The field of machine learning is moving towards improving model calibration and active learning techniques. Researchers are exploring new approaches to ensure that models produce confidence scores that accurately reflect the true likelihood of their predictions being correct. This includes developing calibration-aware data augmentation methods and leveraging manual labeling with beneficial noise. Active learning is also gaining attention, with a focus on efficient data utilization and model performance enhancement. Techniques such as client heterogeneity-aware data selection and enhanced sample selection with confidence tracking are being proposed to overcome challenges in federated active learning and noisy data. Noteworthy papers include:

• Beyond One-Hot Labels: Semantic Mixing for Model Calibration, which introduces a novel framework for generating training samples with mixed class characteristics and annotates them with distinct confidence scores.
• CHASe: Client Heterogeneity-Aware Data Selection for Effective Federated Active Learning, which proposes a client heterogeneity-aware data selection method for federated active learning.