Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Qijun Huang has a total of 107 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • DRL-Enabled Polymorphic Acceleration Framework for Flexible and Energy-Efficient Hybrid-Float Deep Learning Inference in Mobile Computing

    by Qunkang Meng, Jie Hu, Zhihan Zhang, Zhan Li, Xingyu Shi, Jun Luo, Qijun Huang, Hao Wang, Sheng Chang

    Hybrid-float quantization has emerged as a promising solution for efficient deep neural network inference on mobile platforms, but its practical deployment is still limited by three challenges: compute–transmission imbalance, rapidly growing mapping complexity, and the tradeoff between intermediate-data movement and hardware cost under tight memory constraints. To address these issues, we propose a DRL-enabled polymorphic acceleration framework for hybrid-float inference in mobile computing. The framework integrates three tightly coupled modules: a Graph Builder that constructs a compute–transmission aware parameterized DAG to expose workload imbalance, a ScaleCore Designer that provides a retargetable hybrid-float hardware template with flexible low-bit floating-point support and a three-level memory hierarchy, and a DRL Scheduler that jointly optimizes deployment and hardware-parallelism decisions under feasibility constraints and normalized multi-objective costs. Experiments on three FPGA platforms across complex CNNs and Bert show that the proposed framework achieves 16.36%–20.25% lower EPF or 19.92%–25.44% higher energy efficiency than state of-the-art accelerators. These results demonstrate a practical path toward flexible, energy-efficient, and scalable hybrid-float deployment for future mobile intelligence systems.

  • TriFusNet: A Triple-Fusion Convolutional Network for Multivariate Time Series Classification

    by Wenhan Liu, Shurong Pan, Sheng Chang, Qijun Huang, Nan Jiang

    Multivariate time series classification (MTSC) plays a critical role in a wide range of real-world applications, such as healthcare, finance, and industrial monitoring. This paper proposes a triple-fusion network (TriFusNet), a novel convolutional network designed to address the challenges of MTSC. TriFusNet employs a specialized architecture that captures both variable-specific features and features shared across variables through the parallel use of standard, depth-wise, and shared-kernel convolutions. A hierarchical triple-fusion strategy is introduced to enhance representation learning across three stages: input-level fusion transforms raw variables, intermediate-level fusion integrates heterogeneous features, and output-level fusion improves decision robustness. Extensive experiments on 26 benchmark datasets show that TriFusNet outperforms 15 competitive baselines, achieving the best average rank (4.3077) with a Win/Draw/Loss of 5/4/17. The effectiveness of its architectural design and parameter settings is empirically validated, and a qualitative theoretical discussion is conducted to support the proposed fusion strategy. These results highlight TriFusNet's strong potential for real-world applications involving complex and high-dimensional time series data.


  • Single-Step Hardware-Aware Neural Network Quantization With Mixed Precision

    by Jie Hu, Zhihan Zhang, Zhan Li, Qunkang Meng, Xingyu Shi, Qijun Huang, Hao Wang, Sheng Chang

    Quantization is a neural network compression technique that effectively improves the deployment performance on inference hardware. Fixed-point quantization methods use the same bit-width for all layers in the network, which leads to difficulties in balancing compression rate and accuracy loss. Therefore, mixed-precision quantization has recently been proposed. The major challenge of the mixed-precision quantization is to select the quantization bit-widths of each layer in network to simultaneously meet the requirements of minimizing accuracy loss and hardware resource consumption. In this paper, we present a Single-Step Hardware-Aware Quantization (SHQ) method. It can calculate the resource consumption of hardware such as Field Programmable Gate Arrays (FPGAs) before actual deployment and find effective quantization schemes by just single step, different from common software-hardware two-step approaches with large workload and time-consuming. Genetic algorithm is combined with SHQ to search quantization schemes with low hardware resource usage and high accuracy. In addition, the correlation between hardware resource cost and qualitative indicators of proxy signal has been analyzed to bring computer-aided design insights of neural network accelerators. Experiments of full-pipeline accelerator deployment on FPGAs platforms show that our approach saves 39% of Digital Signal Processors (DSP) and 59% of Block Random Access Memory (BRAM) usage on MobileNet compared to full 8-bit quantization, while the accuracy drops by only 0.56%. The source code about our method can be found at this link: https://github.com/hujie369/SHQ.


