arXivDaily arXiv每日学术速递 周一至周五更新

期刊&会议

NeurIPS

Conference on Neural Information Processing Systems · 会议 · Machine Learning

2026-06-03 至 2026-06-03 共收录 12
2606.02765 2026-06-03 cs.LG cs.AI

Representational Capacity: Geometric Limits on Feature Representation in Transformer Language Models

表示能力:Transformer语言模型中特征表示的几何限制

Alexander Guha

机构 * Arizona State University(亚利桑那州立大学)

AI总结 基于线性表示和叠加假设,通过嵌入矩阵的余弦相似度分布估计模型可支持的近正交方向数量,推导出容量公式,并发现容量对偏差ε指数敏感。

Comments 22 pages, 10 figures. Submitted to NeurIPS 2026. This is a condensed version of thesis: https://hdl.handle.net/2286/R.2.N.204857

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.02663 2026-06-03 cs.LG cs.AI

AdaWeather: Adaptively Mixing Probabilistic Weather Forecasts with Logarithmic Regret

AdaWeather: 自适应混合概率天气预报与对数遗憾

Saptarishi Dhanuka, Sarvesh Iyer, Manmeet Singh, Mihir More, Rushil Gupta, Dhruman Gupta, Parthasarathi Mukhopadhyay, Sandeep Juneja

机构 * Ashoka University(阿什oka大学) Western Kentucky University(西方肯塔基大学)

AI总结 提出 AdaWeather 自适应框架,通过结合机器学习和专家混合方法融合多个概率天气预报,实现对数遗憾界,并在温度预测上取得改进。

Comments 36 pages, 16 figures. Submitted to arXiv. Forecast aggregation for probabilistic weather prediction using offline supervised learning and online prediction with expert advice. Includes theoretical regret guarantees and empirical evaluation on temperature forecasting. Submitted to NeurIPS 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.02779 2026-06-03 cs.LG

Optimal Rates for Generalization of Gradient Descent for Deep ReLU Classification

深度ReLU分类中梯度下降泛化的最优速率

Yuanfan Li, Yunwen Lei, Zheng-Chu Guo, Yiming Ying

机构 * School of Mathematical Sciences, Zhejiang University(浙江大学数学科学学院) Department of Mathematics, The University of Hong Kong(香港大学数学系) School of mathematics and statistics, University of Sydney(悉尼大学数学与统计学学院)

AI总结 针对深度ReLU网络,通过权衡优化与泛化误差,在NTK可分离假设下证明了梯度下降的泛化误差率为~O(L^6/(nγ^2)),与SVM最优率仅差深度相关因子,关键技术是控制参考模型附近的激活模式以得到更紧的Rademacher复杂度界。

Comments Published in NeurIPS 2025

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.01698 2026-06-03 cs.IR cs.MM cs.SD eess.AS

TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling

TalkPlay-Tools: 基于大语言模型工具调用的对话式音乐推荐

Seungheon Doh, Keunwoo Choi, Juhan Nam

机构 * KAIST(韩国科学技术院) talkpl.ai

AI总结 提出一种基于LLM工具调用的统一检索-重排序流水线,通过布尔过滤、稀疏检索、稠密检索和生成式检索的组合,实现端到端的对话式音乐推荐。

Comments Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.03316 2026-06-03 cs.CV cs.AI cs.LG

The View From Space: Navigating Instrumentation Differences with EOFMs

从太空视角:利用EOFMs导航仪器差异

Ryan P. Demilt, Nicholas LaHaye, Karis Tenneson

机构 * Spatial Informatics Group(空间信息组)

AI总结 本研究通过分析地球观测基础模型(EOFMs)对传感器架构的敏感性,揭示了当前模型设计的缺陷,并为模型开发者、用户和遥感科学社区指明了前进方向。

Journal ref https://neurips.cc/virtual/2025/loc/san-diego/122891

详情

展开后加载摘要…

URL PDF HTML 收藏
2502.08006 2026-06-03 cs.LG cs.AI stat.ML

Greed is Good: A Unifying Perspective on Guided Generation

贪婪即美德:引导生成的统一视角

Zander W. Blasingame, Chen Liu

AI总结 本文通过将后验引导视为端到端引导的贪婪策略,统一了两种梯度引导方法,并提出了在计算与精度之间权衡的插值方法,在逆图像问题和分子生成任务上验证了有效性。

Comments Accepted at NeurIPS 2025

详情

展开后加载摘要…

URL PDF HTML 收藏
1207.2940 2026-06-03 stat.ML cs.LG cs.SY eess.SY

Expectation Propagation in Gaussian Process Dynamical Systems: Extended Version

高斯过程动态系统中的期望传播:扩展版

Marc Peter Deisenroth, Shakir Mohamed

AI总结 本文提出基于期望传播的消息传递算法用于高斯过程动态系统的近似推理,通过前向后向平滑迭代获得更精确的潜在结构后验分布,提升预测性能,并统一了现有GPDS平滑器。

