skip to main content
US FlagAn official website of the United States government
dot gov icon
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
https lock icon
Secure .gov websites use HTTPS
A lock ( lock ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

Explore Research Products in the PAR
It may take a few hours for recently added research products to appear in PAR search results.


This content will become publicly available on November 16, 2025

Title: Scalable Fine-tuning from Multiple Data Sources: A First-Order Approximation Approach
We study the problem of fine-tuning a language model (LM) for a target task by optimally using the information from n auxiliary tasks. This problem has broad applications in NLP, such as targeted instruction tuning and data selection in chain-of-thought fine-tuning. The key challenge of this problem is that not all auxiliary tasks are useful to improve the performance of the target task. Thus, choosing the right subset of auxiliary tasks is crucial. Conventional subset selection methods, such as forward & backward selection, are unsuitable for LM fine-tuning because they require repeated training on subsets of auxiliary tasks. This paper introduces a new algorithm to estimate model fine-tuning performances without repeated training. Our algorithm first performs multitask training using the data of all the tasks to obtain a meta initialization. Then, we approximate the model fine-tuning loss of a subset using functional values and gradients from the meta initialization. Empirically, we find that this gradient-based approximation holds with remarkable accuracy for twelve transformer-based LMs. Thus, we can now estimate fine-tuning performances on CPUs within a few seconds. We conduct extensive experiments to validate our approach, delivering a speedup of 30× over conventional subset selection while incurring only 1% error of the true fine-tuning performances. In downstream evaluations of instruction tuning and chain-of-thought fine-tuning, our approach improves over prior methods that utilize gradient or representation similarity for subset selection by up to 3.8%.  more » « less
Award ID(s):
2412008
PAR ID:
10595812
Author(s) / Creator(s):
; ; ;
Publisher / Repository:
Association for Computational Linguistics
Date Published:
Journal Name:
Findings of the Association for Computational Linguistics: EMNLP 2024
ISSN:
0736-587X
Page Range / eLocation ID:
5608-5623
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. ; ; ; (, Proceedings of the IEEE International Conference on Data Mining (ICDM 2020))
    null (Ed.)
    The problem of learning to generalize on unseen classes during the training step, also known as few-shot classification, has attracted considerable attention. Initialization based methods, such as the gradient-based model agnostic meta-learning (MAML) [1], tackle the few-shot learning problem by “learning to fine-tune”. The goal of these approaches is to learn proper model initialization so that the classifiers for new classes can be learned from a few labeled examples with a small number of gradient update steps. Few shot meta-learning is well-known with its fast-adapted capability and accuracy generalization onto unseen tasks [2]. Learning fairly with unbiased outcomes is another significant hallmark of human intelligence, which is rarely touched in few-shot meta-learning. In this work, we propose a novel Primal-Dual Fair Meta-learning framework, namely PDFM, which learns to train fair machine learning models using only a few examples based on data from related tasks. The key idea is to learn a good initialization of a fair model’s primal and dual parameters so that it can adapt to a new fair learning task via a few gradient update steps. Instead of manually tuning the dual parameters as hyperparameters via a grid search, PDFM optimizes the initialization of the primal and dual parameters jointly for fair meta-learning via a subgradient primal-dual approach. We further instantiate an example of bias controlling using decision boundary covariance (DBC) [3] as the fairness constraint for each task, and demonstrate the versatility of our proposed approach by applying it to classification on a variety of three real-world datasets. Our experiments show substantial improvements over the best prior work for this setting. 
    more » « less
  2. ; ; ; ; (, Proceedings of the 40th International Conference on Machine Learning)
    It has become standard to solve NLP tasks by fine-tuning pre-trained language models (LMs), especially in low-data settings. There is minimal theoretical understanding of empirical success, e.g., why fine-tuning a model with $10^8$ or more parameters on a couple dozen training points does not result in overfitting. We investigate whether the Neural Tangent Kernel (NTK)—which originated as a model to study the gradient descent dynamics of infinitely wide networks with suitable random initialization—describes fine-tuning of pre-trained LMs. This study was inspired by the decent performance of NTK for computer vision tasks (Wei et al., 2022). We extend the NTK formalism to Adam and use Tensor Programs (Yang, 2020) to characterize conditions under which the NTK lens may describe fine-tuning updates to pre-trained language models. Extensive experiments on 14 NLP tasks validate our theory and show that formulating the downstream task as a masked word prediction problem through prompting often induces kernel-based dynamics during fine-tuning. Finally, we use this kernel view to propose an explanation for the success of parameter-efficient subspace-based fine-tuning methods. 
    more » « less
  3. ; (, The IEEE/CVF Conference on Computer Vision and Pattern Recognition 2024: The 7th Workshop on Efficient Deep Learning for Computer Vision (ECV))
    Meta-learning methods typically learn tasks under the assumption that all tasks are equally important. However, this assumption is often not valid. In real-world applications, tasks can vary both in their importance during different train- ing stages and in whether they contain noisy labeled data or not, making a uniform approach suboptimal. To address these issues, we propose the Data-Efficient and Robust Task Selection (DERTS) algorithm, which can be incorporated into both gradient and metric-based meta-learning algo- rithms. DERTS selects weighted subsets of tasks from task pools by minimizing the approximation error of the full gra- dient of task pools in the meta-training stage. The selected tasks are efficient for rapid training and robust towards noisy label scenarios. Unlike existing algorithms, DERTS does not require any architecture modification for training and can handle noisy label data in both the support and query sets. Analysis of DERTS shows that the algorithm follows similar training dynamics as learning on the full task pools. Experiments show that DERTS outperforms existing sam- pling strategies for meta-learning on both gradient-based and metric-based meta-learning algorithms in limited data budget and noisy task settings. 
    more » « less
  4. ; ; ; ; ; ; (, ICLR 2023)
    A common approach to transfer learning under distribution shift is to fine-tune the last few layers of a pre-trained model, preserving learned features while also adapting to the new task. This paper shows that in such settings, selectively fine-tuning a subset of layers (which we term surgical fine-tuning) matches or outperforms commonly used fine-tuning approaches. Moreover, the type of distribution shift influences which subset is more effective to tune: for example, for image corruptions, fine-tuning only the first few layers works best. We validate our findings systematically across seven real-world data tasks spanning three types of distribution shifts. Theoretically, we prove that for two-layer neural networks in an idealized setting, first-layer tuning can outperform fine-tuning all layers. Intuitively, fine-tuning more parameters on a small target dataset can cause information learned during pre-training to be forgotten, and the relevant information depends on the type of shift. 
    more » « less
  5. ; (, Frontiers in Big Data)
    Dynamic transfer learning refers to the knowledge transfer from a static source task with adequate label information to a dynamic target task with little or no label information. However, most existing theoretical studies and practical algorithms of dynamic transfer learning assume that the target task is continuously evolving over time. This strong assumption is often violated in real world applications, e.g., the target distribution is suddenly changing at some time stamp. To solve this problem, in this paper, we propose a novel meta-learning framework L2S based on a progressive meta-task scheduler for dynamic transfer learning. The crucial idea of L2S is to incrementally learn to schedule the meta-pairs of tasks and then learn the optimal model initialization from those meta-pairs of tasks for fast adaptation to the newest target task. The effectiveness of our L2S framework is verified both theoretically and empirically. 
    more » « less
  • Citation Formats
  • MLA
  • APA
  • Chicago
  • BibTeX
  • Save / Share this Record
  • Send to Email