  • Multilayer Field-Free Magnetic Tunnel Junction With Interlayer Dzyaloshinsky–Moriya Interaction

    by Rui Zhou, Haiyang Zhang, Jin He, Qijun Huang, Hao Wang, Sheng Chang

    Magnetic tunnel junctions (MTJs) are of great significance for the next generation ultrafast spintronic memories due to their non-volatility and nanosecond-level writing characteristics. Effective spin manipulation is the key to realizing high-speed field-free magnetic random access memories (MRAMs). In this paper, the performance of applying the interlayer Dzyaloshinsky-Moriya interaction (DMI) to the conventional MTJ structure is investigated. By replacing the perpendicular free layer with a composite structure of out-of-plane free layer/coupling layer/in-plane free layer, the rapid switching of MTJs is achieved. We analyze the working conditions of the device through macrospin simulations and verify it by micromagnetic simulations. This structure significantly improves the speed at which the magnetic moment recovers to stability after the removal of external effects(voltage and current), reaching 56.25% of the traditional spin-orbit torque (SOT) MTJ. We believe that our work may promote the research and development of high-speed and field-free MRAMs in the future.

  • Open Access

    S2M2ECG: Spatio-temporal bi-directional State Space Model Enabled Multi-branch Mamba for ECG

    by Huaicheng Zhang, Ruoxin Wang, Chenlian Zhou, Jiguang Shi, Yue Ge, Zhoutong Li, Sheng Chang, Hao Wang, Jin He, Qijun Huang

    As one of the most effective methods for cardiovascular disease (CVD) diagnosis, multi-lead Electrocardiogram (ECG) signals present a characteristic multi-sensor information fusion challenge that has been continuously researched in deep learning domains. Despite the numerous algorithms proposed with different DL architectures, maintaining a balance among performance, computational complexity, and multi-source ECG feature fusion remains challenging. Recently, state space models (SSMs), particularly Mamba, have demonstrated remarkable effectiveness across various fields. Their inherent design for high-efficiency computation and linear complexity makes them particularly suitable for low-dimensional data like ECGs. This work proposes S2M2ECG, an SSM architecture featuring three-level fusion mechanisms: (1) Spatio-temporal bi-directional SSMs with segment tokenization for low-level signal fusion, (2) Intra-lead temporal information fusion with bi-directional scanning to enhance recognition accuracy in both forward and backward directions, (3) Cross-lead feature interaction modules for spatial information fusion. To fully leverage the ECG-specific multi-lead mechanisms inherent in ECG signals, a multi-branch design and lead fusion modules are incorporated, enabling individual analysis of each lead while ensuring seamless integration with others. Experimental results reveal that S2M2ECG achieves superior performance in the rhythmic, morphological, and clinical scenarios. Moreover, its lightweight architecture ensures it has nearly the fewest parameters among existing models, making it highly suitable for efficient inference and convenient deployment. Collectively, S2M2ECG offers a promising alternative that strikes an excellent balance among performance, computational complexity, and ECG-specific characteristics, paving the way for high-performance, lightweight computations in CVD diagnosis.


  • Open Access

    Towards Hallucination-Free Music: A Reinforcement Learning Preference Optimization Framework for Reliable Song Generation

    by Huaicheng Zhang, Wei Tan, Guangzheng Li, Yixuan Zhang, Hangting Chen, Shun Lei, Chenyu Yang, Zhiyong Wu, Shuai Wang, Qijun Huang, et al.

