通用视觉自监督研究报
A Daily Digest of Visual Self-Supervised Learning
2026-05-31 图像表征 · VFM · JEPA · 视频预训练
data scaling · 扫读 · 2026-05-31

How Much Is a Dataset：它和通用视觉自监督的关系在于：用 ImageNet-scale 子集验证数据多样性/谱函数价值，作为视觉预训练数据筛选背景扫读

中低相关；详见方法、贡献和实验边界。

How Much Is a Dataset Worth? Scaling Laws, the Vendi Score, and Matrix Spectral Functions　arXiv + data scaling　原文链接

编号2605.29448 优先级扫读 类别data scaling 会议arXiv + data scaling 方法用 ImageNet-scale 子集验证数据多样性/谱函数价值，作为视觉预训练数据筛选背景扫读 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：用 ImageNet-scale 子集验证数据多样性/谱函数价值，作为视觉预训练数据筛选背景扫读。 中低相关；详见方法、贡献和实验边界。

(c) Side-by-side comparison of ImageNet-1K test performance for the highly valued green (l
(c) Side-by-side comparison of ImageNet-1K test performance for the highly valued green (l(c) Side-by-side comparison of ImageNet-1K test performance for the highly valued green (left) points created using classbalanced partition matroid constrained submodular maximization on the facility location, red (middle) created using stratified random sampling to achieve class balance, and blue (right) class-balanced heuristic partition matroid constrained submodular minimization again on the same facility location function. These three plots show the same information as in Figure 27a but is expanded to three side-by-side plots for clarity. Figure 27: The results (top row) are the same as those presented in Figure 1 and the bottom row is expanded out to three side-by-side plots to show more details more clearly.这张图/表用于判断 How Much Is a Dataset 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。
(b) Low-value (low facility location score) subset (blue X in Figure 1 just for accordion)
(b) Low-value (low facility location score) subset (blue X in Figure 1 just for accordion)(b) Low-value (low facility location score) subset (blue X in Figure 1 just for accordion) Figure 20: Accordion. Images in the high-value (high facility location score) subset (above left 8 × 8 grid, and green X in Figure 1) are canonical examples in which an accordion is clearly the, or at least a, central subject. In contrast, the low-value (low facility location score) subset (above right 8 × 8 grid, and blue X in Figure 1) contains many much more atypical images with confounding objects. For example the image in row 6 column 1, contains several other instruments and the accordion is hardly visible. In row 5, column 1, we see an entire band which includes guitar and people, and a better single label for that image might be “band” or “musical group” (which are not ImageNet-1K classes). Other images, such as the ones shown in row 8 column 7, row 8 column 8, and row 1 column 5, appear to not contain anything resembling the totality of the accordion “concept” at all and may simply be instances of label noise, or might be extreme closeups of the buttons in a button accordion, which are not visually similar to the more canonical or typical examples of what defines an accordion visually.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 How Much Is a Dataset 的方法或实验，请结合正文精读段落一起看。

核心问题

它和通用视觉自监督的关系在于：用 ImageNet-scale 子集验证数据多样性/谱函数价值，作为视觉预训练数据筛选背景扫读。

方法拆解

用 ImageNet-scale 子集验证数据多样性/谱函数价值，作为视觉预训练数据筛选背景扫读

主要贡献

中低相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。