[Paper Review] Why Do Pretrained Language Models Help in Downstream Tasks? An Analysis of Head and Prompt Tuning

Colin Wei, Sang Michael Xie+1 more|arXiv (Cornell University)|Jun 17, 2021

Topic ModelingComputer Science28 references33 citations

TL;DR

The paper analyzes why pretrained language models aid downstream tasks by examining head tuning and prompt tuning under latent-variable generative models (HMMs and memory-augmented HMMs), proving recovery guarantees under various non-degeneracy conditions and validating findings with synthetic experiments.

ABSTRACT

Pretrained language models have achieved state-of-the-art performance when adapted to a downstream NLP task. However, theoretical analysis of these models is scarce and challenging since the pretraining and downstream tasks can be very different. We propose an analysis framework that links the pretraining and downstream tasks with an underlying latent variable generative model of text -- the downstream classifier must recover a function of the posterior distribution over the latent variables. We analyze head tuning (learning a classifier on top of the frozen pretrained model) and prompt tuning in this setting. The generative model in our analysis is either a Hidden Markov Model (HMM) or an HMM augmented with a latent memory component, motivated by long-term dependencies in natural language. We show that 1) under certain non-degeneracy conditions on the HMM, simple classification heads can solve the downstream task, 2) prompt tuning obtains downstream guarantees with weaker non-degeneracy conditions, and 3) our recovery guarantees for the memory-augmented HMM are stronger than for the vanilla HMM because task-relevant information is easier to recover from the long-term memory. Experiments on synthetically generated data from HMMs back our theoretical findings.

Motivation & Objective

• Link pretraining and downstream tasks via an underlying latent-variable generative model of text.
• Show that downstream labels can be recovered using simple heads or prompts when conditioned on posterior latent variables.
• Compare head tuning and prompt tuning under vanilla and memory-augmented HMMs to understand robustness of task-relevant information recovery.
• Demonstrate that prompt tuning relaxes non-degeneracy conditions and strengthens recovery guarantees.
• Provide empirical validation on synthetically generated data supporting the theoretical claims.

Proposed method

• Model data with latent-variable generative models: HMM and memory-augmented HMM.
• Prove downstream labels can be recovered by a linear head on exact conditional token probabilities under non-degeneracy (Theorem 3.3).
• Show that soft prompt tuning relaxes non-degeneracy to enable recovery (Theorem 3.6).
• Extend analysis to memory-augmented HMMs where an attention-based head recovers labels under weaker conditions (Theorem 4.3).
• Provide further recovery guarantees for prompt tuning in memory-augmented settings (Theorem 4.6).
• Empirically validate with language-models pretrained on synthetically generated HMM data.

Experimental results

Research questions

• RQ1 Can a linear head on pretrained model outputs recover downstream labels in an HMM-based data-generating process?
• RQ2 Does prompt tuning relax non-degeneracy requirements and improve downstream recoverability compared to head tuning?
• RQ3 Do memory-augmented HMMs enhance recoverability of task-relevant information, and can attention heads outperform linear heads?
• RQ4 How do prompt-tuning guarantees extend to memory-augmented models?

Key findings

• Under vanilla HMMs, a linear head can recover downstream labels given strong non-degeneracy of token emissions.
• Soft prompt tuning relaxes non-degeneracy requirements and enables downstream recovery with weaker conditions.
• Memory-augmented HMMs yield stronger recovery guarantees than vanilla HMMs, as memory facilitates extracting task-relevant information via attention heads.
• Attention-based heads can recover ground-truth labels under weaker assumptions than linear heads in memory-augmented settings.
• Prompt tuning demonstrates superior empirical performance over head tuning on synthetically generated data, aligning with theoretical guarantees.