A Theory of Contrastive Learning with Natural Images
- 类型:arxiv
- 标识:2607.07470
- 链接:https://arxiv.org/abs/2607.07470
- 主分类:multimodal
- 形态:method
- 被引:0
- 被引来源:Semantic Scholar + OpenAlex
- S2被引:0
- OpenAlex被引:0
- 影响力被引:0
- TLDR:Analytically computing the optimal representation in terms of a contrastive loss for a range of basic augmentations and any image dataset with stationary statistics shows that for certain augmentations the optimum can be attained by a CNN whose first layer filters are sinusoids.
- OpenAlex ID:W7167835395
- OpenAlex DOI:10.48550/arxiv.2607.07470
- DOI:10.48550/arxiv.2607.07470
- DOI来源:OpenAlex
- 开放获取:green
- 开放获取链接:https://doi.org/10.48550/arxiv.2607.07470
- OpenAlex更新:2026-07-19
- 待LLM分类:否
- 标题中文:自然图像对比学习的一种理论
- TLDR中文:针对一系列基本增广与任意具有平稳统计量的图像数据集,以解析方式根据对比损失计算最优表示,结果表明对于某些增广,最优解可由第一层滤波器为正弦函数的 CNN 实现。
- 成熟度:research
- 场景:contrastive learning、representation learning、theory
- 来源文件:
- /inbox/tom/_candidates/2026-07-15-agent-rag-longcontext-candidates.json
- [S2 enrich]
- [OpenAlex backfill]