[论文解读] What and How does In-Context Learning Learn? Bayesian Model Averaging, Parameterization, and Generalization

Yufeng Zhang, Fengzhuo Zhang+2 more|arXiv (Cornell University)|May 30, 2023

Topic Modeling被引用 4

一句话总结

本文将大语言模型中的上下文学习（ICL）框架化为通过Transformer注意力机制隐式实现的贝叶斯模型平均（BMA），建立了ICL性能的O(1/T)遗憾边界，并证明了预训练误差随模型深度呈指数衰减，随数据集大小呈次线性衰减。该研究提供了一个统一的理论框架，将注意力机制、预训练与ICL泛化能力联系起来。

ABSTRACT

In this paper, we conduct a comprehensive study of In-Context Learning (ICL) by addressing several open questions: (a) What type of ICL estimator is learned by large language models? (b) What is a proper performance metric for ICL and what is the error rate? (c) How does the transformer architecture enable ICL? To answer these questions, we adopt a Bayesian view and formulate ICL as a problem of predicting the response corresponding to the current covariate, given a number of examples drawn from a latent variable model. To answer (a), we show that, without updating the neural network parameters, ICL implicitly implements the Bayesian model averaging algorithm, which is proven to be approximately parameterized by the attention mechanism. For (b), we analyze the ICL performance from an online learning perspective and establish a $\mathcal{O}(1/T)$ regret bound for perfectly pretrained ICL, where $T$ is the number of examples in the prompt. To answer (c), we show that, in addition to encoding Bayesian model averaging via attention, the transformer architecture also enables a fine-grained statistical analysis of pretraining under realistic assumptions. In particular, we prove that the error of pretrained model is bounded by a sum of an approximation error and a generalization error, where the former decays to zero exponentially as the depth grows, and the latter decays to zero sublinearly with the number of tokens in the pretraining dataset. Our results provide a unified understanding of the transformer and its ICL ability with bounds on ICL regret, approximation, and generalization, which deepens our knowledge of these essential aspects of modern language models.

研究动机与目标

• 理解大语言模型在无参数更新情况下，ICL实现的是何种估计器。
• 为ICL定义合适的性能度量指标，并量化其误差率。
• 解释Transformer架构如何通过提示工程与预训练共同实现ICL。
• 基于贝叶斯与在线学习框架，提供ICL、预训练与泛化能力的统一理论分析。

提出的方法

• 将ICL建模为贝叶斯推断问题，其中样本来自由隐藏概念z*参数化的潜在变量模型。
• 证明无参数更新的ICL实现了贝叶斯模型平均（BMA），且注意力机制近似编码了BMA过程。
• 基于在线学习引入ICL遗憾度量，证明在完美预训练模型下，遗憾边界为O(1/T)。
• 应用PAC-Bayes框架分析预训练误差，将其分解为近似误差与泛化误差两部分。
• 证明近似误差随Transformer深度增加呈指数衰减，泛化误差随预训练数据集token数量呈次线性衰减。
• 利用集中不等式与矩阵范数界分析注意力机制的稳定性与泛化能力。

实验结果

研究问题

• RQ1 在无参数更新的情况下，大语言模型隐式学习到的ICL估计器属于何种类型？
• RQ2 适用于ICL的合适性能度量是什么？其理论误差率如何？
• RQ3 Transformer架构如何在预训练与提示过程中共同实现ICL？
• RQ4 预训练中的近似误差与泛化误差如何影响下游ICL性能？

主要发现

• 无参数更新的ICL隐式实现了贝叶斯模型平均（BMA），且注意力机制近似编码了该过程。
• 在提示中使用T个样本后，ICL的遗憾边界为O(1/T)，其上界受预训练模型统计误差的限制。
• 预训练模型中的近似误差随Transformer深度增加呈指数级快速衰减。
• 泛化误差随预训练数据集中的token数量呈次线性衰减。
• 标准softmax注意力机制近似编码了BMA，尤其在提示长度增加时更为显著。
• 该理论框架统一了ICL、预训练与泛化能力，为三者均提供了理论边界。