Composed Fine-Tuning: Freezing Pre-Trained Denoising Autoencoders for Improved Generalization

TL;DR

This paper introduces composed fine-tuning, a method that preserves the output structure learned from unlabeled data by freezing a pre-trained denoiser, leading to better generalization in structured prediction tasks.

Contribution

The paper proposes composed fine-tuning, which maintains output structure by freezing a pre-trained denoiser during predictor fine-tuning, improving generalization especially on out-of-distribution data.

Findings

Composed fine-tuning outperforms standard fine-tuning on pseudocode-to-code datasets.

It significantly improves generalization on out-of-distribution examples.

Theoretical analysis shows reduced predictor complexity with composed fine-tuning.

Abstract

We focus on prediction problems with structured outputs that are subject to output validity constraints, e.g. pseudocode-to-code translation where the code must compile. While labeled input-output pairs are expensive to obtain, "unlabeled" outputs, i.e. outputs without corresponding inputs, are freely available (e.g. code on GitHub) and provide information about output validity. We can capture the output structure by pre-training a denoiser to denoise corrupted versions of unlabeled outputs. We first show that standard fine-tuning after pre-training destroys some of this structure. We then propose composed fine-tuning, which fine-tunes a predictor composed with the pre-trained denoiser, which is frozen to preserve output structure. For two-layer ReLU networks, we prove that composed fine-tuning significantly reduces the complexity of the predictor, thus improving generalization. Empirically, we show that composed fine-tuning improves over standard fine-tuning on two pseudocode-to-code translation datasets (3% and 6% relative). The improvement from composed fine-tuning is magnified on out-of-distribution (OOD) examples (4% and 25% relative).