    Recent advances in audio-based generative language models have accelerated AI-driven lyric-to-song generation. However, these models frequently suffer from content hallucination, producing outputs misaligned with the input lyrics and undermining musical coherence. Current supervised fine-tuning (SFT) approaches, limited by passive label-fitting, exhibit constrained self-improvement and poor hallucination mitigation. To address this core challenge, we propose a novel reinforcement learning (RL) framework leveraging preference optimization for hallucination control. Our key contributions include: (1) Developing a robust hallucination preference dataset constructed via phoneme error rate (PER) computation and rule-based filtering to capture alignment with human expectations; (2) Implementing and evaluating three distinct preference optimization strategies within the RL framework: Direct Preference Optimization (DPO), Proximal Policy Optimization (PPO), and Group Relative Policy Optimization (GRPO). DPO operates off-policy to enhance positive token likelihood, achieving a significant 7.4% PER reduction. PPO and GRPO employ an on-policy approach, training a PER-based reward model to iteratively optimize sequences via reward maximization and KL-regularization, yielding PER reductions of 4.9% and 4.7%, respectively. Comprehensive objective and subjective evaluations confirm that our methods effectively suppress hallucinations while preserving musical quality. Crucially, this work presents a systematic, RL-based solution to hallucination control in lyric-to-song generation. The framework's transferability also unlocks potential for music style adherence and musicality enhancement, opening new avenues for future generative song research.

  • MetaAccel: A High-Performance and Agile Accelerator Design Framework With Multi Clock Domain Optimization for Complex CNN

    by Zhan Li, Yuxian Jiang, Zhihan Zhang, Jiahui Huang, Qunkang Meng, Xingyu Shi, Hao Wang, Qijun Huang, Sheng Chang

    Edge computing for artificial intelligence (AI) has become a new focus today. At the edge, the growing complexity and diversity of AI models has made FPGA, which has shorter development cycles, a good choice. Traditional AI accelerators on FPGA are mainly based on the Compute Engine (CE) architecture, suffering from low resource utilization and suboptimal speed. In contrast, the pipeline architecture achieves higher performance through its algorithm-structure-aware feature and fully on-chip data flow. However, customized designs and large bandwidth demands bring new challenges to its development agility and memory utilization, while high-performance acceleration for complex neural networks is still hard. In this article, we proposed MetaAccel, a novel fully pipelined accelerator. It uses two clock domains to manage data scheduling and calculation, significantly improving computing resource efficiency and on-chip memory utilization. Besides that, we built a hyperparameter-driven resource estimation model that can match the most appropriate design solutions for specific network structures. Based on this architecture, processing method for networks with complex branch structures and various operations is given, which makes MetaAccel suitable for Convolutional Neural Networks (CNNs) in different fields, such as image classification, object detection, and image segmentation. For typical networks, MetaAccel can achieve a throughput of more than 0.7TOPS and a DSP efficiency of up to 2.0GOPS/DSP and outperforms previous FPGA work in other metrics, showing its advantages in complex CNNs’ acceleration.


  • JAMC: A jigsaw-based autoencoder with masked contrastive learning for cardiovascular disease diagnosis

    by Yue Ge, Huaicheng Zhang, Jiguang Shi, Deyu Luo, Sheng Chang, Jin He, Qijun Huang, Hao Wang


  • Direct Lead Assignment: A Simple and Scalable Contrastive Learning Method for ECG and Its IoMT Applications

    by Wenhan Liu, Shurong Pan, Sheng Chang, Qijun Huang, Nan Jiang

    Nowadays, applying deep learning (DL) to electrocardiogram (ECG) analysis has become a significant topic in intelligent healthcare. DL models heavily rely on large-scale labeled ECGs in supervised learning, while labeling ECGs is a costly and time-consuming process. This paper proposes a simple self-supervised learning (SSL) method to pretrain models using unlabeled ECGs, improving model performances in a low-data regime. It is termed direct lead assignment (DLA). In pretraining, DLA employs multi-lead and single-lead encoders to interact between global and lead-specific representations. The pretrained encoders can constitute scalable models, which can be deployed on Internet of Medical Things (IoMT) devices for ECG monitoring with different leads. According to the experiments, DLA outperforms existing SSL methods for ECGs and reduces the reliance on labels by 4~8×. In other words, DLA can make the model obtain better results when labeled data are scarce than the one trained from scratch, as the pretraining helps the model recognize critical ECG patterns in a low-data regime. For IoMT applications, the models are deployed on a Raspberry Pi 2 W with an ARM Cortex-A53 processor. It can run the models in real time. The maximum running time is about 388 ms/10-second record. Thus, DLA shows excellent potential for automatic ECG analysis based on IoMT, aiding cardiologists in diagnosing cardiovascular diseases.