On the computational cost of Stochastic Gradient Langevin Dynamics
研究了SGLD与EM方法在不同参数下的计算成本，通过理论分析和数值实验比较了两种方法的优劣，特别是在大数据集和小批量情况下的表现。
0 citationsRead paper
Selected work
Representative Papers
Towards a Connected Heterogeneous All-Medium Integrated Network (CHAIN) for Converged Connectivity Across Land, Sea, Air, and Space
论文提出CHAIN框架，通过全光网络骨干和AI驱动的跨域控制层整合海陆空天各领域通信，解决下一代连接中的跨介质问题。
0 citationsRead paper
Pretraining and Distillation Matter More Than Architecture Family for Label-Free Single-Cell Classification
研究通过控制实验比较了不同深度学习架构在无标签单细胞分类中的表现，发现预训练和知识蒸馏比架构选择更重要。
0 citationsRead paper
Beyond Aggregate VARs: A Bayesian Benchmark for HANK Models
本文提出了一种贝叶斯基准方法，用于同时建模宏观经济总量和微观层面数据的边际分布，以解决HANK模型缺乏经验基准的问题。
0 citationsRead paper
A Rig of Transformations
This work addresses the gap between high-level abstractions and efficient or compatible binary layouts in functional languages, where users lack precise control over the low-level representation of data types. It formalizes layout transformations of finite algebraic data types as isomorphisms in a commutative rig (a ring without additive inverses), leveraging rig equations to characterize data layouts and their conversions. The approach introduces partial isomorphisms to accommodate type embeddings such as bit padding. Grounded in the categorical theory of rig algebras and isomorphisms, this framework uniformly handles both total and partial representation mappings, ensuring correctness and composability of layout transformations within the type system. Consequently, it provides an expressive, verifiable, and efficient mechanism for controlling low-level data representations.
0 citationsRead paper
Recent publications
Latest Papers
On the computational cost of Stochastic Gradient Langevin Dynamics
研究了SGLD与EM方法在不同参数下的计算成本，通过理论分析和数值实验比较了两种方法的优劣，特别是在大数据集和小批量情况下的表现。
0 citationsRead paper
Towards a Connected Heterogeneous All-Medium Integrated Network (CHAIN) for Converged Connectivity Across Land, Sea, Air, and Space
论文提出CHAIN框架，通过全光网络骨干和AI驱动的跨域控制层整合海陆空天各领域通信，解决下一代连接中的跨介质问题。
0 citationsRead paper
Pretraining and Distillation Matter More Than Architecture Family for Label-Free Single-Cell Classification
研究通过控制实验比较了不同深度学习架构在无标签单细胞分类中的表现，发现预训练和知识蒸馏比架构选择更重要。
0 citationsRead paper
Beyond Aggregate VARs: A Bayesian Benchmark for HANK Models
本文提出了一种贝叶斯基准方法，用于同时建模宏观经济总量和微观层面数据的边际分布，以解决HANK模型缺乏经验基准的问题。
0 citationsRead paper
A Rig of Transformations
This work addresses the gap between high-level abstractions and efficient or compatible binary layouts in functional languages, where users lack precise control over the low-level representation of data types. It formalizes layout transformations of finite algebraic data types as isomorphisms in a commutative rig (a ring without additive inverses), leveraging rig equations to characterize data layouts and their conversions. The approach introduces partial isomorphisms to accommodate type embeddings such as bit padding. Grounded in the categorical theory of rig algebras and isomorphisms, this framework uniformly handles both total and partial representation mappings, ensuring correctness and composability of layout transformations within the type system. Consequently, it provides an expressive, verifiable, and efficient mechanism for controlling low-level data representations.
0 citationsRead paper