Addressing Bias Through Ensemble Learning and Regularized Fine-Tuning
Ahmed Radwan, Layan Zaafarani, Jetana Abudawood, Faisal AlZahrani, Fares Fourati
TL;DR
The paper tackles bias in AI models when labeled data are scarce and pretrained models may encode biases. It introduces a pipeline that creates multiple models from a biased pretrained model via data-splitting and regularized fine-tuning, ensembles their predictions with logit summation to reduce bias, and then distills the ensemble into a single unbiased network for efficient inference. The core objective during regularized fine-tuning is $ \min _{\\theta} \\mathbb{E}_{(x, y) \\sim \\mathcal{D}} [ \\mathcal{L}(\\theta, x, y) + \\lambda \\|\\theta - \\theta^*\\|^2 ] $, optionally with $ \\beta \\|\\theta\\|^2 $; experiments on CIFAR-10 and HAM10000 using DenseNet-121 show improved performance for underrepresented classes and strong overall accuracy via ensembles, though distillation with logits-based loss can be challenging. The work provides a practical framework for unbiased deployment under data constraints and highlights the trade-offs between regularization strength, ensemble gains, and distillation feasibility.
Abstract
Addressing biases in AI models is crucial for ensuring fair and accurate predictions. However, obtaining large, unbiased datasets for training can be challenging. This paper proposes a comprehensive approach using multiple methods to remove bias in AI models, with only a small dataset and a potentially biased pretrained model. We train multiple models with the counter-bias of the pre-trained model through data splitting, local training, and regularized fine-tuning, gaining potentially counter-biased models. Then, we employ ensemble learning for all models to reach unbiased predictions. To further accelerate the inference time of our ensemble model, we conclude our solution with knowledge distillation that results in a single unbiased neural network. We demonstrate the effectiveness of our approach through experiments on the CIFAR10 and HAM10000 datasets, showcasing promising results. This work contributes to the ongoing effort to create more unbiased and reliable AI models, even with limited data availability.