Journal ref Advances in Neural Information Processing Systems 25 (NIPS), pp. 2609-2617, 2012

详情

展开后加载摘要…

URL PDF HTML 收藏
1102.5597 2026-06-03 math.NA cs.LG cs.NA

Fast and Faster: A Comparison of Two Streamed Matrix Decomposition Algorithms

快速与更快:两种流式矩阵分解算法的比较

Radim Řeh{ů}řek

AI总结 本文比较了单遍分布式算法和两遍流式随机算法在恒定内存下处理大规模矩阵分解的性能与精度,以英文维基百科为数据集进行潜在语义分析实验。

Journal ref NIPS Workshop on Low-Rank Methods for Large-Scale Machine Learning, 2010

详情

展开后加载摘要…

URL PDF HTML 收藏
1009.5055 2026-06-03 math.OC cs.NA cs.SY eess.SY math.NA

The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices

增广拉格朗日乘子法用于精确恢复被破坏的低秩矩阵

Zhouchen Lin, Minming Chen, Yi Ma

AI总结 针对鲁棒主成分分析问题,提出基于增广拉格朗日乘子法的可扩展快速算法,实现被任意破坏的低秩矩阵的精确恢复,并证明其最优性和收敛速率。

Comments Please cite "Zhouchen Lin, Risheng Liu, and Zhixun Su, Linearized Alternating Direction Method with Adaptive Penalty for Low Rank Representation, NIPS 2011." (available at arXiv:1109.0367) instead for a more general method called Linearized Alternating Direction Method This manuscript first appeared as University of Illinois at Urbana-Champaign technical report #UILU-ENG-09-2215 in October 2009 Zhouchen Lin, Risheng Liu, and Zhixun Su, Linearized Alternating Direction Method with Adaptive Penalty for Low Rank Representation, NIPS 2011. (available at http://arxiv.org/abs/1109.0367)

详情

展开后加载摘要…

URL PDF HTML 收藏
1206.1156 2026-06-03 math.OC cs.NA math.NA

A quasi-Newton proximal splitting method

拟牛顿近端分裂方法

Stephen Becker, M. Jalal Fadili

AI总结 本文推导了凸分析中关于特定缩放范数下邻近算子计算的新结果,并利用对偶问题的分段线性特性实现了高效邻近计算,进而提出一种优雅的拟牛顿加速方法,在信号处理、稀疏恢复和机器学习等领域优于现有技术。

Journal ref NIPS (2012), pp. 2627--2635

详情

展开后加载摘要…

URL PDF HTML 收藏
1010.3043 2026-06-03 math.NA cs.NA stat.CO stat.ME

Making Tensor Factorizations Robust to Non-Gaussian Noise

使张量分解对非高斯噪声鲁棒

Eric C. Chi, Tamara G. Kolda

AI总结 针对CP张量分解在非高斯噪声下敏感的问题,提出基于1-范数的损失函数,并设计交替最小化-最大化算法进行拟合。

Comments Contributed presentation at the NIPS Workshop on Tensors, Kernels, and Machine Learning, Whistler, BC, Canada, December 10, 2010

详情

展开后加载摘要…

URL PDF HTML 收藏
math/0506090 2026-06-03 math.NA cs.NA math.PR

Diffusion Maps, Spectral Clustering and Eigenfunctions of Fokker-Planck operators

扩散映射、谱聚类与Fokker-Planck算子的特征函数

Boaz Nadler, Stephane Lafon, Ronald R. Coifman, Ioannis G. Kevrekidis

AI总结 本文提出基于扩散过程的谱聚类和降维算法的概率解释,通过扩散距离和Fokker-Planck算子特征函数,为使用归一化图拉普拉斯特征向量的方法提供数学依据。

Comments submitted to NIPS 2005

详情

展开后加载摘要…

URL PDF HTML 收藏