Rapid Performance Gain through Active Model Reuse

Feng Shi, Yu-Feng LiIJCAI 2019 · OpenReview

Abstract

Model reuse aims at reducing the need of learning resources for a newly target task. In previous model reuse studies, the target task usually receives labeled data passively, which results in a slow performance improvement. However, learning models for target tasks are often required to achieve good enough performance rapidly for practical usage. In this paper, we propose the AcMR (Active Model Reuse) method for the rapid performance improvement problem. Firstly, we construct queries through pre-trained models to facilitate the active learner when labeled examples are insufficient in the target task. Secondly, we consider that pre-trained models are able to filter out not-very-necessary queries so that AcMR can save considerable queries compared with direct active learning. Theoretical analysis verifies that AcMR requires fewer queries than direct active learning. Experimental results validate the effectiveness of AcMR.

Introduction

In the era of big data, while well-trained models for various tasks are abundant, obtaining labeled data for specific new "target" tasks remains a significant bottleneck. Traditional approaches typically follow one of two paradigms: Model Reuse, which leverages pre-trained models but often assumes passive data acquisition, leading to slow gains; or Active Learning, which selectively queries informative data points but ignores existing knowledge in pre-trained models, incurring high initial costs.

This paper addresses the "rapid performance improvement" problem. In practical applications, developers often need a model to reach a high performance threshold as quickly as possible. Neither passive reuse nor model-blind active learning is optimized for this speed. We propose AcMR (Active Model Reuse) to bridge this gap, accelerating learning by combining pre-trained models with active learning.

The AcMR Framework

AcMR integrates pre-trained models into the active learning loop to both guide query selection and reduce human labeling effort. The framework consists of three primary mechanisms:

  1. Query Filtering via Pre-trained Models: Standard active learning sends every "uncertain" point to a human expert (Oracle). AcMR uses an ensemble of pre-trained models as a "Pre-Oracle" filter. If these models are highly confident and align with the current active learner, the system trusts their prediction and skips the human query, significantly saving costs.
  2. Dynamic Weighting: Since pre-trained models vary in relevance to a new task, AcMR employs an error-driven update rule based on Kullback–Leibler divergence. As ground-truth labels are acquired, the system dynamically adjusts the weights of these "experts."
  3. Facilitated Active Learner: The final prediction combines a local active learner (trained on target task labels) with the weighted ensemble of pre-trained models.
The AcMR framework workflow, illustrating the interaction between unlabeled data, pre-trained models, and the human Oracle.

Experimental Evaluation

The effectiveness of AcMR was evaluated across three domains: Text Classification (20 Newsgroups), Sentiment Analysis (Amazon reviews), and Spam Detection (ECML PAKDD).

Key Results

AcMR consistently outperformed baselines such as Query-by-Committee (QBC), Random sampling, and "Safer" model reuse.

  • Acceleration of Learning: In the 20 Newsgroups tasks, AcMR reached approximately 90% accuracy with only 30 queries, whereas standard active learning (QBC) required 60 to 90 queries to achieve comparable results.
  • Robustness: The framework showed superior performance regardless of the underlying sampling strategy (QBC or Random), demonstrating the robustness of the reuse layer. Even in sentiment analysis tasks with significant domain shifts, AcMR successfully leveraged pre-trained knowledge to accelerate performance gains.
Accuracy vs. Number of query samples, showing the rapid performance gain of the proposed approach (Expectation) compared to standard active learning (QBC).

Discussion and Conclusion

AcMR creates a virtuous cycle: active learning improves the local model, which in turn helps the system better evaluate pre-trained models. This leads to more accurate query filtering and faster learning. Theoretical analysis confirms that AcMR's sampling error and query rates are bounded and generally lower than direct active learning.

While AcMR assumes a pool of relevant pre-trained models, its dynamic weighting mechanism provides a safety net against irrelevant models. Future work includes extending this framework to deep learning architectures and more complex tasks like regression or structured prediction.

Summary updated

Abstract

Model reuse aims at reducing the need of learning resources for a newly target task. In previous model reuse studies, the target task usually receives labeled data passively, which results in a slow performance improvement. However, learning models for target tasks are often required to achieve good enough performance rapidly for practical usage. In this paper, we propose the AcMR (Active Model Reuse) method for the rapid performance improvement problem. Firstly, we construct queries through pre-trained models to facilitate the active learner when labeled examples are insufficient in the target task. Secondly, we consider that pre-trained models are able to filter out not-very-necessary queries so that AcMR can save considerable queries compared with direct active learning. Theoretical analysis verifies that AcMR requires fewer queries than direct active learning. Experimental results validate the effectiveness of AcMR.
BibTeX
@inproceedings{shi2019rapid,
  author = {Feng Shi and Yu-Feng Li},
  title = {{Rapid Performance Gain through Active Model Reuse}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=fH8Dbh5kYrl},
}
Figure