Bingjie Yan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Bingjie Yan has a total of 49 co-authors. Here are the top co-authors by number of shared publications:
- Yiqiang Chen
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
4 shared publications - Chenlong Gao
Institute of Computing Technology, Chinese Academy of Sciences, University of Chinese Academy of Sciences, Beijing, China
3 shared publications - Qian Chen
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
3 shared publications - Danli Shi
School of Optometry, The Hong Kong Polytechnic University, Hong Kong, China
2 shared publications - Teng Zhang
Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences & University of Chinese Academy of Sciences, Beijing, China
2 shared publications - Wuliang Huang
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
2 shared publications
Latest Publications
- Open Access
- 7 August 2026
- Published by Springer Nature in Research Square
UWF-FM: A Foundation Model for Comprehensive Ultra-Widefield Retinal Assessment
Ultra-widefield (UWF) retinal imaging enables comprehensive visualization of retinal pathology but introduces substantial heterogeneity across diseases, clinical tasks, and imaging devices, limiting the applicability of existing retinal foundation models. Here we present UWF-FM, a foundation model pretrained using self-supervised learning on 632,627 unlabeled UWF retinal images. We evaluated UWF-FM on 27,641 labeled image–task pairs and 10,686 external images spanning multiple levels of retinal assessment, including characterization of 28 retinal findings involving both posterior-pole and peripheral pathology, diabetic retinopathy grading, and diagnosis of 12 retinal diseases and tumors. Across diverse tasks, patient cohorts, and imaging systems, UWF-FM consistently outperformed foundation models pretrained on non-UWF data. Performance gains were maintained across external datasets, clinically relevant operating thresholds, and cross-vendor evaluation (Optos to Zeiss) without additional adaptation. In a prospectively registered randomized crossover multi-reader study (ClinicalTrials.gov: NCT07651943; 600 UWF images; four ophthalmologists), AI assistance increased clinician sensitivity with minimal loss of specificity. During retrospective and prospective silent deployment involving 37,665 consecutive examinations, UWF-FM achieved a 100% processing success rate, a median inference time of approximately 1.1 seconds per examination, and no workflow disruption. These findings establish UWF-FM as the first foundation model for UWF retinal imaging and demonstrate that UWF-specific foundation modeling provides a scalable framework for robust, generalizable, and clinically deployable retinal decision-support systems.
- 7 August 2026
- Published by Springer Nature in Research Square
Open AccessUWF-FM: A Foundation Model for Comprehensive Ultra-Widefield Retinal Assessment
Ultra-widefield (UWF) retinal imaging enables comprehensive visualization of retinal pathology but introduces substantial heterogeneity across diseases, clinical tasks, and imaging devices, limiting the applicability of existing retinal foundation models. Here we present UWF-FM, a foundation model pretrained using self-supervised learning on 632,627 unlabeled UWF retinal images. We evaluated UWF-FM on 27,641 labeled image–task pairs and 10,686 external images spanning multiple levels of retinal assessment, including characterization of 28 retinal findings involving both posterior-pole and peripheral pathology, diabetic retinopathy grading, and diagnosis of 12 retinal diseases and tumors. Across diverse tasks, patient cohorts, and imaging systems, UWF-FM consistently outperformed foundation models pretrained on non-UWF data. Performance gains were maintained across external datasets, clinically relevant operating thresholds, and cross-vendor evaluation (Optos to Zeiss) without additional adaptation. In a prospectively registered randomized crossover multi-reader study (ClinicalTrials.gov: NCT07651943; 600 UWF images; four ophthalmologists), AI assistance increased clinician sensitivity with minimal loss of specificity. During retrospective and prospective silent deployment involving 37,665 consecutive examinations, UWF-FM achieved a 100% processing success rate, a median inference time of approximately 1.1 seconds per examination, and no workflow disruption. These findings establish UWF-FM as the first foundation model for UWF retinal imaging and demonstrate that UWF-specific foundation modeling provides a scalable framework for robust, generalizable, and clinically deployable retinal decision-support systems.
