按年被提到
20242024 · 12 次2026
视觉、语音与图像生成›视觉语言与自监督表征·arXiv:2404.04125 · cs.CV

No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance

没有指数级数据就没有「零样本」:预训练概念频率决定多模态模型表现
Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh, Yash Sharma, Philip H. S. Torr, Adel Bibi, Samuel Albanie, Matthias Bethge · 2024

多模态模型零样本表现随概念频率呈对数线性增长

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 11 人首次 2024-04-08 · 最近 2026-02-09
11不同的人
6活跃月份
13讨论串
22被引(OpenAlex)

论文在 34 个模型和 5 个标准预训练数据集上研究下游概念在预训练数据中出现的频率如何影响多模态模型表现,发现所谓零样本泛化其实遵循样本效率很低的对数线性趋势:性能线性提高需要指数级更多的数据。该趋势在控制样本级相似度、以及在纯合成分布上测试时依然成立;作者还据此构建长尾测试集 Let it Wag!,各模型在上面普遍表现不佳。

HN 为什么提它

评论者用它说明多模态和生成模型的进步需要指数级增长的数据而非算力,常作为「LLM 能力增长可能趋于平缓」的论据,也有人用它说明生成模型尚未真正做到零样本。

实证研究难度 3 / 5约 50 分钟经得起时间
读之前最好知道
1CLIP 与多模态预训练
2零样本(zero-shot)评测
3缩放定律(scaling law)
4图像生成模型基础(Stable Diffusion)

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
推荐sweezyjeezy · 2024-07

“Personally, I rate this paper”

个人看好这篇论文:零样本表现取决于预训练中概念出现的频率
Vision language models are blind原评论 ↗
引用baryphonic · 2024-05

“researchers concluded that multi-modal models require exponentially more data for each extra "mode" being added to the model.”

引用结论:多模态模型每多一种模态,所需数据呈指数增长
Geoffrey Hinton "godfather of AI" says that AI will surpass human intelligence原评论 ↗
引用intended · 2024-06

“It appears that LLM capability growth is likely to plateau, although this is to be confirmed. No "Zero-Shot" Without Exponential Data”

引用它说明 LLM 能力增长很可能趋于平台期,但有待证实
Mathematical Optimization for Cargo Ships原评论 ↗
引用thesz · 2024-06

“This is not about compute, but about data.”

借它指出问题不在算力而在数据:零样本泛化需要指数级数据
OpenAI and Apple Announce Partnership原评论 ↗
引用smartician · 2024-06

“Research seems to suggest we need exponential training data volume increases to see meaningful performance gains: No "Zero-Shot" Without Exponential Data”

引用研究:要获得有意义的性能提升,训练数据量需指数级增长
Ask HN: When will we hit a limit on LLM performance?原评论 ↗
引用sweezyjeezy · 2024-06

“Some research to the contrary [1] - tldr is that they didn't find evidence that generative models really do zero shot well at all yet”

指出有相反的研究:没找到证据表明生成模型真能零样本处理未见概念
Testing Generative AI for Circuit Board Design原评论 ↗
看全部 9 条 →
引用welkinSL · 2024-07

“shows that these pre-trained few-shots learners exhibit a log-linear relationship with”

引用它说明这些预训练小样本学习者的表现与概念频率呈对数线性关系
Ask HN: Are there videos of people using AI for non-trivial coding?原评论 ↗
引用pton_xd · 2024-11

“EDIT: here's the paper No "Zero-Shot" Without Exponential Data”

补上论文:准确率随数据指数增长而线性提升
OpenAI, Google and Anthropic are struggling to build more advanced AI原评论 ↗
引用thesz · 2026-02

“These diminishing returns are exponentially diminishing returns [1].”

引用它说明收益递减本身还是指数级的递减
(AI) Slop Terrifies Me原评论 ↗
被提交为 story · 4 次