Baosheng Yu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Baosheng Yu has a total of 124 co-authors. Here are the top co-authors by number of shared publications:
- Dacheng Tao
Nanyang technological university, Singapore, Singapore
30 shared publications - Yibing Zhan
Wuhan University
20 shared publications - Jianping Gou
Southwest University
9 shared publications - Dapeng Tao
Yunnan University
7 shared publications - Liang Ding
University of Sydney
6 shared publications - Bo Du
School of Computer Science, Wuhan University, Wuhan, China
5 shared publications
Latest Publications
- Open Access
- 14 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
Federated Distillation: A Survey
Federated learning (FL) enables collaborative model training without centralizing clients’ raw data, but parameter exchange can impose substantial communication costs and usually requires compatible model architectures. Federated distillation (FD) instead transfers predictive or intermediate knowledge through knowledge distillation (KD), which can support heterogeneous models and reduce communication in suitable settings. These benefits are not automatic: their magnitude depends on the transmitted knowledge, reference data, communication schedule, and threat model. This survey provides a challenge-centered synthesis of FD. We first formalize a representative public-data, response-based FD protocol and then organize existing methods by the FL challenges they address, including data, system, and model heterogeneity, client drift, catastrophic forgetting, communication cost, and privacy risk. We further review applications in industrial engineering, computer vision, natural language processing, and healthcare, including recent federated adaptation of large language models. Finally, we identify five research priorities: richer transferred knowledge, joint treatment of interacting forms of heterogeneity, communication-aware distillation, threat-model-driven privacy protection, and standardized theory, benchmarks, and evaluation.
- 14 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
Open AccessFederated Distillation: A Survey
Federated learning (FL) enables collaborative model training without centralizing clients’ raw data, but parameter exchange can impose substantial communication costs and usually requires compatible model architectures. Federated distillation (FD) instead transfers predictive or intermediate knowledge through knowledge distillation (KD), which can support heterogeneous models and reduce communication in suitable settings. These benefits are not automatic: their magnitude depends on the transmitted knowledge, reference data, communication schedule, and threat model. This survey provides a challenge-centered synthesis of FD. We first formalize a representative public-data, response-based FD protocol and then organize existing methods by the FL challenges they address, including data, system, and model heterogeneity, client drift, catastrophic forgetting, communication cost, and privacy risk. We further review applications in industrial engineering, computer vision, natural language processing, and healthcare, including recent federated adaptation of large language models. Finally, we identify five research priorities: richer transferred knowledge, joint treatment of interacting forms of heterogeneity, communication-aware distillation, threat-model-driven privacy protection, and standardized theory, benchmarks, and evaluation.
- Open Access
- 8 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 22 (10) , 1-20
Dual-Prior-Guided Cross-Mamba Network for Hyperspectral and Multispectral Image Fusion
Recently, hyperspectral and multispectral image fusion algorithms based on local and long-range connectivity learning (e.g., CNNs, Transformer, and Mamba) have produced good results. However, designing effective hybrid architectures of CNNs, transformer, and Mamba, guided by rich prior information, that efficiently learn complementary spatial-spectral features in complex hyperspectral and multispectral images, thereby further improving image fusion performance, remains a challenge. In this article, we propose a dual-prior-guided cross-Mamba network (DCMN) for hyperspectral and multispectral image fusion. Specifically, we first introduce an efficient cross-modal adjustment nonlocal module (CANL) with low computational cost to enhance the spatial feature representation of low-resolution hyperspectral images by recalculating attention using multispectral images. We then employ a stack of proposed hybrid CNN-Transformer and Mamba blocks (CTMBs), which integrate multimodal semantic and edge structure prior information, to learn complementary spatial-spectral features at local, regional, and global scales. Extensive experiments are conducted on three datasets from both indoor and outdoor settings, featuring hyperspectral images from agricultural and urban areas. The results show that the proposed DCMN outperforms recent state-of-the-art hyperspectral image fusion methods across different settings, highlighting its effectiveness for hyperspectral and multispectral image fusion.
