yields substantial communication efficiency gains and runs up to 73% faster than existing algorithms on a 32-GPU cluster.

Most face-inpainting methods perform well in face repair. However, these methods can only complete a single face image per input. Although existing various image-inpainting methods can achieve pluralistic image inpainting, they typically produce faces with distorted structures or the same texture. To resolve these shortcomings and achieve high-quality diverse face inpainting, we propose PFTANet, a two-stage pluralistic face-inpainting network that transforms attribute information. In the first stage, the face-parsing network is fine-tuned to obtain semantic facial region information. In the second stage, a generator consisting of SNBlock, CF_ShiftBlocks, and CF_MergeBlock, which ensures that high-quality pluralistic face results are generated, is used. Specifically, CF_ShiftBlocks completes pluralistic face generation by transforming the attribute information from the conditional face extracted by the attribute extractor and ensuring the consistency of the attribute information between the conditional and generated faces. CF_MergeBlock ensures structural consistency between the masked and background regions of the generated face using facial region semantic information. A multi-patch discriminator is used to enhance facial detail generation. Experimental results for the CelebA and CelebA-HQ datasets indicated that PFTANet achieved pluralistic and visually realistic face inpainting.

Blind face deblurring aims to recover a sharper face from its unknown degraded version (i.e., different motion blur, noise). However, most previous works typically rely on degradation facial priors extracted from low-quality inputs, which generally leads to unlifelike deblurring results. In this paper, we propose a multi-scale progressive face-deblurring generative adversarial network (MPFD-GAN) that requires no facial priors to generate more realistic multi-scale deblurring results by one feed-forward process. Specifically, MPFD-GAN mainly includes two core modules: the feature retention module and the texture reconstruction module (TRM). The former can capture non-local similar features by full advantage of the different receptive fields, which facilitates the network to recover the complete structure. The latter adopts a supervisory attention mechanism that fully utilizes the recovered low-scale face to refine incoming features at every scale before propagating them further. Moreover, TRM extracts the high-frequency texture information from the recovered low-scale face by the Laplace operator, which guides subsequent steps to progressively recover faithful face texture details. Experimental results on the CelebA, UTKFace and CelebA-HQ datasets demonstrate the effectiveness of the proposed network, which achieves better accuracy and visual quality against state-of-the-art methods.

It is always a hot issue in the intelligence analysis field to predict the trend of news description by pre-trained language models and graph neural networks. However, there are several problems in the existing research: (1) there are few Chinese data sets on this subject in academia and industry; and (2) using the existing pre-trained language models and graph classification algorithms cannot achieve satisfactory results. The method described in this paper can better solve these problems. (1) We built a Chinese news database predicted by more than 9000 annotated news time trends, filling the gaps in this database. (2) We designed an improved method based on the pre-trained language model and graph neural networks pooling algorithm. In the graph pooling algorithm, the Graph U-Nets Pooling method and self-attention are combined, which can better solve the analysis of the problem of forecasting the development trend of news events. The experimental results show that the effect of this method compared with the baseline graph classification algorithm is improved, and it also solves the shortcomings of the pre-trained language model that cannot handle very long texts. Therefore, it can be concluded that our research has strong processing capabilities for analyzing and predicting the development trend of Chinese news events.