Representation learning of human cortical folding to reveal long lasting neurodevelopmental signatures
研究通过自监督学习框架Champollion从结构MRI中学习皮质折叠的局部表示,以揭示神经发育特征,并在多项任务上优于现有模型。
研究通过自监督学习框架Champollion从结构MRI中学习皮质折叠的局部表示,以揭示神经发育特征,并在多项任务上优于现有模型。
Current 6D pose estimation benchmarks oversimplify visual ambiguities—such as symmetry and occlusion—as global object symmetries, neglecting image-level, viewpoint-dependent visibility variations, thereby misrepresenting real-world pose uncertainty. To address this, we propose a novel image-level pose distribution evaluation paradigm: (1) the first automatic pose distribution annotation method grounded in single-image surface visibility; (2) BOP-Dist, the first pose distribution benchmark tailored to realistic images; and (3) a symmetry-aware sampling strategy coupled with a distribution-aware accuracy/recall evaluation framework. After re-annotating all BOP datasets with pose distributions, we observe substantial corrections to the performance ranking of state-of-the-art single-solution methods—revealing their rankings to be highly sensitive to annotation granularity. This work establishes a physically interpretable, reproducible, and quantitative evaluation standard for multi-solution pose estimation.
Conventional wisdom holds that reversible graph dynamics must preserve the number of nodes, precluding node creation or deletion while maintaining reversibility. Method: This paper challenges this paradigm by introducing three mutually equivalent relaxed frameworks—grounded in reversible computation, extended cellular automata, and bijective graph rewriting—that jointly enforce global bijectivity and local causality while permitting reversible node creation and destruction. Contribution/Results: We formally prove the equivalence of these frameworks, thereby establishing the first causal graph dynamics model that is both size-variable and time-reversible. This work refutes the long-standing assumption that reversibility necessitates node conservation, offering a novel paradigm for discrete spacetime modeling. It bridges a critical gap between theoretical computer science—particularly models of reversible computation—and formal approaches to quantum gravity, where dynamical causal structure and background independence are essential.
研究通过Graph-Guided Token Merging(G2TM)方法减少Vision Transformers的计算成本,证明其有效性主要取决于编码器而非解码器。
研究解决了GPU上处理大量微小线性系统的挑战,通过比较不同LU分解求解器,在NVIDIA GPU上实现了最高17.7倍的加速。
研究通过Graph-Guided Token Merging(G2TM)方法减少Vision Transformers的计算成本,证明其有效性主要取决于编码器而非解码器。
研究解决了GPU上处理大量微小线性系统的挑战,通过比较不同LU分解求解器,在NVIDIA GPU上实现了最高17.7倍的加速。
本文提出DynEoMT方法,通过在线查询增强视频分割模型以预测区域动态性,解决了无法仅从语义推断物体是否独立于相机移动的问题。
为解决大脑皮层沟回标注数据稀缺导致的过拟合问题,提出几何到语义球面迁移学习框架,利用大量无标签数据预训练模型并结合解剖线信息。
研究通过比较人类和不同模型在预测人机交互意图上的表现,发现即使是最先进的视觉-语言模型也难以匹敌人类的社会直觉。