Evaluating a neonatal public health education and empowerment programme to reduce infant mortality risks: insights from parents and facilitators in the midlands, UK.

The vast repertoire of high-performing Deep Learning (DL) architectures has established transfer learning as the standard approach across most research fields. For most computer vision problems, the advantages of bypassing architectural design to focus on data preprocessing and model retraining far outweigh the advantages of creating custom models. However, despite its widespread popularity, transfer learning often fails to yield statistically significant improvements over training from scratch on medical images (; ; ). The efficacy of knowledge transfer to domains that differ substantially from natural images has long been questioned, with some authors suggesting that observed performance gains stem solely from the over-parameterization of pretrained architectures rather than meaningful feature transfer (; ). To compensate for suboptimal architectural design, data augmentation is widely employed to strike a balance between generating the highest diversity possible and maintaining sufficient data quality (). However, expanding the data distribution typically necessitates lengthier training, as models must adapt to more complex tasks rather than relying on memorization. More importantly, inadequate transformations can introduce excessive noise, widening the gap between training samples and the real-world distribution (; ). Ultimately, the efficacy of a certain augmentation scheme depends on both the nature of the data and the model’s capacity to capture its nuances. Thus, data augmentation cannot replace appropriate model selection. Alternatively, Deep Ensemble Learning (DEL) is employed to mitigate suboptimal architectural design by significantly reducing both bias and variance in the final predictions (). Similar to traditional ensemble methods (), multiple models, referred to as learners, are trained on data subsets to leverage the unique features learned by each one of them. This modularity allows for the use of shallower architectures focused on specific sub-tasks rather than a single, deeper model which, while powerful, is more prone to overfitting. Nevertheless, training multiple models is more computationally expensive than fine-tuning a single architecture. Additionally, achieving learner diversity is difficult when identical architectures share the same inductive biases. Even if the ensembles are composed of diverse architectures, some models may not contribute value or may even degrade overall predictions. Ensemble pruning, also known as selective ensemble methods, is then required to mitigate these negative effects. These methods are based on metrics such as validation error, kappa measure, or diversity (; ; ).