Visualization of attention matrices on the METR-LA dataset.

<p>(a) Raw attention matrix before sparsification, showing dense and noisy correlations across nodes. (b) Sparse attention matrix after the adaptive Top-<i>k</i> operation, where high-weight connections capture both local and long-range dependencies beyond the physical topology.<...

وصف كامل

محفوظ في:

التفاصيل البيبلوغرافية
المؤلف الرئيسي:	Zhouyuan Zhang (22615290) (author)
مؤلفون آخرون:	Xin Wang (91924) (author), Xu Tan (193786) (author), Jiatian Pi (22615293) (author)
منشور في:	2025
الموضوعات:	Medicine Ecology Infectious Diseases Environmental Sciences not elsewhere classified Astronomical and Space Sciences not elsewhere classified Biological Sciences not elsewhere classified Mathematical Sciences not elsewhere classified Information Systems not elsewhere classified world datasets demonstrate weak directional interpretability reduce spatial complexity physically grounded paradigm often decouple spatial intelligent transportation systems gated dilated convolutions temporal convolutional networks temporal convolutional block adaptive spatial module dynamically evolving spatio combines sparse attention diffusion graph convolutions based graph convolution temporal coupling mechanism existing graph wise coupling temporal patterns temporal learning temporal dependencies adaptive sparsity unified framework tightly integrated six real road topology results indicate redundant computation preserve parallelism informative neighbors improved performance hop propagation computational efficiency comprehensive experiments based approaches alternating spatio
الوسوم:	إضافة وسم لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!

Visualization of attention matrices on the METR-LA dataset.

مواد مشابهة