-
2026混合精度训练落地教程：解决精度丢失，提速3倍无损模型效果
混合精度训练是通过FP32、FP16、FP8等不同位宽浮点格式搭配使用，在不损失模型精度的前提下，大幅降低深度学习训练显存占用、提升训练与推理速…
-
把「Salem Prize now accepting nominations for 2025」写成可执行的阅读路径
The Salem prize was established in 1968 and named in honor of Raphaël Sa…
-
「Rough numbers between consecutive primes」：数学直觉如何落到可检验的步骤
First things first: due to an abrupt suspension of NSF funding to my hom…
-
NeRF 原理落地解析：体渲染积分、位置编码与分层采样实操避坑
NeRF（神经辐射场）能够依靠一组多视角 2D 图像重建完整 3D 场景，其整套能力的根基是可微体渲染：通过对相机射向每个像素的光线做空间积分，…
-
「SLMath announces new research programs」：数学直觉如何落到可检验的步骤
The Simons-Laufer Mathematical Sciences institute, or SLMath (formerly t…
-
读懂「A crowdsourced project to link up erdosproblems.」：先把定义、反例和适用范围钉死
Thomas Bloom’s erdosproblems.com site hosts nearly a thousand questions …
-
「Smooth numbers and max-entropy」笔记：问题从哪来、证明卡在哪
Given a threshold , a -smooth number (or -friable number) is a natural n…
-
神经网络剪枝：泰勒展开估算参数重要度落地路径与避坑要点
神经网络剪枝中，直接通过 “参数置零前后损失差值平方” 计算参数重要度，理论上最贴合真实影响，但对大模型而言计算成本高到几乎不可用。采用一阶泰勒…
-
「Mathematical exploration and discovery at scale」：数学直觉如何落到可检验的步骤
Bogdan Georgiev, Javier Gómez-Serrano, Adam Zsolt Wagner, and I have upl…
-
「New Nikodym set constructions over finite fields」：数学直觉如何落到可检验的步骤
I have uploaded to the arXiv my paper “New Nikodym set constructions ove…