- 8 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 22 (10) , 1-20
Open AccessDual-Prior-Guided Cross-Mamba Network for Hyperspectral and Multispectral Image Fusion
Recently, hyperspectral and multispectral image fusion algorithms based on local and long-range connectivity learning (e.g., CNNs, Transformer, and Mamba) have produced good results. However, designing effective hybrid architectures of CNNs, transformer, and Mamba, guided by rich prior information, that efficiently learn complementary spatial-spectral features in complex hyperspectral and multispectral images, thereby further improving image fusion performance, remains a challenge. In this article, we propose a dual-prior-guided cross-Mamba network (DCMN) for hyperspectral and multispectral image fusion. Specifically, we first introduce an efficient cross-modal adjustment nonlocal module (CANL) with low computational cost to enhance the spatial feature representation of low-resolution hyperspectral images by recalculating attention using multispectral images. We then employ a stack of proposed hybrid CNN-Transformer and Mamba blocks (CTMBs), which integrate multimodal semantic and edge structure prior information, to learn complementary spatial-spectral features at local, regional, and global scales. Extensive experiments are conducted on three datasets from both indoor and outdoor settings, featuring hyperspectral images from agricultural and urban areas. The results show that the proposed DCMN outperforms recent state-of-the-art hyperspectral image fusion methods across different settings, highlighting its effectiveness for hyperspectral and multispectral image fusion.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 38 (9) , 1-18
The Other Side of the Coin: Exploring Fairness in Retrieval-Augmented Generation
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 38 (9) , 1-18
The Other Side of the Coin: Exploring Fairness in Retrieval-Augmented Generation
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 13038-13045
FingerVeinSyn-5M: A Million-Scale Dataset and Benchmark for Finger Vein Recognition
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 13038-13045
Open AccessFingerVeinSyn-5M: A Million-Scale Dataset and Benchmark for Finger Vein Recognition
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 9 (6) , 4051-4065
Fine-Tuning LLMs for Anesthesiology via Compositional Data Generation
Data scarcity presents a significant challenge in developing medical large language models (LLMs), especially in specialized fields like anesthesiology. Leveraging advanced LLMs like ChatGPT to synthesize high-quality Q&A data is a promising solution to address the data scarcity issue. However, limited research exists on methods to produce such datasets specifically for anesthesiology, despite their importance for training robust LLMs. In this paper, we explore a compositional approach to generating anesthesia data for fine-tuning medical LLMs, and we refer to the proposed compositional data generation method as CDGen. Specifically, we first introduce a compositional augmentation strategy to enhance the diversity and reliability of collected medical records. We then employ a self-talk approach to generate anesthesia Q&A data from augmented medical records. We conduct extensive experiments to demonstrate the effectiveness of our approach using common metrics and GPT-4 evaluation, and experimental results demonstrate that the proposed data generation approach can synthesize high-quality Q&A data and effectively improve the performance of LLMs in Q&A tasks related to anesthesia.
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 9 (6) , 4051-4065
Fine-Tuning LLMs for Anesthesiology via Compositional Data Generation
Data scarcity presents a significant challenge in developing medical large language models (LLMs), especially in specialized fields like anesthesiology. Leveraging advanced LLMs like ChatGPT to synthesize high-quality Q&A data is a promising solution to address the data scarcity issue. However, limited research exists on methods to produce such datasets specifically for anesthesiology, despite their importance for training robust LLMs. In this paper, we explore a compositional approach to generating anesthesia data for fine-tuning medical LLMs, and we refer to the proposed compositional data generation method as CDGen. Specifically, we first introduce a compositional augmentation strategy to enhance the diversity and reliability of collected medical records. We then employ a self-talk approach to generate anesthesia Q&A data from augmented medical records. We conduct extensive experiments to demonstrate the effectiveness of our approach using common metrics and GPT-4 evaluation, and experimental results demonstrate that the proposed data generation approach can synthesize high-quality Q&A data and effectively improve the performance of LLMs in Q&A tasks related to anesthesia.
- 10 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47 (7) , 5445-5459
Learning to Explore Sample Relationships
Despite the great success achieved, deep learning technologies usually suffer from data scarcity issues in real-world applications, where existing methods mainly explore sample relationships in a vanilla way from the perspectives of either the input or the loss function. In this paper, we propose a batch transformer module, BatchFormerV1, to equip deep neural networks themselves with the abilities to explore sample relationships in a learnable way. Basically, the proposed method enables data collaboration, e.g., head-class samples will also contribute to the learning of tail classes. Considering that exploring instance-level relationships has very limited impacts on dense prediction, we generalize and refer to the proposed module as BatchFormerV2, which further enables exploring sample relationships for pixel-/patch-level dense representations. In addition, to address the train-test inconsistency where a mini-batch of data samples are neither necessary nor desirable during inference, we also devise a two-stream training pipeline, i.e., a shared model is first jointly optimized with and without BatchFormerV2 which is then removed during testing. The proposed module is plug-and-play without requiring any extra inference cost. Lastly, we evaluate the proposed method on over ten popular datasets, including 1) different data scarcity settings such as long-tailed recognition, zero-shot learning, domain generalization, and contrastive learning; and 2) different visual recognition tasks ranging from image classification to object detection and panoptic segmentation.