- Top Cited PapersOpen Access
- 24 November 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Intelligent Systems and Technology
- Vol. 16 (6) , 1-27
Survey on Knowledge Distillation for Large Language Models: Methods, Evaluation, and Application
Large Language Models (LLMs) have showcased exceptional capabilities in various domains, attracting significant interest from both academia and industry. Despite their impressive performance, the substantial size and computational demands of LLMs pose considerable challenges for practical deployment, particularly in environments with limited resources. The endeavor to compress language models while maintaining their accuracy has become a focal point of research. Among the various methods, knowledge distillation has emerged as an effective technique to enhance inference speed without greatly compromising performance. This article presents a thorough survey from three aspects: method, evaluation, and application, exploring knowledge distillation techniques tailored specifically for LLMs. Specifically, we divide the methods into white-box KD and black-box KD to better illustrate their differences. Furthermore, we also explored the evaluation tasks and distillation effects between different distillation methods and proposed directions for future research. Through in-depth understanding of the latest advancements and practical applications, this survey provides valuable resources for researchers, paving the way for sustained progress in this field.
- 24 November 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Intelligent Systems and Technology
- Vol. 16 (6) , 1-27
Top Cited PapersOpen AccessSurvey on Knowledge Distillation for Large Language Models: Methods, Evaluation, and Application
Large Language Models (LLMs) have showcased exceptional capabilities in various domains, attracting significant interest from both academia and industry. Despite their impressive performance, the substantial size and computational demands of LLMs pose considerable challenges for practical deployment, particularly in environments with limited resources. The endeavor to compress language models while maintaining their accuracy has become a focal point of research. Among the various methods, knowledge distillation has emerged as an effective technique to enhance inference speed without greatly compromising performance. This article presents a thorough survey from three aspects: method, evaluation, and application, exploring knowledge distillation techniques tailored specifically for LLMs. Specifically, we divide the methods into white-box KD and black-box KD to better illustrate their differences. Furthermore, we also explored the evaluation tasks and distillation effects between different distillation methods and proposed directions for future research. Through in-depth understanding of the latest advancements and practical applications, this survey provides valuable resources for researchers, paving the way for sustained progress in this field.
- 27 August 2025
- Published by BMJ in British Journal of Ophthalmology
- Vol. 110 (1) , 1-7
From visual question answering to intelligent AI agents in ophthalmology
Ophthalmic practice involves the integration of diverse clinical data and interactive decision-making, posing challenges for traditional artificial intelligence (AI) systems. Visual question answering (VQA) addresses this by combining computer vision and natural language processing to interpret medical images through user-driven queries. Evolving from VQA, multimodal AI agents enable continuous dialogue, tool use and context-aware clinical decision support. This review explores recent developments in ophthalmic conversational AI, spanning theoretical advances and practical implementations. We highlight the transformative role of large language models (LLMs) in improving reasoning, adaptability and task execution. However, key obstacles remain, including limited multimodal datasets, absence of standardised evaluation protocols, and challenges in clinical integration. We outline these limitations and propose future research directions to support the development of robust, LLM-driven AI systems. Realising their full potential will depend on close collaboration between AI researchers and the ophthalmic community.
- 27 August 2025
- Published by BMJ in British Journal of Ophthalmology
- Vol. 110 (1) , 1-7
From visual question answering to intelligent AI agents in ophthalmology
Ophthalmic practice involves the integration of diverse clinical data and interactive decision-making, posing challenges for traditional artificial intelligence (AI) systems. Visual question answering (VQA) addresses this by combining computer vision and natural language processing to interpret medical images through user-driven queries. Evolving from VQA, multimodal AI agents enable continuous dialogue, tool use and context-aware clinical decision support. This review explores recent developments in ophthalmic conversational AI, spanning theoretical advances and practical implementations. We highlight the transformative role of large language models (LLMs) in improving reasoning, adaptability and task execution. However, key obstacles remain, including limited multimodal datasets, absence of standardised evaluation protocols, and challenges in clinical integration. We outline these limitations and propose future research directions to support the development of robust, LLM-driven AI systems. Realising their full potential will depend on close collaboration between AI researchers and the ophthalmic community.
