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

arXiv: https://arxiv.org/abs/2006.16205 · HTML: https://mlsanity.com/paper/2006.16205/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 29 Jun 2020.

## Built on by (most cited first, 9 of 9)

- [Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution](https://mlsanity.com/paper/2202.10054/) (2022)
- [Joint Embedding vs Reconstruction: Provable Benefits of Latent Space Prediction for Self Supervised Learning](https://mlsanity.com/paper/2505.12477/) (2025)
- [Preserving Pre-trained Features Helps Calibrate Fine-tuned Language Models](https://mlsanity.com/paper/2305.19249/) (2023)
- [Fast Trainable Projection for Robust Fine-Tuning](https://mlsanity.com/paper/2310.19182/) (2023)
- [Do we really have to filter out random noise in pre-training data for language models?](https://mlsanity.com/paper/2502.06604/) (2025)
- [Backdoor Learning for NLP: Recent Advances, Challenges, and Future Research Directions](https://mlsanity.com/paper/2302.06801/) (2023)
- [Knowledge Transfer for Pseudo-code Generation from Low Resource Programming Language](https://mlsanity.com/paper/2303.09062/) (2023)
- [Open Domain Generalization with a Single Network by Regularization Exploiting Pre-trained Features](https://mlsanity.com/paper/2312.05141/) (2023)
- [Scaling Automatic Extraction of Pseudocode](https://mlsanity.com/paper/2406.04635/) (2024)

## Abstract (arXiv)

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).