- 10 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47 (7) , 5445-5459
Learning to Explore Sample Relationships
Despite the great success achieved, deep learning technologies usually suffer from data scarcity issues in real-world applications, where existing methods mainly explore sample relationships in a vanilla way from the perspectives of either the input or the loss function. In this paper, we propose a batch transformer module, BatchFormerV1, to equip deep neural networks themselves with the abilities to explore sample relationships in a learnable way. Basically, the proposed method enables data collaboration, e.g., head-class samples will also contribute to the learning of tail classes. Considering that exploring instance-level relationships has very limited impacts on dense prediction, we generalize and refer to the proposed module as BatchFormerV2, which further enables exploring sample relationships for pixel-/patch-level dense representations. In addition, to address the train-test inconsistency where a mini-batch of data samples are neither necessary nor desirable during inference, we also devise a two-stream training pipeline, i.e., a shared model is first jointly optimized with and without BatchFormerV2 which is then removed during testing. The proposed module is plug-and-play without requiring any extra inference cost. Lastly, we evaluate the proposed method on over ten popular datasets, including 1) different data scarcity settings such as long-tailed recognition, zero-shot learning, domain generalization, and contrastive learning; and 2) different visual recognition tasks ranging from image classification to object detection and panoptic segmentation.
- 14 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 9 (4) , 3057-3071
Benchmarking Medical LLMs on Anesthesiology: A Comprehensive Dataset in Chinese
With the recent success of large language models (LLMs), interest in developing them for medical domains has increased. However, due to the lack of benchmark datasets, evaluating the capabilities of medical LLMs remains challenging, particularly in highly specialized fields such as anesthesiology. To address this gap, we introduce a comprehensive anesthesiology benchmark dataset in Chinese, known as the Chinese Anesthesiology Benchmark (CAB). This benchmark facilitates the evaluation of medical LLMs for anesthesiology across three crucial dimensions: knowledge, application, and safety. Specifically, the CAB provides more than 8 k questions collected from examinations and books for knowledge-level evaluation; more than 2 k questions collected from online anesthesia consultations and hospitals for application-level evaluation; and 136 tests from seven anesthesia medical care scenarios for safety-level evaluation. With the proposed CAB dataset, we conducted a thorough evaluation of six medical LLMs, such as Bianque-2 and HuatuoGPT-13B, and eleven general LLMs, such as Qwen-7B-Chat and GPT-4. The evaluation results revealed that there are still clear gaps in the capacities of medical LLMs for anesthesiology compared with those of medical students in the field of anesthesia. We hope that the proposed CAB dataset can facilitate the development of medical LLMs for anesthesiology.
- 14 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 9 (4) , 3057-3071
Benchmarking Medical LLMs on Anesthesiology: A Comprehensive Dataset in Chinese
With the recent success of large language models (LLMs), interest in developing them for medical domains has increased. However, due to the lack of benchmark datasets, evaluating the capabilities of medical LLMs remains challenging, particularly in highly specialized fields such as anesthesiology. To address this gap, we introduce a comprehensive anesthesiology benchmark dataset in Chinese, known as the Chinese Anesthesiology Benchmark (CAB). This benchmark facilitates the evaluation of medical LLMs for anesthesiology across three crucial dimensions: knowledge, application, and safety. Specifically, the CAB provides more than 8 k questions collected from examinations and books for knowledge-level evaluation; more than 2 k questions collected from online anesthesia consultations and hospitals for application-level evaluation; and 136 tests from seven anesthesia medical care scenarios for safety-level evaluation. With the proposed CAB dataset, we conducted a thorough evaluation of six medical LLMs, such as Bianque-2 and HuatuoGPT-13B, and eleven general LLMs, such as Qwen-7B-Chat and GPT-4. The evaluation results revealed that there are still clear gaps in the capacities of medical LLMs for anesthesiology compared with those of medical students in the field of anesthesia. We hope that the proposed CAB dataset can facilitate the development of medical LLMs for anesthesiology.
- 21 November 2024
- Published by Springer Nature
- p. 251-272
Visual Relationship Transformation
- 4 November 2024
- Published by Springer Nature
- p. 139-151
Self-Distillation via Intra-Class Compactness