- Open Access
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 2272-2281
Correlation-Driven Multi-Modality Graph Decomposition for Cross-Subject Emotion Recognition
Multi-modality physiological signal-based emotion recognition has attracted increasing attention as its capacity to capture human affective states comprehensively. Due to multi-modality heterogeneity and cross-subject divergence, practical applications struggle with generalizing models across individuals. Effectively addressing both issues requires mitigating the gap between multimodal signals while acquiring generalizable representations across subjects. However, existing approaches often handle these dual challenges separately, resulting in suboptimal generalization. This study introduces a novel framework, termed Correlation-Driven Multi-Modality Graph Decomposition (CMMGD). The proposed CMMGD initially captures adaptive cross-modal correlations. It connects each unimodal graph to a multimodal mixed graph. To simultaneously address the dual challenges, it incorporates a correlation-driven graph decomposition module that decomposes the mixed graph into concordant and discrepant subgraphs based on the correlations. The decomposed concordant subgraph encompasses consistently activated features across modalities and subjects during emotion elicitation, unveiling a generalizable subspace. Additionally, we design a Multi-Modality Graph Regularized Transformer (MGRT) backbone specifically tailored for multimodal physiological signals. The MGRT can alleviate the over-smoothing issue and mitigate over-reliance on any single modality. Extensive experiments demonstrate that CMMGD outperforms the state-of-the-art methods by 1.79% and 2.65% on DEAP and MAHNOB-HCI datasets, respectively, under the leave-one-subject-out cross-validation strategy.
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 2272-2281
Open AccessCorrelation-Driven Multi-Modality Graph Decomposition for Cross-Subject Emotion Recognition
Multi-modality physiological signal-based emotion recognition has attracted increasing attention as its capacity to capture human affective states comprehensively. Due to multi-modality heterogeneity and cross-subject divergence, practical applications struggle with generalizing models across individuals. Effectively addressing both issues requires mitigating the gap between multimodal signals while acquiring generalizable representations across subjects. However, existing approaches often handle these dual challenges separately, resulting in suboptimal generalization. This study introduces a novel framework, termed Correlation-Driven Multi-Modality Graph Decomposition (CMMGD). The proposed CMMGD initially captures adaptive cross-modal correlations. It connects each unimodal graph to a multimodal mixed graph. To simultaneously address the dual challenges, it incorporates a correlation-driven graph decomposition module that decomposes the mixed graph into concordant and discrepant subgraphs based on the correlations. The decomposed concordant subgraph encompasses consistently activated features across modalities and subjects during emotion elicitation, unveiling a generalizable subspace. Additionally, we design a Multi-Modality Graph Regularized Transformer (MGRT) backbone specifically tailored for multimodal physiological signals. The MGRT can alleviate the over-smoothing issue and mitigate over-reliance on any single modality. Extensive experiments demonstrate that CMMGD outperforms the state-of-the-art methods by 1.79% and 2.65% on DEAP and MAHNOB-HCI datasets, respectively, under the leave-one-subject-out cross-validation strategy.
- Open Access
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 2775-2785
Buffalo: Biomedical Vision-Language Understanding with Cross-Modal Prototype and Federated Foundation Model Collaboration
Federated learning (FL) enables collaborative learning across multiple biomedical data silos with multimodal foundation models while preserving privacy. Due to the heterogeneity in data processing and collection methodologies across diverse medical institutions and the varying medical inspections patients undergo, modal heterogeneity exists in practical scenarios, where severe modal heterogeneity may even prevent model training. With privacy considerations, data transfer cannot be permitted, restricting knowledge exchange among different clients. To trickle these issues, we propose a cross-modal prototype imputation method for visual-language understanding (Buffalo) with only a slight increase in communication cost, which can improve the performance of fine-tuning general foundation models for downstream biomedical tasks. We conducted extensive experiments on medical report generation and biomedical visual question-answering tasks. The results demonstrate that Buffalo can fully utilize data from all clients to improve model generalization compared to other modal imputation methods in three modal heterogeneity scenarios, approaching or even surpassing the performance in the ideal scenario without missing modality.
