arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.17762cs.CLcs.LG
收藏

大语言模型中的巨量激活(Massive Activations)

Massive Activations in Large Language Models

Carnegie Mellon University(卡内基梅隆大学);Meta AI Research(元人工智能研究院);Bosch Center for AI(博世人工智能中心)

Mingjie Sun, Xinlei Chen, J. Zico Kolter, Zhuang Liu

AI总结:

该研究发现大语言模型中存在数值远大于其他激活的巨量激活,证实其广泛存在、不随输入变化、充当偏置项并影响注意力分布,还拓展研究了视觉Transformer中的同类现象,提供了相关代码。

摘要:

我们在大语言模型(LLMs)中观察到一种经验现象:极少数激活值显著大于其他激活值(例如大100,000倍),我们将其称为巨量激活(massive activations)。首先,我们证明了巨量激活在各类LLMs中广泛存在,并刻画了它们的位置特征。其次,我们发现其数值基本不随输入变化而改变,在LLMs中充当不可或缺的偏置项。第三,这些巨量激活会导致注意力概率向其对应token集中,进而在自注意力输出中形成隐式偏置项。最后,我们还研究了Vision Transformers中的巨量激活。代码可在https://github.com/locuslab/massive-activations获取。

英文摘要:

We observe an empirical phenomenon in Large Language Models (LLMs) -- very few activations exhibit significantly larger values than others (e.g., 100,000 times larger). We call them massive activations. First, we demonstrate the widespread existence of massive activations across various LLMs and characterize their locations. Second, we find their values largely stay constant regardless of the input, and they function as indispensable bias terms in LLMs. Third, these massive activations lead to the concentration of attention probabilities to their corresponding tokens, and further, implicit bias terms in the self-attention output. Last, we also study massive activations in Vision Transformers. Code is available at https://github.com/locuslab/massive-activations.

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。

↑