Han Zhang

Han Zhang

Auto-generated profile, based on publicly available information

Han Zhang

School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen, China

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
11 December 2025 in Information Processing & Management
Verified
Journal Article
Closed access
YZYuanzhao Zhai
Yuanzhao Zhai
YLYu Lei
Yu Lei
HZHan Zhang
Han Zhang
YYYue Yu
Yue Yu
KXKele Xu
Kele Xu
HWHuaimin Wang
Huaimin Wang

17 August 2022 in IEEE/ACM Transactions on Audio, Speech, and Language Processing
Verified
Journal Article
Closed access
HZHan Zhang
Han Zhang
BLBin Liang
Bin Liang
MYMin Yang
Min Yang
HWHui Wang
Hui Wang
Ruifeng XuRuifeng Xu
Ruifeng Xu
Continual relation extraction (CRE) is an important task of continual learning, which aims to learn incessantly emerging new relations between entities from texts. To avoid catastrophically forgetting old relations, some existing research efforts have focused on exploring memory replayed methods by storing typical historical learned instances or embedding all observed relations as prototypes in the episodic memory and replaying them in the subsequent training process. However, they generally fail to exploit the relation knowledge contained in the pre-trained language model (PLM), which could provide enlightening information to the representations of new relations from the known ones. To this end, we investigate the CRE from a novel perspective by generating knowledge-infused relation prototypes to leverage the relational knowledge from PLM with prompt tuning. Specifically, based on the typical samples collected from the historical learned instances with K-means algorithm, we devise novel relational knowledge-infused prompts to elicit relational knowledge from PLM for generating knowledge-infused relation prototypes. Then the prototypes are used to refine the typical examples embedding and calculate the stability-plasticity balance score for adjusting the memory replayed progress. The experimental results show that our method outperforms the state-of-the-art baseline models in CRE. The further extensive analysis presents that the proposed method is robust to memory size, task order, length of the task sequence, and the number of training instances.

5 June 2022 in arXiv
Verified
Preprint
Open access
PDF file
JYJIAHUI YU
JIAHUI YU
XLXin Li
Xin Li
JKJing Yu Koh
Jing Yu Koh
HZHan Zhang
Han Zhang
RPRuoming Pang
Ruoming Pang
Yonghui WuYonghui Wu
Yonghui Wu
Pretraining language models with next-token prediction on massive text corpora has delivered phenomenal zero-shot, few-shot, transfer learning and multi-tasking capabilities on both generative and discriminative language tasks. Motivated by this success, we explore a Vector-quantized Image Modeling (VIM) approach that involves pretraining a Transformer to predict rasterized image tokens autoregressively. The discrete image tokens are encoded from a learned Vision-Transformer-based VQGAN (ViT-VQGAN). We first propose multiple improvements over vanilla VQGAN from architecture to codebook learning, yielding better efficiency and reconstruction fidelity. The improved ViT-VQGAN further improves vector-quantized image modeling tasks, including unconditional, class-conditioned image generation and unsupervised representation learning. When trained on ImageNet at \(256\times256\) resolution, we achieve Inception Score (IS) of 175.1 and Fr'echet Inception Distance (FID) of 4.17, a dramatic improvement over the vanilla VQGAN, which obtains 70.6 and 17.04 for IS and FID, respectively. Based on ViT-VQGAN and unsupervised pretraining, we further evaluate the pretrained Transformer by averaging intermediate features, similar to Image GPT (iGPT). This ImageNet-pretrained VIM-L significantly beats iGPT-L on linear-probe accuracy from 60.3% to 73.2% for a similar model size. VIM-L also outperforms iGPT-XL which is trained with extra web image data and larger model size.

12 March 2026 in IEEE Transactions on Mobile Computing
Verified
Journal Article
Closed access
LJLiwen Jing
Liwen Jing
TYTingting Yang
Tingting Yang
HZHan Zhang
Han Zhang
YSYuxuan Shi
Yuxuan Shi
CZChi Zhang
Chi Zhang
BZBowen Zhang
Bowen Zhang
Artificial intelligence is expected to play a central role in enabling future 6 G networks. Developing foundation models that support a wide range of downstream tasks is critical for advancing 6 G standardization. This paper proposes a general framework for compressing wireless channel state information (CSI) using pretrained wireless foundation models. The foundation model is pre-trained using self-supervised learning with a masked reconstruction objective, achieving a normalized mean square error on the order of $10^{-3}$ during pretraining. The model is evaluated on a range of wireless communication and sensing tasks, including classification tasks where compressed CSI is directly used for prediction, and regression tasks that require full CSI reconstruction. For regression tasks such as massive MIMO CSI feedback, the pre-compressed output from the foundation model is used as an auxiliary input to the downstream compressor, effectively enhancing the reconstruction quality. Compared to the Type-I codebook with comparable number of feedback bits, our method improves SGCS by 16.22% and reduces NMSE by 93.24%. Additionally, it achieves comparable SGCS performance to the Type-II codebook while using only 29% of the feedback bits. Comparison with existing research further confirms the contribution of the foundation model's compressed output in improving CSI compression performance. For classification tasks such as WiFi-based human activity recognition and human identification, the compressed representations produced by the foundation model can be directly utilized without additional fine-tuning. These representations achieve over 97% accuracy, outperforming conventional AI-based methods even under higher compression ratios. These findings demonstrate that leveraging a pretrained wireless foundation model consistently enhances performance across both classification and regression tasks, underscoring its versatility and potential in wireless CSI processing.

6 June 2025 in IEEE Transactions on Audio, Speech and Language Processing
Verified
Journal Article
Closed access
HZHan Zhang
Han Zhang
YLYu Lei
Yu Lei
BLBin Liang
Bin Liang
HWHui Wang
Hui Wang
LGLin Gui
Lin Gui
Ruifeng XuRuifeng Xu
Ruifeng Xu
Continual text classification has played an increasingly important role in real-world applications, which focus on classifying the constantly emerging texts over time. Although many efforts have been made to overcome the major challenge of continual learning (CL), i.e., catastrophic forgetting (CF), the existing CL methods ignore the underlying interference among different tasks and still suffer from the CF issue. To get a deeper insight into CF, in this paper, we make the first attempt to study the relationship between forgetting and representation similarity in continual text classification tasks. Our thorough analyses of both real-world and pseudo datasets, for the pre-trained language model (PLM) based methods, demonstrate that forgetting can be effectively alleviated by reducing the similarity of the representations of different datasets. Based on our findings, we develop a novel Orthogonality-based Dual-memory Framework (Org-Frame), which is able to minimize the interference among different tasks via a two-stage learning process analogous to the human memory system. Org-Frame can also be combined with the LoRA fine-tuning method to enhance the continual learning capability of large language models (LLMs). Extensive experiments conducted on large-scale benchmarks using BERT and LLaMA demonstrate the effectiveness and robustness of our proposed framework in continual text classification tasks.