arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.04875cs.LGcs.CLstat.ML
收藏

关于可证明的长度与组合泛化

On Provable Length and Compositional Generalization

FAIR (Meta)(FAIR(Meta));EPFL and Mila-Quebec AI Institute(EPFL和Mila-Quebec人工智能研究所)

Kartik Ahuja, Amin Mansouri

AI总结:

本文首次为Deep Sets、Transformers等常见序列到序列模型提供了长度泛化与组合泛化的可证明保证,证明在训练分布足够多样时其有限容量版本可实现双重泛化,并发现思维链监督能促使更高容量模型实现长度泛化。

摘要:

序列到序列模型的分布外泛化能力可以从两种关键泛化形式的视角进行研究:长度泛化——即泛化到比训练期间所见序列更长的序列的能力,以及组合泛化——即泛化到训练期间未见的 token 组合的能力。在本研究中,我们首次为常见的序列到序列模型——Deep Sets、Transformers、状态空间模型和循环神经网络——在通过最小化预测误差进行训练时,提供了关于长度泛化和组合泛化的可证明保证。我们证明,只要训练分布具有足够的多样性,这些不同架构的“有限容量”版本就能同时实现长度泛化和组合泛化。在第一部分中,我们研究了不同架构的结构化有限容量变体,并在对训练分布的多样性要求有限的条件下,得出了泛化保证。在第二部分中,我们研究了结构假设较少的有限容量变体,同样得出了泛化保证,但对训练分布的多样性要求更高。此外,我们还表明,思维链监督能够在我们所研究的不同架构的更高容量对应模型中实现长度泛化。

英文摘要:

Out-of-distribution generalization capabilities of sequence-to-sequence models can be studied from the lens of two crucial forms of generalization: length generalization -- the ability to generalize to longer sequences than ones seen during training, and compositional generalization: the ability to generalize to token combinations not seen during training. In this work, we provide first provable guarantees on length and compositional generalization for common sequence-to-sequence models -- deep sets, transformers, state space models, and recurrent neural nets -- trained to minimize the prediction error. We show that \emph{limited capacity} versions of these different architectures achieve both length and compositional generalization provided the training distribution is sufficiently diverse. In the first part, we study structured limited capacity variants of different architectures and arrive at the generalization guarantees with limited diversity requirements on the training distribution. In the second part, we study limited capacity variants with less structural assumptions and arrive at generalization guarantees but with more diversity requirements on the training distribution. Further, we also show that chain-of-thought supervision enables length generalization in higher capacity counterparts of the different architectures we study.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