Diffusion Alignment Beyond KL: Variance Minimisation as Effective Policy Optimiser
扩散对齐超越KL:方差最小化作为有效的策略优化器
机构 * Imperial College London(帝国理工学院伦敦分校) ; Samsung R&D Institute UK(三星英国研发中心)
AI总结 本文提出方差最小化策略优化方法,通过最小化对数重要权重的方差来实现扩散对齐,超越传统KL优化,提供新的设计方向。
高校专区
扩散对齐超越KL:方差最小化作为有效的策略优化器
机构 * Imperial College London(帝国理工学院伦敦分校) ; Samsung R&D Institute UK(三星英国研发中心)
AI总结 本文提出方差最小化策略优化方法,通过最小化对数重要权重的方差来实现扩散对齐,超越传统KL优化,提供新的设计方向。
基于变压器的深度核融合
机构 * Imperial College London(帝国理工学院伦敦分校) ; University of Cambridge(剑桥大学)
AI总结 本研究提出DeepFusionKernel,通过深度融合内核减少HBM流量并提高缓存复用,实现大语言模型在长上下文推理中的性能提升。
自动化胎儿脑MRI分割与生物测量的进展:来自FeTA 2024挑战的见解
机构 * organization= Department of Radiology, Lausanne University Hospital ; University of Lausanne , city= Lausanne , country= Switzerland ; organization= CIBM Center for Biomedical Imaging , city= Lausanne , country= Switzerland ; organization= Department of Early Life Imaging, School of Biomedical Engineering \& Imaging Sciences, King’s College London , city= London , country= UK ; organization= Smart Imaging Lab, University Hospital Erlangen , city= Erlangen , country= Germany ; organization= Center for MR-Research, University Children’s Hospital Zurich, University of Zurich , city= Zurich , country= Switzerland ; organization= Neuroscience Center Zurich, University of Zurich , city= Zurich , country= Switzerland ; organization= National Heart \& Lung Institute, Imperial College London , city= London , country= UK ; organization= University of California, San Francisco ; UCSF Benioff Children’s Hospital , city= San Francisco , state= California , country= USA ; organization= Department of Quantitative Biomedicine, University of Zurich , city= Zurich , country= Switzerland ; organization= Department of Informatics, Technical University of Munich , city= Munich , country= Germany ; organization= Boston Children’s Hospital, Harvard Medical School , city= Boston , state= Massachusetts , country= USA ; organization= Neuroimaging Unit, Scientific Institute IRCCS E. Medea , city= Bosisio Parini , country= Italy ; organization= Department of Informatics, Systems ; Communication, University of Milano Bicocca , city= Milan , country= Italy ; organization= Research Institute of Computer Vision ; organization= BCN MedTech, Department of Engineering, Universitat Pompeu Fabra , city= Barcelona , country= Spain ; organization= Department of Information Engineering, University of Padova , city= Padova , country= Italy ; organization= Institut Pasteur, Université Paris Cité, CNRS UMR 3571, Decision ; organization= Inria, HeKA, PariSantéCampus , city= Paris , country= France ; organization= L. D. College of Engineering , city= Gujarat , country= India ; organization= Medical Faculty Heidelberg, Heidelberg University , addressline= Pattern Analysis ; Learning Group, Department of Radiation Oncology, Heidelberg University Hospital , city= Heidelberg , country= Germany ; organization= Canon Medical Systems (China) Co., Ltd , city= , country= China ; organization= Faculty of Electrical Engineering, University of Ljubljana , city= Ljubljana , country= Slovenia ; organization= Department of Radiology, Seoul National University Hospital , city= Seoul , country= South Korea ; organization= School of Mechanical ; Electrical Engineering, University of Electronic Science ; organization= School of Computer Science, Wuhan University , city= Wuhan , country= China ; Developmental Science Center, Boston Children’s Hospital, Harvard Medical School , city= Boston , state= Massachusetts , country= USA ; organization= Hawkes Institute, Department of Computer Science, University College London , city= London , country= UK ; organization= Laboratory for Computational Neuroimaging, Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital/Harvard Medical School , city= Charlestown , state= Massachusetts , country= USA ; organization= Department of Biomedical Imaging ; Image-guided Therapy, Computational Imaging Research Lab (CIR), Early Life Image Analysis Group, Medical University of Vienna , city= Vienna , country= Austria ; organization= University Research Priority Project Adaptive Brain Circuits in Development ; Learning (AdaBD), University of Zurich , city= Zurich , country= Switzerland ; organization= Sagol Brain Institute, Tel Aviv Sourasky Medical Center ; School of EE, Tel-Aviv University , city= Tel-Aviv , country= Israel ; organization= Department of Medical Imaging Sciences, The Faculty of Social Welfare ; Health Sciences, University of Haifa , city= Haifa , country= Israel ; Faculty of Medicine ; Sagol School of Neuroscience, Tel-Aviv University , city= Tel-Aviv , country= Israel ; organization= Department Woman-Mother-Child, CHUV , city= Lausanne , country= Switzerland ; organization= BCNatal Fetal Medicine Research Center (Hospital Clínic ; Hospital Sant Joan de Déu), Universitat de Barcelona , city= Barcelona , country= Spain ; organization= German Cancer Research Center (DKFZ) Heidelberg, Division of Medical Image Computing , city= Heidelberg , country= Germany ; organization= Helmholtz Imaging, German Cancer Research Center (DKFZ) , city= Heidelberg , country= Germany ; organization= Faculty of Mathematics ; Computer Science, Heidelberg University , city= Heidelberg , country= Germany ; organization= University of Zurich , city= Zurich , country= Switzerland ; organization= Croatian Institute for Brain Research, School of Medicine, University of Zagreb , city= Zagreb , country= Croatia ; organization= Department of Biomedical Engineering, School of Biomedical Engineering \& Imaging Sciences, King’s College , city= London , country= United Kingdom ; Musculoskeletal Radiology, Medical University of Vienna , city= Vienna , country= Austria ; organization= Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School , city= Boston , state= Massachusetts , country= USA ; organization= Department of Radiology, Boston Children’s Hospital, Harvard Medical School , city= Boston , state= Massachusetts , country= USA
AI总结 FeTA 2024挑战通过引入生物测量预测和低场MRI数据,推动了胎儿脑MRI分割与生物测量的自动化进展,揭示了拓扑差异和成像系统对分割性能的影响。
知更鸟的回声:审计大语言模型生成合成文本的隐私风险
机构 * Imperial College London(帝国理工学院伦敦分校) ; Microsoft(微软公司) ; National University of Singapore(新加坡国立大学)
AI总结 本文提出通过设计具有分布内前缀和高困惑度后缀的知更鸟,提高基于数据的MIAs的威力,以更准确评估LLM生成合成数据的隐私风险。
Comments 42nd International Conference on Machine Learning (ICML 2025)
Journal ref Proc. Mach. Learn. Res. 267 (2025) 43557-43580