Papers14:00
Chinese contentDataset Distillation via Drifting：不用分类器，直接在潜空间逼近真实分布搞数据集蒸馏的人可以看看这篇：不靠分类器优化，直接在潜空间里把合成样本往真实分布上漂移，分布覆盖比老方法好。还顺带送一个模型可视化解释的用法，图像之外的模态也能跑。
#分布匹配
2 stories · 1 in 7 days · 1 in 30 days
Stories, products and papers tagged "分布匹配", newest first.
Nov 13
Sep 21
Beyond Gaussian Worlds: Latent Geometry Matters for JEPAs
This matters to AI researchers and developers because it challenges a fundamental assumption in JEPA models and suggests that choosing the right latent geometry could lead to more accurate and stable representations.