HXHao Xu

Hao Xu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Hao Xu has a total of 19 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Async-Checkpoint: Redesigning Gradient Checkpointing with Asynchronous Execution

    by Changhao Ma, Suping Wu, Jiawei Dai, Jie Yang, Hao Xu, Dong Qiang

    Activation recomputation (gradient checkpointing) reduces the memory footprint of intermediate activations from O(n) to \(O(\sqrt {n})\) by selectively discarding and recomputing them during the backward pass. However, a fundamental limitation persists: all existing implementations follow a synchronous blocking paradigm, forcing recomputation to fully occupy the critical path of backpropagation and typically introducing 30%–40% additional training overhead. This paper proposes Async-Checkpoint, an asynchronous activation recomputation algorithm. Our core insight is that the serial dependency between the recomputation and gradient computation of adjacent blocks is not required by data flow. Based on this, we redesign the execution logic—replacing the serial “recompute-then-gradient” loop with an asynchronous pipeline that overlaps the two tasks in separate CUDA streams, thereby removing recomputation from the critical path. Experiments on BERT, ResNet, U-Net, and a recurrent MVS model demonstrate generality across linear chains, DAGs with skip connections, and dynamic recurrent structures. Async-Checkpoint consistently reduces iteration time compared to the PyTorch synchronous implementation while maintaining memory consumption on par with synchronous checkpointing.

  • HiAvatar: High-Fidelity Animatable Head Avatar

    by Qiang Ma, Suping Wu, Sheng Yang, Weibin Qiu, Feng Liu, Zhaocheng Jin, Hao Xu

  • ZC-MVSNet: Zero-Sum Convolution and Prior Fusion for Multi-view Stereo

    by Weibin Qiu, Suping Wu, Hao Xu, Jie Yang, Xiang Zhang

  • Enhancing Distributed Source Coding With Encoder-Centric Frequency Adaptation and Spatial Transformation

    by Hao Xu, Bin Tan, Yihao Chen, Die Hu, Jun Wu

    Current methodologies in distributed source coding have predominantly investigated decoder-focused strategies, emphasizing the alignment and exploitation of side information. This study introduces a paradigm shift by presenting an encoder-centric algorithm that conducts proactive optimization in the frequency domain. This shift is motivated by the current deep learning models' tendency to passively extract high-frequency elements, such as contours and content in the spatial domain at the encoder side, without considering the frequency characteristics of these spatial components. Unlike current trends, the proposed scheme actively selects the essential frequency components directly in the frequency domain by introducing an adaptive self-learning filter, enabling the encoder to discern and retain critical frequency components effectively and precisely. Furthermore, we align the side information in the spatial domain before feature extraction and implement an affine transformation-based alignment strategy to utilize the side information better. By leveraging the shared frequency domain components of the image pairs, the proposed algorithm adeptly learns affine coefficients to accomplish precise spatial alignment. This dual strategy of proactive encoder optimization and decoder alignment via affine transformations is highly efficient, outperforming existing state-of-the-art methods in distributed source coding when tested across two diverse datasets by an average of 0.5 dB in PSNR.

  • Multi-Space Point Geometry Compression With Progressive Relation-Aware Transformer

    by Wenxiang Shen, BaoYe Zhang, Hao Xu, XiaoHan Li, Jun Wu

    Deep network-based point cloud geometry compression is becoming more crucial and attractive due to constantly expanding 3D applications. The current strategy employing holistic point clouds as input imposes limitations on the compressed point cloud size and results in the loss of local geometry information, simplifying the repetition of local point cloud patterns. Nevertheless, employing point cloud patches as input results in a fresh set of erroneous points within the spliced point cloud. This occurrence can be attributed to the loss of global information and the relationship among the point cloud patches. Hence, this paper introduces a novel framework for progressive downsampling point cloud compression that follows the principles of two distinct methodologies. Our strategy still uses the autoencoder structure. Specifically, the encoder reduces the point cloud size and learns features incorporating local patch structure information and global semantic information. At the same time, the decoder upsamples the quantized and entropy-encoded features to reconstruct the original point cloud. More specifically, the geometrical details inside the patch and the relationship between patches are encoded to obtain the point cloud's global semantic and local geometry information. Concurrently, the computational burden is reduced by partitioning the complete point cloud input into segments and conducting continuous downsampling. Furthermore, we introduced an attention-based point cloud deconvolution module to address the localized repeating concentrations in reconstructed point clouds resulting from linear interpolation. This module samples the parent node and its neighboring relationships in the multi-space domain to improve the characteristics of the parent nodes to be sampled. It then uses the deconvolution to create sub-node characteristics that are more varied than the linear interpolation. Empirical evidence demonstrates that the proposed methodology achieves a effective equilibrium between the compression quality and ratio.


  • Open Access

    A Neural Network Empowered Belief Propagation Algorithm Optimized for Short-Cycles in Tanner Graph

    by Hao Xu, Ying Li, Bin Tan, Jun Wu, Die Hu

    Short-cycles in Tanner graphs have a direct impact on the accuracy and effectiveness of the belief propagation (BP) algorithm, as they diverge the BP algorithm by disrupting the independency of message transmission. In this paper, we present a Neural Network Empowered BP (NNE-BP) algorithm as an alternative to the conventional BP algorithm, which can approach the optimum form of BP in shortcycles. The earlier deep learning (DL) based decoder treated the entire decoding process as a classification issue, making it effective only in short block lengths and with limited generalization. In the proposed NNE-BP, we design two neural networks, namely the horizontal network and vertical network, to learn, respectively, the optimized horizontal and vertical processes for single nodes in the presence of cycles. These two networks operate in an iterative fashion, allowing them to approach a theoretical message propagation equation without assuming independent message transmission. Compared to the classification based DL decoder, the proposed network topology is no longer bundled with any specific Tanner graph, and the number of network parameters are greatly reduced. The obtained experimental results show that the proposed NNE-BP outperforms the conventional BP decoder, especially in the case of short-cycles.