Performance in Chaos: How Ensemble Learning Conquers Label Noise

In this paper, we investigate the impact of label noise on the performance of machine learning models using the CIFAR-100N dataset. The study evaluates the behaviour of deep neural networks in handling noisy human-labeled data by experimenting with several model configurations. We explore a wide range of architectures (e.g., ResNet, EfficientNet), assessing the robustness and adaptability of different approaches while incorporating regularization techniques and noise-handling strategies. Our proposed method, EVONet, introduces a novel soft-voting ensemble technique, which achieves a test accuracy of over 76%, outperforming state-of-the-art solutions. The results highlight the promising capabilities of ensemble-based methods in handling noisy datasets, an unavoidable aspect of real-world applications, and underscore the importance of developing robust models that can extract meaningful information even from imperfect data.