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 2775-2785
Open AccessBuffalo: Biomedical Vision-Language Understanding with Cross-Modal Prototype and Federated Foundation Model Collaboration
Federated learning (FL) enables collaborative learning across multiple biomedical data silos with multimodal foundation models while preserving privacy. Due to the heterogeneity in data processing and collection methodologies across diverse medical institutions and the varying medical inspections patients undergo, modal heterogeneity exists in practical scenarios, where severe modal heterogeneity may even prevent model training. With privacy considerations, data transfer cannot be permitted, restricting knowledge exchange among different clients. To trickle these issues, we propose a cross-modal prototype imputation method for visual-language understanding (Buffalo) with only a slight increase in communication cost, which can improve the performance of fine-tuning general foundation models for downstream biomedical tasks. We conducted extensive experiments on medical report generation and biomedical visual question-answering tasks. The results demonstrate that Buffalo can fully utilize data from all clients to improve model generalization compared to other modal imputation methods in three modal heterogeneity scenarios, approaching or even surpassing the performance in the ideal scenario without missing modality.
- Open Access
- 12 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 1072-1081
Im2col-Winograd: An Efficient and Flexible Fused-Winograd Convolution for NHWC Format on GPUs
Compared to standard convolution, Winograd algorithm has lower time complexity and can accelerate the execution of convolutional neural networks. Previous studies have utilized Winograd to implement 2D convolution on GPUs, mainly using 2D Winograd, and arranging tensors in NCHW or CHWN format instead of NHWC to make data access coalesced. Due to the higher space complexity of Winograd and limited hardware resources, these implementations are usually confined to small filters. To provide an efficient and flexible fused-Winograd convolution for NHWC format on GPUs, we propose Im2col-Winograd. This algorithm decomposes an ND convolution into a series of 1D convolutions to utilize 1D Winograd, thereby reducing space complexity and data-access discontinuity. The reduced space complexity makes Im2col-Winograd less restricted by hardware capability, enabling it to accommodate a wider range of filter shapes. Our implementations support 2-9 filter widths and do not use any workspace to store intermediate variables. According to the experiments, Im2col-Winograd achieves a speedup of 0.788 × to 2.05 × over the fastest benchmark algorithm in cuDNN; and shows similar convergence to PyTorch on Cifar10 and ILSVRC2012 datasets. Along with memory efficiency, the more generalized acceleration offered by Im2col-Winograd can be beneficial for extracting features at different convolution scales.
- 12 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 1072-1081
Open AccessIm2col-Winograd: An Efficient and Flexible Fused-Winograd Convolution for NHWC Format on GPUs
Compared to standard convolution, Winograd algorithm has lower time complexity and can accelerate the execution of convolutional neural networks. Previous studies have utilized Winograd to implement 2D convolution on GPUs, mainly using 2D Winograd, and arranging tensors in NCHW or CHWN format instead of NHWC to make data access coalesced. Due to the higher space complexity of Winograd and limited hardware resources, these implementations are usually confined to small filters. To provide an efficient and flexible fused-Winograd convolution for NHWC format on GPUs, we propose Im2col-Winograd. This algorithm decomposes an ND convolution into a series of 1D convolutions to utilize 1D Winograd, thereby reducing space complexity and data-access discontinuity. The reduced space complexity makes Im2col-Winograd less restricted by hardware capability, enabling it to accommodate a wider range of filter shapes. Our implementations support 2-9 filter widths and do not use any workspace to store intermediate variables. According to the experiments, Im2col-Winograd achieves a speedup of 0.788 × to 2.05 × over the fastest benchmark algorithm in cuDNN; and shows similar convergence to PyTorch on Cifar10 and ILSVRC2012 datasets. Along with memory efficiency, the more generalized acceleration offered by Im2col-Winograd can be beneficial for extracting features at different convolution scales.
- Open Access
FedEYE: A scalable and flexible end-to-end federated learning platform for ophthalmology