Optimizing Active Learning in Vision-Language Models via Parameter-Efficient Uncertainty Calibration
This repository will host the code for the paper titled "Optimizing Active Learning in Vision-Language Models via Parameter-Efficient Uncertainty Calibration"
Stay tuned! The code will be released soon.
Active Learning (AL) has emerged as a powerful approach for minimizing labeling costs by selectively sampling the most informative data for neural network model development. Effective AL for large-scale vision-language models necessitates addressing challenges in uncertainty estimation and efficient sampling given the vast number of parameters involved. In this work, we introduce a novel parameter-efficient learning methodology that incorporates uncertainty calibration loss within the AL framework. We propose a differentiable loss function that promotes uncertainty calibration for effectively selecting fewer and most informative data samples for fine-tuning. Through extensive experiments across several datasets and vision backbones, we demonstrate that our solution can match and exceed the performance of complex feature-based sampling techniques while being computationally very efficient. Additionally, we investigate the efficacy of Prompt learning versus Low-rank adaptation (LoRA) in sample selection, providing a detailed comparative analysis of these methods in the context of efficient AL.
If you find this work useful, please consider citing our previous works:
@article{narayanan2024parameter,
title={Parameter-Efficient Active Learning for Foundational models},
author={Narayanan, Athmanarayanan Lakshmi and Krishnan, Ranganath and Machireddy, Amrutha and Subedar, Mahesh},
journal={arXiv preprint arXiv:2406.09296},
year={2024}
}
Details about the license will be provided upon release.